Using ensemble learning and explainable AI to predict bank marketing customer subscription Authors: Wenyue Wang, Qi Liu, Sufeng Li, Hongyan Feng Published: 2026 ABSTRACT To address class imbalance and distribution shifts in bank marketing tasks, this paper presents an approach based on CatBoost ensemble and explainable AI, referred to as CBE-XAI (CatBoost Ensemble with eXplainable AI), to predict term deposit subscription intentions. The approach integrates five heterogeneous CatBoost base learners using a post-training dynamic weighting mechanism based on validation performance. It also employs a hierarchical SHapley Additive exPlanations (SHAP) system to aggregate local attributions for micro-to-macro feature importance analysis. In addition, the approach includes a pre-deployment adaptive fine-tuning (AFT) strategy with a composite loss function for cross-environment model calibration. Experimental results show that CBE-XAI achieves an Area under the receiver operating characteristic curve (AUROC) of 0.949 and an F1-Score of 0.621 on public datasets, demonstrating performance comparable to several standard benchmarks. External validation on a Chinese city commercial bank dataset confirms that the AFT strategy improves AUROC from 0.842 to 0.931 (p<0.001), enhancing cross-domain generalization. Furthermore, hierarchical SHAP analysis reveals key decision pathways for features such as call duration, ensuring consistency with banking business logic. CBE-XAI provides a potential technical solution for balancing accuracy, interpretability, and generalization in bank marketing. Introduction Currently, digital transformation is driving profound changes in the financial services industry. As a cornerstone of the economic system, the digital transformation of the banking sector is crucial for enhancing service efficiency and enabling business model innovation1. However, due to interest rate liberalization and intensified interbank competition, many banks still rely on traditional, experience-based marketing, resulting in low customer acquisition efficiency2. A key challenge is the lack of a data-driven predictive framework and data-driven decision-making mechanisms, making it difficult for most banks to accurately identify conversion opportunities from customer behavior. With the widespread application of Artificial Intelligence (AI) in the financial sector3, constructing marketing prediction models based on data-driven and AI technologies has become a vital direction for overcoming these challenges and achieving sustainable competitiveness. This study focuses on bank marketing tasks. It uses customer profiles and macroeconomic indicators to capture and analyze customers’ subscription intentions in complex marketing environments. Although AI has shown promise in financial forecasting, three practical challenges remain in developing predictive models that balance precision, interpretability, and robustness. This study addresses these challenges by using ensemble learning and explainable AI. First, the issue of weight adaptation in ensemble learning under imbalanced data. Early machine learning primarily utilized single models such as Logistic Regression and Decision Trees. However, when dealing with class-imbalanced data, prediction results are prone to being dominated by the majority class. Ensemble learning improves prediction robustness by combining multiple base learners4,5, with Random Forest, XGBoost, LightGBM, and CatBoost being widely applied due to their high efficiency6,7. Nevertheless, existing ensemble methods mostly employ serial iteration or parallel independent training. Serial ensembles train base learners sequentially, where subsequent models depend on the errors of previous ones, often leading to error accumulation and rigid weight distribution. In parallel training, simple averaging or voting is applied after independent training. This approach fails to dynamically adjust weights based on the performance variance of base learners on imbalanced data, which limits the model’s ability to capture minority class instances. Consequently, a weight allocation mechanism that is decoupled from the training process and directly reflects performance on an imbalanced validation set is worth exploring. Second, the lack of interpretability in the decision-making process. While ensemble methods enhance prediction accuracy, their high complexity and non-parametric nature render the decision process opaque, leading to skepticism among users. Furthermore, many studies overlook interpretability analysis8. The SHapley Additive exPlanations (SHAP) method, based on the Shapley value from game theory, provides fair attribution and clarifies the contribution and influence direction of each feature9,10, thereby improving model transparency11. However, most existing research focuses on single models. For ensemble models composed of multiple heterogeneous base learners, further investigation is required. Specifically, it is necessary to aggregate the local explanations from each base learner into a globally coherent interpretation, while also quantifying the contribution of each learner. Third, the adaptation to distribution shifts during cross-environment model deployment. Missing values are prevalent in bank customer data, and while existing studies utilize imputation methods12,13, the process can introduce estimation bias. Moreover, when models are deployed across regions or face changes in the macroeconomic environment, distribution shifts in customer demographics and economic indicators often lead to degraded predictive performance. Traditional retraining schemes are prone to overfitting on small samples and are costly. Therefore, it is important to develop an efficient model adaptation mechanism that leverages a small amount of local historical data to adapt to new environments, thereby ensuring model reliability in dynamic banking settings. In summary, current AI methods for bank marketing prediction face three primary limitations: (1) under imbalanced data, base learner training and weight allocation in ensemble learning remain relatively rigid, lacking a post-training weighting mechanism based on imbalanced performance; (2) the internal decision mechanisms of multi-base-learner ensemble models are opaque, and hierarchical attribution explanations require refinement; and (3) a lack of efficient and robust self-adaptive fine-tuning strategies to handle distribution shifts during cross-environment deployment. To address these issues, this paper presents an approach called CBE-XAI (CatBoost Ensemble with eXplainable AI) based on ensemble learning and explainable AI. This approach aims to mitigate the high-investment and low-conversion problems in bank term deposit marketing by enhancing prediction accuracy while exploring hierarchical interpretability and adaptation strategies for cross-environment shifts. The main technical components of this work are as follows: A heterogeneous CatBoost ensemble with post-training weighting. To address the class imbalance inherent in bank marketing data, five CatBoost base learners with heterogeneous configurations are used and trained independently. Following training, a post-training weighting mechanism is implemented to assign ensemble weights based on each learner’s weighted AUC and PR-AUC scores on an independent validation set. This approach facilitates a more flexible decoupling of base learner fitting from weight allocation, allowing the model to adaptively reflect actual performance on skewed data distributions.A hierarchical weighted aggregation method for SHAP interpretability. To enhance transparency in multi-model decisions, we use a hierarchical interpretability system based on Shapley values. Local SHAP values from individual CatBoost learners are first calculated to quantify feature marginal contributions at the micro-level. Subsequently, these attributions are linearly aggregated using the derived performance weights to estimate ensemble-level feature impacts. This design provides a structured path to trace global feature importance back to individual model sensitivities.A pre-deployment adaptive fine-tuning method for cross-environment calibration. To mitigate performance degradation caused by distribution shifts during deployment, we use a lightweight fine-tuning strategy. This approach utilizes a composite loss function designed to balance the retention of global knowledge with the capture of local domain characteristics. This strategy offers an efficient technical option for institutions to calibrate pre-trained models using limited local historical data, facilitating robust model lifecycle management in dynamic environments. A heterogeneous CatBoost ensemble with post-training weighting. To address the class imbalance inherent in bank marketing data, five CatBoost base learners with heterogeneous configurations are used and trained independently. Following training, a post-training weighting mechanism is implemented to assign ensemble weights based on each learner’s weighted AUC and PR-AUC scores on an independent validation set. This approach facilitates a more flexible decoupling of base learner fitting from weight allocation, allowing the model to adaptively reflect actual performance on skewed data distributions. A heterogeneous CatBoost ensemble with post-training weighting. To address the class imbalance inherent in bank marketing data, five CatBoost base learners with heterogeneous configurations are used and trained independently. Following training, a post-training weighting mechanism is implemented to assign ensemble weights based on each learner’s weighted AUC and PR-AUC scores on an independent validation set. This approach facilitates a more flexible decoupling of base learner fitting from weight allocation, allowing the model to adaptively reflect actual performance on skewed data distributions. A hierarchical weighted aggregation method for SHAP interpretability. To enhance transparency in multi-model decisions, we use a hierarchical interpretability system based on Shapley values. Local SHAP values from individual CatBoost learners are first calculated to quantify feature marginal contributions at the micro-level. Subsequently, these attributions are linearly aggregated using the derived performance weights to estimate ensemble-level feature impacts. This design provides a structured path to trace global feature importance back to individual model sensitivities. A hierarchical weighted aggregation method for SHAP interpretability. To enhance transparency in multi-model decisions, we use a hierarchical interpretability system based on Shapley values. Local SHAP values from individual CatBoost learners are first calculated to quantify feature marginal contributions at the micro-level. Subsequently, these attributions are linearly aggregated using the derived performance weights to estimate ensemble-level feature impacts. This design provides a structured path to trace global feature importance back to individual model sensitivities. A pre-deployment adaptive fine-tuning method for cross-environment calibration. To mitigate performance degradation caused by distribution shifts during deployment, we use a lightweight fine-tuning strategy. This approach utilizes a composite loss function designed to balance the retention of global knowledge with the capture of local domain characteristics. This strategy offers an efficient technical option for institutions to calibrate pre-trained models using limited local historical data, facilitating robust model lifecycle management in dynamic environments. A pre-deployment adaptive fine-tuning method for cross-environment calibration. To mitigate performance degradation caused by distribution shifts during deployment, we use a lightweight fine-tuning strategy. This approach utilizes a composite loss function designed to balance the retention of global knowledge with the capture of local domain characteristics. This strategy offers an efficient technical option for institutions to calibrate pre-trained models using limited local historical data, facilitating robust model lifecycle management in dynamic environments. Literature review Applications of artificial intelligence in bank marketing As a technological engine driving transformation in the digital era, AI has brought about a fundamental shift in innovation tools14. Initially, AI research primarily focused on single-task processing, where individual models such as decision trees15 and linear regression16 were employed in sectors like bank marketing and retail analytics. To enhance model performance, scholars subsequently proposed hybrid algorithms leveraging the strengths of different models17, applied intelligent improvements to models like Support Vector Machine (SVM)18, and conducted model comparisons to identify optimal solutions19. However, the generalization ability of single models remains limited when addressing complex tasks. Consequently, ensemble learning, which combines multiple base learners, has been widely adopted for classification and prediction tasks. Sequential ensemble methods, which rely on the iterative chain logic of Boosting, were among the first to be developed. To achieve high performance, multi-model comparison remains a common approach20. Furthermore, researchers have re-optimized these methods to meet practical demands, proposing enhanced XGBoost algorithms for fraud prevention21,22, extending AdaBoost for fair classification23, and constructing hybrid ensemble models based on Random Forest and CatBoost to detect Bitcoin transactions24. Nevertheless, sequential ensembles are often constrained by low training efficiency and sensitivity to noise. In response, parallel ensemble methods such as Random Forest, Extra Trees, and Bagging have gained prominence, seeing extensive application in risk assessment25, insurance fraud detection26, and feature classification27. However, many standard ensemble implementations encounter limitations in weight allocation, and few provide adaptive mechanisms explicitly tailored to base learner performance on imbalanced validation data. With the expansion of data scales and the advancement of computational power, deep learning has been widely implemented due to its ability to automatically learn hierarchical features. To address the deficiencies of traditional deep learning models in information integration and redundancy elimination, the Multi-Granularity Spatio-temporal Fusion Transformer was proposed28. Building on this, scholars have combined deep learning with heterogeneous ensemble surrogate models and improved multi-objective optimization algorithms, thereby reducing the computational burden in high-dimensional search spaces29. Research has further deepened into more complex time-series forecasting. For instance, the MetaGNSDformer framework integrates base learners with meta-learners to achieve point and interval prediction30. Currently, AI is demonstrating notable potential within the financial domain. At the macroeconomic level, machine learning can predict economic trends through text analysis31 and uncover correlations between the NASDAQ index and various industry-specific cryptocurrency indices32. As a pivotal hub of the financial sector, the banking industry has introduced AI and data analytics to internally enhance return on assets33, efficiently coordinate regulatory compliance and risk management34, prevent credit risk8, and predict bank failure rates35, while simultaneously measuring changes in bank capital requirements resulting from corporate defaults36. Externally, machine learning models are utilized to predict banks’ own credit ratings37. Meanwhile, machine learning models have been extensively applied to stock market return forecasting38–40. Hybrid LSTM-GRU models, by combining Long Short-Term Memory units with gating mechanisms, can better resolve short-term fluctuations and long-term trends in stock prices41. Additionally, deep learning models such as GANs and Transformers have been widely deployed42. To combat the prevalent issue of overfitting, the Deep Generative Adversarial Network approach not only improves marketing efficiency but also minimizes prediction error rates43. Interpretability of ensemble models Ensemble learning is extensively utilized for complex tasks, yet its interpretability often remains limited. Stakeholders frequently maintain a skeptical attitude toward these ‘black-box’ models, which can restrict their practical application44. Since 2018, there has been a surge in research regarding Explainable Artificial Intelligence (XAI)45, which is typically categorized into local and global explanations46. Among these, SHAP has received considerable academic attention, as it not only quantifies the contribution of individual features to prediction outcomes but also demonstrates versatility in both supervised and unsupervised environments47. Nevertheless, the specific needs of different stakeholders are often overlooked. Consequently, some scholars have begun constructing interpretability frameworks across three levels—developers, domain experts, and end-users—aiming to bridge the gap between technical refinement and practical application48. Currently, SHAP exhibits high levels of acceptance and application potential in fields such as finance and healthcare. In the financial sector, it is most widely used in credit management, financial distress prediction, stock price forecasting, and fraud detection49. To enhance the interpretability of credit risk ratings for lending platforms, researchers have developed similarity network models based on Shapley values to achieve traceability and explanation of underlying explanatory variables50. Other studies have integrated SHAP with LASSO or LIME algorithms for mortgage default prediction51 and credit risk management. Subsequent research applied SHAP to financial distress prediction52 and further improved interpretability by introducing voluntary disclosure models53. Additionally, some studies have merged XAI with deep learning to predict real-time, sentiment-driven stock prices54, or combined SHAP with autoencoder-based label generation to propose a fully unsupervised framework for detecting credit card fraud55. Furthermore, combining unsupervised SHAP feature selection with unsupervised labeling methods has proven effective in investigating fraudulent behavior in highly imbalanced datasets56. However, a practical challenge in XAI is how to transition from local SHAP explanations of individual base learners to a hierarchically aggregated and weighted global interpretability framework. This issue warrants further investigation. Handling of missing data and model adaptation Missing data is an inevitable challenge in empirical research57, and imputation has become the prevailing approach. However, errors introduced during the imputation process can undermine model predictive performance. To address the imprecisions inherent in both completion and classification, optimal evidence classification methods based on the Dempster-Shafer theory have proven effective58. Furthermore, imputation frameworks based on feature importance can address the limitations of existing studies, which often treat all features as equally significant59. Concurrently, scholars have proposed diverse enhancement strategies: for small datasets, self-representative random forests offer a viable solution60; for large-scale mixed-type missing data, imputation can be achieved through generalized factor models61, or by combining clustering methods for mixed types with aggregation strategies62. Given that financial data is predominantly composed of time series, researchers have utilized interpolation attention mechanisms and adaptive graph convolutional networks63, as well as models built on Bi-LSTM and Extreme Learning Machines to achieve both classification and missing value recovery64. As practical requirements grow increasingly complex, innovation in completion methods continues to emerge. This includes conditional autoregressive models based on next-scale prediction65, as well as semi-supervised Generative Adversarial Networks that leverage supervisory signals for high-precision imputation66. Despite the gradual maturation of missing value imputation techniques in terms of accuracy, model deployment in authentic cross-environment settings still faces severe challenges. Traditional methods typically assume a stationary distribution of test data, which struggles to cope with non-stationary data streams in dynamic scenarios; this necessitates algorithms with robust online adaptive capabilities67. To address the dependency of unsupervised domain adaptation on source data, source similarity measurement mechanisms can be employed to facilitate source-free domain adaptation68. Furthermore, as existing fine-tuning methods are largely tailored for the computer vision domain, some scholars have introduced local fine-tuning specifically for tabular structured data to enhance generalization69. When test environments migrate frequently, the model adaptation process often occurs on edge devices with limited memory, imposing dual constraints on efficiency and stability70. Consequently, in the face of distributional shifts caused by environmental transitions, developing efficient and adaptive fine-tuning frameworks tailored for tabular data remains a significant challenge in practical model deployment71. Methods Bank marketing prediction faces three core challenges: class imbalance, “black-box” decision-making, and distribution shifts during cross-environment deployment. To address these challenges, this paper presents an approach called CBE-XAI. As illustrated in Fig. 1, the approach consists of three primary components. The process begins with data collection from the Bank Marketing Dataset and the CACB external validation dataset (Fig. 1A), followed by a data preparation pipeline involving encoding, scaling, and an 8:1:1 dataset split (Fig. 1B). To tackle the class imbalance inherent in bank marketing data, we use five heterogeneous CatBoost base learners with differentiated loss functions and structural constraints. These learners are trained independently and in parallel on source domain data. Upon completion of training, a validation-set performance-aware weighting mechanism is employed, where ensemble weights are calculated based on the weighted combined performance of AUC and PR-AUC. Finally, the optimal decision boundary is determined through threshold optimization (Fig. 1C). To unravel the internal decision-making mechanism of the ensemble model, we implement a hierarchical SHAP analysis system based on Shapley values. Initially, the pre-trained weights of the five base learners are loaded to calculate local feature attribution for each individual learner. Subsequently, these individual SHAP values are linearly aggregated using the weights derived from the first module to produce the comprehensive SHAP values at the ensemble level. The overall impact of each feature on the prediction is then quantified via the global mean absolute SHAP values, achieving global decision synthesis (Fig. 1D). To address performance degradation caused by distribution shifts during cross-environment deployment, this study uses a pre-deployment adaptation method. This method allows target institutions to perform lightweight calibration of the pre-trained model using a small amount of local historical labeled data. First, missing values in local data are processed using a type-sensitive imputation protocol consistent with the inference phase. Then, the model parameters are updated through a composite loss function that balances global knowledge retention with local distribution adaptation. This step is optional: if no local data is available, the original model can be directly utilized for inference; if local historical data exists, low-cost adaptation is enabled, facilitating comprehensive model lifecycle management (Fig. 1E). Fig. 1Overall architecture and validation workflow of the proposed approach (CBE-XAI). (A) Data collection: Bank Marketing public dataset and CACB external validation dataset; (B) Data preparation pipeline: Encoding, scaling, and dataset split; (C) Heterogeneous CatBoost ensemble: Parallel prediction of five heterogeneous base learners, followed by ensemble fusion via validation-set performance-aware weighting and threshold optimization; (D) Hierarchical SHAP analysis: Loading five base learner weights for local feature attribution and global decision synthesis; (E) Pre-deployment Adaptation: Optional calibration of the pre-trained model using a small amount of local historical labeled data to address cross-environment distribution shift. Overall architecture and validation workflow of the proposed approach (CBE-XAI). (A) Data collection: Bank Marketing public dataset and CACB external validation dataset; (B) Data preparation pipeline: Encoding, scaling, and dataset split; (C) Heterogeneous CatBoost ensemble: Parallel prediction of five heterogeneous base learners, followed by ensemble fusion via validation-set performance-aware weighting and threshold optimization; (D) Hierarchical SHAP analysis: Loading five base learner weights for local feature attribution and global decision synthesis; (E) Pre-deployment Adaptation: Optional calibration of the pre-trained model using a small amount of local historical labeled data to address cross-environment distribution shift. Heterogeneous CatBoost ensemble construction Problem description and notations Given an original dataset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:D=\left\{\right({x}_{i},{y}_{i}){\}}_{i=1}^{N}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}\in\:{\mathbb{R}}^{d}$$\end{document} denotes a d-dimensional feature vector and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{i}\in\:\left\{\mathrm{0,1}\right\}$$\end{document} is the class label (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{i}=1$$\end{document} for a subscriber, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{i}=0$$\end{document} for non-subscriber). The imbalance ratio is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\rho\:={N}_{0}/{N}_{1}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{N}_{0}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{N}_{1}$$\end{document} are the numbers of majority and minority samples, respectively. We combine five heterogeneous base learners \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{m}\left(x\right)$$\end{document} (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m=1,\dots\:,5$$\end{document}) into an ensemble mapping \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F\left(x\right)\to\:\widehat{y}$$\end{document} to improve robustness under skewed class distributions. Heterogeneous base learner design Single models often bias toward the majority class when processing imbalanced data, leading to low recognition rates for the minority class. To address this limitation, we employ five CatBoost base learners with distinct bias-variance characteristics. By using differentiated loss functions and regularization strategies, each learner captures minority class patterns from different perspectives, ensuring complementarity in the feature space. Focal Loss Model: To address the issue of the majority class dominating the loss function, this model introduces a modulating factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\left(1-{p}_{i}\right)}^{\gamma\:}$$\end{document}. This factor increases focus on hard-to-classify minority samples. 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_{FL}(p_i)=\begin{cases}-\alpha(1-p_i)^{\gamma}\log(p_i), & \text{if } y_i = 1 \\-(1-\alpha)p_i^{\gamma}\log(1-p_i), & \text{if } y_i = 0\end{cases}$$\end{document} Focal Loss Model: To address the issue of the majority class dominating the loss function, this model introduces a modulating factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\left(1-{p}_{i}\right)}^{\gamma\:}$$\end{document}. This factor increases focus on hard-to-classify minority samples. Focal Loss Model: To address the issue of the majority class dominating the loss function, this model introduces a modulating factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\left(1-{p}_{i}\right)}^{\gamma\:}$$\end{document}. This factor increases focus on hard-to-classify minority samples. where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{i}$$\end{document} is the predicted probability for sample i, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\gamma\:$$\end{document} is the focusing parameter (set to 2), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:$$\end{document} balances the loss contribution: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:={N}_{1}/N$$\end{document} for positive samples and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1-\alpha\:$$\end{document} for negative samples. (2)Balanced Weighted Model: This model explicitly introduces class weight coefficients \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{0}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{1}$$\end{document} into the cross-entropy loss to achieve rebalancing of the data distribution: (2)Balanced Weighted Model: This model explicitly introduces class weight coefficients \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{0}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{1}$$\end{document} into the cross-entropy loss to achieve rebalancing of the data distribution: Balanced Weighted Model: This model explicitly introduces class weight coefficients \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{0}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{1}$$\end{document} into the cross-entropy loss to achieve rebalancing of the data distribution: 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{bal}=-\frac{1}{N}\sum\limits_{i=1}^{N}\:\left[{w}_{0}\right(1-{y}_{i})\mathrm{l}\mathrm{o}\mathrm{g}(1-{p}_{i})+{w}_{1}{y}_{i}\mathrm{l}\mathrm{o}\mathrm{g}({p}_{i}\left)\right].$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{i}$$\end{document} is the predicted probability, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{0}=N/\left(2{N}_{0}\right)$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{1}=N/\left(2{N}_{1}\right)$$\end{document}. (3)Conservative Model: This model adopts a shallow structure (depth d=6) and applies combined \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{2}$$\end{document} regularization terms \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varOmega\:\left(f\right)$$\end{document} to the leaf node weights: 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varOmega\:\left(f\right)={\lambda\:}_{1}\parallel\:w{\parallel\:}_{1}+{\lambda\:}_{2}\parallel\:w{\parallel\:}_{2}^{2}.$$\end{document} (3)Conservative Model: This model adopts a shallow structure (depth d=6) and applies combined \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{2}$$\end{document} regularization terms \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varOmega\:\left(f\right)$$\end{document} to the leaf node weights: Conservative Model: This model adopts a shallow structure (depth d=6) and applies combined \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{2}$$\end{document} regularization terms \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varOmega\:\left(f\right)$$\end{document} to the leaf node weights: where w denotes the vector of leaf node weights, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\parallel\:\cdot\:{\parallel\:}_{1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\parallel\:\cdot\:{\parallel\:}_{2}^{2}$$\end{document} are the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{1}$$\end{document} and squared\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\:L}_{2}$$\end{document} norms, respectively. We set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{1}=0.01$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{2}=10.0$$\end{document}, and enforce a minimum of 100 samples per leaf. (4)Fast Model: Aimed at capturing globally significant features, this model utilizes an extremely shallow structure ( d=4) and a high learning rate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:=0.1$$\end{document}). Its regularization term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Omega\:}}_{fast}\left({f}_{m}\right)$$\end{document} is defined as: (4)Fast Model: Aimed at capturing globally significant features, this model utilizes an extremely shallow structure ( d=4) and a high learning rate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:=0.1$$\end{document}). Its regularization term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Omega\:}}_{fast}\left({f}_{m}\right)$$\end{document} is defined as: Fast Model: Aimed at capturing globally significant features, this model utilizes an extremely shallow structure ( d=4) and a high learning rate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:=0.1$$\end{document}). Its regularization term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Omega\:}}_{fast}\left({f}_{m}\right)$$\end{document} is defined as: 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Omega\:}}_{fast}\left({f}_{m}\right)=\frac{1}{2}\lambda\:\sum\limits_{j=1}^{{T}_{h}}\:{w}_{j}^{2}.$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{h}$$\end{document} is the number of leaf nodes, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{j}$$\end{document} represents the leaf node weights, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:=1.0$$\end{document}. (5)Deep Model: To capture high-order nonlinear interaction features, this model allows the decision tree depth to reach d=12 and employs a dynamic regularization coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{dynamic}$$\end{document} to balance fitting capability with generalization risk: (5)Deep Model: To capture high-order nonlinear interaction features, this model allows the decision tree depth to reach d=12 and employs a dynamic regularization coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{dynamic}$$\end{document} to balance fitting capability with generalization risk: Deep Model: To capture high-order nonlinear interaction features, this model allows the decision tree depth to reach d=12 and employs a dynamic regularization coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{dynamic}$$\end{document} to balance fitting capability with generalization risk: 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{dynamic}={\lambda\:}_{0}\left(1+\frac{m}{K}\right).$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{0}$$\end{document} is the initial regularization coefficient, m is the current iteration, and K is the total number of iterations. As training progresses, the regularization intensity gradually increases to suppress overfitting. Validation set performance aware weighting strategy Upon completing the independent training of the five base learners, their respective ensemble weights must be determined. Distinct from the iterative weighting in Boosting or uniform averaging in Bagging, our post-training mechanism assigns weights based on each learner’s actual performance on an imbalanced validation set. First, the comprehensive score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{s}}_{\mathrm{m}}$$\end{document} for each base learner is evaluated on an independent validation set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{D}_{val}\:$$\end{document}:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{m}=\alpha\:\cdot\:{AUC}_{m}+(1-\alpha\:)\cdot\:{PR-AUC}_{m}.$$\end{document} where the coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:=0.7$$\end{document}. To achieve a robust weight distribution, the normalized ensemble weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{m}$$\end{document}are derived via a Softmax function with a temperature parameter T:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{m}=\frac{\mathrm{e}\mathrm{x}\mathrm{p}({s}_{m}/T)}{\sum\:_{j=1}^{5}\:\mathrm{e}\mathrm{x}\mathrm{p}({s}_{j}/T)}.$$\end{document} The Softmax mechanism is employed to achieve a smooth mapping from model performance to the weight space. This design nonlinearly amplifies the contributions of high-performing base learners, while the temperature parameter T regulates the entropy of the weight distribution. This ensures that the ensemble system maintains higher robustness against individual model errors by balancing weight concentration and diversity (set to T=1 in this study). The final ensemble prediction probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:P\left(y=1|x\right)$$\end{document} is calculated as:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:P(y=1|x)=\sum\limits_{m=1}^{5}\:{w}_{m}{f}_{m}(x).$$\end{document} Threshold optimization and decision criteria The ensemble model outputs continuous probability values\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:P(y=1|x)\in\:[0\mathrm{,}1]$$\end{document} derived from Eq. (8), which must be converted into binary classification decisions. This study determines the optimal threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\tau\:}^{*}$$\end{document} by identifying the probability cutoff that maximizes the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{1}$$\end{document} score on the validation set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{D}_{val}$$\end{document} :9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\tau\:}^{*}=\mathrm{a}\mathrm{r}\mathrm{g}\underset{\tau\:\in\:\left[\mathrm{0,1}\right]}{max}\:\frac{2\cdot\:P\left(\tau\:\right)\cdot\:R\left(\tau\:\right)}{P\left(\tau\:\right)+R\left(\tau\:\right)}.$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:P\left(\tau\:\right)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:R\left(\tau\:\right)$$\end{document} denote the precision and recall at threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\tau\:$$\end{document}, respectively. For a given sample \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{test}$$\end{document} in the final test set, let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{test}=P(y=1|{x}_{test})$$\end{document} be the ensemble probability. The discrimination logic is defined as:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widehat{y}_{test}=\left\{\begin{array}{ll}1\mathrm{,} & \text{if } P_{test} \geq \tau^* \\0\mathrm{,} & \text{if } P_{test} < \tau^*\end{array}\right.$$\end{document} SHAP interpretability analysis module Despite their accuracy, the “black-box” nature of ensemble models often limits their adoption in high-stakes fields like financial risk management. To address this, this study constructs a hierarchical weighted interpretability system based on Shapley additivity. The system reveals feature contributions at two granularities: the base learner level (micro-level) and the ensemble level (macro-level). Local feature attribution For each CatBoost base learner \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{m}\left(x\right)$$\end{document} with distinct structural constraints, the contribution of feature i to the model’s prediction is measured by its corresponding SHAP value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\phi\:}_{i,\:m}$$\end{document}. According to the Shapley definition from game theory, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\phi\:}_{i,\:m}$$\end{document} represents the average marginal contribution of feature i across all possible feature subsets:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\phi\:}_{i,\:m}\left(x\right)=\sum\limits_{S\subseteq\:\mathcal{F}\setminus\:\left\{i\right\}}\:\frac{\left|S\right|!(d-|S|-1)!}{d!}\left[{f}_{m,\:S\cup\:\left\{i\right\}}(x)-{f}_{m,\:S}(x)\right].$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{F}$$\end{document} is the complete set of features, d is the total dimension of features, S is a subset of features excluding feature i, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{m\mathrm{,}\text{}S}\left(x\right)$$\end{document} represents the expected prediction output of the model given only the feature subset S. Global decision synthesis Based on the linearity property of Shapley values, since the proposed ensemble model is a linear weighted combination of base learners, the contribution attribution of the ensemble is mathematically equivalent to the weighted average of the attribution values from each base learner. Consequently, utilizing the performance weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{m}$$\end{document} derived during the validation phase, the local SHAP values of the five base learners are linearly fused to obtain the comprehensive impact \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varPhi\:}_{i}$$\end{document} of feature i on the final ensemble output:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varPhi\:}_{i}^{ensemble}\left(x\right)=\sum\limits_{m=1}^{5}\:{w}_{m}\cdot\:{\phi\:}_{i,m}\left(x\right).$$\end{document} Furthermore, the global importance score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{I}_{i}$$\end{document} of feature i is defined as the mean absolute magnitude of its comprehensive SHAP values:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{I}_{i}=\frac{1}{N}\sum\limits_{j=1}^{N}\:\left|{\varPhi\:}_{i}^{ensemble}\right({x}_{j}\left)\right|.$$\end{document} This hierarchical design addresses the limitations of existing SHAP applications, which predominantly focus on single models and lack weighted aggregation methods for multi-model ensembles. When a feature contributes significantly to the final decision, it allows for traceability to determine which base learners dominated that specific judgment. Pre-deployment adaptive fine-tuning strategy In real-world business scenarios, the performance of a successfully pre-trained model may degrade when deployed in a new environment due to distribution shifts in customer demographics and economic indicators. To mitigate this, we use a pre-deployment adaptive fine-tuning (AFT) method. This strategy allows target institutions to calibrate the pre-trained ensemble model using a small amount of local historical labeled data\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{D}_{local}=\left\{\right({x}_{j},{\:y}_{j}){\}}_{j=1}^{M}$$\end{document}before the model is officially commissioned. This strategy is optional: if no local data is available, the pre-trained weights can be used directly for inference; if local historical data exists, this strategy enables low-cost adaptation. Target domain data preprocessing and imputation To address potential missing values in the target domain, a type-sensitive imputation protocol is executed. Let the original input features be \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}^{{\prime\:}}$$\end{document}, and the imputed features be \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:x={x}^{{\prime\:}}+ϵ$$\end{document}, where ϵ represents the estimation bias introduced by imputation. The imputation logic for each feature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{j}$$\end{document} is defined as:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{l}{x}_{j}=\left\{\begin{array}{lll}{\mu\:}_{j}=\frac{1}{N}\sum\:_{i=1}^{N}\:{x}_{ij}\mathrm{,}&\:if\:{x}_{j}^{{\prime\:}}\in\:{F}_{num}\text{ and is missing}&\:\\\:{mode}_{j}=arg\underset{v\in\:{V}_{j}}{max}\:\sum\:_{i=1}^{N}\:I({x}_{ij}=v)\mathrm{,}&\:if\:{x}_{j}^{{\prime\:}}\in\:{F}_{cat}\text{ and is missing}&\:\\\:{x}_{j}^{{\prime\:}}\mathrm{,}&\mathrm{otherwise}&\:\end{array}\right.\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{num}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{cat}$$\end{document} are the sets of numerical and categorical features, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{j}$$\end{document} is the mean of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{j}$$\end{document} in the training set, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:mod{e}_{j}$$\end{document} is the mode, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:I(\cdot\:)$$\end{document} is the indicator function. Knowledge retention fine-tuning via composite loss function To mitigate the risk of catastrophic forgetting on small-sample local data, the design of the composite loss function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{APT}$$\end{document} follows the principles of the regularization path and knowledge transfer. By incorporating the source domain empirical risk (the first term) as a structural constraint, the framework preserves global knowledge while allowing for efficient local distribution adaptation through the target domain risk term. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\theta\:$$\end{document} denote the model parameters, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:L(\cdot\:,\cdot\:)$$\end{document} the loss function, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\sim}{x}}_{j}$$\end{document} the imputed feature vector of the j-th local sample after applying the type-sensitive imputation protocol. Then:15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{APT}\left(\theta\:\right)=\beta\:\cdot\:\frac{1}{N}\sum\limits_{i=1}^{N}\:L\left(f\left({x}_{i}\text{; }\theta\:\right),{y}_{i}\right)+(1-\beta\:)\cdot\:\frac{1}{{N}_{Loc}}\sum\limits_{j=1}^{{N}_{Loc}}\:L\left(f\left({\stackrel{\sim}{x}}_{j}\text{; }\theta\:\right),{y}_{j}\right).$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:\in\:\left[0\mathrm{,}\text{}1\right]$$\end{document} is the balancing coefficient, typically set to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:\in\:\left[0.6\mathrm{,}\text{}0.8\right]$$\end{document} in practical applications to retain more global knowledge. Here, N is the number of source domain training samples, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{N}_{Loc}$$\end{document} is the number of local labeled samples. The first term represents the global empirical risk constraint, aiming to lock the general knowledge acquired by the pre-trained model. The second term represents the target domain risk constraint, aiming to capture domain-specific distributions. The fine-tuning process starts from the pre-trained optimal parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{*}$$\end{document} and achieves smooth parameter migration via iterative updates:16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t+1}={\theta\:}_{t}-{\eta\:}_{t}{\nabla\:}_{\theta\:}{\mathcal{L}}_{APT}\left({\theta\:}_{t}\right).$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\eta\:}_{t}$$\end{document} is the learning rate, set to a small fixed value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{10}^{-4}$$\end{document}. For the ensemble model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F\left(x\right)=\sum\:_{m=1}^{5}{w}_{m}{f}_{m}\left(x\mathrm{;}\:{\theta\:}_{m}\right)$$\end{document} the AFT process is executed independently for each base learner. While keeping the ensemble weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{m}$$\end{document} constant, an adapted ensemble decision system is generated:17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{APT}\left(x\right)=\sum\limits_{m=1}^{5}\:{w}_{m}{f}_{m}\left(x\mathrm{;}\text{}{\theta\:}_{m}^{APT}\right).$$\end{document} This strategy ensures that when the model is deployed to a new environment, it inherits the global discriminative power learned from the source domain while achieving low-cost and high-efficiency adaptation to the local data characteristics of the target institution. Results Datasets Two bank marketing datasets were employed for experimental validation in this study. The public Bank Marketing Dataset from a Portuguese bank was used for model pre-training and benchmark comparison. Additionally, the China A City Commercial Bank (CACB) dataset was collected for external validation to examine the model’s generalization capability and the effectiveness of fine-tuning strategies in the tested cross-institutional scenario. Figure 2 illustrates the class distribution of both datasets and key features of the CACB dataset. Bank marketing dataset The Bank Marketing Dataset was publicly donated by a Portuguese banking institution, covering the period from May 2008 to November 2010. It contains a total of 41,188 telemarketing records for term deposits. As shown in Fig. 2A, the dataset exhibits a significant class imbalance, with a ratio of subscribers to non-subscribers of approximately 1:8. The dataset comprises 21 features, encompassing customer profiles such as age, job, marital status (marital), and education level (education); credit status including credit in default (default), housing loan (housing), and personal loan (loan); marketing interactions such as contact communication type (contact), last contact month of year (month), last contact day of the week (day of week), last contact duration (duration), and number of contacts performed during this campaign (campaign); previous campaign attributes including number of days since last contact (pdays), number of contacts performed before this campaign (previous), and outcome of the previous marketing campaign (poutcome); and macroeconomic variables such as employment variation rate (emp.var.rate), consumer price index (cons.price.idx), consumer confidence index (cons.conf.idx), Euribor 3-month rate (euribor3m), and number of employees (nr.employed). China A City Commercial Bank Dataset The CACB dataset was obtained from a city commercial bank in China, totaling 8426 raw telemarketing records from the year 2024. To extract effective interaction patterns and minimize non-marketing noise, a rigorous cleaning protocol was implemented. We specifically excluded records with a call duration of less than 10 s, which represents invalid or failed contacts, and those with severe missingness of key information. Additionally, inactive accounts with no basic capital transactions in the past six months were removed to ensure the validity of historical behavioral features. This process yielded 1973 high-quality samples, prioritizing data relevance and predictive reliability over raw sample volume. As illustrated in Fig. 2B, the dataset consists of 196 subscribers and 1777 non-subscribers, resulting in a positive-to-negative sample ratio of approximately 1:9, exhibiting an imbalance characteristic similar to the Portuguese bank dataset. This dataset comprises 13 features, encompassing customer profiles (age, job, marital status, and education level) and marketing interaction information (contact type, call duration, previous marketing outcome, the number of days since last contact, etc.). Figure 2C and D illustrate several key features driving customer subscription in the CACB dataset. Based on the distribution of call durations in Fig. 2C, the median call duration of subscribers is significantly higher than that of non-subscribers, suggesting its potential as a strong discriminative feature for predicting customer purchase intention. According to the outcomes of the previous marketing campaign in Fig. 2D, the current conversion rate for customers with a successful historical marketing outcome reached as high as 69%, indicating that customer conversion behavior exhibits a distinct characteristic of continuity. Fig. 2Feature visualization analysis of the datasets. (A) Target variable distribution in the Bank Marketing Dataset; (B) Target variable distribution in the CACB dataset; (C) Association between call duration and subscription status in the CACB dataset; (D) Impact of poutcome on subscription rate in the CACB dataset. Feature visualization analysis of the datasets. (A) Target variable distribution in the Bank Marketing Dataset; (B) Target variable distribution in the CACB dataset; (C) Association between call duration and subscription status in the CACB dataset; (D) Impact of poutcome on subscription rate in the CACB dataset. Experimental environment and parameter configuration Experiments were conducted on a high-performance workstation equipped with an NVIDIA GeForce RTX 4090 GPU (24GB VRAM), with hardware acceleration implemented via CUDA 11.8. To ensure the statistical reliability of the evaluation, a stratified 5-fold cross-validation strategy was employed. For each fold, the dataset was partitioned into training, validation, and test sets according to an 8:1:1 ratio, dedicated to model fitting, hyperparameter optimization, and final performance assessment, respectively. To ensure a rigorous and fair comparison, all baseline models were standardized under a unified preprocessing framework. The specific configurations are as follows: Logistic Regression: Configured with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{2}$$\end{document} regularization and the L-BFGS optimization algorithm; the maximum number of iterations was set to T=1000, and inverse-frequency weighting was applied to balance loss contributions across classes. Ensemble and Tree-based Models: The Random Forest comprised 100 decision trees. The Decision Tree utilized the CART algorithm, with the maximum depth constrained to d=10 to prevent overfitting. For XGBoost, the learning rate was set to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:=0.1$$\end{document}, and the scale factor for positive/negative samples \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\rho\:\approx\:8.0$$\end{document} was dynamically calculated based on the training set to adjust misclassification penalties. K-Nearest Neighbors: Configured with k=5, neighbors using distance-based weighting. All baseline models adhered to their respective library-default convergence criteria. The random seed for all experiments was fixed at 42 to ensure reproducibility. Furthermore, beyond internal model weight adjustments, the Synthetic Minority Over-sampling Technique (SMOTE) was uniformly integrated to dynamically reconstruct the training sets, increasing the minority class proportion to over 30%. For the CBE-XAI framework, its five base learners were optimized with an initial learning rate of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:5\times\:{10}^{-2}$$\end{document} in full-batch mode. A patience value of 100 was set for early stopping, with a maximum of 2000 training epochs. Evaluation metrics To comprehensively evaluate the effectiveness of the CBE-XAI in predicting bank term deposit subscriptions, five key performance metrics were selected. Given the inherent class imbalance in bank marketing data, emphasis was placed on metrics with high sensitivity toward the minority class (positive samples). The specific definitions are as follows: (1) Area under the receiver operating characteristic curve The Area under the receiver operating characteristic curve (AUROC) measures the global classification performance across all possible decision thresholds. It is defined as the integral of the True Positive Rate ( TPR) against the False Positive Rate ( FPR):18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:AUROC={\int\:}_{0}^{1}\:TPR\left(FPR\right)d\left(FPR\right).$$\end{document} where TPR represents the model’s ability to capture positive instances, and FPR denotes the probability of false alarms for negative instances. An AUROC value closer to 1 indicates superior discriminative power in distinguishing between potential subscribers and no subscribers. (2) Precision and Recall In unbalanced data scenarios, a single metric often fails to capture performance on the minority class. Precision measures the proportion of true positives among all instances predicted as positive, while Recall (Sensitivity) quantifies the extent to which the model covers all actual positive instances:19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Precision=\frac{TP}{TP+FP}\mathrm{;}\:Recall=\frac{TP}{TP+FN}$$\end{document} where TP, TN, FP, and FN represent True Positives, True Negatives, False Positives, and False Negatives, respectively. (3) F1-Score The F1-score is the harmonic mean of Precision and Recall, providing a balanced assessment of the trade-off between precision and completeness. It serves as a core metric for evaluating imbalanced classification tasks:20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F1-Score=2\times\:\frac{Precision\times\:Recall}{Precision+Recall}.$$\end{document} (4) Specificity Specificity measures the model’s proficiency in identifying negative instances. It reflects the model’s performance in avoiding over-marketing and preserving the customer experience:21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Specificity=\frac{TN}{TN+FP}.$$\end{document} (5) Accuracy Accuracy reflects the overall proportion of correctly predicted instances:22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Accuracy=\frac{TP+TN}{TP+TN+FP+FN}.$$\end{document} (6) Statistical significance assessment To ensure the robustness of the experimental results, all aforementioned metrics are reported as the mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\bar{X}$$\end{document} across the 5-fold cross-validation, accompanied by their 95% Confidence Intervals (95% CI). The confidence interval is calculated as:23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$95\%\ \mathrm{CI} = \bar{X} \pm t_{\alpha/2,\ n-1} \times \frac{s}{\sqrt{n}}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{s}$$\end{document} is the standard deviation of the 5-fold results, and n=5 denotes the number of folds. For a 95% confidence level, the critical value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{t}_{0.025,\:4}$$\end{document} is 2.776. This statistical approach reveals the fluctuation range of model performance and validates its stability across different data subsets. Comparison of CBE-XAI with baseline methods To comprehensively evaluate the predictive performance of CBE-XAI, five classic machine learning models: Logistic Regression (LR), Random Forest (RF), Decision Tree (DT), K-Nearest Neighbors (KNN), and XGBoost were selected as benchmarks. All models were trained and evaluated under an identical experimental environment and preprocessing pipeline. Performance metrics are reported as the mean and 95% Confidence Interval (CI) derived from 5-fold cross-validation. Additionally, the DeLong test was employed to assess the statistical significance of the differences in AUROC between CBE-XAI and each baseline model. The performance comparison of the different models is summarized in Table 1. CBE-XAI achieved competitive results across four key metrics in this experimental setting: AUROC (0.949), Accuracy (0.877), F1-Score (0.621), and Sensitivity (0.899) (Fig. 3A, B). In comparison, the AUROC values for the baseline models were 0.943 for XGBoost and 0.942 for RF. The DeLong test results indicated that CBE-XAI significantly outperformed LR, DT, and KNN (p<0.001). However, its improvement over strong ensemble baselines like XGBoost (p=0.050) and RF (p=0.072) was marginally significant or not statistically significant. Nevertheless, CBE-XAI demonstrated more consistent performance in metrics sensitive to class imbalance, such as F1-Score and Sensitivity, suggesting its potential practical utility in mitigating the cost of missed conversion opportunities. In this study, Platt Scaling was utilized to calibrate the output probabilities of CBE-XAI. Post-calibration, the model yielded an Expected Calibration Error (ECE) of 0.008 and a Brier score of 0.056, indicating reasonable probability calibration performance (Fig. 3C). Furthermore, Decision Curve Analysis (DCA) (Fig. 3D) illustrated the potential practical utility of the model. In the bank marketing context, the model achieves the maximum net benefit at a threshold probability of 0.51. Within the decision threshold range of 0.3 to 0.7, its net benefit consistently outperformed the two extreme strategies: “Market All” and “Market None”. Table 1Performance comparison of CBE-XAI and various machine learning models on the Bank Marketing Dataset test set.MethodAUROC(95%CI)Accuracy(95%CI)F1-Score(95%CI)Sensitivity(95%CI)Specificity(95%CI)p-valueLR0.932(0.928–0.935)0.858(0.852–0.864)0.578(0.567–0.590)0.864(0.854–0.873)0.857(0.850–0.864)< 0.001RF0.942(0.936–0.949)0.874(0.845–0.899)0.618(0.597–0.639)0.664(0.644–0.684)0.905(0.897–0.924)0.072DT0.902(0.892–0.912)0.871(0.869–0.883)0.595(0.571–0.619)0.799(0.768–0.829)0.870(0.861–0.890)< 0.001KNN0.865(0.852–0.879)0.858(0.850–0.867)0.544(0.523–0.565)0.750(0.729–0.772)0.872(0.864–0.880)< 0.001XGBoost0.943(0.939–0.948)0.851(0.843–0.859)0.582(0.568–0.596)0.880 (0.870–0.910)0.841(0.833–0.850)0.050CBE-XAI0.949(0.945–0.953)0.877(0.871–0.882)0.621(0.612–0.631)0.899(0.882–0.916)0.874(0.863–0.882)N/AResults report the mean and 95% confidence interval from 5-fold cross-validation, with the best values highlighted in bold. The p-values (DeLong’s test) indicate the significance of the difference in AUROC between each model and CBE-XAI (N/A: not applicable). Table 1Performance comparison of CBE-XAI and various machine learning models on the Bank Marketing Dataset test set.MethodAUROC(95%CI)Accuracy(95%CI)F1-Score(95%CI)Sensitivity(95%CI)Specificity(95%CI)p-valueLR0.932(0.928–0.935)0.858(0.852–0.864)0.578(0.567–0.590)0.864(0.854–0.873)0.857(0.850–0.864)< 0.001RF0.942(0.936–0.949)0.874(0.845–0.899)0.618(0.597–0.639)0.664(0.644–0.684)0.905(0.897–0.924)0.072DT0.902(0.892–0.912)0.871(0.869–0.883)0.595(0.571–0.619)0.799(0.768–0.829)0.870(0.861–0.890)< 0.001KNN0.865(0.852–0.879)0.858(0.850–0.867)0.544(0.523–0.565)0.750(0.729–0.772)0.872(0.864–0.880)< 0.001XGBoost0.943(0.939–0.948)0.851(0.843–0.859)0.582(0.568–0.596)0.880 (0.870–0.910)0.841(0.833–0.850)0.050CBE-XAI0.949(0.945–0.953)0.877(0.871–0.882)0.621(0.612–0.631)0.899(0.882–0.916)0.874(0.863–0.882)N/AResults report the mean and 95% confidence interval from 5-fold cross-validation, with the best values highlighted in bold. The p-values (DeLong’s test) indicate the significance of the difference in AUROC between each model and CBE-XAI (N/A: not applicable). Performance comparison of CBE-XAI and various machine learning models on the Bank Marketing Dataset test set. 0.932 (0.928–0.935) 0.858 (0.852–0.864) 0.578 (0.567–0.590) 0.864 (0.854–0.873) 0.857 (0.850–0.864) 0.942 (0.936–0.949) 0.874 (0.845–0.899) 0.618 (0.597–0.639) 0.664 (0.644–0.684) 0.905 (0.897–0.924) 0.902 (0.892–0.912) 0.871 (0.869–0.883) 0.595 (0.571–0.619) 0.799 (0.768–0.829) 0.870 (0.861–0.890) 0.865 (0.852–0.879) 0.858 (0.850–0.867) 0.544 (0.523–0.565) 0.750 (0.729–0.772) 0.872 (0.864–0.880) 0.943 (0.939–0.948) 0.851 (0.843–0.859) 0.582 (0.568–0.596) 0.841 (0.833–0.850) 0.949 (0.945–0.953) 0.877 (0.871–0.882) 0.621 (0.612–0.631) 0.899 (0.882–0.916) 0.874 (0.863–0.882) Results report the mean and 95% confidence interval from 5-fold cross-validation, with the best values highlighted in bold. The p-values (DeLong’s test) indicate the significance of the difference in AUROC between each model and CBE-XAI (N/A: not applicable). Fig. 3Performance evaluation of the CBE-XAI model on the test set of the Bank Marketing dataset. (A) Precision-Recall curve; (B) Receiver operating characteristic curve; (C) Calibration curve; (D) Decision curve analysis curve. Performance evaluation of the CBE-XAI model on the test set of the Bank Marketing dataset. (A) Precision-Recall curve; (B) Receiver operating characteristic curve; (C) Calibration curve; (D) Decision curve analysis curve. Hierarchical SHAP interpretability analysis To unravel the internal decision-making mechanism of the CBE-XAI ensemble model, we construct a hierarchical weighted interpretability system based on Shapley values. Analyses were conducted at both the base learner and ensemble levels, as illustrated in Fig. 4. Figure 4A presents a heatmap of the normalized mean absolute SHAP values for the five heterogeneous CatBoost base learners across the top 15 original features. Distinct base learners exhibit significant variations in feature sensitivity. The Fast Model, proficient in capturing global prominent patterns, shows extreme sensitivity to the macroeconomic variable emp.var.rate (SHAP = 0.963), while also being sensitive to discrete features such as default and job. The Focal Loss Model, focusing on hard-to-classify samples, pays more attention to raw numerical signals, assigning higher weights to continuous features like duration and euribor3m. In contrast, the Balanced Weighted Model exhibits the lowest overall sensitivity, with most feature SHAP values trending toward zero, suggesting a potential decline in its ability to discriminate feature variances. The sensitivity distributions of the Deep Model and Conservative Model fall between these extremes. This divergence in feature sensitivity illustrates the complementary nature of the heterogeneous ensemble design. Figure 4B displays the global feature importance ranking after weighted aggregation. Based on the mean absolute SHAP values, the top five important features are duration (0.518), emp.var.rate (0.396), euribor3m (0.320), nr.employed (0.161), and cons.price.idx (0.156). Call duration ranks first, indicating that deeper communication between the bank and the customer—leading to a fuller understanding of the product—correlates with a higher subscription probability. The five macroeconomic variables follow closely, suggesting an influence of the external economic environment on customers’ term deposit decisions. Conversely, static attributes such as job and education rank lower, suggesting that in bank marketing scenarios, customer behavioral features and macroeconomic indicators possess greater predictive power than demographic characteristics. The SHAP beeswarm plot in Fig. 4C further reveals the direction of influence for each feature. The SHAP value distribution for duration is the widest, signifying its dominant impact on prediction results, with a strong positive correlation observed. Macroeconomic features such as emp.var.rate and euribor3m exhibit a negative impact on subscription intention. When the economic environment improves, the availability of high-yield wealth management products increases, and capital flows toward physical consumption. This reduces the attractiveness of term deposits. For campaign, high-value samples are largely dispersed in the negative zone, indicating that frequent contact with customers may trigger negative spillover effects. Features like loan and education have SHAP values concentrated near zero with narrow distributions, reflecting their relatively minor contributions to the prediction. Fig. 4Hierarchical feature interpretability analysis in CBE-XAI. (A) Feature sensitivity heatmap of the five base learners; (B) Global feature importance bar plot of CBE-XAI; (C) SHAP summary beeswarm plot of CBE-XAI. Hierarchical feature interpretability analysis in CBE-XAI. (A) Feature sensitivity heatmap of the five base learners; (B) Global feature importance bar plot of CBE-XAI; (C) SHAP summary beeswarm plot of CBE-XAI. To extend the explanation from collective patterns to individual decisions, six cases (three correct and three incorrect predictions) were randomly selected for SHAP force plot decomposition, as shown in Fig. 5. In the correct prediction cases (Fig. 5A), the CBE-XAI shows how multiple features jointly contribute to predictions. As seen in Case (3), the model does not rely solely on high call duration but infers the outcome by combining it with a downward macroeconomic cycle and convenient contact methods. In Case (2), even with a professional background in management, the customer faced challenges such as zero previous marketing successes. This illustrates that the weighted ensemble mechanism can reflect customer interaction patterns in heterogeneous environments. In the incorrect cases (Fig. 5B), prediction failures typically stem from the model’s misjudgment of term deposit decision drivers, reflecting limitations when handling atypical or extreme samples. In Case (3), an extreme call duration may produce a feature overflow effect, where its marginal contribution excessively overrides the negative impacts of features like education, resulting in a false positive. Furthermore, the model’s “stereotypes” regarding groups such as housemaid or admin in Cases (1) and (2) may lead it to overlook the subscription intentions inherent in features like education level. This effect occurs to some extent. Fig. 5Interpretation of individual prediction samples via SHAP force plots. (A) Three representative cases with correct predictions; (B) Three representative cases with incorrect predictions. Interpretation of individual prediction samples via SHAP force plots. (A) Three representative cases with correct predictions; (B) Three representative cases with incorrect predictions. Performance of CBE-XAI on the CACB dataset To verify the effectiveness of the adaptive fine-tuning strategy in addressing cross-environment distribution shifts, this study uses a real-world dataset from a CACB for external validation. This dataset differs from the Portuguese bank dataset in terms of time period, geographical distribution, and customer structure, making it ideal for simulating cross-environment model deployment scenarios. The CACB dataset was stratified into a fine-tuning set and a test set at a 7:3 ratio, containing 1380 and 593 records, respectively. All missing values were processed using type-sensitive imputation based on the statistics of the Portuguese training set. Three deployment strategies were compared: (A) Direct prediction, where the original pre-trained model predicts directly; (B) Local retraining, where a CatBoost model is retrained from scratch using the fine-tuning set (with an internal 8:2 split for training and validation sub-sets); and (C) CBE-XAI with adaptive fine-tuning (Fine-tuning), the strategy used in this paper. The validation results are presented in Table 2; Fig. 6A, B. According to Strategy (A), the original pre-trained model achieves an AUROC of 0.842 on the CACB test set, representing an 11.3% decline compared to its original performance on the Portuguese test set (AUROC 0.949). This confirms a significant distribution shift between the European and local Chinese banking customer bases. This shift is reflected in demographic attributes such as occupation and age. It is also reflected in how different macroeconomic environments influence the predictive logic of customer subscription behavior. With Strategy (B), the AUROC improved to 0.895 with an F1-Score of 0.564. However, constrained by the small local sample size and extreme class imbalance, this model underperformed in capturing minority class instances, yielding a Sensitivity of only 0.812. This suggests that relying solely on limited local historical data is insufficient to construct an accurate decision logic. In contrast, Strategy (C) proposed in this study achieved the best results among the three strategies across all evaluation metrics. Specifically, the AUROC reached 0.931, the F1-Score increased to 0.612, and Sensitivity reached 0.886—representing improvements of 10.6%, 27.0%, and 17.5%, respectively, over Strategy A. As shown in the calibration curves in Fig. 6C, the ECE for the three strategies was 0.035 for Direct Prediction, 0.017 for Local Retraining, and 0.028 for fine-tuning, indicating that the fine-tuned model shows competitive probability calibration in this setting. This improvement can be attributed to the synergy of three mechanisms: Knowledge Retention and Transfer: The pre-trained model inherits the robust decision logic embedded in the large-scale source domain data.Composite Loss Function: This allows for distributional calibration of model weights during lightweight fine-tuning with local labeled data.Type-sensitive Imputation Protocol: This effectively handles approximately 15% of natural missing values in the CACB dataset, ensuring high consistency between preprocessing logic and the inference phase. Knowledge Retention and Transfer: The pre-trained model inherits the robust decision logic embedded in the large-scale source domain data. Knowledge Retention and Transfer: The pre-trained model inherits the robust decision logic embedded in the large-scale source domain data. Composite Loss Function: This allows for distributional calibration of model weights during lightweight fine-tuning with local labeled data. Composite Loss Function: This allows for distributional calibration of model weights during lightweight fine-tuning with local labeled data. Type-sensitive Imputation Protocol: This effectively handles approximately 15% of natural missing values in the CACB dataset, ensuring high consistency between preprocessing logic and the inference phase. Type-sensitive Imputation Protocol: This effectively handles approximately 15% of natural missing values in the CACB dataset, ensuring high consistency between preprocessing logic and the inference phase. The external validation experiments suggest that the CBE-XAI is feasible for cross-environment deployment. The pre-deployment adaptive fine-tuning strategy effectively utilizes small amounts of local historical data to adapt pre-trained models to new environments, offering a potential technical reference for the implementation of cross-regional marketing models in the banking sector. Table 2Performance comparison of different deployment strategies on the CACB external dataset. Results report the mean and 95% confidence interval, with the best values highlighted in bold. The p-values (DeLong’s test) indicate the significance of the difference in AUROC between each strategy and the proposed fine-tuning strategy (N/A: not applicable).Deployment strategyAUROC (95% CI)Accuracy (95% CI)F1-Score (95% CI)Sensitivity (95% CI)Specificity (95% CI)p-valueDirect prediction0.842 (0.831–0.853)0.825 (0.810–0.840)0.482 (0.465–0.499)0.754 (0.730–0.778)0.832 (0.815–0.849)< 0.001Local retraining0.895 (0.886–0.904)0.865 (0.854–0.876)0.564 (0.548–0.580)0.812 (0.795–0.829)0.871 (0.858–0.884)< 0.001Fine-tuning0.931 (0.924–0.938)0.881 (0.875–0.887)0.612 (0.601–0.623)0.886 (0.872–0.900)0.882 (0.870–0.894)N/A Table 2Performance comparison of different deployment strategies on the CACB external dataset. Results report the mean and 95% confidence interval, with the best values highlighted in bold. The p-values (DeLong’s test) indicate the significance of the difference in AUROC between each strategy and the proposed fine-tuning strategy (N/A: not applicable).Deployment strategyAUROC (95% CI)Accuracy (95% CI)F1-Score (95% CI)Sensitivity (95% CI)Specificity (95% CI)p-valueDirect prediction0.842 (0.831–0.853)0.825 (0.810–0.840)0.482 (0.465–0.499)0.754 (0.730–0.778)0.832 (0.815–0.849)< 0.001Local retraining0.895 (0.886–0.904)0.865 (0.854–0.876)0.564 (0.548–0.580)0.812 (0.795–0.829)0.871 (0.858–0.884)< 0.001Fine-tuning0.931 (0.924–0.938)0.881 (0.875–0.887)0.612 (0.601–0.623)0.886 (0.872–0.900)0.882 (0.870–0.894)N/A Performance comparison of different deployment strategies on the CACB external dataset. Results report the mean and 95% confidence interval, with the best values highlighted in bold. The p-values (DeLong’s test) indicate the significance of the difference in AUROC between each strategy and the proposed fine-tuning strategy (N/A: not applicable). Fig. 6Performance comparison of different deployment strategies on the CACB dataset. (A) Grouped bar chart of three strategies across five evaluation metrics; (B) ROC curve comparison of three strategies; (C) Calibration curve comparison of three strategies. Performance comparison of different deployment strategies on the CACB dataset. (A) Grouped bar chart of three strategies across five evaluation metrics; (B) ROC curve comparison of three strategies; (C) Calibration curve comparison of three strategies. Subgroup analysis and prediction robustness assessment To evaluate the generalization capability and predictive stability of the CBE-XAI across diverse demographic characteristics and behavioral attributes, this study performed subgroup validation on the top five SHAP-identified features: occupation, age, number of previous marketing contacts, previous marketing outcomes, and education level. Table 3 summarizes the comprehensive performance metrics for each subgroup on the test set. The experimental results show that CBE-XAI achieves competitive discriminative performance across several subgroups, though with notable variations. The model performs particularly well among core customer segments with sufficient sample sizes and regular behavioral patterns. For instance, in the prime-age group (30–45 years), the model achieved its highest AUROC (0.954, 95% CI: 0.950–0.958), maintaining a balanced level of both Sensitivity and Specificity above 0.89. Furthermore, in occupational subgroups such as Skilled workers and High-income professionals, the AUROC reached 0.955 and 0.946, respectively. The narrow confidence intervals further suggest reasonable prediction consistency of the model when processing subgroups with distinct feature distributions. Despite the strong overall performance, performance degradation was observed in specific extreme or complex subgroups, revealing potential challenges in the prediction task. In the Elderly group (> 60 years), the AUROC declined to 0.827, with a notably lower Sensitivity (0.482), indicating a higher risk of “missed detections” regarding the subscription intentions of this demographic. This may be attributed to the unique response patterns of elderly populations to digital marketing and insufficient learning of the feature space due to sample sparsity in this subgroup. In the Positive history (successful previous subscription) and Multiple contacts subgroups, the model exhibited significant prediction bias, with Sensitivity dropping to 0.316 and 0.474, respectively. This phenomenon suggests that when customers have complex marketing histories, simple linear logic struggles to capture their multifaceted decision-making psychology. In certain scenarios, historical success may act as noise in the prediction, leading the model toward conservative estimations. These subgroup disparities provide practical insights for precision marketing in banking. The high accuracy of CBE-XAI among mainstream professional segments supports the feasibility of marketing automation. However, for the elderly and customers with complex historical interactions, model predictions should serve only as auxiliary references. It is recommended that in practical deployment, manual review mechanisms be introduced for these low-performance subgroups. Alternatively, further optimization of feature representation for minority subgroups through advanced ensemble learning or oversampling techniques could enhance the fairness and coverage of the model across the entire customer base. Table 3Subgroup prediction performance of CBE-XAI based on key features in the Bank Marketing dataset.FeatureSubgroupAUROC(95% CI)Accuracy(95% CI)F1-Score(95% CI)Sensitivity(95% CI)Specificity(95% CI)ageYoung adults(18–30)0.934 (0.926–0.942)0.641 (0.624–0.658)0.917 (0.894–0.939)0.830 (0.814–0.846)0.844 (0.831–0.856)Prime working age(30–45)0.954 (0.950–0.958)0.612 (0.604–0.620)0.889 (0.871–0.907)0.893 (0.885-0.900)0.892 (0.886–0.899)Mid career(45–60)0.953 (0.947–0.958)0.593 (0.572–0.614)0.881 (0.841–0.922)0.887 (0.880–0.894)0.886 (0.881–0.892)Senior(>60)0.827 (0.787–0.866)0.692 (0.644–0.741)0.947 (0.931–0.964)0.482 (0.418–0.546)0.666 (0.618–0.714)educationBasic education0.955 (0.949–0.961)0.606 (0.584–0.628)0.889 (0.845–0.932)0.900 (0.888–0.913)0.899 (0.889–0.910)Higher education0.938 (0.934–0.942)0.626 (0.610–0.642)0.903 (0.882–0.924)0.843 (0.826–0.861)0.851 (0.839–0.864)Secondary education0.952 (0.945–0.958)0.624 (0.612–0.635)0.897 (0.875–0.919)0.879 (0.872–0.885)0.881 (0.873–0.888)Unknown education0.945 (0.935–0.955)0.646 (0.596–0.695)0.935 (0.890–0.981)0.836 (0.805–0.867)0.850 (0.827–0.873)jobHigh income professional0.946 (0.941–0.952)0.629 (0.604–0.655)0.910 (0.892–0.929)0.858 (0.838–0.879)0.865 (0.848–0.882)Retired0.897 (0.876–0.919)0.662 (0.627–0.696)0.935 (0.903–0.967)0.701 (0.675–0.727)0.759 (0.736–0.783)Self employed0.948 (0.937–0.959)0.575 (0.533–0.617)0.877 (0.822–0.932)0.877 (0.851–0.903)0.877 (0.856–0.898)Service industry0.955 (0.948–0.961)0.597 (0.540–0.653)0.872 (0.832–0.912)0.902 (0.890–0.914)0.900 (0.887–0.913)Skilled worker0.955 (0.953–0.958)0.606 (0.593–0.618)0.877 (0.853–0.901)0.905 (0.900-0.909)0.902 (0.899–0.906)Unknown0.927 (0.909–0.945)0.611 (0.485–0.738)0.921 (0.832–1.011)0.860 (0.812–0.907)0.866 (0.820–0.913)Vulnerable groups0.915 (0.901–0.928)0.659 (0.625–0.694)0.928 (0.902–0.955)0.747 (0.718–0.776)0.787 (0.764–0.810)poutcomeNegative history0.913 (0.906–0.920)0.578 (0.571–0.585)0.886 (0.850–0.922)0.804 (0.780–0.828)0.816 (0.800-0.832)No history0.949 (0.946–0.951)0.581 (0.570–0.592)0.877 (0.855–0.899)0.890 (0.883–0.896)0.889 (0.884–0.893)Positive history0.795 (0.765–0.826)0.837 (0.812–0.862)0.986 (0.982–0.991)0.316 (0.272–0.359)0.751 (0.717–0.785)previousMultiple contacts0.870 (0.832–0.907)0.783 (0.764–0.802)0.983 (0.965-1.000)0.474 (0.444–0.504)0.728 (0.711–0.745)No history0.949 (0.946–0.951)0.581 (0.570–0.592)0.877 (0.855–0.899)0.890 (0.883–0.896)0.889 (0.884–0.893)Once contacted0.927 (0.920–0.935)0.681 (0.651–0.711)0.925 (0.910–0.941)0.788 (0.761–0.814)0.817 (0.798–0.835)Results report the mean and 95% confidence interval from 5-fold cross-validation. Table 3Subgroup prediction performance of CBE-XAI based on key features in the Bank Marketing dataset.FeatureSubgroupAUROC(95% CI)Accuracy(95% CI)F1-Score(95% CI)Sensitivity(95% CI)Specificity(95% CI)ageYoung adults(18–30)0.934 (0.926–0.942)0.641 (0.624–0.658)0.917 (0.894–0.939)0.830 (0.814–0.846)0.844 (0.831–0.856)Prime working age(30–45)0.954 (0.950–0.958)0.612 (0.604–0.620)0.889 (0.871–0.907)0.893 (0.885-0.900)0.892 (0.886–0.899)Mid career(45–60)0.953 (0.947–0.958)0.593 (0.572–0.614)0.881 (0.841–0.922)0.887 (0.880–0.894)0.886 (0.881–0.892)Senior(>60)0.827 (0.787–0.866)0.692 (0.644–0.741)0.947 (0.931–0.964)0.482 (0.418–0.546)0.666 (0.618–0.714)educationBasic education0.955 (0.949–0.961)0.606 (0.584–0.628)0.889 (0.845–0.932)0.900 (0.888–0.913)0.899 (0.889–0.910)Higher education0.938 (0.934–0.942)0.626 (0.610–0.642)0.903 (0.882–0.924)0.843 (0.826–0.861)0.851 (0.839–0.864)Secondary education0.952 (0.945–0.958)0.624 (0.612–0.635)0.897 (0.875–0.919)0.879 (0.872–0.885)0.881 (0.873–0.888)Unknown education0.945 (0.935–0.955)0.646 (0.596–0.695)0.935 (0.890–0.981)0.836 (0.805–0.867)0.850 (0.827–0.873)jobHigh income professional0.946 (0.941–0.952)0.629 (0.604–0.655)0.910 (0.892–0.929)0.858 (0.838–0.879)0.865 (0.848–0.882)Retired0.897 (0.876–0.919)0.662 (0.627–0.696)0.935 (0.903–0.967)0.701 (0.675–0.727)0.759 (0.736–0.783)Self employed0.948 (0.937–0.959)0.575 (0.533–0.617)0.877 (0.822–0.932)0.877 (0.851–0.903)0.877 (0.856–0.898)Service industry0.955 (0.948–0.961)0.597 (0.540–0.653)0.872 (0.832–0.912)0.902 (0.890–0.914)0.900 (0.887–0.913)Skilled worker0.955 (0.953–0.958)0.606 (0.593–0.618)0.877 (0.853–0.901)0.905 (0.900-0.909)0.902 (0.899–0.906)Unknown0.927 (0.909–0.945)0.611 (0.485–0.738)0.921 (0.832–1.011)0.860 (0.812–0.907)0.866 (0.820–0.913)Vulnerable groups0.915 (0.901–0.928)0.659 (0.625–0.694)0.928 (0.902–0.955)0.747 (0.718–0.776)0.787 (0.764–0.810)poutcomeNegative history0.913 (0.906–0.920)0.578 (0.571–0.585)0.886 (0.850–0.922)0.804 (0.780–0.828)0.816 (0.800-0.832)No history0.949 (0.946–0.951)0.581 (0.570–0.592)0.877 (0.855–0.899)0.890 (0.883–0.896)0.889 (0.884–0.893)Positive history0.795 (0.765–0.826)0.837 (0.812–0.862)0.986 (0.982–0.991)0.316 (0.272–0.359)0.751 (0.717–0.785)previousMultiple contacts0.870 (0.832–0.907)0.783 (0.764–0.802)0.983 (0.965-1.000)0.474 (0.444–0.504)0.728 (0.711–0.745)No history0.949 (0.946–0.951)0.581 (0.570–0.592)0.877 (0.855–0.899)0.890 (0.883–0.896)0.889 (0.884–0.893)Once contacted0.927 (0.920–0.935)0.681 (0.651–0.711)0.925 (0.910–0.941)0.788 (0.761–0.814)0.817 (0.798–0.835)Results report the mean and 95% confidence interval from 5-fold cross-validation. Subgroup prediction performance of CBE-XAI based on key features in the Bank Marketing dataset. Young adults (18–30) Mid career (45–60) Senior (>60) Results report the mean and 95% confidence interval from 5-fold cross-validation. Ablation study of base learners and the CBE-XAI ensemble To evaluate the specific contributions of each component within the CBE-XAI, five base learners were used to represent model variants after removing specific strategies. The experimental results are summarized in Table 4. Among the individual base learners, the Deep Model achieved the highest performance, with an AUROC of 0.938, an Accuracy of 0.868, and an F1-Score of 0.603. Its performance among the five single strategies suggests that hierarchical structures help capture feature interactions. The Focal Loss Model followed, yielding an AUROC of 0.934 and a notably high Sensitivity of 0.912, demonstrating that the focal loss function effectively enhances the model’s ability to identify minority class samples. The AUROCs for the Balanced Weighted Model and the Conservative Model were 0.928 and 0.920, respectively, representing moderate performance levels. The Fast Model exhibited the lowest performance, with an AUROC of only 0.910 and an Accuracy of 0.842, reflecting how an overemphasis on training efficiency can constrain the model’s representational capacity. Compared to any single strategy, the CBE-XAI ensemble model achieved the highest values across three core metrics: AUROC (0.949), Accuracy (0.877), and F1-Score (0.621). Although the Sensitivity of CBE-XAI (0.899) was slightly lower than that of the Focal Loss Model (0.912), its Specificity (0.874) outperformed the latter (0.860). This indicates that the ensemble model achieves a different balance between the recognition capabilities for both positive and negative classes compared to individual learners. In summary, the five heterogeneous base learners exhibit complementary patterns in the feature space. Through post-training performance-aware weighting, CBE-XAI achieves better results than individual strategies in this experimental setting, suggesting potential benefits of the multi-strategy design. Table 4Performance comparison of individual base learners versus the CBE-XAI ensemble.Base learnerAUROC (95% CI)Accuracy (95% CI)F1-Score (95% CI)Sensitivity (95% CI)Specificity(95% CI)p-valueFocal Loss Model0.934 (0.929–0.939)0.863 (0.857–0.869)0.591 (0.581–0.601)0.912 (0.894–0.930)0.860 (0.851–0.869)< 0.001Balanced Weighted Model0.928 (0.923–0.933)0.859 (0.853–0.865)0.583 (0.573–0.593)0.901 (0.882–0.920)0.857 (0.848–0.866)< 0.001Conservative Model0.920 (0.915–0.925)0.851 (0.845–0.857)0.570 (0.560–0.580)0.889 (0.870–0.908)0.849 (0.840–0.858)< 0.001Fast Model0.910 (0.905–0.915)0.842 (0.836–0.848)0.552 (0.542–0.562)0.875 (0.856–0.894)0.840 (0.831–0.849)< 0.001Deep Model0.938 (0.933–0.943)0.868 (0.862–0.874)0.603 (0.593–0.613)0.908 (0.890–0.926)0.866 (0.857–0.875)0.018CBE-XAI0.949 (0.945–0.953)0.877 (0.871–0.882)0.621 (0.612–0.631)0.899 (0.882–0.916) 0.874 (0.863–0.882) N/AResults report the mean and 95% confidence interval, with the best values highlighted in bold. The p-values (DeLong’s test) indicate the significance of the difference in AUROC between each base learner and CBE-XAI (N/A: not applicable). Table 4Performance comparison of individual base learners versus the CBE-XAI ensemble.Base learnerAUROC (95% CI)Accuracy (95% CI)F1-Score (95% CI)Sensitivity (95% CI)Specificity(95% CI)p-valueFocal Loss Model0.934 (0.929–0.939)0.863 (0.857–0.869)0.591 (0.581–0.601)0.912 (0.894–0.930)0.860 (0.851–0.869)< 0.001Balanced Weighted Model0.928 (0.923–0.933)0.859 (0.853–0.865)0.583 (0.573–0.593)0.901 (0.882–0.920)0.857 (0.848–0.866)< 0.001Conservative Model0.920 (0.915–0.925)0.851 (0.845–0.857)0.570 (0.560–0.580)0.889 (0.870–0.908)0.849 (0.840–0.858)< 0.001Fast Model0.910 (0.905–0.915)0.842 (0.836–0.848)0.552 (0.542–0.562)0.875 (0.856–0.894)0.840 (0.831–0.849)< 0.001Deep Model0.938 (0.933–0.943)0.868 (0.862–0.874)0.603 (0.593–0.613)0.908 (0.890–0.926)0.866 (0.857–0.875)0.018CBE-XAI0.949 (0.945–0.953)0.877 (0.871–0.882)0.621 (0.612–0.631)0.899 (0.882–0.916) 0.874 (0.863–0.882) N/AResults report the mean and 95% confidence interval, with the best values highlighted in bold. The p-values (DeLong’s test) indicate the significance of the difference in AUROC between each base learner and CBE-XAI (N/A: not applicable). Performance comparison of individual base learners versus the CBE-XAI ensemble. Results report the mean and 95% confidence interval, with the best values highlighted in bold. The p-values (DeLong’s test) indicate the significance of the difference in AUROC between each base learner and CBE-XAI (N/A: not applicable). Conclusion This study has developed and validated a bank marketing prediction method based on ensemble learning and explainable AI. Experimental results show that the method is practically useful for handling highly imbalanced marketing data. On the Portuguese bank public dataset, CBE-XAI achieves an AUROC of 0.949 and an F1-score of 0.621. On the CACB real-world dataset, the adaptive fine-tuning strategy effectively restores predictive performance to 0.931. This demonstrates the potential for cross-regional domain adaptation. In addition, hierarchical SHAP analysis provides interpretability support for model decisions. In summary, by integrating existing mature techniques, CBE-XAI offers a feasible technical reference path for digital marketing in banking. Nevertheless, this study has several limitations. First, methodological contribution. The core of this work is the engineering integration and empirical validation of existing techniques, rather than the proposal of fundamentally new machine learning theory. Therefore, the value of this paper lies primarily in practical deployability, interpretability, and cross-environment robustness, rather than in a methodological breakthrough. Second, experimental comparison. This study mainly focuses on widely recognized ensemble models such as Random Forest and XGBoost. Although the improvement in AUROC over XGBoost is only marginally significant from a statistical perspective (p = 0.050), CBE-XAI shows relatively better performance on business metrics directly related to marketing costs and conversion rates, such as F1-score and sensitivity. In addition, practical deployment places high demands on model lightweighting and maintainability. Therefore, this study does not include computationally expensive deep tabular learning models (e.g., TabNet, FT-Transformer) as baselines. Third, explainability. The high contribution of the “duration” feature identified by the model reflects its role as an important indicator of customer engagement and intention depth. Due to the post-hoc nature of this explanation, the business value of this feature is mainly reflected in two aspects: post-call attribution analysis and real-time interaction monitoring. For pre-call customer screening scenarios where call duration is not yet available, the model can still provide effective classification guidance based on prior features such as customer profiles and macroeconomic indicators. If pure pre-call prediction is required in practice, the duration feature should be removed and the performance re-evaluated. Fourth, external validation. Approximately 75% of the raw records in the CACB dataset were excluded during preprocessing. This quality control operation may introduce selection bias. Furthermore, the current external validation is limited to a single institution and is based on cross-sectional data, failing to capture dynamic changes across time and institutions. To address the above limitations, future research will focus on the following five directions. (1) Introduce bootstrap and multiple comparison corrections to perform more rigorous statistical testing of model performance. (2) Conduct ablation experiments without the duration feature to establish a performance baseline for pure pre-call prediction scenarios. (3) Collect cross-regional, multi-period, and multi-center dynamic datasets to validate model robustness against temporal drift and cross-institution variations. (4) Explore the integration of causal inference methods with XAI. This would go beyond correlation analysis and gain a deeper understanding of the logic behind customer decisions. (5) Evaluate real-time deployment performance and business value in controlled production environments to assess the model’s feasibility in actual operations.