An explainable hybrid deep learning framework for multi-source student employment prediction Authors: Huanhuan Zheng, Jiayang Wu, Yingjiao Zhang Published: 2026 ABSTRACT Accurately predicting student employment outcomes remains a significant challenge in educational data mining, particularly given the increasing diversity of student backgrounds and the dynamic nature of labor market demands. This study proposes an explainable hybrid deep learning framework that integrates multi-source heterogeneous data, including academic records, demographic profiles, financial attributes, and engagement in scientific or organizational activities, to perform multi-class employment prediction. The framework employs recursive feature elimination for precise feature selection, a bi-directional long short-term memory network with attention mechanisms to capture temporal academic patterns, and a tree-structured Parzen estimator-optimized XGBoost classifier to model complex feature interactions. To enhance model interpretability, SHAP values are utilized to quantify the contribution of each feature to the final prediction. Extensive experiments on a real-world vocational college dataset demonstrate that the proposed model consistently surpasses competitive baselines across multiple evaluation metrics, including accuracy, macro-F1, AUC, and Cohen’s kappa. The results confirm the framework’s capability to effectively leverage both longitudinal educational trajectories and static characteristics for accurate and interpretable prediction of employment outcomes. Introduction With the ongoing fluctuations in the global economy, the employment situation for graduates of higher vocational colleges has become an increasingly pressing issue for educational institutions and policymakers alike. In this context, higher vocational colleges–serving as critical training grounds for applied and skilled talents–are confronted with the urgent challenge of how to align educational outputs with labor market demands. Bridging the persistent gap between academic training and real-world requirements has thus emerged as a core problem that necessitates timely and effective solutions1. Recent advances in educational data mining (EDM) offer promising tools to tackle this challenge. By extracting actionable patterns from educational datasets, EDM can not only optimize student learning outcomes but also significantly reduce both students’ learning costs and instructors’ teaching burdens2. Beyond academic performance prediction, these data-driven technologies facilitate proactive interventions in student development pathways, enabling institutions to anticipate and improve employment outcomes through informed decision-making3. Among the various student-related outcomes, predicting employment destinations holds particular importance. This task is inherently a multi-class classification problem, where each student must be accurately categorized into one of four distinct outcomes: pursuing further studies domestically, gaining successful employment, remaining unemployed, or going abroad. Effectively addressing this problem requires integrating multi-source heterogeneous information (e.g., academic records, scholarship history, demographic background) and modeling sequential patterns such as GPA trajectories across semesters. To this end, various methods have been proposed. For instance, Tan4 employed a grey forecasting model enhanced by an automated weighting optimization mechanism to predict college students’ entrepreneurial and employment trajectories. A six-dimensional variable system was constructed, incorporating academic, psychological, and physical development indicators. Dong5 leveraged big data technologies to simulate individual employment behavior, incorporating contextual variables such as living expenses and personal interests, and utilizing edge-fog computing to associate individuals into predictive clusters. Similarly, Gemilang et al.6 explored the use of machine learning algorithms in employment readiness assessments and found that the XGBoost model significantly outperformed traditional multinomial logistic regression approaches in predictive accuracy. However, existing studies exhibit several notable limitations: first, sequential features–particularly semester-level academic progression–are often underutilized or treated simplistically; second, model interpretability, which is crucial for educational applications, is frequently overlooked; and third, the integration of advanced machine learning techniques tailored to higher vocational education contexts remains insufficient. To address these challenges, this paper makes the following contributions:We formulate student employment prediction as a heterogeneous static–temporal fusion problem and construct a unified framework that jointly models longitudinal GPA trajectories and multi-source static educational attributes. Unlike existing approaches that primarily rely on either static tabular data or sequential academic records, the proposed framework explicitly captures their complementary predictive information.We design a hierarchical fusion architecture in which attention-enhanced temporal representations are integrated with RFE-selected educational features and subsequently optimized through a TPE-guided gradient boosting classifier. This design enables effective interaction learning between academic progression patterns and student background characteristics.We provide a comprehensive empirical and explainability analysis on a large-scale educational dataset, demonstrating both predictive effectiveness and interpretable decision mechanisms through SHAP-based feature attribution. We formulate student employment prediction as a heterogeneous static–temporal fusion problem and construct a unified framework that jointly models longitudinal GPA trajectories and multi-source static educational attributes. Unlike existing approaches that primarily rely on either static tabular data or sequential academic records, the proposed framework explicitly captures their complementary predictive information. We formulate student employment prediction as a heterogeneous static–temporal fusion problem and construct a unified framework that jointly models longitudinal GPA trajectories and multi-source static educational attributes. Unlike existing approaches that primarily rely on either static tabular data or sequential academic records, the proposed framework explicitly captures their complementary predictive information. We design a hierarchical fusion architecture in which attention-enhanced temporal representations are integrated with RFE-selected educational features and subsequently optimized through a TPE-guided gradient boosting classifier. This design enables effective interaction learning between academic progression patterns and student background characteristics. We design a hierarchical fusion architecture in which attention-enhanced temporal representations are integrated with RFE-selected educational features and subsequently optimized through a TPE-guided gradient boosting classifier. This design enables effective interaction learning between academic progression patterns and student background characteristics. We provide a comprehensive empirical and explainability analysis on a large-scale educational dataset, demonstrating both predictive effectiveness and interpretable decision mechanisms through SHAP-based feature attribution. We provide a comprehensive empirical and explainability analysis on a large-scale educational dataset, demonstrating both predictive effectiveness and interpretable decision mechanisms through SHAP-based feature attribution. Related works The rapid advancement of educational data mining and machine learning technologies has paved new technical pathways for accurately predicting students’ employment status, thereby enhancing the quality of talent cultivation and the alignment between education and employment. In the field of employment prediction for college students, existing research has explored various methodologies, including the optimization of grey models based on multi-dimensional variable systems, big data cluster simulations, and performance validation of ensemble learning algorithms such as XGBoost. However, specialized research focusing specifically on employment prediction for higher vocational graduates remains relatively scarce. Fortunately, the key technologies supporting such predictions are becoming increasingly mature, laying a solid foundation for in-depth exploration of employment prediction in the higher vocational education sector. Among the core techniques underpinning the construction of predictive models, feature selection plays a pivotal role in identifying key influencing factors. Recursive Feature Elimination (RFE), recognized for its ability to automatically screen for optimal feature subsets, achieves this by recursively removing less important features. However, its implementation, particularly Support Vector Machine based RFE (SVM-RFE), often faces challenges related to model selection difficulties and the complexity of generalization error estimation. In response to this, Ding et al.7 proposed an approximation method for evaluating the generalization error in linear SVM-RFE, designed a new parameter tuning criterion, and introduced multiple alpha-seeding techniques to significantly reduce computational costs. The effectiveness and broad applicability of RFE have been validated across multiple domains. For instance, in turbocharger rotor fault diagnosis, Chunyu et al.8 combined improved Hu invariant moments with SVM-RFE for feature ranking and selection; the resultant feature subset, used with a probabilistic neural network, enabled highly accurate and time-efficient fault classification. In recyclable container classification tasks, Ba et al.9 applied RFE based on LightGBM (RFE-LightGBM). By recursively eliminating low-importance features, classification accuracy was significantly enhanced to 95%, outperforming results using the original feature set or PCA. Addressing the computational efficiency of RFE in hyperspectral image classification, Song et al.10 proposed the ReliefF-RFE method. This approach first employs ReliefF for preliminary screening to narrow down the feature space, followed by RFE for recursive elimination within the subset. This method maintained classification accuracy while reducing the average runtime to just 75% of standard RFE’s duration. These mature RFE applications provide strong support and valuable insights for its use in screening key influencing factors and enhancing model performance (including accuracy and generalization ability) within vocational graduate employment prediction models. Simultaneously, deep learning architectures have demonstrated significant potential for modeling temporal data problems characteristic of student development. Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) along with their variants (e.g., BiGRU - Bidirectional Gated Recurrent Units) have proven effective in capturing long-term dependencies within sequential data like academic performance, thereby optimizing predictive performance. Adefemi et al.11 developed an architecture fusing CNN and BiGRU, leveraging CNN for local feature extraction and BiGRU to process long sequential information, effectively enhancing academic performance prediction capabilities. Sharif et al.12 explored novel applications of deep reinforcement learning combined with multimodal data integration in educational analytics; their case study confirmed the framework’s effectiveness in providing personalized feedback. Mou et al.13 applied deep learning based on Autoencoders to analyze the status and predict outcomes in college sports psychology teaching, constructing models that offer technical support for teaching reforms. Notably, a key limitation of deep learning applications is their frequent reliance on large amounts of labeled data and substantial computational resources. Yong et al.14 addressed this from a resource-sharing perspective, proposing a collective intelligence-based deep learning solution for education, offering a promising approach to the challenges of data collection and resource utilization. Ultimately, Liyanage et al.15 integrated diverse techniques including logistic regression, support vector machines, ARIMA time series forecasting, deep neural networks, and RNNs, applying them to optimize online education. This framework evolved into an integrated Learning Management System (LMS), serving as a central hub for both students and instructors. As the complexity of predictive models increases, explainability becomes a critical requirement. Attribution methods like SHAP (SHapley Additive exPlanations) have consequently been introduced into the educational prediction domain to quantify feature contributions and provide transparent bases for decision-making. Representative research includes: Ujkani et al.16 utilizing SHAP to explain student success prediction models based on Open University datasets, identifying at-risk students to enable targeted interventions by educators; Kar et al.17,within a Technology-Enhanced Learning (TEL) context, employing XAI (eXplainable AI) algorithms like SHAP, LIME, and FAMeX to analyze key factors influencing online adaptation across different educational levels, revealing counter-intuitive insights; Afrin et al.18 using SHAP to investigate models predicting learning outcomes based on social media, identifying sensitive features model reliance to scrutinize potential fairness issues; Hu et al.19, in digital reading research, applying SHAP after SVM had identified key contextual factors to analyze individual factor contributions, guiding the formulation of personalized teaching strategies. This body of work demonstrates that SHAP not only enhances model transparency but also facilitates the extraction of deep educational insights from result interpretation.Table 1Technical Characteristics of Representative Educational Prediction Models.AuthorRFEDLSeqModMultimodalEffOptFusionStatic–Temporal IntegrationExplainabilitySHAPDing7\checkmark\checkmarkChunyu8\checkmark\checkmarkBa9\checkmark\checkmark\checkmarkSong10\checkmark\checkmark\checkmarkAdefemi11\checkmark\checkmark\checkmarkSharif12\checkmark\checkmark\checkmarkMou13\checkmark\checkmarkYong14\checkmark\checkmarkLiyanage15\checkmark\checkmark\checkmark\checkmark\checkmark\checkmark\checkmark\checkmarkUjkani16\checkmark\checkmarkKar17\checkmark\checkmarkAfrin18\checkmark\checkmarkHu19\checkmark\checkmarkOurs\checkmark\checkmark\checkmark\checkmark\checkmark\checkmark\checkmark\checkmark\checkmarkNote: Abbreviations used in the header row – RFE = Recursive Feature Elimination; DL = Deep Learning; SeqMod = Sequence Modeling (e.g., RNN, BiGRU); Multimodal = Multimodal Data Fusion; EffOpt = Efficiency Optimization (e.g., runtime, tuning); Fusion = Architecture or Feature Fusion; Explainability = Explainable Modeling; SHAP = SHapley Additive exPlanations used. Table 1Technical Characteristics of Representative Educational Prediction Models.AuthorRFEDLSeqModMultimodalEffOptFusionStatic–Temporal IntegrationExplainabilitySHAPDing7\checkmark\checkmarkChunyu8\checkmark\checkmarkBa9\checkmark\checkmark\checkmarkSong10\checkmark\checkmark\checkmarkAdefemi11\checkmark\checkmark\checkmarkSharif12\checkmark\checkmark\checkmarkMou13\checkmark\checkmarkYong14\checkmark\checkmarkLiyanage15\checkmark\checkmark\checkmark\checkmark\checkmark\checkmark\checkmark\checkmarkUjkani16\checkmark\checkmarkKar17\checkmark\checkmarkAfrin18\checkmark\checkmarkHu19\checkmark\checkmarkOurs\checkmark\checkmark\checkmark\checkmark\checkmark\checkmark\checkmark\checkmark\checkmarkNote: Abbreviations used in the header row – RFE = Recursive Feature Elimination; DL = Deep Learning; SeqMod = Sequence Modeling (e.g., RNN, BiGRU); Multimodal = Multimodal Data Fusion; EffOpt = Efficiency Optimization (e.g., runtime, tuning); Fusion = Architecture or Feature Fusion; Explainability = Explainable Modeling; SHAP = SHapley Additive exPlanations used. Technical Characteristics of Representative Educational Prediction Models. Note: Abbreviations used in the header row – RFE = Recursive Feature Elimination; DL = Deep Learning; SeqMod = Sequence Modeling (e.g., RNN, BiGRU); Multimodal = Multimodal Data Fusion; EffOpt = Efficiency Optimization (e.g., runtime, tuning); Fusion = Architecture or Feature Fusion; Explainability = Explainable Modeling; SHAP = SHapley Additive exPlanations used. Although existing studies have explored feature selection, deep learning, sequence modeling, and explainable learning independently, several limitations remain. Most educational prediction approaches focus either on static student attributes or on temporal academic records, while the interaction between these two complementary information sources is rarely modeled explicitly. In addition, existing hybrid frameworks generally emphasize algorithmic combination rather than the joint exploitation of longitudinal academic evolution and heterogeneous educational characteristics. As shown in Table 1, only a limited number of studies simultaneously consider temporal modeling, feature fusion, and interpretability within a unified framework. Therefore, this study formulates student employment prediction as a static–temporal heterogeneous fusion problem and proposes an explainable hybrid architecture that jointly captures semester-level academic trajectories and multi-source student attributes. The methodological contribution lies not in introducing a new standalone learning component, but in establishing an integrated prediction framework that effectively leverages the complementary strengths of temporal and static educational information. Proposed method Overall framework To address the complex challenge of predicting student employment outcomes, this study introduces a comprehensive hybrid deep learning framework that integrates multi-source data with advanced temporal modeling. The framework is designed to overcome the limitations of static-feature-based approaches by capturing the dynamic nature of students’ educational trajectories, thereby improving both prediction accuracy and interpretability. It consists of four key components: multi-source data integration and preprocessing, feature selection and engineering, hybrid model construction, and interpretable prediction output. The data preprocessing stage involves aligning heterogeneous datasets–including semester-wise GPA, demographic information, financial aid records, and categorical descriptors–using unified identifiers and temporal ordering, followed by standard procedures such as missing value imputation, outlier removal, z-score normalization, and one-hot encoding. A recursive feature elimination (RFE) algorithm is employed to select a core set of 13 informative features, enhancing model efficiency and generalizability. For classification, a hybrid architecture is constructed by combining static and sequential data: static features are processed by a TPE-optimized XGBoost classifier, while semester-wise GPA sequences are encoded using a BiLSTM network with an attention mechanism that highlights salient academic periods. The resulting temporal embeddings are fused with static features and fed into the final classifier to predict one of four employment outcomes–domestic further education, successful employment, unemployment, or overseas advancement. To ensure interpretability, SHAP (SHapley Additive exPlanations) values are applied to attribute prediction outcomes to specific features, enabling transparent decision-making and targeted educational interventions. Collectively, this framework offers a robust, data-driven solution for employment prediction, balancing model complexity with interpretability and practical applicability. Multi-source data integration and feature engineering In this study, we formalize the multi-source data integration and feature engineering pipeline as follows. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {D}$$\end{document} denote the raw dataset comprising n students:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathscr {D} = \left\{ (x_i, y_i) \right\} _{i=1}^{n} \end{aligned}$$\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} is the concatenated feature vector of student i before preprocessing, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_i \in \{1, 2, 3, 4\}$$\end{document} corresponds to one of the four employment categories: 1 = domestic further education, 2 = successful employment, 3 = unemployment, 4 = overseas advancement. For each numerical feature j that is missing, we employ mean imputation:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} x_{ij} \leftarrow \mu _j \end{aligned}$$\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}$$x_{ij}$$\end{document} is the j-th feature of student i, and \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 empirical mean of feature j. Define the standardized score (z-score) for each entry:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} z_{ij} = \frac{x_{ij} - \mu _j}{\sigma _j} \end{aligned}$$\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}$$\mu _j$$\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}$$\sigma _j$$\end{document} are the mean and standard deviation of feature j after imputation. Observations satisfying \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|z_{ij}|> \tau$$\end{document} (with threshold τ = 3) are considered outliers and are excluded from subsequent analysis. All numerical features are normalized via z-score scaling:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \tilde{x}_{ij} = \frac{x_{ij} - \mu _j}{\sigma _j} \end{aligned}$$\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}$$\tilde{x}_{ij}$$\end{document} is the normalized value; this ensures zero mean and unit variance for feature j. For each categorical feature k taking values in a finite set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {C}_k$$\end{document}, we apply one-hot encoding:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \phi _k(c) = L_{\{c = v\}} \quad \text {for } v \in \mathscr {C}_k \end{aligned}$$\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}$$L_{\{\cdot \}}$$\end{document} is the indicator function, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|\mathscr {C}_k| = m_k$$\end{document}. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {F}$$\end{document} denote the full set of preprocessed features (after encoding), where D is the new dimensionality. We iteratively train an estimator f (here, XGBoost) and compute an importance score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s_j$$\end{document} 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}$$j \in \mathscr {F}$$\end{document}. At each iteration t, the least important feature6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} j^* = \arg \min _{j \in \mathscr {F}_t} s_j \end{aligned}$$\end{document}is removed, yielding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {F}_{t+1} = \mathscr {F}_t \setminus \{j^*\}$$\end{document}. The process continues until t = T, where T is the final iteration and the resulting \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {F}_T$$\end{document} contains the selected feature indices. Through the above steps–with rigorous symbol definitions for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {D}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {F}$$\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}$$\mathscr {F}_T$$\end{document}–we obtain a compact, high-quality feature set that serves as the foundation for the hybrid deep learning and tree-based predictive model. The overall preprocessing procedure is summarized in Algorithm 1. Algorithm 1Multi-source Data Integration and Feature Engineering Multi-source Data Integration and Feature Engineering Hybrid model architecture with BiLSTM and attention To address the challenge of jointly modeling temporal academic evolution and heterogeneous student attributes, we propose a hierarchical fusion framework that integrates attention-enhanced sequential representations with discriminative static features, and further optimizes classification via a TPE-guided gradient boosting model. The core idea of the proposed method is to explicitly learn complementary interactions between academic progression dynamics and student background characteristics through a structured fusion pipeline, where temporal dependency modeling, feature selection, and nonlinear classification are jointly optimized. Each student \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i \in \{1,\dots ,N\}$$\end{document} is represented by a variable-length GPA sequence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{g}_i = \{g_{i,1}, g_{i,2}, \dots , g_{i,T_i}\}$$\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}$$T_i$$\end{document} denotes the number of observed semesters. Each scalar GPA value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$g_{i,t} \in \mathbb {R}$$\end{document} is mapped into a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_e$$\end{document}-dimensional embedding space:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{e}_{i,t} = \textbf{W}_e g_{i,t} + \textbf{b}_e, \end{aligned}$$\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}$$\textbf{W}_e \in \mathbb {R}^{d_e \times 1}$$\end{document} is the embedding projection matrix and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{b}_e \in \mathbb {R}^{d_e}$$\end{document} is the bias term. The embedding dimension \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_e$$\end{document} is a hyperparameter controlling representation capacity. To capture temporal dependencies in academic performance evolution, we employ a bidirectional LSTM that models both past and future contextual information. The forward and backward hidden states are defined as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \overrightarrow{\textbf{h}}_{i,t}&\in \mathbb {R}^{d_h} = \text {LSTM}_{\rightarrow }(\textbf{e}_{i,t},\overrightarrow{\textbf{h}}_{i,t-1}),\\ \overleftarrow{\textbf{h}}_{i,t}&\in \mathbb {R}^{d_h} = \text {LSTM}_{\leftarrow }(\textbf{e}_{i,t},\overleftarrow{\textbf{h}}_{i,t+1}), \end{aligned}$$\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}$$d_h$$\end{document} is the hidden state dimension shared by both directions. The concatenated temporal representation is:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{h}_{i,t} = [\overrightarrow{\textbf{h}}_{i,t}; \overleftarrow{\textbf{h}}_{i,t}] \in \mathbb {R}^{2d_h}. \end{aligned}$$\end{document}To enhance interpretability and identify temporally discriminative semesters, we introduce an attention mechanism that learns a normalized importance distribution over all time steps. The intermediate representation is:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{u}_{i,t} = \tanh (\textbf{W}_a \textbf{h}_{i,t} + \textbf{b}_a), \end{aligned}$$\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}$$\textbf{W}_a \in \mathbb {R}^{d_a \times 2d_h}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{b}_a \in \mathbb {R}^{d_a}$$\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}$$d_a$$\end{document} is the attention hidden dimension. The attention weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _{i,t} \in (0,1)$$\end{document} is computed as:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \alpha _{i,t} = \frac{\exp (\textbf{v}_a^{\top } \textbf{u}_{i,t})}{\sum _{j=1}^{T_i} \exp (\textbf{v}_a^{\top } \textbf{u}_{i,j})}, \end{aligned}$$\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}$$\textbf{v}_a \in \mathbb {R}^{d_a}$$\end{document} is the attention scoring vector. The resulting context vector aggregates temporally informative states:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{c}_i = \sum _{t=1}^{T_i} \alpha _{i,t} \textbf{h}_{i,t}, \end{aligned}$$\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}$$\textbf{c}_i \in \mathbb {R}^{2d_h}$$\end{document} summarizes the entire academic trajectory of student i. In parallel, we construct a static feature vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{s}_i \in \mathbb {R}^{d_s}$$\end{document} (with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_s=13$$\end{document} before selection), consisting of demographic attributes, enrollment characteristics, and aggregated academic indicators. To reduce redundancy and improve signal-to-noise ratio, Recursive Feature Elimination (RFE) is applied to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{s}_i$$\end{document}, yielding a refined feature subset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{\textbf{s}}_i \in \mathbb {R}^{d_r}$$\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}$$d_r \le d_s$$\end{document} is determined empirically. The final student representation is obtained via hierarchical fusion:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{x}_i = [\tilde{\textbf{s}}_i; \textbf{c}_i] \in \mathbb {R}^{d_r + 2d_h}. \end{aligned}$$\end{document}Importantly, this fusion is not a simple concatenation, but a structured representation alignment mechanism: static attributes encode cross-sectional student heterogeneity, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{c}_i$$\end{document} encodes longitudinal behavioral evolution, enabling the classifier to learn cross-domain interactions between stable factors and temporal dynamics. For classification, we adopt XGBoost as the base learner. Let C denote the number of classes. The model produces a score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_k(\textbf{x}_i)$$\end{document} for each class k, and converts it into a probability via:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} P(y_i = k \mid \textbf{x}_i) = \frac{\exp (f_k(\textbf{x}_i))}{\sum _{j=1}^{C} \exp (f_j(\textbf{x}_i))}. \end{aligned}$$\end{document}To improve generalization and avoid overfitting, hyperparameters Θ (including learning rate, tree depth, subsampling ratio, and number of estimators) are optimized using Tree-structured Parzen Estimator (TPE). TPE models the conditional density \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p(\mathscr {L} \mid \Theta )$$\end{document} over validation loss \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {L}$$\end{document} and iteratively selects configurations that maximize expected improvement. Overall, the proposed framework constitutes a hierarchical hybrid learning architecture that unifies temporal sequence modeling, adaptive representation weighting, and TPE-optimized gradient boosting within a single fusion paradigm, enabling effective interaction learning between longitudinal academic dynamics and static student heterogeneity for robust employment outcome prediction. The full hybrid prediction pipeline combining sequential and static features is detailed in Algorithm 2. Algorithm 2Hybrid Modeling with BiLSTM, Attention, and XGBoost Hybrid Modeling with BiLSTM, Attention, and XGBoost Experimental analysis Ethical approval and informed consent All methods were carried out in accordance with relevant guidelines and regulations. All experimental protocols and the use of student data were approved by the Ethics Committee of Zhejiang Business College. Furthermore, we confirm that informed consent was obtained from all subjects and/or their legal guardians prior to the collection of the original records used for generating the study’s framework. Experimental setup To validate the effectiveness of the proposed hybrid model for predicting student employment outcomes, we designed a comprehensive experimental framework based on a self-constructed dataset, competitive baseline methods, and widely accepted performance metrics. The dataset was collected from a vocational university, covering student cohorts from eight academic years. It includes a rich set of multi-source features that span demographic attributes, academic performance, engagement in scientific and organizational activities, as well as financial background. Sequential academic features such as semester-wise GPA (from the first to eighth term) form the core temporal input, while static features including gender, admission type, scholarship history, and graduation honors are selected using Recursive Feature Elimination (RFE) for their strong predictive value. For classification, each student is assigned one of four labels representing post-graduation outcomes: domestic further education, successful employment, unemployment, or overseas advancement. Specifically, the dataset comprises 1,000,000 synthetic records spanning admission cohorts from 2015 to 2022. Employment outcome serves as the target variable with clear category proportions. Successful employment accounts for 68.5%, domestic further education reaches 18.2%, unemployment takes up 11.5% and overseas advancement occupies 1.8%. Before data preprocessing, roughly 4.3% of samples carried missing values, and these vacancies were filled through empirical mean imputation. In addition, all samples with standardized Z-scores higher than 3 were marked as outliers and eliminated. After standardized data cleaning, 987,652 valid records were finally retained for model training and performance evaluation. To achieve reliable assessment of the prediction framework and effectively prevent cohort leakage, chronological splitting was adopted in place of random data division. We adopted data from admission cohorts between 2015 and 2020 for model training, which accounts for about 80% of the total data volume. Data collected from the subsequent 2021 and 2022 cohorts that had not been involved in training was independently applied to model testing. This time-sequenced division achieves strict temporal separation, while maintaining stable class distribution and ensuring complete longitudinal rationality for evaluation work. To mitigate the potential bias caused by the highly imbalanced class distribution–where successful employment accounts for 68.5% versus overseas advancement at only 1.8%–we introduced a class-weighted Focal Loss during the XGBoost training phase. This dynamically scales the loss based on prediction confidence, forcing the model to focus on the hard-to-classify minority classes. Combined with our use of Macro-F1 and Cohen’s Kappa as primary evaluation metrics, this optimization ensures the model maintains equitable and unbiased performance across all four employment outcomes. The model processes both the RFE-selected static features and the GPA sequence using the BiLSTM+Attention structure, followed by feature fusion and classification via XGBoost, whose hyperparameters are optimized using the Tree-structured Parzen Estimator (TPE). To benchmark the model, we implement four recent methods for comparison: ARW2 (Attention-RNN-Word2) from Ma and Ding20, which utilizes attention-based RNNs for talent cultivation quality prediction, and OSVM-Match (Optimized SVM for Employment Matching) by Lv21, which employs kernel-tuned SVMs for personalized job matching based on multi-dimensional profiles; FT-Transformer (Feature Tokenizer + Transformer) from Gorishniy et al.22, which is a highly competitive baseline specially designed for heterogeneous tabular data modeling; and LSTM-FCN (Long Short-Term Memory Fully Convolutional Network) from Karim et al.23, which is a robust hybrid model with outstanding efficacy in multivariate time-series classification. These methods serve as strong baselines for static modeling, sequential modeling, tabular data learning and time-series modeling tasks respectively, enabling a comprehensive and fair performance comparison. The experiments are conducted under an 80/20 stratified train-test split to preserve class balance. Evaluation metrics include overall Accuracy, Macro-averaged F1 Score, Area Under the ROC Curve (AUC), and Cohen’s Kappa, ensuring that both predictive correctness and class balance are comprehensively assessed. Baseline performance comparison Table 2Performance Comparison of Different Employment Prediction Models.ModelAccuracyMacro-F1AUCCohen’s KappaOSVM-Match0.781 ± 0.0050.743 ± 0.0060.802 ± 0.0050.692 ± 0.007ARW20.806 ± 0.0040.769 ± 0.0050.823 ± 0.0040.728 ± 0.006LSTM-FCN0.822 ± 0.0040.795 ± 0.0040.841 ± 0.0030.748 ± 0.005FT-Transformer0.835 ± 0.0030.810 ± 0.0030.850 ± 0.0020.765 ± 0.004Ours (BiLSTM-Attn-XGB)0.857 ± 0.0020.831 ± 0.0030.876 ± 0.0020.794 ± 0.003 Table 2Performance Comparison of Different Employment Prediction Models.ModelAccuracyMacro-F1AUCCohen’s KappaOSVM-Match0.781 ± 0.0050.743 ± 0.0060.802 ± 0.0050.692 ± 0.007ARW20.806 ± 0.0040.769 ± 0.0050.823 ± 0.0040.728 ± 0.006LSTM-FCN0.822 ± 0.0040.795 ± 0.0040.841 ± 0.0030.748 ± 0.005FT-Transformer0.835 ± 0.0030.810 ± 0.0030.850 ± 0.0020.765 ± 0.004Ours (BiLSTM-Attn-XGB)0.857 ± 0.0020.831 ± 0.0030.876 ± 0.0020.794 ± 0.003 Performance Comparison of Different Employment Prediction Models. To ensure the robustness of our evaluation, all models were evaluated across 10 independent runs using different random seeds. Table 2 reports the mean performance and standard deviation for each metric. Furthermore, a paired t-test was conducted between our proposed BiLSTM-Attn-XGB model and the strongest baseline, FT-Transformer. The results indicate that the performance improvements achieved by our framework are statistically significant across all metrics (p < 0.01), confirming the reliability of our static-temporal fusion strategy. Table 2 presents the performance comparison of all evaluated models. Traditional machine learning approaches, including OSVM-Match and ARW2, achieve relatively lower performance across all metrics. Deep learning models, represented by LSTM-FCN and FT-Transformer, demonstrate improved predictive capability, indicating the effectiveness of learning complex patterns from educational data. Among all baseline methods, FT-Transformer achieves the strongest performance, with an accuracy of 0.835 and a Macro-F1 score of 0.810. The proposed BiLSTM-Attn-XGB framework consistently achieves the best results, obtaining an accuracy of 0.857, a Macro-F1 score of 0.831, an AUC of 0.876, and a Cohen’s Kappa coefficient of 0.794. Compared with FT-Transformer, the strongest baseline, the proposed method improves accuracy by 2.2 percentage points and Macro-F1 by 2.1 percentage points. These results demonstrate that the integration of temporal academic trajectories and static student attributes provides more discriminative representations, leading to superior prediction performance. The observed performance improvement can be attributed to three key factors. First, the Bidirectional Long Short-Term Memory network effectively captures long-range dependencies embedded within students’ grade trajectories, while the attention mechanism automatically identifies academically critical periods that contribute most strongly to future outcomes. Second, the fusion of temporal representations and static attributes allows the model to jointly characterize academic performance, demographic characteristics, scholarship achievements, and extracurricular engagement, thereby providing a more comprehensive description of student development patterns. Third, XGBoost serves as a powerful nonlinear decision layer capable of exploiting complex interactions among heterogeneous features, leading to superior classification performance and stronger generalization capability. To evaluate the practical efficiency of the proposed framework, we analyze its computational complexity and empirical runtime. Theoretically, the time complexity of the BiLSTM encoder is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(T \cdot d_h^2)$$\end{document}, where T is the sequence length and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_h$$\end{document} is the hidden state dimension. The TPE-optimized XGBoost classifier introduces a complexity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(K \cdot D \cdot N \log N)$$\end{document}, where K is the number of trees, D is the maximum depth, and N is the number of samples. Empirically, training the entire pipeline on our dataset (987,652 valid records) requires approximately 45 minutes on a single NVIDIA RTX 4090 GPU, with an inference time of merely 12 milliseconds per student sample. This demonstrates that despite the integration of deep learning and tree-based components, the model remains computationally efficient for large-scale educational deployment. Ablation study Fig. 1Ablation study demonstrating the incremental impact of model components. Ablation study demonstrating the incremental impact of model components. Figure 1 depicts the results of an ablation study that quantifies the contribution of individual architectural components to the final model performance. Starting from the baseline XGBoost classifier trained solely on the selected static features, the model achieves an accuracy of 78.5% and a macro-F1 of 0.742. Incorporating a BiLSTM encoder to process GPA sequences results in a substantial performance gain, increasing accuracy to 81.5% and macro-F1 to 0.778, which confirms the importance of modeling academic progression as a temporal sequence. The inclusion of the attention mechanism further improves the model to 84.2% accuracy and 0.803 macro-F1, underscoring the value of dynamic weighting in identifying critical learning phases. The full model, with additional TPE-based hyperparameter optimization, reaches the highest accuracy of 85.7% and macro-F1 of 0.831, reflecting the compounded effect of structural and parametric refinement.Fig. 2Impact of temporal depth on prediction accuracy. Impact of temporal depth on prediction accuracy. Figure 2 evaluates how the model’s accuracy evolves with increasing sequence length, i.e., the number of semesters used in training the BiLSTM. While both BiLSTM-only and BiLSTM+Attention show upward trends as sequence length grows from 4 to 8, the attention-enhanced variant consistently yields superior results. At 8 semesters, the BiLSTM-only model attains 81.4%, whereas BiLSTM+Attention achieves 85.7%, highlighting the attention mechanism’s ability to extract salient temporal information even when sequence depth increases. These findings collectively demonstrate that each model component–temporal modeling, attention-based feature weighting, and automated hyperparameter tuning–plays a distinct and complementary role in enhancing predictive performance. The attention mechanism, in particular, contributes by allowing the model to focus on high-impact semesters, such as those near graduation, which are often more indicative of eventual employment outcomes. Furthermore, TPE fine-tuning allows the final classifier to more effectively leverage fused features from heterogeneous sources, improving generalization across varied student profiles. Robustness experiment Fig. 3Comparison of predictive accuracy across varying temporal depths (number of GPA semesters used). Comparison of predictive accuracy across varying temporal depths (number of GPA semesters used). Figure 3 compares the predictive accuracy of three models–OSVM-Match, ARW2, and the proposed hybrid architecture–across varying temporal depths, measured as the number of GPA semesters used as input. As shown, model performance improves as the sequence length increases, yet the rate of improvement and the final accuracy plateau vary significantly across methods. The OSVM-Match model, which operates purely on static features without temporal modeling capabilities, exhibits minimal gains from incorporating longer GPA sequences. Its accuracy increases modestly from 73.8% with 2 semesters to 76.2% at 10 semesters, indicating that the model is not effectively utilizing temporal dependencies. In contrast, ARW2–a recurrent model equipped with an attention mechanism–achieves more substantial improvements, reaching 80.7% accuracy at 10 semesters. This demonstrates its capacity to model sequential academic signals, although its performance tends to saturate beyond 8 semesters. The proposed model significantly outperforms both baselines across all depths, starting from 76.4% at 2 semesters and rising consistently to 86.4% at 10 semesters. Notably, the performance advantage over ARW2 widens with deeper sequences, illustrating the robustness and scalability of our architecture. This superiority can be attributed to three core factors. First, the BiLSTM encoder effectively captures long-range dependencies in academic progression, while the attention mechanism dynamically focuses on semantically salient semesters (e.g., final-year performance or rapid changes in GPA trajectory). Second, the fusion of static features with temporal embeddings allows the model to integrate contextual background with evolving academic patterns, which is crucial for accurately modeling heterogeneous employment outcomes. Third, the TPE-optimized XGBoost classifier ensures that the fused representation is leveraged by a high-capacity decision function with strong generalization ability. Explainability analysis via SHAP Based on cooperative game theory, SHAP quantifies the marginal contribution of each feature to the final prediction and establishes a unified framework for interpreting complex machine learning models. Given that XGBoost serves as the final decision layer of the proposed framework, the TreeSHAP algorithm was adopted to compute exact SHAP values efficiently in the log-odds space.Fig. 4SHAP summary plot illustrating global feature importance and directional impact for the “Successful Employment” class. SHAP summary plot illustrating global feature importance and directional impact for the “Successful Employment” class. Figure 4 presents the SHAP summary plot, which characterizes both the overall importance of input variables and their directional influence on prediction outcomes. The results indicate that academic performance indicators from the final stages of study, particularly \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$gpa\_sem8$$\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}$$gpa\_sem7$$\end{document}, exhibit the largest mean absolute SHAP values, highlighting their dominant influence on employment outcome prediction. Beyond feature ranking, the distribution of SHAP values reveals substantial heterogeneity in feature effects across different samples. Variables reflecting student engagement, including \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$scientific\_activities$$\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}$$scholarship\_frequency$$\end{document}, demonstrate a pronounced positive contribution to the prediction of the domestic further-education category. In contrast, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$total\_credits\_attempted$$\end{document} provides a relatively stable contribution across outcome categories, suggesting that cumulative academic participation serves as a foundational indicator but possesses limited discriminative capability compared with senior-year academic performance.Fig. 5SHAP local waterfall plot detailing the decision trajectory for a correctly classified “Successful Employment” instance. SHAP local waterfall plot detailing the decision trajectory for a correctly classified “Successful Employment” instance. In addition to global feature analysis, SHAP enables instance-level interpretation and facilitates a detailed examination of individual prediction trajectories. Figure 5 illustrates the local waterfall plot for a correctly classified sample belonging to the successful employment category. Starting from the expected prediction baseline derived from the training data, the model output is progressively adjusted by the combined effects of feature-specific contributions. Positive contributions are primarily associated with a strong \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$cumulative\_gpa\_entry$$\end{document} and sustained participation in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$organizational\_activities$$\end{document}, indicating that solid academic foundations and active social engagement substantially increase the likelihood of successful employment. Conversely, relatively limited involvement in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$scientific\_activities$$\end{document} and the influence of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$high\_impact\_major\_courses$$\end{document} reduce the propensity toward the employment category and instead shift the prediction toward academically oriented pathways, particularly further education. The incorporation of SHAP substantially enhances the transparency of the proposed hybrid framework by revealing the heterogeneous feature interactions that govern multi-class student outcome prediction. The resulting interpretations provide statistically grounded evidence for understanding model behavior and offer valuable insights for educational decision-making, thereby supporting the development of targeted intervention strategies and personalized career guidance programs. Conclusion and future work This study introduces an interpretable and high-performing hybrid framework for student employment prediction, addressing key limitations in existing methods by fusing temporal and static data streams. By leveraging recursive feature elimination for dimensionality reduction, BiLSTM+Attention for modeling longitudinal academic performance, and TPE-optimized XGBoost for robust classification, the framework not only achieves superior predictive accuracy but also provides actionable insights into feature-level contributions via SHAP analysis. The model’s ability to dynamically attend to critical academic periods while integrating heterogeneous student profiles offers a nuanced understanding of employability patterns in higher vocational education. Future work will extend this research in several directions. First, we plan to incorporate external data sources such as internship records, socioeconomic background, and local employment statistics to further enhance predictive granularity. Additionally, we aim to explore more advanced explainability techniques and dynamic graph-based architectures to capture peer influence and institutional effects. Finally, translating this predictive framework into real-world educational practice requires robust system architectures and strict data governance. Future deployment strategies will focus on developing RESTful APIs for seamless integration with existing university Learning Management Systems (LMS), alongside building intuitive, interactive dashboards that translate SHAP-based interpretations into actionable guidance for academic advisors. Furthermore, to address the ethical imperatives of data utilization, we will explore the integration of differential privacy mechanisms during the model training phase. This will ensure that sensitive student trajectories, financial statuses, and demographic attributes remain cryptographically protected while fully preserving the framework’s analytical utility.