Optimize a chain convenience store location prediction model by using MTS-machine learning methodology Authors: Tsung-Yin Ou, Hsin-Pin Fu, Mei-Zhen Wu Published: 2025 ABSTRACT AbstractIn the past, convenience store location selection required substantial time, cost and effort for field research, and the location selection decision was often limited by the subjective experience of the decision-maker. The rapid development of machine learning (ML) in recent years has enabled many studies of model training. This study introduces a novel hybrid approach integrating Mahalanobis–Taguchi System (MTS) for variable reduction with ML algorithms to address the challenges of large-scale, multi-variable retail location selection. The proposed model reduces computational requirements and supports data-driven decision-making for convenience store chains. To achieve this, a hybrid approach combining the Mahalanobis–Taguchi System (MTS) and machine learning algorithms—XGBoost, Random Forest (RF), and support vector machine (SVM)—were employed to develop a location prediction model for chain convenience stores. Using real-world location data from Taiwanese convenience stores, MTS analysis mainly performed features selection, reducing the number of variables from nine to five. The research results show that the prediction ability of the MTS-XGBoost algorithm can be accurately maintained at more than 75% under different training set proportions, and it has more accurate prediction results than MTS-RF and MTS-SVM. This approach enables more efficient and accurate store location decisions, potentially transforming strategic planning in the retail sector.Keywords: Convenience store, Location selection, Mahalanobis–Taguchi system (MTS), XGBoost, Random forest (RF), SVM, Features selectionSubject terms: Engineering, Mathematics and computing Introduction From 2019 to 2023, the number of convenience stores in Taiwan grew at annual rates of 3.4% to 5.5%. By the end of 2023, Taiwan had 13,706 convenience stores, an average of one for every 1,703 people. Taiwan’s density of convenience stores is second only to South Korea’s1. Ever-changing market conditions and fierce competition in the retail industry mean wise decision-making about opening new stores is crucial. Failure to occupy good business locations not only wastes capital costs but seriously impedes expansion plans. A poor location has both strategic and long-term effects, reducing the facility’s profitability and operability for an extended period of time given the conditions of the market, environment and population change2. Convenience store location selection is complex because it is affected by many factors, such as demographic and socioeconomic characteristics, availability and cost of space, business ecosystems, competition, customer behavior and preferences, legal and fiscal factors, accessibility, availability of supplies, and climate and environmental factors3. Previous studies of differentiation in the competitive market structure of physical retail formats mostly focused on relative location and spatial competition4, finding that an accessible location and large store size are associated with decreased likelihood of store closure. Even in an increasingly digital world in which online shopping is burgeoning, location is crucial to the survival of physical stores5. The convenience store site selection process requires substantial time, cost, and effort to conduct research and collect relevant data for evaluation. The rapid development of machine learning (ML) and Artificial Intelligence (AI) technologies and algorithms in recent years has boosted research on location prediction6–8. However, ML methods require substantial data, in turn increasing cost and the computing power required. In practical applications, large amounts of data and many variables may create noise and decrease accuracy. Zouhri et al.9 assessed the effectiveness of dimensionality reduction on the interpretability of machine learning-based detection systems which achieved 99.99% accuracy with a reduction from 80 to 6 features. In this study, a model designed to predict the success of new convenience stores was established by reducing variables and applying ML. Convenience store location selection often involves various interacting nonlinear variables. Recent studies have highlighted the robustness of the Mahalanobis–Taguchi System (MTS) in addressing complex, nonlinear problems. Specifically, MTS has been shown to perform effectively in nonlinear multiclass classification10, high-dimensional small-sample analysis11, optimized feature selection12, the identification and screening of critical variables13, and abnormal observation analysis14. A comprehensive review further confirms its broad applicability across diverse contexts15. Building on this foundation, our study employs MTS as a method for variable screening and deletion, leveraging its strength in handling multiple nonlinear variables while preserving interpretability. Case data were validated using XGBoost16, Random Forest (RF)17, and Support Vector Machine (SVM)18–20, which are well suited for retail location prediction due to their complementary strengths. XGBoost leverages gradient boosting with regularization and parallelization to deliver efficient, high-accuracy predictions, while RF reduces variance through ensemble learning and provides interpretable feature importance measures useful for managerial decisions. SVM, by maximizing classification margins and employing kernel functions, effectively captures nonlinear patterns and performs robustly with moderate sample sizes. Collectively, these algorithms are recognized for their ability to model complex, high-dimensional data with resistance to overfitting, strong interpretability, and computational efficiency, ensuring both reliable predictive accuracy and actionable insights for location decision-making. This study makes twofold contributions. Practically, it equips convenience store operators with a systematic, data-driven tool that reduces reliance on subjective judgment, lowers evaluation costs and time, and enhances prediction accuracy by identifying key factors such as visibility, competitiveness, and residential density. Academically, it advances retail location analytics by integrating the Mahalanobis–Taguchi System with XGBoost, a hybrid approach that addresses noise and dimensionality challenges in real-world datasets while preserving accuracy and interpretability. Using real convenience store data, the study verifies the model’s effectiveness for nonlinear, multi-variable decision-making, thereby extending the applicability of MTS to a novel retail context. Literature review Convenience store location selection Technologies for location selection analysis have existed for more than 50 years, yet most traditional retailers have historically relied on intuition, experience, and “common sense” rather than systematic models21. While such heuristic approaches are practical and low-cost, they lack rigor and are prone to subjective bias, resulting in inconsistent or suboptimal site decisions. With the advancement of analytical methods, researchers began adopting quantitative models. For instance, Lin et al.22 applied a spatial model to examine agglomeration among convenience store chains, analyzing how inter-store distance affects revenue. Although spatial models effectively capture geographic competition, they typically oversimplify complex consumer behaviors and external market dynamics. Similarly, Lienbacher et al.23 measured the distances between customers’ residential addresses, suppliers, and supermarkets, and found strong associations with supermarket location decisions. While distance-based models highlight the importance of accessibility, they neglect nonlinear interactions between demographic, economic, and competitive variables that often drive retail performance. Some scholars extended the scope of location analysis to resource-constrained environments. For example, Song et al.24 applied integer programming to optimize the placement of electric vehicle charging stations, while25 used data envelopment analysis to improve logistics center siting. These optimization-based methods are rigorous and computationally efficient; however, their applicability to convenience store site selection is limited because they treat location as a purely operational problem, without incorporating consumer demand heterogeneity or retail-specific competitive pressures. Recent studies have integrated big data and data mining techniques to enhance retail location decisions. Formánek and Sokol26 proposed a methodology that combines linear mixed-effects models with cluster analysis to assess and plan store locations. While such hybrid statistical approaches improve segmentation and pattern detection, they often remain sensitive to noise and require large, clean datasets that are rarely available in practice. Other studies explored machine learning techniques such as neural networks, least absolute shrinkage and selection operator (LASSO), and random forests, which demonstrated strong performance on error metrics like RMSE and MAPE27. Nevertheless, these models tend to operate as “black boxes,” providing high accuracy but limited interpretability, which constrains their usefulness for managerial decision-making. Further, Brei et al.28 introduced a novel approach leveraging satellite imagery and ML to predict retail performance. Although innovative, image-based methods require extensive data preprocessing. Finally, some authors have argued for broader adoption of artificial intelligence in retail site selection29,30. While these studies underscore the potential of AI, they often present conceptual arguments or exploratory applications without addressing the challenges of data redundancy, variable noise, and the balance between predictive accuracy and managerial interpretability. Prior research demonstrates diverse methodological advances but also reveals important limitations: spatial and optimization models often oversimplify consumer–market dynamics, while ML and AI techniques, despite their predictive power, struggle with noise reduction, interpretability, and transferability across contexts. These shortcomings highlight the need for approaches that are both computationally efficient and practically interpretable. The Mahalanobis–Taguchi system The MTS is a multivariate diagnostic prediction method that combines quality engineering, Mahalanobis distance (MD), signal-to-noise (SN) ratio, and orthogonal array (OA) methods. The main purpose of MTS is to identify and screen variables that contribute to abnormality, thereby enhancing the accuracy of classification and diagnosis31. Compared with traditional dimensionality reduction methods, MTS does not rely on distributional assumptions, making it suitable for diverse and complex datasets. However, its sensitivity to abnormal data and reliance on carefully constructed reference groups remain potential limitations. Several studies have demonstrated the effectiveness of MTS in different domains. For instance, Zhang et al.32 highlighted its role in simplifying analysis and improving efficiency in site selection tasks. While the study illustrates MTS’s ability to streamline decision-making, it offers limited discussion of how the method performs under conditions of noisy, high-dimensional retail data. Similarly, Sikder et al.33 proposed a synergistic MTS-based framework for process quality control in manufacturing. This work confirmed MTS’s robustness in industrial applications, yet its focus on controlled environments limits generalizability to service industries such as retail. Wang et al.34 demonstrated that MTS, as a big data classification and reduction method, could support fault diagnosis and maintenance modeling without assuming data distributions. This highlights MTS’s versatility, though the study overlooked challenges of interpretability in managerial contexts. Similarly, Wang et al.35 showed that optimized MTS improves classification for imbalanced datasets, Mao et al.36 applied MTS to hypertension prediction, and Dashdondov and Kim37 applied MTS to evaluate EV battery consistency with strong results. While these works confirm MTS’s adaptability to medical and imbalance problems, their findings are highly domain-specific and less transferable to consumer-oriented sectors. In summary, while MTS has proven effective in fields such as manufacturing and healthcare, prior applications are largely limited to controlled settings and lack integration with modern machine learning. This study addresses these gaps by combining MTS with advanced classifiers to deliver interpretable and practical insights for convenience store location decisions. Research positioning and expected contributions Over the past decades, research on retail location selection has evolved from classical spatial and optimization approaches to more advanced data-driven methods. Traditional studies often employed spatial competition models to capture agglomeration and accessibility3,26 or relied on optimization techniques such as integer programming and data envelopment analysis24,25. While insightful, these methods struggle with nonlinear interactions among multiple variables, a hallmark of modern retail contexts. More recently, machine learning (ML) and artificial intelligence (AI) have expanded the methodological toolkit. For example, Lu et al.27 applied ML algorithms for retail store screening, and28 used satellite imagery with ML to predict performance. Despite their predictive power, such studies often depend on large datasets and face issues of redundancy and noise, limiting interpretability and practical application9. This study addresses these gaps by introducing a hybrid approach that integrates the Mahalanobis–Taguchi System (MTS) with advanced ML classifiers (XGBoost, Random Forest, SVM). MTS offers interpretable variable screening via signal-to-noise ratios, retaining factors with genuine managerial significance32,36. This is vital for decision-making, where transparency of key factors—such as visibility, supermarket competitiveness, and residential density—is as important as predictive accuracy. Prior works combining ML have focused on dimensionality reduction and computational cost. Table 1 compares the differences of previous research as follows: Table 1Comparison of computational cost and scalability of previous studies.Domain/dataDimensionality reductionML model(s)Computational costScalabilityManufacturing fault diagnosis10MTSSVM, RFReportedLimited domain scopeRetail logistics optimization24LASSOGradient BoostingPartial (runtime)Small-scale testsLogistics & facility placement25PCARandom Forest, ANNNoSimulation-scale onlyRetail location screening; geospatial and demographic data27PCAXGBoost, SVMNoRegional dataset onlyRetail performance using satellite imagery28AutoencoderCNN, Random ForestNo explicit discussionSingle-country setting Comparison of computational cost and scalability of previous studies. This study employs the Mahalanobis–Taguchi System (MTS) for variable screening due to its robustness in handling noisy, high-dimensional datasets. Unlike Correlation-based Feature Selection (CFS), which primarily captures linear correlations, MTS leverages Mahalanobis distance and signal-to-noise ratios to retain variables with genuine discriminative power32,36. Prior research has further validated MTS in nonlinear classification and abnormal observation detection10,14. By integrating MTS with advanced machine learning classifiers (XGBoost, RF, SVM), this study addresses gaps in earlier works that focused on supermarkets, logistics centers, or social supermarkets23,25, enhancing interpretability, reducing computational burden, and delivering accurate, actionable insights for convenience store site selection. Research methodology The experimental process of this study can be divided into the following four processes: Data collection, MTS analysis, classification prediction using different ML algorithms. Final process is conclusions and suggestions, as shown in Fig. 1 below. Research flow chart. Mahalanobis–Taguchi system procedures Mahalanobis distance is a statistical measure of the correlation between data points. It is widely used in classification methods. It considers not only the distance between data points and the centers of group but the groups’ distributions, improving the accuracy of classifying unknown points into reference groups. Hence, MD is an efficient way to overcome statistical problems such as the failure of Euclidean distance to consider the distribution of data38,39. MD is the result of scaling the traditional Euclidean distance by using the joint within-group covariate matrix, as shown in Eq. 1: where Xi: (xi1, xi2 ,…, xik) is the vector formed by the k characteristic variables of the nth sample; T is the transposed matrix; is the matrix inverse of the joint within-group covariate matrix. When the multivariable system has k characteristic variables and contains n samples, using the scaled MD is appropriate. The calculation is shown in Eq. 2. where Zij = (Z1j, Z2j, …,Zkj), which denotes the standard vector for the standardization of Xij (i = 1…k); Zij = (Xij − i)/Si; Xij is the value of the ith characteristic variable of the jth sample; i is the average value of the ith characteristic variable; Si is the standard deviation of the ith characteristic variable; T is the transposed vector; C−1 is the matrix inverse of correlation coefficients; k is the characteristic variable. In MTS, when the actual severity level of the abnormal sample is unknown, the use of the larger-the-better SN ratio is appropriate. The dataset used in this research included d abnormal samples; the MD corresponding to each experiment was calculated using the characteristics of the OA, and the square root of the MD of d abnormal samples was taken as the measurement value (y), that is, yi = , j = 1,2,…,d. In this study, it was expected that the larger the MD of the abnormal samples, the better, thus being in line with the larger-the-better characteristic, and the SN ratio formula was as shown in Eq. 3. After calculating each SN ratio in the OA, the variable effect of each characteristic variable must be calculated. 1+ is the average of the SN ratios of all experiments applying variable Xi, and 1− is the average of the SN ratios of all experiments without applying variable Xi. The effects of each variable are calculated as follows, the calculation is shown in Eq. 4: The MTS method has three major advantages over Back Propagation Neural Network (BPNN) and principal components analysis (PCA), its ability to deal with multiple variables, ability to reduce noise, and its cost effectiveness14. Machine learning algorithms Machine learning is a subfield of AI and is commonly used in data mining. It involves applying algorithms to parse data, learn from it, and make decisions based on what is learned40. AI plays a crucial role in both front-end and back-end processes41. In the former area, AI is employed for recommendation systems and in-store customer experience management. When considering back-end processes, AI improves inventory optimization, price optimization, and the anticipation of customer demand, benefiting retailers’ planning decisions6. AI allows retailers to operate more efficiently. XGBoost XGBoost is an ML algorithm with high classification accuracy42–44. In the context of this study, one of the reasons for opting for XGBoost is its advanced optimization capability. It offers a fast and scalable tree-boosting system equipped with parallel processing, pruning, regularization, and missing-value handling. XGBoost has been successfully applied in diverse scientific domains to effectively address anomaly detection challenges45,46. Due to its low computational cost, the XGBoost algorithm can be used for hyperparameter tuning in order to find the best hyperparameter combination47. Different studies contribute to depression classification tasks or diabetes prediction and diagnosis with several machine learning algorithms; among these classifiers, the performance of XGBoost is found to be the best48,49. XGBoost is an optimized distributed gradient boosting library, which belongs to an ML algorithm under the gradient boosting framework. where is the sum of the first derivatives of all samples; is the first derivative of the loss function; represents the true value; is the predicted value of Round t − 1; j = 1 T is the number of sets. The calculation is shown in Eq. 5. where is the sum of the second derivatives of all samples; is the second derivative of the loss function. When the loss function is determined, XGBoost performs supervised learning, with the following objective function is as follows, the calculation is shown in Eq. 6: : L2 regular term, which is a hyperparameter, where is the adjustment function; represents the number of leaves in the decision tree (Eq. 7). Randon forest Randon forest (RF) is an ensemble machine learning algorithm that uses multiple decision trees to improve prediction accuracy and reduce overfitting. Each tree in the forest is built using a random subset of the data and features. The split criterion for classification is Gini Index shown in Eq. 8: where is the proportion of samples belonging to class i. In the Bootstrap Aggregation (Bagging) stage, each tree is trained on a random sample (with replacement) of the training data shown in Eq. 9: where is the dataset, and n is the size of the sample. For classification, predictions are based on majority voting from all trees (Eq. 10). where is the prediction from the ith tree. These equations outline the core principles of RF: random sampling of data, random feature selection, and ensemble predictions. The combination of randomness and aggregation results in a robust model with high predictive accuracy. These advantages are well demonstrated in power load and building energy forecasting50,51. Support vector machine Support vector machine (SVM) is a supervised learning algorithm used for classification and regression. It aims to find the optimal hyperplane that separates data points of different classes with the maximum margin. The decision boundary in SVM is a hyperplane shown in Eq. 11: where is the weight vector (normal to the hyperplane), is the bias (offset of the hyperplane), is the input feature vector. To classify a point is shown in Eq. 12. SVM maximizes the margin, defined as the distance between the hyperplane and the nearest data points (support vectors). The margin is shown in Eq. 13. The optimization goal is to minimize (maximize margin) subject to correctly classifying all points. For linearly separable data is shown in Eq. 14. For data that is not perfectly separable, SVM introduces slack variables to allow some misclassification (Eq. 15). where is the penalty parameter controlling the trade-off between margin size and misclassification. When data is not linearly separable, SVM maps input data to a higher-dimensional space using a kernel function (Eq. 16). where is a mapping function. The optimization problem is reformulated using Lagrange multipliers is shown in Eq. 17. Subject to These equations summarize the key principles of SVM, which focus on finding the hyperplane that maximizes the margin and uses kernel functions to handle nonlinear data effectively. SVM has excellent performance in judging system risk problems and various classification problems52,53. This stody applied XGBoost, RF and SVM to improve the accuracy of convenience store location selection. These algorithms offer excellent prediction accuracy, high-dimensional data processing, resistance to overfitting, interpretability, and efficiency, features that make it suitable for dealing with complex decision problems. Case study and analysis results Case description and selection of characteristic variables Data collection The dataset used in this study was provided by developers of a major convenience store chain in Taiwan. It covers the period from 2021 to 2024 and includes information from both residential and commercial districts, the two areas with the highest concentration of convenience stores. In addition to general site characteristics, the dataset incorporates investment-related information, including expected and actual revenues recorded before and after store establishment. To support systematic evaluation, the dataset contains criteria for store location decisions, encompassing nine key variables summarized in Table 2. Table 2Nine store selection variables.CodeNameVariable typesABusiness district typeBinaryBConvenience store competitivenessCategoricalCSupermarket competitivenessCategoricalDTraffic volumeContinuousEMain customer typeCategoricalFNumber of residentsContinuousGStore locationCategoricalHVisibilityCategoricalIStore areaCategorical Nine store selection variables. Each factor can be rated with different scores. For example, if no supermarket is situated within X meters of the proposed new store location, the value of variable is 100; if more than four supermarkets are within this radius, in its vicinity, it is zero. Business districts are either residential or commercial areas, so variable A is binary, respectively represented by 1 or 0 (respectively). Data for 96 stores were analyzed, as shown in Table 3. Table 3Selection data for 96 convenience stores in Taiwan.NoABCDEFGHINoA1BCDEFGHI111006015.1610017.1680501.149110010014.588021.3280501.1212010014.838030.52501001.150110207.531003.28801001.131655512.838014.4480501.151110010013.4405.4701001.141101005.541008.56801001.1521201009.1310015.48501001.151100400.05802.881001001.153110010010.071006.48801001.16110010010.3010028.92801001.154110011.418013.88801001.1702010011.76806.7580501.1551651006.30807.96801001.18115512.731003.68501001.1561100507.89805.96501001.1911001001.548001001001.157120756.848026.8880501.11010505.44804.125001.1581301009.491004.28501001.11111001004.06803.1250501.15914010012.701007.0450501.11213010014.808028.280501.1601601006.6910050.0880501.113110506.158032.0420501.161130759.0310013.7250501.11411001003.21804.165001.16215755.358019.44501001.11512510010.501004.850501.1631101007.108010.880501.11612510024.8010027.48801001.1641101005.63804.9680501.1171103514.4810013.4480501.16511001006.918020.6880500.9181206012.5910019.92801001.1661101005.488014.28501001.119110010011.058017.850500.96711510013.4910014.72100501.12001010010.0810013.25501001.16811006010.541006201001.12111001006.5010013.24501001.16911001003.911006.92801001.1221101006.061005.2480501.17010456.7110010.9670501.123130754.56052.480501.17112010010.178036.9650501.124120802.811002.08501001.1721201004.381008.56801001.1251101009.2010013.4880501.1731401004.74015.6880501.126110010024.4810025.88801001.174120555.51025.76801001271301008.808022.56801001.1751401008.658023.280501.128140556.178011501001.1761201008.758016.76501001.129120557.4310014.32801001.177110659.47042201001.1301201002.6210010.64801001.17812010010.318019.72801001.13111001007.241005.44501001.17911001006.918022.728001.1321301008.32808.5250501.1800253510.2710027.95201001.133110603.68803.1280501.1811010013.068015.4850501.1341101009.828017.6480100182020553.3510012.4201001.13511510020.451001.441001001.18311001004.7504.24801001.1361151003.8610027.82010018411510014.9510017.1650501.1371401004.151007.92501001.18511001006.28809.0450501.13811001001.21037.0850501.18611009013.038015.88501001.13913010015.838017.8580501.1871151008.33041801001.140140752.5210029.620501.188110856.28801.4501001.1411201009.2110010801001.1891301007.881009.28100501.142110010022.02805.84505019014010018.7610018.7280501.1431203512.96804.7280501.191140508.648028.76801001.14411010015.928012.72501001.192155558.808017.3280501.14511001007.188013.8450501.1931304020.8710029.680501.1461301009.448016.16501001.19411010025.23803.2801001.1470204018.978024.121001001.19511001007.068039.92801001.14815107.9810014.68801001.19611008012.8110015.8850501.1 Selection data for 96 convenience stores in Taiwan. Mahalanobis–Taguchi system analysis The full dataset analyzed in this research included the data used to select the 96 store locations, their predicted turnover, and their average monthly turnover six months after the opening of each new store50. The dependent variable in this study is defined as a binary indicator capturing the success or failure of a candidate store location. A site is classified as successful if its actual turnover exceeds 90% of the predicted turnover, and as unsuccessful otherwise. Based on this criterion, 53 stores were identified as successful, while the remaining locations—those with turnover equal to or below 90% of the predicted value—were classified as unsuccessful (abnormal data). Under this classification, the baseline prediction accuracy of store outcomes was 55% (Table 4). Table 4Stores’ predicted actual turnover, and successful rate.StorePredictActualDifferenceRatioSuccessStorePredictActualDifferenceRatioSuccess170,40055,02315,3770.784952,79047,00757830.89260,30046,92313,3770.785047,50037,81996810.80343,04039,37536650.9115160,80040,15220,6480.66449,70047,21424860.9515273,20046,71026,4900.64565,00039,66925,3310.615369,60061,93076700.89687,20091,00438041.0415433,50067,50734,0072.021732,55042,91410,3641.3215563,00064,05610561.021834,70053,74319,0431.5515639,02036,37926410.931947,05043,51335370.9215754,60035,05019,5500.641032,90044,32211,4221.3515844,20030,78013,4200.701155,00056,08110811.0215939,10037,79913010.9711268,00048,00719,9930.716085,20046,56438,6360.551333,00035,87228721.0916149,00055,18461841.1311452,60039,85112,7490.766243,80039,19646040.891540,70033,12175790.816332,52070,66738,1472.1711649,08052,61935391.0716450,00042,15278480.841749,00040,46085400.836546,75050,03332831.0711859,10048,15510,9450.816641,50046,38448841.1211950,00053,14031401.0616743,30080,20536,9051.8512045,00053,56185611.1916851,10041,51795830.812161,60033,41028,1900.546963,20040,17523,0250.642243,80044,6228221.0217038,50043,78852881.1412374,50037,86136,6390.517159,10044,29814,8020.752442,50047,95554551.1317255,20062,97677761.1412545,10043,36017400.9617345,70030,05115,6490.662673,70053,27820,4220.727441,80045,71139111.0912768,40067,5088920.9917563,20047,53515,6650.752844,10040,35237480.9217651,00036,74514,2550.722956,10042,92113,1790.777739,50048,99694961.2413051,30057,30860081.1217850,70073,13822,4381.4413152,10049,88322170.9617964,80061,39634040.9513270,10050,05520,0450.718044,50020,76323,7370.473340,80051,54310,7431.2618140,70043,10124011.0613433,87052,79118,9211.5618240,00058,01518,0151.4513564,20042,55421,6460.668342,00051,49594951.2313624,80051,12726,3272.0618443,60048,26846681.1113744,00050,13061301.1418552,60056,03234321.0713873,60044,73428,8660.618645,46045,8053451.0113957,20048,73384670.858760,10052,03680640.874039,10056,20517,1051.4418834,90037,71728171.0814143,30038,25750430.888959,30051,20180990.864236,28045,63593551.2619066,30039,76326,5370.604348,00043,64343570.9119160,90065,75748571.0814435,60022,77412,8260.649262,00045,23316,7670.734571,00047,49923,5010.679340,10044,46343631.1114646,00040,23057700.879451,10055,101-40011.0814755,00074,89919,8991.3619596,90052,97343,9270.554843,80050,33965391.1519644,32039,23150890.89RMSE15,791Success53MAE12,155Success rate55% Stores’ predicted actual turnover, and successful rate. The MTS was employed to reduce the number of variables to improve the ability to predict successful stores. Previous research suggests that when the training set accounts for 60%–80% of the total dataset, and the test set accounts for 20%–40%, the performance of the model can be fully demonstrated, and the model can have good and broad interpretability54,55. As shown in Table 4, 53 stores were defined as successful stores and 43 as unsuccessful stores. To match these figures, successful and unsuccessful stores were randomly assigned to the set used to train and establish the MTS model, and likewise to the test set (Table 5). Table 5The training and test sets.Training set (n, %)Test set (n, %)TotalSuccessful store39 (73.6%)14 (26.4%)53Unsuccessful store33 (76.7%)10 (23.2%)43Total72 (75%)24 (25%)96 The training and test sets. The four main stages of MTS application were as follows. Stage 1: Calculation of the Mahalanobis space. In this stage, the main goal was to identify the k variables used to determine the normal sample when its operational performance reaches the standard turnover. The original data for the 53 successful stores were used as the benchmark, and their mean value and standard deviation (Table 6) calculated to establish the Mahalanobis space (MS). Table 6Original data of successful stores (normal group).VariablesABCDE…ISamples11655512.8380…1.121101005.54100…1.13110010010.3100…1.1402010011.7680…1.15115512.73100…1.1⋮⋮⋮⋮⋮⋮⋮⋮5311010025.2380…1.1Average0.92452838.1132179.716989.1543483.77358…1.084906Standard error0.26415135.979928.374445.42675322.67293…0.045126 Original data of successful stores (normal group). Then, the original data of these 53 successful stores were standardized (Table 7), and the correlation coefficient matrix (R) and the matrix inverse (R−1) of the normal group were calculated using the standardized data (Eqs. 18 and 19). Table 7Standardized data and MD of successful stores (normal group).VariablesABCDE…ISMDSamples10.2857140.747273-0.87110.677322-0.16644…0.3344970.4286520.285714-0.781360.714834-0.666020.715673…0.3344970.44398230.2857141.7200380.7148340.2111130.715673…0.3344971.2806344-3.5-0.503430.7148340.480151-0.16644…0.3344971.84470650.285714-0.64239-2.633250.6588950.715673…0.3344971.361734⋮⋮⋮⋮⋮⋮⋮⋮⋮530.285714-0.781360.7148342.962298-0.16644…0.3344971.793922 Standardized data and MD of successful stores (normal group). Finally, the MD was calculated from data in the normal group and the inverse matrix. The MD forms the MS and serves as a reference for the complete measurement scale. Data are shown in the rightmost column of Table 7, calculated using Eqs. 20 and 21. Stage 2: Determination of the complete measurement scale framework. In this stage, the k variable values for each abnormal sample were used to calculate MDs. The mean and standard deviations of MS established using the normal samples were used to standardize values for the abnormal samples. Then, the MDs of these abnormal samples were calculated in accordance with the inverse correlation matrix. Finally, the threshold of MTS classification was set to distinguish normal samples from abnormal samples. According to the procedure described above, the mean value and standard deviation of the MS constructed in Stage 1 were used to standardize the data for the 33 unsuccessful stores in the training set, and the matrix inverse (R−1) was used to calculate the MDs. This calculation was performed to verify the accuracy of the measurement scale; if the measurement scale established in Stage 1 is accurate, the MD of the unsuccessful stores is relatively large. The MD distribution for the 39 successful stores and 33 unsuccessful stores in the training set is shown in Fig. 2. MD of the training set (complete model). To validate the reduced model measurement scale, the probabilistic thresholding method can be used to determine the threshold as the separation standard10,56. In this study, the mean MD of data from 39 successful stores was 1.03 (close to 1); most were close to 0.60, with only a few exceeding 0.74. Among the 33 unsuccessful stores, 18 MDs were less than 0.75. Therefore, as in previous studies, the probabilistic thresholding method was employed to set the threshold at 0.741 to obtain an improved performance indicator. As shown in Fig. 1, the MD of many unsuccessful stores was higher than the MD of successful stores; the separation between successful and unsuccessful stores is poor. In the follow-up study, variables were screened to identify those with minimal effect on distinguishing successful stores from unsuccessful stores. Then, the MDs of the test set were calculated in the same way as described for the training set, and their distribution was plotted (Fig. 3). The MDs of the unsuccessful stores were closer than those of the successful stores. MDs of the test set (complete model). Stage 3: Evaluation and screening of the impact of variables on each variable. In this stage, k characteristic variables are used as control factors, and two levels are set for each factor. Moreover, d or d groups of abnormal samples are used. The MD is calculated on the basis of the characteristic variables used in the orthogonal array, and the square root of the MD is taken as the measured value (y). Next, the SN ratio is calculated by selecting the larger-the-better characteristic and is used as the basis for each response value in the orthogonal array. Finally, which characteristic variables are important is determined in accordance with the variable impact of the SN ratio. Variables with positive impact are retained, and variables with negative impact are reduced. In accordance with the above results, 16 experiments are conducted by using the L16(29) orthogonal array, and each variable is divided into two levels. One variable is used at Level 1 and not used at Level 2. In each experiment, an MS is established on the basis of 43 unsuccessful stores. Then, the SN ratio is calculated by using these MSs (see Table 8). In this study, it was expected that the greater the distance between unsuccessful stores, the better. Therefore, the SN ratio adopts the larger-the-better characteristic. With Experiment 1 used as an example, the SN ratio was calculated as follows (Eq. 22): Table 8Configuration of the orthogonal array and calculation of the SN ratio.ExperimentABC…IMD1MD2…MD43SN ratio1111…10.964650.76334…0.74855-1.883972111…21.082960.78578…0.79218-2.127053111…11.174540.40168…0.93144-3.104234111…21.458950.60366…0.94561-3.804445122…10.504690.45382…0.35708-4.007636122…20.607480.43177…0.35331-8.214567122…10.394320.87976…0.31995-3.006998122…20.271841.10434…0.19083-3.990099212…21.267531.08346…1.05421-2.0139210212…11.371461.01735…1.13205-4.9105711212…21.236420.35348…1.20622-1.7531412212…11.225580.21442…1.24307-3.9521313221…20.690680.91690…0.47907-1.9476214221…10.486950.75929…0.34790-3.4464515221…20.504001.14107…0.33944-1.9120916221…10.316791.15093…0.17882-3.74339 Configuration of the orthogonal array and calculation of the SN ratio. After the SN ratio calculation, the variable impact of each variable is analyzed: The SN ratio of each experiment in Table 8 and the configuration of the orthogonal array are used to calculate , and of each variable. The impact of variable A was calculated as follows (Eqs. 23–25): Given that indicates that the use of this variable can reduce the impact of noise on the model, positive effect is taken as the screening standard in this study (such as variable B, C, F, G and H). If a variable increases noise, then it is excluded. Results for each variable are shown in Table 9. The impact gap is clear; the impact ranking of the variables is H > C > G > B > F > E > I > D > A. At the end of this screening stage, five variables were retained: B, C, F, G, and H. Table 9Variable Impact.VariablesEffectKeepA-3.76737-2.95991-0.80746B-2.94368-3.78360.839923▲C-2.74616-3.981131.234972▲D-3.56897-2.94864-0.62033E-3.44925-3.27804-0.17121F-2.94851-3.778770.830264▲G-2.76343-3.963861.200426▲H-2.4537-4.273581.819884▲I-3.50692-3.22036-0.28656 Variable Impact. Stage 4: Future diagnosis and prediction. In this stage, the measurement scale of the reduced model was confirmed. Then, the variables determined to be important were reconstructed, and the MDs of the abnormal samples used to confirm the scale. In addition, a cluster test was conducted to verify whether the reduced model had sufficient judgment ability. The new predicted turnover obtained through the test set was compared with the actual turnover. A predicted turnover of at least 90% of the actual turnover was regarded as correct, and the model’s overall prediction accuracy was calculated. Using the five variables retained after screening in the previous stage, a new MS was calculated on the basis of the data of successful stores in Stage 1 and their inverse matrix. The reduced measurement scale was established and used to confirm the data of 33 unsuccessful stores (i.e., the MD of the training set was calculated retrospectively). Then, the threshold for prediction of a successful store was set to 1 using the probabilistic thresholding method. According to the MD distribution diagram of the reduced training set (Fig. 4), the MDs of normal samples were all below 2.5 after the variables were reduced, although the MDs of the reduced model still partially overlapped with those of abnormal samples. Moreover, the MDs of abnormal samples were larger than those of the complete model, which improves the degree of distinguishing anomaly in part. In Fig. 4, the x-axis is labeled “Store Samples,” the y-axis is labeled “Mahalanobis Distance (MD),” and the legend differentiates Successful Stores (Normal) from Unsuccessful Stores (Abnormal). A dashed line indicates the threshold (MD = 1). MDs of the training set (reduced model). As shown in the MD distribution diagram of the test set (Fig. 5), the reduced model distinguished successful stores from unsuccessful stores more clearly than the complete model, and identified abnormal states more accurately. The model’s effect was maintained even after variables with low impact were removed. In Fig. 5, the same axis labels and legend are applied, showing clearer separation between groups and confirming that the reduced model improves anomaly detection while preserving predictive accuracy. MDs of the testing set (reduced model). Classification prediction using machine learning algorithms XGBoost, RF and SVM were used to establish a classification prediction model, verified with the training set and the test set, accounting for 60–80% and 20–40% of the total dataset, respectively54,55. This study employed grid search combined with tenfold cross-validation to explore parameter spaces systematically. This ensures that selected hyperparameters generalize well beyond the training dataset and reduces the risk of overfitting. The classification prediction results of MTS-XGBoost are shown in Table 10. The accuracy of complete variable models ranged from 61.83 to 68.75%, while the accuracy of filtered variables ranged from 70.00 to 76.25%. Examination of the results of 12 different experimental combinations revealed that the accuracy of each variable after screening is higher than that of the complete variable; the maximum improvement was 20.22%. Table 10Classification prediction results of MTS-XGBoost.Training set proportionCorrect number/sample numberComplete variable model prediction accuracy (A) (%)RMSECorrect number/sample numberScreening Variable model prediction accuracy (B) (%)RMSEAccuracy improvement percentage [(B)–(A)]/(B) (%)75%160/24066.6713,627181/24075.4212,06413.12162/24067.5012,815178/24074.1712,2919.88159/24066.2512,711180/24075.0012,11413.21158/24065.8312,751183/24076.2511,84215.8370%178/29061.3813,591214/29073.7912,71320.22189/29065.1713,300210/29072.4112,73311.11196/29067.5913,896221/29076.2111,91412.75174/29060.0013,032202/29070.0013,36516.6767%213/32066.5613,624242/32075.6312,06413.63207/32064.6914,005240/32075.0012,51215.94220/32068.7514,002231/32072.1912,1445.00213/32066.5612,785229/32071.5612,6927.51 Classification prediction results of MTS-XGBoost. The classification prediction results of MTS-RF are shown in Table 11. The accuracy of complete variable models ranged from 64.83 to 72.92%, while the accuracy of filtered variables ranged from 70.08 to 77.19%. Examination of the results of 12 different experimental combinations revealed that the accuracy of each variable after screening is higher than that of the complete variable; the maximum improvement was 15.95%. Table 11Classification prediction results of MTS-RF.Training set proportionCorrect number/sample numberComplete variable model prediction accuracy (A) (%)RMSECorrect number/sample numberScreening variable model prediction accuracy (B) (%)RMSEAccuracy improvement percentage [(B)–(A)]/(B) (%)75%171/24071.2511,545174/24072.5012,3751.75175/24072.9212,775178/24074.1712,8241.71165/24068.7512,572184/24076.6712,27211.52173/24072.0811,171173/24072.0812,8130.0070%204/29070.3412,931212/29073.4512,0004.42215/29074.1412,619221/29076.2112,0982.79188/29064.8311,830218/29075.1712,07115.95206/29071.0312,536223/29076.9013,1088.2667%228/32071.2512,071232/32072.5012,3921.75224/32070.0011,756240/32075.0012,5817.14214/32066.8812,552232/32072.5012,1748.40220/32068.7512,595247/32077.1912,73812.28 Classification prediction results of MTS-RF. The classification prediction results of MTS-SVM are shown in Table 12. The accuracy of complete variable models ranged from 62.45 to 71.22%, while the accuracy of filtered variables ranged from 71.56 to 74.88%. Examination of the results of 12 different experimental combinations revealed that the accuracy of each variable after screening is higher than that of the complete variable; the maximum improvement was 15.82%. Table 12Classification prediction results of MTS-SVM.Training set proportionCorrect number/sample numberComplete variable model prediction accuracy (A) (%)RMSECorrect number/sample numberScreening variable model prediction accuracy (B) (%)RMSEAccuracy improvement percentage [(B)–(A)]/(B) (%)75%165/24068.3112,341179/24073.6513,6957.25163/24070.1212,561181/24072.8414,0443.73161/24071.2212,352180/24074.2013,6484.02162/24069.5312,055176/24075.1513,5567.4870%191/29066.5113,201213/29074.8814,08511.18202/29068.3212,954220/29072.5413,9925.82197/29067.2112,223219/29073.4814,3688.53206/29062.4512,877218/29074.1914,76715.8267%218/32068.5512,567238/32071.5613,2344.21217/32067.9312,672241/32073.9813,5458.18217/32069.7313,269239/32075.4512,9877.58219/32065.9012,741233/32072.6512,9529.29 Classification prediction results of MTS-SVM. After using MTS to reduce the number of variables, the classification prediction accuracy of the XGBoost algorithm improved substantially. When training with 75% of the dataset, the prediction accuracy of decision trees improved by 9.88–15.83%; with 70% data, prediction accuracy improved by 11.11–20.22%; with 67% data, prediction accuracy improved by 5.00–15.94%. These results imply that the five variables identified in this study have significant influence on prediction accuracy. Moreover, after using the MTS to reduce the number of variables, the reduced model has a practical distinguishing effect, verifying that the MTS-XGBoost is an appropriate and effective method of obtaining a parsimonious location selection prediction model. Tables 13 show that MTS-XGBoost achieves the best performance. This is because MTS removes noise variables, allowing the model to focus on key location attributes, while XGBoost’s boosting and regularization are able to capture complex nonlinear patterns without overfitting. As a result, this model’s predictions are more reliable than those of RF and SVM. For decision makers, this means that location selection can be guided by clear evidence of key factors, thereby reducing risk and improving investment decisions. Table 13Comparative performance of MTS-XGBoost, MTS-RF, and MTS-SVM.AlgorithmAccuracy (range)Accuracy improvement percentage (average) (%)RMSE (average)Execution timeDiscussion of accuracy and computational complexityMTS-XGBoost70.00–76.25%12.9112,371ModerateSuperior accuracy and efficiency; best suited for nonlinear, high-dimensional data. Execution time longer than RFMTS-RF72.08–77.19%6.3312,454LowestStrong stability and interpretability; Execution time is better than XGBoost and SVMMTS-SVM71.56–75.45%7.7613,739HighestRobust in small-to-medium samples; accuracy competitive but highest execution time than XGBoost and RF Comparative performance of MTS-XGBoost, MTS-RF, and MTS-SVM. Discussion and conclusions In the case study, the MTS was used to analyze nine influencing factors of convenience store location selection to optimize the location selection process. The factors were ranked, in descending order of influence, as follows: visibility (H) > supermarket competitiveness (C) > store location (G) > convenience store competitiveness (B) > number of residents (F) > main customer type (E) > store area (I) > traffic volume (D) > business district type (A). Entrepreneurs and convenience store chains can use this ranking to improve store location decisions, or rely on the reduced set of five factors identified as most important in this study (visibility, supermarket competitiveness, store location, convenience store competitiveness, number of residents), improving the efficiency of store location selection. The prediction model of new store location selection established in this study can reduce interference from subjective factors. From a strategic perspective, this ranking helps retailers prioritize investments in highly visible sites and systematically evaluate competitive intensity before considering secondary criteria such as store area or traffic flow. By focusing on the top-ranked factors, firms can reduce data collection costs, accelerate market entry, and align expansion strategies with consumer demand, thereby minimizing risks associated with subjective judgment or trial-and-error decision-making. This study incorporates machine learning algorithms and verifies the effectiveness of MTS-XGBoost, MTS-RF, and MTS-SVM in the convenience store location selection process. Results show that after variable screening and reduction by MTS-XGBoost, the accuracy can be improved by up to 20.22%. When the proportion of training data is 75%, 70% or 67%, the accuracy of the classification model reaches 76.25%, 76.21% and 75.63% respectively. Obviously, after MTS performed variable screening, the predictive ability of the model did not decrease, but was steadily improved. The MTS performs well when dealing with numerous variables, and can effectively identify variables with insignificant impact on the prediction target. It optimizes the store location selection process and resource allocation, saves time in searching and analysis, reduces the cost of data collection and processing, and minimizes economic losses due to improper store location selection. The method offers substantial benefits for convenience store developers. Limitations and future development Although the MTS method performed well in this study of convenience store location selection in Taiwan, its sensitivity to abnormal values and its stability under extreme market conditions require further investigation. On the basis of the results presented herein, the following research directions are suggested. Cross-regional application. Expand the scope of research to other countries and markets. Variations in shopping culture and habits mean research results for different regions must be compared to test the applicability and stability of the MTS method under diverse conditions. In markets with different retail density, mobility patterns, or regulatory contexts, key variables may shift, so the model’s accuracy and rankings require region-specific calibration. Cross-regional application. Expand the scope of research to other countries and markets. Variations in shopping culture and habits mean research results for different regions must be compared to test the applicability and stability of the MTS method under diverse conditions. In markets with different retail density, mobility patterns, or regulatory contexts, key variables may shift, so the model’s accuracy and rankings require region-specific calibration. Technology integration. Researchers should explore ways to integrate the MTS method with other advanced data analysis technologies, such as deep learning or reinforcement learning, to achieve further improvement in the accuracy and efficiency of location selection decisions. Technology integration. Researchers should explore ways to integrate the MTS method with other advanced data analysis technologies, such as deep learning or reinforcement learning, to achieve further improvement in the accuracy and efficiency of location selection decisions. Application extension. Researchers should extend the methods presented here to other types of retail businesses, such as department stores and specialty stores, to evaluate the effect and potential of the MTS method in these retail models. Application extension. Researchers should extend the methods presented here to other types of retail businesses, such as department stores and specialty stores, to evaluate the effect and potential of the MTS method in these retail models. Model transferability. The model’s transferability is limited, as it does not account for market differences, external shocks, or spatial variables, requiring further calibration for broader applicability. Model transferability. The model’s transferability is limited, as it does not account for market differences, external shocks, or spatial variables, requiring further calibration for broader applicability. Acknowledgements This research was funded by the National Science and Technology Council, Taiwan, R.O.C. Author contributions Tsung-Yin is the main author of this article and made the greatest contribution. Hsin-Pin provided important concepts and many suggestions for this article and was an important provider of innovative ideas for this study. Mei-Zhen assisted in validating the research process of this paper and completed many figures. Data availability The datasets used and/or analyzed during the current study available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets used and/or analyzed during the current study available from the corresponding author on reasonable request.