Publication: Data mining for service industry: branch classification of a malaysia retail bank
No Thumbnail Available
Date
2023-01-01
Authors
Chong Siu Hou
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The on-going customers’ behavioural shifts from assessing physical banking services to virtual banking services have made retail banks continually review their physical banking network strategies. Such exercise allows the retail banks to maintain their competitive advantages over their competitors while balancing their business portfolios between online banking service needs, optimizing physical clientele coverage, and minimizing costs in running a physical branch network. The research objective is to develop a suitable classification model in decision-making that can mimic experts’ opinions on retail bank branch operational status using a case study on a Malaysian retail bank. Seventy-four branch attributes were used to build a classification model in which experts’ opinions are included as an attribute to allow the data mining (DM) application to emulate experts’ decision-making. The research methodology involves collecting the experts’ opinion through an expert survey to determine the feasibility of a branch remaining open or closing permanently as part of the bank’s network rationalization strategy. The value obtained from the survey was converted into a variable termed “aggregate closure possibility” in this study and then discretized into target class datasets. Then ten classification models were constructed, including ZeroR, Decision Tree, Random Forest, Naïve Bayes etc. Next, three experiments were conducted. The first experiment examined the effect of hyperparameter tuning on classification accuracy. The second experiment studied the impact of changes in the number of target classes on classifier accuracy and then highlighted the best-performing classifier. When the number of target classes was optimized, the accuracy of classifiers improved drastically, of which, three classifiers (Decision Table, J48, and JRip) were the most accurate classifiers. The third experiment studied the effect of attribute selection and instance reduction in optimizing classification performance. It was observed that the accuracy of most classifiers improved proportionate to the reduction of numbers of attributes and instances where Decision Table, J48, BayesNet, and JRip were the best performing classifiers in the third experiment. The result shows that JRip was the best performing classifier in all three experiments. Attribute selection and instance reduction optimized the classification accuracy in encompassing efficient datasets. The study also showed that the DM application could emulate expert opinion in decision-making in retail banking sector. In conclusion, the study has successfully achieved its research objectives. The study contributes to research knowledge on how DM capabilities could improve management decision-making in a real-world application.