Decision tree and rule-based classification for predicting online purchase behavior in Malaysia / Maslina Abdul Aziz, Nurul Ain Mustakim and Shuzlina Abdul Rahman

In Malaysia, fast growth in e-commerce speeds a business need to understand and predict consumer online behavior in order to be more competitive. While the whole world is embracing big data analytics, many businesses in Malaysia, particularly those in the ecommerce sector, find it hard to harness these technologies to their benefit. The absence of specific predictive models and the complexity of socio-cultural diversity further complicate the efforts toward understanding consumer preferences. Therefore, this research tries to fill in some of the gaps by applying decision tree and rule-based algorithms to classify online purchasing behavior amongst Malaysian consumers. The study looks into the data from an online survey comprising 560 respondents with a view to demographic, factors influences, and purchasing behaviour. The performance of six machine learning models comprising J48, Random Tree, REPTree representing decision trees and JRip, PART, and OneR as rule-based algorithms was assessed. Feature selection, pre-processing, and SMOTE were applied in order to balance class inequalities of the dataset. The result indicated that the highest accuracy of 89.34% was achieved by the Random Tree algorithm, while the rule-based algorithm PART reached an accuracy of 87.56%. Results of these models open up the possibility of providing very important insights from a business perspective into consumer behaviour and thus offer actionable data which allows them to complete their job of finetuning marketing strategies and engaging customers. The current study contributes to the literature by highlighting decision tree and rule-based classification models as very useful in the Malaysian e-commerce context. These developed predictive models can serve as building blocks where businesses might know more about consumer behavior, personalize marketing, and reach operationally efficient levels. Future research may involve integrating other influencing variables and applying them across industries.