Improving prediction performance of customer behavior

Abstract

Increasing competition in business and newly emerging information technology have led to the development of concepts that focus on nurturing the relationship with customers. The challenge of achieving a good relationship with customers is based on service which is individualized for each customer’s preference, and this naturally requires obtaining “customer information” and developing information-based business strategies. The quality of customer information is a very important factor for developing such customer-related strategies. ‘Completeness’ is one of the fundamental data quality components. Incomplete data can yield biased analysis results and undermine the accuracy of predictive models for customer behavior, which may lead to inappropriate decision-making processes and the development of wrong marketing or risk strategies. This study is organized in two parts and investigates the incomplete data problem and suggests a methodology to improve the prediction performance of customer behavior in two customer-related strategies: Customer Relation Management (CRM) in Part I and Credit Risk Management in Part II. In Part I, we examine the incomplete data problem caused by missing values. To solve the problem, we suggest a novel approach to integrate multiple-imputation (MI) and bootstrap-imputation methods into the conventional methodology. Two experimental works are applied to real cases in electronic customer relationship management (e-CRM) domain. The first predicts the customer purchase likelihood in an online shopping mall with missing covariates and the second builds a collaborative filtering system in the case of missing target values. According to the results, the MI-based predictive models perform better than the traditional approach in both of the case studies, especially in the dataset with a high missing rate compared to the one with a low missing rate. Bootstrap-based approaches also produce better results than the traditional approach. However, t...