Repurchase Prediction Using Survival Ensembles in CRM Systems for Home Appliance Business

In order for company’s promotions to continue to have a beneficial impact on sales, it is important for companies to identify which of the interested buyers can be converted into repeat buyers. By targeting these potential loyal customers, companies can significantly reduce promotional costs and increase return on investment. The existing studies related to repurchase prediction in the e-commerce area have focused on the statistical techniques and more common binary classification models. In this paper, we propose a survival analysis-based machine learning/deep learning model to predict TV repurchase time of customers using home appliance company’s CRM data. The prediction model is verified based on actual operational data such as customer profile, purchase, counseling, and repair history for approximately 1.45 million customers in electronics company’s CRM. As a deep learning method, Algo 6–1 (DeepHit with the feature set selected from Cox regression and preprocessed with multiple imputation) achieved the best performance (c-index 0.828). Algo3 (Random Survival Forest with the feature set selected from Cox regression and preprocessed with multiple imputation), a machine learning method, not only showed similar performance to deep learning (c-index 0.823), but also provided insights in key features that influenced repurchase. In addition, we provided a utility function that provides TV repurchase probability over time so that marketers can cost-effectively determine the timing to provide promotional events or benefits to customers.