An interpretable data-driven approach for customer purchase prediction using cost-sensitive learning

Customer purchase prediction can not only identify potential consumers but also help companies carry out related marketing activities based on the prediction results. In the field of marketing, ‘Buy till you die’ (BTYD) models are perhaps the most representative techniques for customer purchase prediction. Unlike BTYD models that rely more on probability distributions and a small number of input variables, machine learning models are a data-driven approach to extract information from a large number of input variables. To explore whether these two different classes of models can cross-fertilize each other, this paper incorporates Beta Geometric/Negative Binomial Distribution (BG/NBD) model into extreme gradient boosting (XGB) model to obtain the customer purchase predictions. To optimize profit-oriented decision-making, this study develops a profit-oriented XGB (POXGB) approach. This approach involves training a prediction model with the objective of directly maximizing the profit of downstream decision-making tasks. We conduct an empirical assessment of the prediction performance of BTYD modeling and machine learning with a real-world travel service purchase dataset. The integrative model shows a significant improvement over POXGB. This research contributes to the purchase forecasting literature and practical implications.