FCP Filter: A Dynamic Clustering-Prediction Framework for Customer Behavior

Abstract

Customer purchase behavior prediction plays an important role in modern retailing, but the performance of this task is often limited by the randomness of individual historic transaction data. In the meanwhile, Fragmentation and Coagulation Process (FCP), a stochastic partition model, has recently been proposed for identifying dynamic customer groups and modeling their purchase behavior. However, FCP is not able to forecast the purchase behavior because such a data-driven method requires transaction observations to conduct clustering. To tackle this challenge, we propose FCP filter, a clustering-prediction framework based on FCP, which can forecast purchase behavior and filter random noise of individual transaction data. In our model, FCP clusters customers into groups by their temporal interests to filter random noise of individual transaction data. Then a predictor is built on grouped data. The predicted results are also fed to FCP to adjust the parameter for prior knowledge at the next time step. Our model is superior in capturing temporal dynamics and having flexible number of groups. We conduct experiments on both synthetic and real-world datasets, demonstrating that our model is able to discover the latent group of individual customers and provides accurate predictions for dynamic purchase behavior.