The LSTM-EMPG Model for Next Basket Recommendation in E-commerce

Abstract

Personalized recommendations play a crucial role in the modern e-commerce landscape, enabling businesses to meet customers' evolving preferences and boost sales. As customer preferences change, businesses are realizing the importance of suggesting what customers might want to buy next. However, existing approaches face challenges in capturing sequential patterns in user behavior and accurately utilizing previous purchase information. These challenges can be addressed using Long Short-Term Memory Networks (LSTMs). Nevertheless, LSTMs alone may not fully capture users' repetitive purchase behavior or consider the exact timing of purchases. To account for these limitations, Probabilistic Models such as the Modified Poisson Gamma model (MPG) can be employed. The research reported in this paper proposes and investigates a new approach for the next basket recommendation based on the integration of LSTM with an enhanced Modified Poisson Gamma model to enhance next basket recommendation accuracy in e-commerce. The enhanced model (EMPG) includes a refinement of the MPG model to increase its predictive accuracy, and its recommendations are then integrated with an LSTM network to optimize the LSTM’s predictions. The proposed hybrid LSTM-EMPG model has been evaluated on the Instacart dataset and has produced superior results compared to the Multi-period LSTM, the GRU-based model. DREAM (RNN), and DREAM (LSTM) in terms of predictive accuracy, achieving a higher precision and recall.