Confounder Balance in Next Basket Prediction

Next basket prediction (NBP) is essential for online businesses such as grocery shopping and retail, as it aims to learn users' interests based on historical basket sequences and predict a set of items for the next purchase. Existing methods tend to prioritize items with high interaction frequencies in the dataset. In this article, we use causal analysis to theoretically demonstrate that such methods can lead to biased results. Specifically, item interactions in the dataset represent item-specific weights that reflect average user preferences. However, these weights are imprecise due to varying user interest levels, leading to biased learning results and suboptimal predictions. We find that repeated interactions between a user and items can represent the user's level of interest, which can be leveraged to improve user interest modeling by assigning user-specific weights. In addition, different users exhibit varying preferences for items with high interaction frequencies, highlighting the necessity of different degrees of bias mitigation. Making personalized adjustments based on these differences can further refine existing methods. Consequently, we propose a simple yet effective confounder balance prediction (CBP) model to mitigate bias while preserving individual user interests. Specifically, we employ counterfactual inference to construct a counterfactual world in which predictions are influenced solely by either high interaction frequencies or user interests, both of which are potential confounders. This approach enables us to individually assess the impact of these confounders on interaction probabilities. The goal of CBP is to balance the confounders and thereby refine the learning process without resorting to oversimplified unbiased learning. Experiments on four widely used real-world datasets demonstrate the significant advantages of CBP over existing state-of-the-art methods. Our code is available at https://github.com/Hiiizhy/CBP.