Causal factorization machine for robust recommendation

Abstract

Factorization Machines (FMs) are widely used for the collaborative recommendation because of their effectiveness and flexibility in feature interaction modeling. Previous FM-based works have claimed the importance of selecting useful features since incorporating unnecessary features will introduce noise and reduce the recommendation performance. However, previous feature selection algorithms for FMs are proposed based on the i.i.d. hypothesis and select features according to their importance to the predictive accuracy on training data. However, the i.i.d. assumption is often violated in real-world applications, and shifts between training and testing sets may exist. In this paper, we consider achieving causal feature selection in FMs so as to enhance the robustness of recommendation when the distributions of training data and testing data are different. What's more, different from other machine learning tasks like image classification, which usually select a global set of causal features for a predictive model, we emphasize the importance of considering personalized causal feature selection in recommendation scenarios since the causal features for different users may be different. To achieve our goal, we propose a personalized feature selection method for FMs and refer to the confounder balancing approach to balance the confounders for every treatment feature. We conduct experiments on three real-world datasets and compare our method with both representative shallow and deep FM-based baselines to show the effectiveness of our method in enhancing the robustness of recommendations and improving the recommendation accuracy.