Detecting Relevant Feature Interactions for Recommender Systems via Graph Neural Networks

Feature interactions are essential for achieving high accuracy in recommender systems (RS), so they have been taken into consideration in many existing RS, where all feature interactions are modeled. Nevertheless, not all feature interactions have positive effects for RS: modeling the irrelevant feature interactions may introduce noises and degrade the accuracy. To overcome this problem, in this work, we propose a graph neural network-based model, L0-SIGN, to detect the relevance of feature interactions and utilize only the relevant ones for RS, with features as nodes and feature interactions as edges. Generally, our model consists of two components: an L0 regularization based edge prediction model to explicitly detect relevant feature interactions; and a graph classification model, SIGN, to effectively model and aggregate the detected ones for recommendations. These two components positively influence each other to ensure that the most relevant feature interactions will be detected and modeled. In addition, we further prove that the effectiveness of our model is theoretically sound. We first show that our model is a variational approximation of information bottleneck principle, i.e., the detected feature interactions are guaranteed to be most relevant. We then show that our model follows the definition of statistical interactions, proving that the modeling of detected feature interactions in L0-SIGN is effective. Experimental results show that (i) L0-SIGN outperforms existing baselines in terms of accuracy, and (ii) the detected feature interactions are beneficial for performance gain and interpretability.