Spectral-Based Graph Neural Networks for Complementary Item Recommendation

Modeling complementary relationships greatly helps recommender systems toaccurately and promptly recommend the subsequent items when one item ispurchased. Unlike traditional similar relationships, items with complementaryrelationships may be purchased successively (such as iPhone and Airpods Pro),and they not only share relevance but also exhibit dissimilarity. Since the twoattributes are opposites, modeling complementary relationships is challenging.Previous attempts to exploit these relationships have either ignored oroversimplified the dissimilarity attribute, resulting in ineffective modelingand an inability to balance the two attributes. Since Graph Neural Networks(GNNs) can capture the relevance and dissimilarity between nodes in thespectral domain, we can leverage spectral-based GNNs to effectively understandand model complementary relationships. In this study, we present a novelapproach called Spectral-based Complementary Graph Neural Networks (SComGNN)that utilizes the spectral properties of complementary item graphs. We make thefirst observation that complementary relationships consist of low-frequency andmid-frequency components, corresponding to the relevance and dissimilarityattributes, respectively. Based on this spectral observation, we designspectral graph convolutional networks with low-pass and mid-pass filters tocapture the low-frequency and mid-frequency components. Additionally, wepropose a two-stage attention mechanism to adaptively integrate and balance thetwo attributes. Experimental results on four e-commerce datasets demonstratethe effectiveness of our model, with SComGNN significantly outperformingexisting baseline models.