MCCR: Learning Multi-order Convolutional Correlations for Recommendation

Graph Neural Networks (GNNs) has been widely used to address the sparsity and cold start problems in recommendation system. By propagating embeddings from multi-hop neighbor nodes among the interaction graph and update target user and item embeddings, GNNs-based methods can achieve better recommendation performance. But those methods directly concatenate the output of each layer and ignore the different influences between different layers, and they simply use the inner product of the user and item’s embeddings to calculate the similarity and make recommendation based on it, which is insufficient to reveal the complex and nonlinear interactions.In this work, we propose to learn multi-order interactions between users and items and capture correlations between different-order information. We design a new recommendation framework MCCR, which treats each layer’s output as differentorder feature, and propose a multi-order interaction module to represent feature interactions. We adopt a multi-layer 3D CNN module to learn high-order interaction signals between users and items in an explicit approach. Through extensive experiments on three real-world datasets, which shows that MCCR evidently outperforms the state-of-the-art methods consistently.