Geometric Matrix Completion with Collaborative Routing Between Capsules

Abstract

Collaborative filtering (CF) is the most common application of matrix completion, where the user-item rating matrix can be formulated as a bipartite graph. Due to the power of learning on graphs, graph neural networks (GNNs) have been widely adapted to the user-item graph. However, existing graph-based models mainly focus on modeling implicit feedback, while explicit feedback is not fully exploited where the geometric relationships between users and items are latent. Inspired by routing between capsules, we devise a routing algorithm Collaborative Routing for CF to capture these relationships. A novel graph-based matrix completion model based on Collaborative Routing (CRMC) is proposed, which can exploit the user-item graph from both implicit and explicit perspectives to learn better representations, and preserve the relationships between users and items latent in the user-item graph. Experimental results on three real-world datasets demonstrate the effectiveness of our model.