행렬 완성 개선을 위한 점진적 메시지 패싱 기법

Abstract

Matrix completion is an important area of research in recommender systems. Recent methods view a rating matrix as a user-item bipartite graph with labeled edges denoting observed ratings and predict the edges between the user and item nodes by using the graph neural network (GNN). By exploiting the high-order connectivity in the graph, these GNN-based methods have achieved great success. However, they treat each rating type as an independent relation type and thus cannot sufficiently consider the ordinal nature of the ratings. This paper proposes Progressive Message Passing(PMP), a new learning paradigm that effectively models user preference levels by utilizing message passing that takes into account the ordinal nature of ratings. In specific, we decompose the given graph into a set of unweighted graphs so that each decomposed graph includes progressively stronger relations from users’ general interest to high preferences. Then, we apply GNN to each decomposed graph with different preference levels, while enabling them to share the same semantics via interest regularization. Also, we apply self-supervised learning to progressively update the adjacency matrix. Our experiments show that PMP effectively utilizes the rating type information and consistently outperforms all baseline methods, and ablation studies validate the effectiveness of each component.