Graph Neural Networks for Multi-modal Recommendation: A Bipartite Heterogeneous Approach

Abstract In the era of big data and information technology, the exponential growth in data volume has significantly enriched people's lives, meeting their ever-expanding needs. However, this deluge of data has also led to a proliferation of invalid information, posing a considerable challenge in people's daily lives. Graph data has emerged as a focal point in data mining research, offering a powerful framework for organizing and analyzing complex relationships within datasets. Graphs, comprised of nodes and edges, serve as versatile structures for representing diverse datasets. This paper proposes the use of graph neural network techniques to address the challenging cold start problem while preserving the model's ability to capture and understand node content features. In recommender systems, the user-item interaction graph typically takes the form of a bipartite heterogeneous graph, where user nodes and item nodes belong to distinct domains, each with potentially different initial feature dimensions. This inherent heterogeneity necessitates the development of specialized techniques tailored to bipartite heterogeneous graph inference. The paper presents a recommendation method based on graph neural networks and driven by the attention mechanism. By incorporating multidimensional representation vectors of users, items, and ratings, the proposed approach aims to improve recommendation quality, particularly when dealing with sparse rating matrices. Additionally, we introduce subgraph extraction to diversify training graph topology by sampling subgraphs from large graphs. Experimental results demonstrate that the recommendation model proposed in this paper outperforms several mainstream baselines while effectively addressing the cold start problem.