Answers are grounded in this paper and cite their sources.
Applying graph convolutional networks (GCNs) to recommender systems can provide more relevant items to each user. However, it is important to provide not only items that are relevant to each user's preferences but also items that are diverse and novel on e-commerce platforms. Users often prefer a variety of recommendations as well as ordinary recommendations obtained from their own search behaviors. This is achieved by LightGCN, a GCN-based recommender system that utilizes a multi-layer aggregation function and adjacency matrix to learn the latent vectors of users and items. However, LightGCN often provides recommendations without diversity when the number of layers is insufficient. On the other hand, when it is excessive, the accuracy is not acceptable, which is often referred to as the over-smoothing problem. To address this problem, we propose Generalized Personalized PageRank (GPPR) and Adjacency Personalized PageRank (APPR). GPPR is a generalized model of Personalized PageRank that can be applied to any graphs. In addition, APPR enhances Personalized PageRank, which is a specialized version of GPPR. We incorporate our approach into LightGCN and demonstrate that it provides more relevant and diverse recommendations, outperforming LightGCN.
No takes yet. Share an insight, caveat, or question.
Wayama et al. (2025) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: