HIN-based rating prediction in recommender systems via GCN and meta-learning

Rating prediction is a crucial task for recommender systems, but it has the problem of difficulty in quickly capturing user preference transfer and cold-start problem. Thus, this paper proposes the meta-learning-based rating prediction model for heterogeneous information networks (HIN) called Meta-HRP (HIN-based Rating Prediction) to solve these problems. The model first constructs meta-tasks through meta-paths on HIN and then constructs an embedding representation generator based on graph convolutional network (GCN) and attention mechanism to generate embeddings for users and items. Then the proposed rating prediction meta-learner leverages historical interaction data to learn prior knowledge and rapidly adapts to new items based on a few recent user rating records to timely capture user preference transfer and alleviate the cold-start problem. We validate Meta-HRP with extensive experiments, and the proposed model reduces root mean square error by at least 8.49%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\%$$\end{document} on average over the baselines on two public benchmark datasets. Furthermore, Meta-HRP outperforms the state-of-the-arts in most cold-start cases.