The cold-start problem is a fundamental challenge in recommendation. Heterogeneous information networks (HINs) provide rich side information in addition to sparse user-item interactions, which can be used to alleviate the cold-start problem. However, most existing models based on graph neural networks (GNNs) only consider the user-item interactions as supervision signals, making them unable to effectively exploit the side information. In this paper, we propose a novel pre-training model, named MHGP, for cold-start recommendation in a self-supervised manner. The key idea is to leverage the mixed-order information in a HIN. We first use GNNs with a hierarchical attention mechanism to encode the first-order and high-order structures of a user-item HIN. Then, we pre-train the embeddings of users and items by contrasting the two structure views and maximizing the agreement of positive samples in each view. Afterwards, the embeddings are fine-tuned together with the recommendation model. Experiments show that our model can consistently improve the performance of cold-start recommendation and outperform other state-of-the-art pre-training models.
CITATION STYLE
Sui, W., Jiang, X., Ge, W., & Hu, W. (2023). Mixed-Order Heterogeneous Graph Pre-training for Cold-Start Recommendation. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 13423 LNCS, pp. 182–190). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-3-031-25201-3_14
Mendeley helps you to discover research relevant for your work.