Mixed-Order Heterogeneous Graph Pre-training for Cold-Start Recommendation

Abstract

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.