A personalized recommendation framework through exploiting jump-enhanced random walk based multiple heterogeneous graph neural networks

Abstract

Due to the powerful representation ability to learn the embedding of each node in heterogenous graph (HG), heterogenous graph neural network (HGNN) based personalized recommender can effectively alleviate the notorious issues of user-item interaction sparsity and cold-start in recommendation systems. However, the existing schemes always rely on meta-paths and/or random walks for generating embeddings of nodes in HG. However, the former requires prior domain knowledge to determine the optimal meta-paths, and the latter will bias to the high-degree nodes in HG. To overcome these issues, this paper proposes a novel personalized recommendation framework, MHRec, based on multiple heterogeneous sub-graphs generated by jump-enhanced random walk (JerW). Specifically, our work’s contributions are following. First, the whole HG is explicitly constructed, which not only naturally includes multiple type nodes, i.e., user, item, user attribute, item attribute, and their connections, but also explicitly adds the user-user and item-item edges based on their interactively historical data. Then, starting from each node as ego, JerW is used to construct multiple heterogeneous sub-graphs for the ego, which can balance the distribution of different types of nodes in the formed sub-graphs, and appropriately model the multiple relationships between the ego and its multiple-hop neighboring nodes. Second, on each heterogeneous sub-graph, hierarchical graph representation is designed to formulate the ego’s representation, which is explicitly composed of same-type and cross-type aggregation using GNN with multi-head attention mechanism. Thorough experiments on multiple real-world datasets demonstrate our proposed MHRec outperforms state-of-the-art HGNN based personalized recommendation schemes, in terms of multiple evaluation metrics.