A New Meta-path Monte Carlo Tree Search Algorithm for Heterogeneous Graph Neural Networks

Heterogeneous graph neural networks (HGNN) can capture heterogeneous semantic information in heterogeneous networks, learn the low-dimensional embedding vectors, and use them for downstream tasks. The selection of meta-paths is always the focus of HGNN. Existing HGNN models often employ random selections of meta-paths or utilize all meta-paths with a fixed maximum number of hops, thereby overlooking significant heterogeneous semantic information of graphs and struggling to effectively leverage non-redundant information. To this end, a new Monte Carlo tree search-based heterogeneous graph neural network (MCTS-HGNN) model is developed to search for the appropriate set of meta-paths in heterogeneous graphs automatically, thus overcoming the difficulty of meta-path selection. Subsequently, the meta-path set is decomposed based on aggregation objects and independently applied to a subset of meta-paths by using a customized transformer-based semantic aggregation module, and then the diverse semantic information from meta-paths can be effectively utilized. Furthermore, the information from the meta-path subset is integrated by the graph-level transformer to achieve a comprehensive heterogeneous graph embedding. The learned embedding is evaluated via the downstream task of the heterogeneous graph. Finally, the ablation experiments validate the effectiveness of the module designed for the MCTS-HGNN. The experimental results demonstrate that the MCTS-HGNN outperforms state-of-the-art baselines across all evaluation metrics.