Adaptive heterogeneous graph reasoning for relational understanding in interconnected systems

Traditional graph neural networks (GNNs) have proven effective on homogeneous graph data. However, due to the increased complexity and diversity in the structure and node features of heterogeneous graphs, existing GNNs struggle to effectively capture relationships between different node types, resulting in inefficient learning. Additionally, the semantic information in heterogeneous graphs is often underutilized, limiting their overall performance. To address these challenges, we propose the heterogeneous graph attention joint network (HGAJ) framework, which leverages graph attention neural networks to jointly learn the structure and node features of heterogeneous graphs. We introduce heterogeneous structure learning by assigning attention weights to different types of relational subgraphs, enhancing the fusion and generation of heterogeneous feature subgraphs. To further improve the performance of classification tasks, we employ a bidirectional propagation method to optimize and update parameters. Moreover, we integrate the Galactica large language model (LLM) into the HGAJ framework to utilize its semantic reasoning capabilities in scientific scenarios, assisting in relational reasoning on heterogeneous graphs. We extensively evaluate our framework on the Association for Computing Machinery (ACM), Digital Bibliography & Library Project (DBLP), and Yelp datasets, demonstrating its superiority in relational reasoning on heterogeneous graph data compared to current state-of-the-art (SOTA) models. Code is available at https://github.com/AmbitYuki/HGAJ.