Multi-level Aggregation Heterogeneous Graph Attention Network

In recent years, heterogeneous graph neural networks have garnered increasing attention due to their wide range of real-world applications. However, most existing models rely solely on the information from one meta-path while disregarding the information from other meta-paths or overlooking the significance of intermediate nodes within the meta-path, which play a crucial role in understanding the overall graph structure. To address this issue, we propose an innovative multi-level aggregated heterogeneous graph attention network (MAHGAN) that aims to more effectively integrate information from different levels. Firstly, we utilizes the self-attention mechanism to aggregate target nodes based on various meta-paths and thereby integrate information from their homogeneous neighboring nodes. Secondly, in order to better incorporate diverse information between different types of nodes, we aggregate node features of heterogeneous neighbors (first-order neighbors) for each target node in the original graph. This generates an embedding vector for each target node which is then converged with the node features obtained in the first step. Next, we combine these aggregated embedding vectors under different meta-paths according to learned weights to form a final embedding representation. Finally, through extensive experiments conducted on three datasets, it has been demonstrated that our model exhibits superior performance in node classification tasks.