Heterogeneous Graph Embedding Based on Attribute-guided Semantic Augmentation

Heterogeneous Graph Neural Networks (HGNNs) have demonstrated significant advancements in heterogeneous graph embedding. Most existing methods mainly leverage original graph structures to implement message passing, and node attributes are often only used as node features at the initial layer of HGNNs. This strategy impedes the potential of node attributes for augmenting node connections, leading to limited receptive field for inactive nodes with few or even no neighbors. To address the above issues, we propose a novel model Heterogeneous graph embedding based on Attribute-guided Semantic Augmentation (HASA) to make full use of node attributes for node representation learning. Specifically, we reveal the implicit semantic relationships between nodes by the node similarities based on node attributes. Then, we take the attribute categories of nodes as new nodes to augment the graph, which further exploits the implicit semantic relations between nodes by new nodes participating in the message-passing process in HGNNs. Experiments on real-world heterogeneous graph datasets show the proposed HASA model outperforms the baselines in both node classification and node clustering.