| Abstract: |
| Heterogeneous Graph Neural Networks (HGNNs) have shown powerful performance on heterogeneous graph learning by aggregating information from different types of nodes and edges. However, existing heterogeneous graph related models may confront with three major challenges: (1) Predefined meta-paths are required to capture the semantic relations between nodes from different types. (2) Existing models have to stack too many layers to learn long-range dependencies. (3) Performance degradation and semantic confusion may happen with the growth of the network depth. In this talk, we introduce two models to deal with the above-mentioned challenges. Specifically, we develop an end-to-end Dense connected Heterogeneous Graph Convolutional Network to learn node representations (Dense-HGCN). Dense-HGCN computes the attention weights between different nodes and incorporates the information of previous layers into each layer`s aggregation process via a specific fuse function. Moreover, Dense-HGCN leverages multi-scale information for node classification or other downstream tasks. Furthermore, we develop a Virtual Nodes based Heterogeneous Graph Convolutional Network (VN-HGCN). Virtual nodes are auxiliary nodes that are connected to all the nodes of a certain type in the graph, thus enabling efficient aggregation of long-range information across different types of nodes and edges. |
|