Graph neural architecture search with heterogeneous message-passing mechanisms

In recent years, neural network search has been utilized in designing effective heterogeneous graph neural networks (HGNN) and has achieved remarkable performance beyond manually designed networks. Generally, there are two mainstream design manners in heterogeneous graph neural architecture search (HGNAS). The one is to automatically design a meta-graph to guide the direction of message-passing in a heterogeneous graph, thereby obtaining semantic information. The other learns to design the convolutional operator aiming to enhance message extraction capabilities to handle the diverse information in a heterogeneous graph. Through experiments, we observe a strong interdependence between message-passing direction and message extraction, which has a significant impact on the performance of HGNNs. However, previous HGNAS methods focus on one-sided design and lacked the ability to capture this interdependence. To address the issue, we propose a novel perspective called heterogeneous message-passing mechanism for HGNAS, which enables HGNAS to effectively capture the interdependence between message-passing direction and message extraction for designing HGNNs with better performance automatically. We call our method heterogeneous message-passing mechanisms search (HMMS). Extensive experiments on two popular tasks show that our method designs powerful HGNNs that have achieved SOTA results in different benchmark datasets. Codes are available at https://github.com/HetGNAS/HMMS.