Researchers from several universities in China and UK have jointly developed a new method for Graph Neural Networks (GNNs), known as Edge-Node Attention-based Differentiable Pooling (ENADPool). This method uses hard clustering and incorporates attention mechanisms to compress node features and edge strengths in GNNs. They also introduced the Multi-distance GNN (MD-GNN) model to mitigate over-smoothing and improve graph representation by allowing nodes to gather information from neighbors at different distances. By combining these techniques, the team was able to enhance graph classification performance.
The researchers conducted a thorough review of previous works related to GNNs, which are advanced tools used for graph classification that capture both local and global structure of a graph. General categories of GNNs include spectral-based and spatial-based methods, with the former using the Laplacian matrix while the latter aggregates local node information. However, both types have limitations, such as over-smoothing that several models have attempted to overcome. Similarly, for graph-level classification, hierarchical and global pooling operations become important, but they too come with limitations.
In this context, the ENADPool method presents a distinct advantage. It is a cluster-based hierarchical pooling method that assigns nodes to unique clusters, uses attention mechanisms to gauge node importance and compresses node features and edge connectivity for subsequent layers. This is achieved through a three-step process involving hard node assignment, node-based attention, and edge-based attention.
The researchers compared the performance of their newly introduced ENADPool method and MD-GNN model against other established graph deep learning methods using several benchmark datasets. They found that ENADPool and MD-GNN outperformed the other models, boasting higher average accuracy and more consistent results due to hard node assignment and effective feature representation.
In conclusion, the ENADPool technique successfully compresses node features and edge connectivity into hierarchical structures and uses attention mechanisms to identify the importance of nodes and edges. Additionally, MD-GNN combats over-smoothing by allowing nodes to access information from various neighbor distances. This research is being recognized for its contribution to improving graph classification performance. Investigators credit goes to the researchers who collaborated to make this project successful.
