Masked Graph Auto-Encoder Constrained Graph Pooling

The node drop pooling is a significant type of graph pooling that is required for learning graph-level representations. However, existing node drop pooling models still suffer from the information loss problem, impairing their effectiveness in graph classification. To mitigate the detrimental effect of the information loss, we propose a novel and flexible technique called Masked Graph Auto-encoder constrained Pooling (MGAP), which enables vanilla node drop pooling methods to retain sufficient effective graph information from both node-attribute and network-topology perspectives. Specifically, MGAP reconstructs the original node attributes of the graph using a graph convolutional network and the node degree of the graph (i.e., structural information) using a feedforward neural network with exponential neurons from the pooled (masked) graphs generated by the vanilla node drop pooling models. Notably, MGAP is a plug-and-play technique that can be directly adopted in the current node drop pooling methods. To evaluate the effectiveness of MGAP, we conduct extensive experiments on eleven real-world datasets by applying MGAP to three commonly-used methods, i.e., TopKPool, SAGPool, and GSAPool. The experimental results reveal that MGAP has the capacity to consistently improve the performance of all the three node drop pooling models in the graph classification task.