EDMDPool: An Encoder-Decoder Multi-Dimension Graph Pooling for Graph Classification

A number of graph neural networks have been proposed for processing graph data, but most research has focused on the convolutional layer rather than the pooling layer. Current pooling methods often require the computation of an additional assignment matrix, which is computationally expensive, or use a single score to sort nodes, which leads to biased results. To address these problems, we introduce a novel multi-dimensional graph pooling operator, MDPool. This operator uses an attention mechanism to aggregate node scores across multiple dimensions, providing more robust node rankings. The proposed approach allows for adaptive selection of node sets based on the aforementioned rankings, thereby facilitating the generation of induced subgraphs. By integrating MDPool with the graph neural network module, we have developed an encoder-decoder framework for graph classification, called EDMDPool. Experimental results show that EDMDPool achieves accuracy improvements of 3.06%, 0.98%, 1.24%, and 1.80% over existing methods on four datasets.