A Subgraph Embedded GIN with Attention for Graph Classification

Hyung-Jun Moon, Sung-Bae Cho

Abstract: Graph neural networks (GNNs) have emerged as a powerful tool for analyzing graph data, where data are represented by nodes and edges. However, the conventional methods have limitations in analyzing graphs with diverse attributes and preserving crucial information during the graph embedding. As a result, there is a possibility of losing crucial information during the integration of individual nodes. To address this problem, we propose an attention-based readout with subgraphs for graph embedding that partitions the graph according to unique node attributes. This method ensures that important attributes are retained and prevents dilution of distinctive node features. The adjacency matrices and node feature matrices for the partitioned graphs go into a graph isomorphism network (GIN) to aggregate the features, where the attention mechanism merges the partitioned graphs to construct the whole graph embedding vector. Extensive experiments on six graph datasets demonstrate that the proposed method captures various local patterns and produces superior performance against the state-of-the-art methods for graph classification. Especially, on the challenging IMDB-MULTI dataset, our method achieves a significant performance gain of 27.87%p over the best method called MA-GCNN.