Graph similarity learning for cross-level interactions

Abstract

Graph similarity computation is crucial in fields such as bioinformatics, e.g., identifying compounds with similar biological activities by comparing molecular structural similarities. Traditional methods such as graph edit distance (GED) and maximal common subgraphs suffer from high computational complexity and sensitivity to noise, which limit their practical applications. Existing deep learning methods make it difficult to extract graph features, which affects computational accuracy comprehensively. To address these problems, we propose a new method, CLSim, which improves performance by enhancing feature extraction and improving graph similarity computation. Using the attention mechanism, CLSim first aligns graph pair features to the shared space and aggregates node features into global embeddings. The directionality of the embedding vectors is considered when extracting graph-level features to handle more complex data. In addition, we develop cross-layer feature extraction techniques that combine node-level information with graph-level embeddings to capture detailed node-graph interaction details. Experimental results on three datasets show that CLSim has excellent generalization capabilities and achieves lower error rates compared to the GED approach and the graph neural network baseline. In the worst case, its time complexity remains quadratic. Example query results further validate the effectiveness of the model, providing a more efficient and accurate solutions for graph similarity tasks.