Interpretable Neural Subgraph Matching for Graph Retrieval

Given a query graph and database of corpus graphs, retrieval system aims to deliver the most relevant graphs. Graph based on subgraph matching has wide variety applications, e.g., molecular fingerprint detection, circuit design, software analysis, question answering. In such is graph, if (perfectly or approximately) graph. Existing neural models compare node embeddings query-corpus pairs, compute relevance scores between them. However, may not provide edge consistency Moreover, they predominantly use symmetric scores, which are appropriate in context matching, since underlying score search should be measured using partial order induced by subgraph-supergraph relationship. Consequently, show poor performance matching. response, we propose ISONET, novel interpretable alignment formulation, better able learn edge-consistent mapping necessary for ISONET incorporates new scoring mechanism enforces an asymmetric score, specifically tailored ISONET’s design enables it directly identify given Our experiments diverse datasets that outperforms recent formulations systems. Additionally, can alignments pairs during inference, despite being trained only binary labels whole graphs training, without any fine-grained ground truth information about alignments.