Interpretable Neural Subgraph Matching for Graph Retrieval

Abstract

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

ProvG-Searcher: A Graph Representation Learning Approach for Efficient Provenance Graph Search

“…This is realized by jointly computing a similarity score between a pair of graphs. With the cross-graph matching approach, the reasoning of the relation between two graphs is made by modeling the node-to-node interactions [37,40,60,75,79] or by modeling the graph-to-graph interactions [5]. Since the graph-learning model attends a pair of graphs jointly, cross-graph matching methods are potentially stronger than the graph embedding models, and they can be made more resistant to slight variations between graphs.…”

“…Recently, graph neural networks (GNNs) have achieved significant success in graph representation learning [8] 2 . This led to the development of several learning-based methods for approximate subgraph matching with superior performance [5,37,40,46,60]. At the core of these methods is the learning of an embedding function that maps each graph into an embedding vector encapsulating its key features.…”

Optimizing Subgraph Retrieval and Matching with an Efficient Indexing Scheme

“…With the rapid development of graph neural networks (GNNs), they have been widely applied to tasks related to subgraph isomorphism. Most works focus on graph similarity matching [20][21][22][23] or subgraph counting [24][25][26], which cannot provide exact matching results. However, another subset of studies [27,28] is dedicated to addressing the subgraph containment problem by leveraging GNNs.…”

Efficient algorithms for subgraph counting and enumeration on large graphs

“…In some extreme cases, the running time and the memory costs are not affordable for some real applications. The other approach is to predict the subgraph counts directly via deep learning [35,102,130,157,192]. These models can be trained on some small graphs and applied to large graphs without re-training, which is time and space-efficient when handling large graphs.…”

“…We note that we do not compare our algorithm with those GNN-based approaches based on two considerations. First, there exist several works that target using GNN to capture the graphs' information and predict the subgraph counts, such as [35,102,130,157,192]. However, all these studies (1) target the static graphs, (2) need to store the whole graph which cannot satisfy the constraint where we just use a limited space and…”