Abstract:
Subgraph isomorphism detection is among the most frequently encountered problems in various graph-based applications like Bioinformatics, Social network analysis and Recommendation systems. The problem aims to determine the presence of a query graph in a target graph. It is anNP-complete problem, due to which algorithmic approaches are computationally expensive and non-scalable. Inspired by the success of graph neural networks, we propose SubGIN, a novel neu-ral network architecture for Subgraph Isomorphism detection. A learnable node representation technique is used to map every graph into an embedding matrix. SubGIN uses query graph asa context to down-sample the target graph by implicitly computing and aggregating node-level features of the target w.r.t each query node. We introduce a novel graph interaction mechanism that leverages both intra and inter level graph information of both query and down-sampled tar-get representations using CNN. We have empirically shown the generalizability of our model by experimenting on bioinformatics and social network datasets. Experimental results demonstrate that SubGIN outperforms the previous methods by a significant margin.