SIREN: Subgraph Isomorphism via Reinforcement Enhanced Graph Neural Networks

ICLR 2026 Conference Submission17918 Authors

Keywords: Subgraph Isomorphism, Reinforcement Learning, Graph Neural Network, Partial Order

TL;DR: Solving subgraph isomorphism problem via deep reinforcement learning and partial order-aware GNN model

Abstract: The subgraph isomorphism problem comprises two distinct objectives: (a) Existence determination: Verifying whether an input graph contains a subgraph isomorphic to another input graph; and (b) Complete solution enumeration: Outputting the exhaustive set of all isomorphic mappings when they exist. Solving this problem serves as a fundamental requirement for numerous application domains. However, as an NP-complete problem, existing mainstream solvers primarily rely on heuristic techniques, demonstrating limited efficiency when handling large-scale input graphs. To address this challenge, we propose SIREN - a graph neural network enhanced with deep reinforcement learning for subgraph isomorphism resolution. SIREN establishes graph embeddings through partial order-aware GNNs, while employing Deep Q-Networks with bidomain-based pruning to accelerate the graph matching process. Experimental results on real-world datasets demonstrate that SIREN achieves 100% precision with modest computational time, outperforming AI-based approximate matching methods. Compared to state-of-the-art exact solvers, SIREN delivers 52% faster execution than leading AI approaches and 21% acceleration over top heuristic methods.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 17918