DIFFCOM: Conditional Discrete Diffusion Model for Community Search

Learning-based community search methods have attracted widespread attention due to their effectiveness. Recently, continuous diffusion models have been introduced for community search, achieving strong performance. However, such continuous diffusion models may introduce boundary ambiguity, making it difficult to convert continuous scores into high-quality discrete community selections. To address these issues, we propose DIF-FCOM, a discrete diffusion model-based method for community search. Our method instantiates a query-conditioned discrete diffusion process directly in the binary membership space, where a conditional graph neural network integrates query semantics and community-level context at each denoising step to provide richer conditioning for the reverse process. It further incorporates an edge-level structural loss to model pairwise connectivity within a community and employs a threshold-free decoding strategy that jointly leverages node and edge confidences to determine the final community. Experiments on real-world and synthetic datasets demonstrate that DIFFCOM achieves state-of-the-art performance on almost all datasets for single-node queries and remains competitive for multi-node queries.