Research map / Graph learning / Graph self-supervised learning
Flow Matching for Denoised Social Recommendation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RecFlow, a flow-matching-based generative denoiser on social graph embeddings, jointly trained with a LightGCN recommender and BPR loss, for social recommendation on Ciao, Yelp and Epinions.
Problem
Noisy (redundant or missing) social graph edges degrade GNN message propagation in social recommendation, and diffusion-based denoisers assume isotropic Gaussian noise that does not match the anisotropic structure of social data.
Approach
Replaces DDPM-style denoising with conditional flow matching that learns a velocity field over social-graph user embeddings, interpolating between Gaussian noise and the socially informed embedding, and adds the resulting reflow adjustment to the BPR-trained user embeddings.
Key results
- On Ciao, RecFlow reaches 0.725 Recall and 0.438 NDCG, versus 0.712 and 0.419 for RecDiff.
- On Yelp, RecFlow reaches 0.618 Recall and 0.341 NDCG, versus 0.597 and 0.308 for RecDiff.
- On Epinions, RecFlow reaches 0.486 Recall and 0.341 NDCG, versus 0.460 and 0.336 for RecDiff.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use RecFlow instead of RecDiff | social recommendation · Ciao | Recall: 0.712 -> 0.725 | Helped |
| use RecFlow instead of RecDiff | social recommendation · Ciao | NDCG: 0.419 -> 0.438 | Helped |
| use RecFlow instead of RecDiff | social recommendation · Yelp | Recall: 0.597 -> 0.618 | Helped |
| use RecFlow instead of RecDiff | social recommendation · Yelp | NDCG: 0.308 -> 0.341 | Helped |
| use RecFlow instead of RecDiff | social recommendation · Epinions | Recall: 0.460 -> 0.486 | Helped |
| use RecFlow instead of RecDiff | social recommendation · Epinions | NDCG: 0.336 -> 0.341 | Helped |
| add the flow-matching module (ablation) | RecFlow · Ciao | Recall: 0.633 -> 0.725 | Helped |
| add conditional learning guidance (ablation) | RecFlow · Ciao | NDCG: 0.401 -> 0.438 | Helped |
| add both the flow module and conditional learning guidance (ablation) | RecFlow · Yelp | Recall: 0.589 -> 0.618 | Helped |
| use the ODE solver instead of Multistep Heun (ablation) | RecFlow sampling · Ciao | Recall: 0.710 -> 0.725 | Helped |
| use the ODE solver instead of RK4 (ablation) | RecFlow sampling · Yelp | NDCG: 0.317 -> 0.341 | Helped |
| use the ODE solver instead of Euler (ablation) | RecFlow sampling · Epinions | Recall: 0.438 -> 0.486 | Helped |
Surprising findings
- Removing the flow module and conditional learning together yields a Recall on Ciao (0.621) lower than removing only conditional learning, while NDCG on Ciao (0.407) is higher than w/o CL (0.401).
Limitations
- Evaluated only on three social recommendation datasets, with scalability left to future work.
- Theoretical efficiency claims are not backed by measured runtime comparisons.
- Some reported details (e.g., gamma selection, hyperparameter tables) are vague or incomplete.
Experimental setup
Datasets: Ciao Yelp Epinions
Benchmarks: Ciao Yelp Epinions
Metrics: Recall@20 NDCG@20 HR@N
Baselines: PMF TrustMF GraphRec DiffNet DGRec NGCF MHCN KCGN SMIN GDMSR DSL RecDiff
Builds on: flow matching LightGCN RecDiff DDPM BPR
Closest papers by meaning
- Robust Graph Based Social Recommendation Through Contrastive Multi-View Learning (AAAI 2025 · similarity 0.86)
- Hgformer: Hyperbolic Graph Transformer for Collaborative Filtering (ICML 2025 · similarity 0.85)
- A Generalized Neural Diffusion Framework on Graphs (AAAI 2024 · similarity 0.85)
- When Edge Independence Fails: Joint Graph Diffusion with Latent Sociability Priors (NeurIPS 2026 · similarity 0.85)
- On Efficiency-Effectiveness Trade-off of Diffusion-based Recommenders (NeurIPS 2025 · similarity 0.84)
- Non-Euclidean Mixture Model for Social Network Embedding (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.