Research map / Graph learning / Graph self-supervised learning
Social Recommendation via Graph-Level Counterfactual Augmentation
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 SR-GCA, a plug-and-play graph-level counterfactual augmentation module that generates counterfactual user-user and item-item links with a GCN backbone and BPR loss to improve social recommendation.
Problem
Social recommendation with graph-based models suffers from insufficient expressiveness and overfitting due to sparse graph data; existing counterfactual methods in recommendation focus on user-item negative sampling rather than graph-level structure.
Approach
Construct counterfactual social and item links by nearest-neighbor matching of treatment-flipped node pairs in embedding space, then learn factual and counterfactual link prediction with an MLP decoder estimating individual and average treatment effects, plus a distribution discrepancy loss, jointly with BPR.
Key results
- On Ciao, SR-GCA reaches HR@N 0.6586 and NDCG 0.4262, beating the strongest baseline DSL (HR 0.6374, NDCG 0.4065).
- On Epinions, SR-GCA reports HR@N 0.4120 and NDCG 0.2455, outperforming DSL on NDCG and MHCN on HR.
- Plugging GCA into DGNN improves Ciao HR@N from 0.5515 to 0.5703; into GDMSR, from 0.5402 to 0.5543.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add counterfactual item links (ablation) | SR-GCA · Ciao and Epinions | HR@N: 0.6203 -> 0.6586; 0.3993 -> 0.4120 | Helped |
| Add counterfactual item links (ablation) | SR-GCA · Ciao and Epinions | NDCG: 0.4023 -> 0.4262; 0.2308 -> 0.2455 | Helped |
| Add counterfactual user links (ablation) | SR-GCA · Ciao and Epinions | HR@N: 0.6329 -> 0.6586; 0.4010 -> 0.4120 | Helped |
| Add counterfactual user links (ablation) | SR-GCA · Ciao and Epinions | NDCG: 0.4102 -> 0.4262; 0.2388 -> 0.2455 | Helped |
| Add counterfactual user and item links (ablation) | SR-GCA · Ciao and Epinions | HR@N: 0.6091 -> 0.6586; 0.3375 -> 0.4120 | Helped |
| Add counterfactual user and item links (ablation) | SR-GCA · Ciao and Epinions | NDCG: 0.3872 -> 0.4262; 0.2012 -> 0.2455 | Helped |
| Add GCA module | GDMSR · Ciao and Epinions | HR@N: 0.5402 -> 0.5543; 0.7219 -> 0.7392 | Helped |
| Add GCA module | DGNN · Ciao and Epinions | HR@N: 0.5515 -> 0.5703; 0.7335 -> 0.7403 | Helped |
Surprising findings
- The paper's text reports a noticeably lower HR@N for SR-GCA on Epinions in the ablation table (0.4220) than in the main table (0.4120), an inconsistency.
Limitations
- Counterfactual links are built by nearest-neighbor matching in embedding space, with a distance threshold gamma and no guarantee of true counterfactual validity.
- Evaluated only on two datasets (Ciao and Epinions) with HR@N and NDCG@N metrics.
- Transferability experiments cover only two backbones (DGNN and GDMSR).
Experimental setup
Datasets: Ciao Epinions
Benchmarks: Ciao Epinions
Metrics: HR@10 NDCG@10
Baselines: PMF TrustMF DiffNet DGRec EATNN NGCF MHCN KCGN SMIN DSL
Builds on: LightGCN BPR Learning from Counterfactual Links for Link Prediction (Zhao et al. 2022)
Closest papers by meaning
- Graph Heterogeneous Multi-Relational Recommendation (AAAI 2021 · similarity 0.86)
- Mitigating the Popularity Bias of Graph Collaborative Filtering: A Dimensional Collapse Perspective (NeurIPS 2023 · similarity 0.86)
- gOCCF: Graph-Theoretic One-Class Collaborative Filtering Based on Uninteresting Items (AAAI 2018 · similarity 0.86)
- Robust Graph Based Social Recommendation Through Contrastive Multi-View Learning (AAAI 2025 · similarity 0.86)
- Multi-Component Graph Convolutional Collaborative Filtering (AAAI 2020 · similarity 0.86)
- Disentangled Modeling of Preferences and Social Influence for Group Recommendation (AAAI 2025 · similarity 0.85)
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.