Research map / Graph learning / Graph self-supervised learning
Sub-Interest-Aware Representation Uniformity for Recommender System
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 SIURec, a graph-based recommender that optimizes uniformity on fine-grained sub-interest representations at every LightGCN layer and sets the uniformity/BPR loss weights automatically via gradient alignment and one-shot (first-epoch) weighting.
Problem
AU-based recommendation methods apply uniformity to the final aggregated user/item embedding, which can conflict with real user interest structure and requires manually tuned alignment/uniformity weights.
Approach
Generate a pseudo-interaction matrix from layer-wise coarse sub-interest similarities to get fine-grained sub-interests, apply uniformity loss to these per-layer fine-grained sub-interests rather than the final embedding, replace the alignment loss with BPR, and automatically weight layers via Pareto-stationarity and gradient-ratio balancing computed once in the first epoch.
Key results
- SIURec achieves the best Recall@20 and NDCG@20 on all four datasets versus eleven baselines, e.g., Gowalla R@20 0.2048 vs 0.1956 for the best baseline (AdaGCL).
- The paper reports an average 4.26% accuracy improvement over eleven SOTA methods.
- Ablation on ML-1M and Beauty shows replacing fine-grained sub-interests with coarse ones or layer-wise uniformity with ultimate uniformity lowers R@40/N@40, and continuous weighting gives comparable results to one-shot weighting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SIURec instead of SimGCL | LightGCN-based recommendation on ML-1M · ML-1M | R@20: 0.1056 -> 0.1099 | Helped |
| Use SIURec instead of LightGCN | LightGCN-based recommendation on ML-1M · ML-1M | N@20: 0.0826 -> 0.0877 | Helped |
| Use SIURec instead of XSimGCL | LightGCN-based recommendation on Gowalla · Gowalla | R@20: 0.1956 -> 0.2048 | Helped |
| Use SIURec instead of XSimGCL | LightGCN-based recommendation on Gowalla · Gowalla | N@20: 0.1165 -> 0.1224 | Helped |
| Use SIURec instead of BOD | LightGCN-based recommendation on Beauty · Beauty | R@20: 0.1348 -> 0.1374 | Helped |
| Use SIURec instead of ProtoAU | LightGCN-based recommendation on Beauty · Beauty | N@20: 0.0758 -> 0.0759 | Helped |
| Use SIURec instead of XSimGCL | LightGCN-based recommendation on Book · Book | R@20: 0.0893 -> 0.0944 | Helped |
| Use SIURec instead of SimGCL | LightGCN-based recommendation on Book · Book | N@20: 0.0682 -> 0.0722 | Helped |
| Add fine-grained sub-interest extraction to LightGCN (ablation) | SIURec on ML-1M · ML-1M | R@40: 0.1785 -> 0.1859 | Helped |
| Add fine-grained sub-interest extraction to LightGCN (ablation) | SIURec on Beauty · Beauty | R@40: 0.1724 -> 0.1896 | Helped |
| Use fine-grained instead of coarse-grained sub-interest extraction (ablation) | SIURec on ML-1M · ML-1M | R@40: 0.1741 -> 0.1859 | Helped |
| Use fine-grained instead of coarse-grained sub-interest extraction (ablation) | SIURec on Beauty · Beauty | R@40: 0.1651 -> 0.1896 | Helped |
Surprising findings
- Excluding Beauty, AU-based methods (DirectAU, GraphAU, ProtoAU, AUPlus) underperform other categories such as contrastive learning methods, which the authors attribute to unrealistic uniformity on the final embedding.
- Solving layer weights every epoch yields almost no gain over solving them once in the first epoch, so the extra cost is unnecessary.
Limitations
- Experiments are limited to LightGCN-based models on four datasets.
- Uniformity loss weights rely on gradient-based heuristics whose theoretical justification is limited.
- Gains over the best baseline are modest on some datasets (e.g., Beauty R@20 0.1374 vs 0.1348 for DirectAU).
Experimental setup
Datasets: MovieLens-1M Gowalla Amazon-Beauty Amazon-Book
Benchmarks: MovieLens-1M Gowalla Amazon-Beauty Amazon-Book
Metrics: Recall@20 Recall@40 NDCG@20 NDCG@40
Baselines: LightGCN SGL SimGCL XSimGCL AdaGCL DCCF BOD DirectAU GraphAU ProtoAU AUPlus
Builds on: LightGCN DirectAU GraphAU Pareto optimization (MGDA)
Closest papers by meaning
- Augmenting Sequential Recommendation with Balanced Relevance and Diversity (AAAI 2025 · similarity 0.85)
- Tree of Preferences for Diversified Recommendation (NeurIPS 2025 · similarity 0.85)
- LGMRec: Local and Global Graph Learning for Multimodal Recommendation (AAAI 2024 · similarity 0.84)
- DivGCL: A Graph Contrastive Learning Model for Diverse Recommendation (AAAI 2025 · similarity 0.84)
- MTSTRec: Multimodal Time-Aligned Shared Token Recommender (ICML 2025 · similarity 0.84)
- Semantic Convergence: Harmonizing Recommender Systems via Two-Stage Alignment and Behavioral Semantic Tokenization (AAAI 2025 · 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.