Research map / Graph learning / Graph self-supervised learning
Negative Sampling From the Ground Up: A Redesign for Recommendation
Yanbang Wang, Jon Kleinberg, Yanhong Wu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add adaptive reweighting by model probability pf(v|u) (adaptive hard negatives) (ablation) | Our negative sampling (Node Embedding) · MovieLens | Recall@20 (%): 7.60 -> 12.77 | Improved |
| add adaptive reweighting by model probability pf(v|u) (ablation) | Our negative sampling (Node Embedding) · Pinterest | Recall@20 (%): 6.54 -> 12.59 | Improved |
| learn propensity score with graph features instead of a constant (ablation) | Our negative sampling (Node Embedding) · Pinterest | Recall@20 (%): 9.38 -> 12.59 | Improved |
| add regularization (first moment, consistency) to propensity estimator (ablation) | Our negative sampling (Node Embedding) · Pinterest | Recall@20 (%): 8.99 -> 12.59 | Improved |
| tune positive class prior instead of setting it to 0.5 (ablation) | Our negative sampling (Node Embedding) · Pinterest | Recall@20 (%): 9.39 -> 12.59 | Improved |
| model observation bias instead of using unbiased-case sampling (ablation) | Our negative sampling (Node Embedding) · Pinterest | Recall@20 (%): 8.01 -> 12.59 | Improved |
| model observation bias instead of using unbiased-case sampling (ablation) | Our negative sampling (Node Embedding) · MovieLens | Recall@20 (%): 10.35 -> 12.77 | Improved |
| use proposed sampling instead of AHNS (strongest baseline), LightGCN | LightGCN · Pinterest | Recall@20 (%): 11.35 -> 17.61 | Improved |
| use proposed sampling on unbiased LastFM split | Node Embedding · LastFM | Recall@20 (%): 7.37 -> 7.38 | No clear effect |
| use proposed sampling instead of full contrastive loss | GRACE · Cora, CiteSeer | accuracy: Cora 83.3 -> 83.6; CiteSeer 72.1 -> 71.6 | Mixed |
| use proposed sampling instead of full contrastive loss | GCE · Wiki-CS, Amazon-Photo | accuracy: 78.30 -> 77.41; 92.49 -> 91.94 | Hurt |
| use proposed sampling instead of RNS | LightGCN · MovieLens | Recall@20 (%): 8.94 -> 17.41 | Improved |
Closest papers by meaning
- Does Negative Sampling Matter? a Review With Insights Into its Theory and Applications (TPAMI 2024 · similarity 0.91)
- Simplify and Robustify Negative Sampling for Implicit Collaborative Filtering (NeurIPS 2020 · similarity 0.90)
- Adaptive Hardness Negative Sampling for Collaborative Filtering (AAAI 2024 · similarity 0.89)
- Unbiased Recommender Learning from Implicit Feedback via Weakly Supervised Learning (ICML 2025 · similarity 0.86)
- Relaxing the Accurate Imputation Assumption in Doubly Robust Learning for Debiased Collaborative Filtering (ICML 2024 · similarity 0.86)
- Neighborhood-Aware Negative Sampling for Student Knowledge and Behavior Modeling (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.
