Research map / Label-efficient and robust learning / Recognition and long-tail
Adaptive Hardness Negative Sampling for Collaborative Filtering
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 adaptive hardness negative sampling (AHNS), instantiated as AHNS_{p<0} that picks, from a random candidate set, the negative whose score is closest to a positive-dependent target, improving implicit CF recommendation over fixed-hardness samplers.
Problem
Existing negative sampling methods for implicit collaborative filtering select negatives of a fixed hardness level, causing false positive and false negative problems that hurt recommendation quality.
Approach
A new paradigm, AHNS, with three criteria (positive-aware, negatively correlated with positive score, adjustable hardness); the instantiation AHNS_{p<0} chooses from M random candidates the item minimizing |e_u^T e_i - beta*(e_u^T e_pos + alpha)^{p+1}|, so negative hardness adapts to each positive pair during training.
Key results
- AHNS_{p=-2} achieves best R@20 of 25.51 on ML-1M and 13.03 on Phones, versus best baseline 25.12 and 12.27 respectively.
- On Tools, AHNS_{p=-2} reports R@20 of 7.42 vs 6.86 for the best baseline DNS(M,N), an 8.2% improvement.
- AHNS_{p<0} reportedly significantly outperforms RNS by about 20% on average, with improvements over the best baseline of 1.6% to 8.2% in R@20.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use adaptive hardness negative sampling (AHNS p=-2) instead of the strongest baseline | MF collaborative filtering · ML-1M, Phones, Sports, and Tools | Recall@20: 25.12 -> 25.51; 12.27 -> 13.03; 7.90 -> 8.52; 6.86 -> 7.42 | Helped |
| Use adaptive hardness negative sampling (AHNS p=-2) instead of the strongest baseline | MF collaborative filtering · ML-1M, Phones, Sports, and Tools | NDCG@20: 37.74 -> 38.77; 6.75 -> 7.14; 4.35 -> 4.61; 3.88 -> 4.05 | Helped |
| Use adaptive hardness negative sampling (AHNS p=-2) instead of the strongest baseline | MF collaborative filtering · ML-1M, Phones, Sports, and Tools | NDCG@50: 39.71 -> 40.57; 8.15 -> 8.74; 5.36 -> 5.81; 4.61 -> 4.92 | Helped |
| Use p=-2 instead of p=-1 (ablation) | AHNS · ML-1M, Phones, Sports, and Tools | Recall@20: 25.17 -> 25.51; 13.02 -> 13.03; 8.42 -> 8.52; 7.27 -> 7.42 | Helped |
| Use p=-2 instead of p=-1 (ablation) | AHNS · ML-1M, Phones, Sports, and Tools | NDCG@20: 37.72 -> 38.77; 7.08 -> 7.14; 4.58 -> 4.61; 4.02 -> 4.05 | Helped |
| Use p=-2 instead of p=-1 (ablation) | AHNS · ML-1M, Phones, Sports, and Tools | NDCG@50: 39.31 -> 40.57; 8.71 -> 8.74; 5.82 -> 5.81; 4.95 -> 4.92 | Mixed |
Surprising findings
- Fixed-hardness hard negative samplers (DNS, DENS) suffer a significant performance drop in later training stages, attributed to false negatives.
Limitations
- Hyperparameters alpha and beta have no universal optimum and must be tuned per dataset and metric.
- Theoretical NDCG bound relies on approximations (e.g., replacing positive score with score plus alpha, and approximating the candidate set by a single negative).
Experimental setup
Datasets: MovieLens-1M Amazon-Phones Amazon-Sports Amazon-Tools
Benchmarks: MovieLens-1M Amazon-Phones Amazon-Sports Amazon-Tools
Metrics: Recall@20 NDCG@20 NDCG@50
Baselines: RNS SSM DNS MixGCF DENS DNS(M,N) CuCo
Builds on: BPR DNS DENS CuCo MixGCF
Closest papers by meaning
- Relaxing the Accurate Imputation Assumption in Doubly Robust Learning for Debiased Collaborative Filtering (ICML 2024 · similarity 0.87)
- ImplicitSLIM and How it Improves Embedding-based Collaborative Filtering (ICLR 2024 · similarity 0.85)
- Unbiased Recommender Learning from Implicit Feedback via Weakly Supervised Learning (ICML 2025 · similarity 0.85)
- Objective Shaping with Hard Negatives: Windowed Partial AUC Optimization for RL-based LLM Recommenders (NeurIPS 2026 · similarity 0.85)
- Neighborhood-Aware Negative Sampling for Student Knowledge and Behavior Modeling (AAAI 2025 · similarity 0.85)
- Sampling for Beyond-Worst-Case Online Ranking (AAAI 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.