Source-linked AI summary
Fairly Adaptive Negative Sampling for Recommendations
Xiao Chen, Wenqi Fan, Jingfan Chen, Haochen Liu, Zitao Liu, Zhaoxiang Zhang, Qing Li
TL;DR
Uniform negative sampling can oversample majority-group items as negatives when item groups are imbalanced, countering item-side fairness. FairNeg adapts group-level sampling during training and mixes fairness-aware and importance-aware sampling. Experiments report reduced group disparity while maintaining user-side recommendation quality.
Problem
With uneven item-group sizes, uniform negative sampling can oversample majority-group items as negatives and counter item-side group fairness.
Source: §1 · p. 2Method
FairNeg adapts group-level negative-sampling probabilities to perceived group performance disparity, using momentum updates and a mixup with importance-aware sampling.
Source: §1 · p. 2, §3.4 · p. 5, §6 · p. 8Results
Across different real-world datasets, FairNeg significantly mitigates item-group performance disparity while maintaining overall user-side recommendation quality.
Source: §1 · p. 2Takeaways & Limitations
The authors conclude that negative-sampling design can jointly consider group fairness and representation learning while pursuing a better fairness–utility tradeoff.
Source: §1 · p. 2, §6 · p. 8The formulation assumes item groups are non-overlapping, with each item belonging to one attribute-defined group.
Source: §2.2 · p. 3
Abstract
from arXiv · showhide
Pairwise learning strategies are prevalent for optimizing recommendation models on implicit feedback data, which usually learns user preference by discriminating between positive (i.e., clicked by a user) and negative items (i.e., obtained by negative sampling). However, the size of different item groups (specified by item attribute) is usually unevenly distributed. We empirically find that the commonly used uniform negative sampling strategy for pairwise algorithms (e.g., BPR) can inherit such data bias and oversample the majority item group as negative instances, severely countering group fairness on the item side. In this paper, we propose a Fairly adaptive Negative sampling approach (FairNeg), which improves item group fairness via adaptively adjusting the group-level negative sampling distribution in the training process. In particular, it first perceives the model's unfairness status at each step and then adjusts the group-wise sampling distribution with an adaptive momentum update strategy for better facilitating fairness optimization. Moreover, a negative sampling distribution Mixup mechanism is proposed, which gracefully incorporates existing importance-aware sampling techniques intended for mining informative negative samples, thus allowing for achieving multiple optimization purposes. Extensive experiments on four public datasets show our proposed method's superiority in group fairness enhancement and fairness-utility tradeoff.
1 INTRODUCTION
Uniform negative sampling can inherit item-group imbalance and oversample majority-group negatives, while fixed fairness adjustments trade off fairness against overall recall. FairNeg adapts group-level sampling during training, and the paper reports improved group fairness while maintaining user-side recommendation quality.
1 INTRODUCTION
- Uniform negative sampling is more likely to select majority-group unobserved items as negatives, which can dampen that group’s recommendation quality.The paper attributes disparate group performance to the sampling distribution inheriting group-wise data bias.Source: §1 · p. 2
- FairStatic equalizes each group’s chance of being sampled as negative and positive, but a fixed distribution can greedily pursue fairness and produce suboptimal recommendation performance.This motivates an adaptive rather than static sampling strategy.Source: §1 · p. 2
- MF/LightGCN+FairStatic reduces Recall@20 disparity between Sci-Fi and Horror groups but sacrifices overall recall compared with UNS.Figure 2 reports the two groups’ Recall@20 for MF and LightGCN under the two sampling strategies on ML1M-2.Source: §1 · p. 2, Figure 2 · p. 2
- FairNeg dynamically adjusts group-level negative-sampling probabilities based on perceived performance disparity and can combine with importance-aware sampling.The proposed approach targets a better utility–fairness tradeoff.Source: §1 · p. 2
- Across real-world datasets, FairNeg significantly mitigates item-group performance disparity while maintaining overall user-side recommendation quality.The authors report this as an outcome of their experiments on different real-world datasets.Source: §1 · p. 2
2 PRELIMINARIES
The preliminaries introduce implicit-feedback pairwise recommendation and define item-oriented group fairness through differences in groups’ recommendation performance.
2 PRELIMINARIES
- The preliminaries frame the problem around top-k recommendation and item-oriented group performance fairness.The paper presents top-k recommendation, BPR loss, and fairness notions before introducing its approach.Source: §2 · p. 2
2.1 Personalized Ranking on Implicit feedback
- BPR encourages higher predicted scores for observed interactions than for unobserved items sampled from each user’s negative set.The setup uses implicit user-item interactions and recommendation-model scores.Source: §2.1 · pp. 2–3
2.2 Problem Definition
- Item-oriented fairness is evaluated through group-wise Recall@k disparity, with lower Recall-Disp@k indicating less disparity.The item groups are attribute-defined, non-overlapping groups, and Recall-Disp@k is based on the standard deviation across group recalls.Source: §2.2 · p. 3
3 THE PROPOSED METHOD
FairNeg addresses item-group unfairness by adjusting negative-sampling probabilities without changing model architecture or training data. Its framework measures group recall disparity, adapts group probabilities, and mixes fairness- and importance-aware sampling.
3.1 An Overview of the Proposed Framework
- FairNeg adjusts each item group’s negative-sampling probability for fairness and incorporates importance-aware sampling to improve negative-sample informativeness.The framework combines item-group fairness perception, adaptive momentum updates, and negative-sampling distribution mixup.Source: §3.1 · p. 3
3.2 Item Group Fairness Perception
- G-BCE uses average classification loss on observed positive interactions as a proxy for group recall; reducing disparities between groups’ G-BCE losses targets item-oriented fairness.The loss provides a differentiable training-phase measure where Recall-Disp@k cannot be used directly.Source: §3.2 · p. 4
3.3 Adaptive Momentum Update of Group Sampling Distribution
- Group probabilities are adjusted according to deviations from average G-BCE: disadvantaged groups are down-weighted, while historical-gradient momentum smooths updates.The update uses current group-wise gradients and historical gradients to mitigate oscillations and gradient-direction instability.Source: §3.3 · pp. 4–5
3.4 Negative Sampling Distribution Mixup
- Mixup addresses equal within-group sampling, which can favor easier negatives: model-scored importance sampling favors higher-relevance candidates, then β mixes it with the fairness-aware distribution.The two distributions respectively support more discriminative feature learning and item-group fairness; β controls fairness-optimization strength.Source: §3.4 · p. 5
3.5 Optimization for Item Group Fairness
- Optimization alternates outer updates to group sampling probabilities against recall disparity with inner BPR updates to model parameters, stopping after p consecutive epochs without performance improvement.The group-wise sampling distribution is the outer optimization variable, and recommendation parameters are updated by the inner ranking algorithm.Source: §3.5 · p. 5
4 EXPERIMENT
Experiments compare FairNeg with sampling and fairness-aware baselines on four datasets using MF and LightGCN. FairNeg consistently improves group-wise recall fairness while preserving good recommendation utility.
4.1.1 Datasets.
- The evaluation uses ML1M-4, ML1M-2, Amazon-4, and Amazon-2, formed from movie genres and product types with binary- or multi-valued attributes.Each dataset is randomly split into 60% training, 20% validation, and 20% test data.Source: §4.1.1 · p. 6
4.1.2 Evaluation Metrics.
- Group fairness is measured with Recall-Disp, Recall-Min, and Recall-Avg across item groups, while recommendation utility uses P@k, R@k, and N@k.Lower Recall-Disp and higher Recall-Min or Recall-Avg indicate better group fairness.Source: §4.1.2 · p. 6
4.1.4 Baselines.
- The comparison includes uniform, popularity-based, and dynamic negative samplers, a static FairNeg variant, and fairness-aware regularization, adversarial, and GAN methods.Reg is reported only for binary-group datasets; FairGAN uses two strictness settings per backbone.Source: §4.1.4 · p. 6, §4.2 · p. 6
4.1 Experiment Settings
- Experiments use MF and LightGCN backbones, average results across five random seeds, and compare top-20 recommendation performance.Top-30 results in the appendix are reported to align with the observations.Source: §4.2 · p. 6
4.2 Performance Comparison of Recommender Systems
- FairNeg consistently outperforms other methods on group-wise recall fairness across two backbones and four datasets while preserving good recommendation utility.The authors characterize the utility loss as minor.Source: Table 2 · p. 7, Table 3 · p. 7, §4.2 · p. 7
4.3 Ablation Study
- The ablation evaluates dynamic sampling, importance-aware sampling, and adaptive momentum by removing each component, with UNS as the baseline.The reported Figure 5 results cover F1@20 and Recall-Disp@20 on ML1M-4 using MF and LightGCN; Amazon-4 results are in the appendix.Source: §4.3 · pp. 7–8, Figure 5 · p. 8
- Ablations find dynamic sampling important for fairness and utility preservation, importance-aware sampling important for utility, and momentum beneficial to fairness stability; FairNeg has the best utility-fairness tradeoff on both backbones.The authors attribute the momentum finding to noisy gradients in an unstable fairness optimization process.Source: §4.3 · p. 8
5 RELATED WORK
Prior work addresses recommendation fairness through exposure allocation or recall parity, while negative-sampling research largely prioritizes utility. FairNeg approaches item-group unfairness through negative sampling.
Fair Recommendations
Negative Sampling
- Static and hard negative sampling support efficient training or mine informative examples [6, 33, 45, 47, 49], but prior approaches mainly optimize overall utility [44] and neglect item-side group bias.The paper frames its contribution as connecting negative-sampling design with item-group fairness optimization.Source: §5 · p. 8
6 CONCLUSION
The conclusion presents FairNeg as an adaptive negative-sampling framework for reducing item-group unfairness, combining fairness perception, momentum updates, and importance-aware mixup. It reports improved fairness and fairness-performance tradeoffs, and identifies broader fairness metrics as future work.
6 CONCLUSION
- FairNeg measures recall disparity and adaptively updates group-wise sampling probabilities with momentum, while mixup combines fairness-aware and importance-related sampling distributions.The stated aim is to jointly consider representation learning and group fairness optimization.Source: §6 · p. 8
- The authors report significant fairness improvements over state-of-the-art debiasing approaches and better fairness-performance tradeoffs.Source: §6 · p. 8
- Future work will investigate compatibility with additional recommendation group fairness metrics.Source: §6 · p. 8
A APPENDIX
The appendix provides information intended to support reproduction of the paper’s insights and experimental results.
A APPENDIX
- The appendix presents information for reproducing the paper’s insights and experimental results.Source: A APPENDIX · p. 10
A.1 Group-wise Statistics of Datasets
The four datasets have imbalanced item-group counts; Amazon is less dense than ML1M because average positive feedback per item group is lower.
A.1 Group-wise Statistics of Datasets
- Item counts are fairly imbalanced across the four datasets.Table 4 presents the group statistics.Source: §A.1 · p. 10, Table 4 · p. 10
- Amazon has lower density than ML1M because average positive feedback per item group is much lower.Source: §A.1 · p. 10
A.2 Implementation Details
Experiments use MF and LightGCN with BPR, and FairNeg is implemented in PyTorch with Adam; figures show FairNeg variants and hyperparameter analyses.
A.2 Implementation Details
- Experiments use MF and LightGCN optimized with BPR; FairNeg is implemented in PyTorch with Adam.Baseline models use TensorFlow with Adam.Source: §A.2 · p. 10
- Figure 6 compares FairNeg and its variants on Amazon-4 using MF and LightGCN backbones.The caption notes validation-set involvement and hidden-dimension search over [10, 80].Source: Figure 6 · p. 10
- The momentum coefficient γ is searched over [0, 0.2] in 0.05 steps, and β over [0.1, 0.9] in 0.1 steps.The outer-optimizer learning rate is 0.1, and the hard-negative probability temperature is 0.4.Source: §A.2 · p. 10
- Figures 7 and 8 plot fairness and utility with respect to γ and β, respectively.Source: Figure 7 · p. 10, Figure 8 · p. 10
A.3 Performance Comparison of Recommender Systems
This section evaluates recommendation utility and item-group fairness using top-30 ranking results.
A.3 Performance Comparison of Recommender Systems
- Methods are evaluated on user-side recommendation utility and item-group fairness using top-30 ranking results.Source: §A.3 · p. 10
A.4 Ablation Study
The ablations find that importance-aware sampling can maintain or improve utility, while momentum and dynamic sampling are important for fairness optimization; hyperparameter analyses examine γ and β.
A.4 Ablation Study
- On Amazon-4, importance-aware sampling can maintain or improve utility, while momentum and dynamic sampling are both essential for item-group fairness optimization.The reported pattern aligns with results on ML1M-4.Source: §A.4 · p. 10
A.5 Hyper-parameter Analysis of FairNeg
- At γ = 0.1, fairness improves without hurting utility in the ML1M-4 LightGCN analysis at K = 20 and 30.This analysis sets β to 0 to isolate the momentum parameter; similar patterns are reported on other datasets and model backbones.Source: §A.5 · pp. 10–11
- As β increases from 0.1 to 0.9, fairness first improves and then worsens, while utility consistently improves as importance-aware sampling probability rises.The authors attribute the utility improvement to more informative negative samples and conclude that only proper incorporation supports a better utility-fairness tradeoff.Source: §A.5 · p. 11