Source-linked AI summary
Fairness in Recommendation Ranking through Pairwise Comparisons
Alex Beutel, Jilin Chen, Tulsee Doshi, Hai Qian, Li Wei, Yi Wu, Lukasz Heldt, Zhe Zhao, Lichan Hong, Ed H. Chi, Cristos Goodrow
TL;DR
Recommender systems raise fairness concerns because pointwise accuracy measures do not adequately characterize the rankings users see, while evaluation is difficult under sparse and biased feedback. The paper introduces pairwise fairness metrics based on randomized experiments and a pairwise regularizer for training pointwise models. It reports significantly improved pairwise fairness in a large-scale production recommender system.
Problem
Pointwise fairness metrics do not guarantee fairness in the resulting recommender ranking, and recommendation evaluation is difficult under sparse, biased, and changing feedback.
Method
The paper measures recommender fairness through pairwise comparisons from randomized experiments and adds a pairwise regularization term that works with pointwise models.
Results
The approach significantly improves pairwise fairness in a large-scale production recommender system.
Takeaways & Limitations
Pairwise experiments provide a tractable way to obtain unbiased measurements of ranking fairness and evaluate systematic under-ranking of item groups.
Takeaways & Limitations
The pairwise fairness metric ignores user engagement and may risk promoting clickbait that users do not ultimately value.
Abstract
from arXiv · showhide
Recommender systems are one of the most pervasive applications of machine learning in industry, with many services using them to match users to products or information. As such it is important to ask: what are the possible fairness risks, how can we quantify them, and how should we address them? In this paper we offer a set of novel metrics for evaluating algorithmic fairness concerns in recommender systems. In particular we show how measuring fairness based on pairwise comparisons from randomized experiments provides a tractable means to reason about fairness in rankings from recommender systems. Building on this metric, we offer a new regularizer to encourage improving this metric during model training and thus improve fairness in the resulting rankings. We apply this pairwise regularization to a large-scale, production recommender system and show that we are able to significantly improve the system's pairwise fairness.
1 Introduction
The paper targets fairness risks in recommender rankings, where pointwise fairness measures do not adequately describe the ranked lists users see. It proposes pairwise fairness metrics and pairwise regularization, applying them in a production recommender system.
- Recommender systems can under-rank groups of items, limiting their visibility to users.
- Fairness metrics for pointwise accuracy do not indicate much about the resulting ranking users see.The paper contrasts pointwise metrics with work on fair unpersonalized rankings.
- Pairwise regularization significantly closes the inter-group pairwise accuracy gap.
- Evaluation is difficult because recommender feedback is sparse, biased by previous systems, and affected by changing users and items.
- Pairwise fairness metrics use randomized experiments to estimate user preferences and correspond directly to ranking performance.
- Pairwise regularization improves fairness for pointwise models and produces significant improvements in a large-scale production recommender system.
2 Related Work
Related work spans recommender systems, machine learning fairness, recommender-system ranking fairness, and fairness optimization. The paper situates its approach amid challenges from sparse, biased, and temporally dynamic recommendation data.
- Recommender-system research includes collaborative filtering, ratings prediction, pairwise user-preference models, and production ranking systems.
- Temporal dynamics and biased training data complicate both recommender-system training and performance evaluation.
- Machine learning fairness research has proposed many classification definitions, with group fairness as a prevalent structure and tensions among definitions.
- Prior recommender-fairness work studies fair rankings, post-processing, statistical parity, diversity, filter bubbles, and feedback loops from different perspectives.
- Fairness optimization methods include post-processing, constraint-based optimization, adversarial learning, and regularization during model training.
3 Pairwise Fairness for Recommendation
The paper defines pairwise fairness for recommendation rankings by comparing whether clicked items are ranked above relevant unclicked items across item groups. It separates overall accuracy into intra-group and inter-group measures, connects these metrics to ranking fairness, and notes important measurement and retrieval-system limitations.
- Pairwise Fairness Metric: Pointwise prediction errors require pairwise comparisons because a prediction can be under-predicted yet still produce the top-ranked item.The relevant question is how predictions compare among items for the same query.
- Measurement: A metric that ignores engagement can risk promoting clickbait, while the framework also assumes the retrieval system supplying relevant items is fair.The authors identify retrieval fairness and bias propagation into the final ranking as open issues.
- Pairwise Fairness Metric: Pairwise fairness compares the probability that a clicked item is ranked above another relevant unclicked item across groups, conditioned on equal engagement.This reframes fairness around relative ordering for the same query rather than pointwise prediction accuracy.
- Pairwise Fairness Metric: The metric splits comparisons into intra-group and inter-group accuracy, defining separate fairness criteria for within-group and cross-group pairs.Overall pairwise accuracy is a weighted sum of these two accuracies, with weights determined by the probability of each pair type under the relevant click and engagement conditions.
- Pairwise Fairness Metric: Intra-group fairness reflects whether the recommender ranks users’ items of interest well, while inter-group fairness reveals whether ranking mistakes disadvantage a group overall.The distinction helps identify systematic group preference that aggregate pairwise accuracy can miss.
- Measurement: The proposed metrics use randomized experiments to obtain unbiased preference estimates despite dynamic users, sparse feedback, and bias from prior recommendations.The framework measures fairness for pointwise recommenders and connects ranker performance to the fairness of resulting rankings.
4 Theoretical Analysis
Pairwise fairness connects recommendation comparisons to ranked position, while calibration and MSE alone do not guarantee fairness in the resulting rankings.
- Ranking interpretation: Pairwise fairness implies that the expected position of a clicked, engaged-with item is the same across groups.
- Ranking interpretation: Pairwise recommender fairness can be interpreted as group-independent average ranked position for clicked and engaged-with items.
- Ranking interpretation: Overall ranked position decomposes into within-group and between-group positions, motivating separate probability-based analyses.
- Relation to pointwise metrics: A calibrated model is insufficient for guaranteeing pairwise ranking fairness when groups have different average labels.
- Relation to pointwise metrics: Equal MSE across groups is insufficient for guaranteeing pairwise ranking fairness because MSE ignores relative ranking and distinguishes neither over- nor under-prediction.
5 Pairwise Regularization to Improve Fairness
The paper introduces pairwise regularization for pointwise recommenders by penalizing group-dependent differences in pairwise prediction residuals.
- The training objective augments the original loss with an absolute-correlation regularizer computed from experimental pair data.
- The pairwise regularizer penalizes models whose ability to predict the clicked item is better for one group than another.
- The experimental pair data are rebalanced so approximately half the examples have clicked items from each group.
- The approach can target intra-group fairness or engagement differences by restricting or bucketing the experimental pairs.
- The method does not provably achieve pairwise fairness, but is used for empirical performance and ease of production deployment.
6 Experiments
The production recommender exhibits subgroup disadvantages in pairwise fairness, especially for inter-group comparisons. Pairwise regularization substantially reduces these gaps while leaving overall engagement metrics neutral in live experiments.
- 6.1 Experimental Setup: The production model compares subgroup items against non-subgroup items in a large-scale ranking system, using randomized relevant-item pairs and separate data for training and evaluation.The subgroup represents approximately 0.2% of all items, and the models share the same architecture.
- 6.2 Baseline Performance: 14.9% was the non-subgroup advantage in intra-group pairwise fairness, partly reflecting the subgroup’s small size and lower diversity.After removing highly similar subgroup comparisons, the passage reports no meaningful performance difference.
- 6.2 Baseline Performance: 35.6% was the non-subgroup advantage in inter-group pairwise fairness, making it the largest observed fairness gap.Subgroup items were significantly under-ranked relative to non-subgroup items across engagement levels.
- 6.3 Fairness Improvements: Pairwise regularization reduced the inter-group non-subgroup advantage from 35.6% to 2.6%.The test model’s non-subgroup pairwise accuracy became approximately on-par with its intra-group pairwise accuracy.
- 6.3 Fairness Improvements: Overall pairwise fairness improved from an 8.3% non-subgroup advantage to 2.5%, while intra-group fairness remained largely unchanged at a 16.7% advantage.Live experiments found overall engagement metrics were neutral relative to the production system.
- 6.4 How are improvements achieved?: The regularizer increased subgroup exposure above its base click-through rate at lower engagement levels, creating a calibration–pairwise-fairness trade-off.The production model’s exposure approximately matched each group’s base click-through rate before regularization.
7 Conclusion
The paper presents randomized pairwise experiments as a tractable way to measure recommender ranking fairness and decomposes systematic group mis-ranking. It also introduces pairwise regularization, which significantly improves fairness metrics in a large-scale production system.
- 7 Conclusion: Randomized pairwise experiments provide unbiased measurements of recommender ranking fairness.The measurements can evaluate and decompose whether a model systematically mis-ranks or under-ranks items from a group.
- 7 Conclusion: The proposed fairness measure aligns with ranking-fairness definitions but is not covered by pointwise fairness measures.
- 7 Conclusion: Pairwise regularization improves recommender fairness during training, with significant fairness-metric gains demonstrated in a large-scale production system.