Abstract
Recent research in recommendation systems has emphasized diversity as a critical component of recommendation quality. However, this work argues that post-processing algorithms designed solely to improve aggregate diversity can lead to significant discrimination among users. We introduce the concept of "user fairness"—a dimension often overlooked in existing literature—and propose metrics to quantify it. Our experiments with two diversification algorithms demonstrate that increasing aggregate diversity consistently results in increased disparity among users.
Introduction: The Hidden Cost of Diversity
While diversifying recommendations is often viewed as a solution to "filter bubbles" and a way to promote niche items, it introduces a hidden cost: user unfairness. Most recommender systems are optimized for accuracy—predicting exactly what a user wants to see next. To counteract the resulting lack of variety and marketplace skew, researchers developed diversification algorithms.
This paper addresses whether the "utility cost" of such diversification is distributed equitably across the user base. If a system achieves high aggregate diversity by maintaining high-quality recommendations for some users while providing lower-quality, experimental items to others, it is fundamentally unfair to the latter group.
Methodology and Metrics
To quantify user disparity, we propose two metrics inspired by the Gini coefficient, a standard economic measure of inequality:
- Score Disparity ($D_S$): Measures the inequality of user satisfaction, defined as the ratio of the preference scores of recommended items to the scores of the optimal items the system could have suggested.
- Recommendation Disparity ($D_R$): Measures the inequality of "user gain" based on the overlap between recommended items and the user’s true top-k preferences.
These metrics were evaluated using the MovieLens dataset (100,000 ratings). Initial predictions were generated using K-Nearest Neighbors (KNN) and Non-negative Matrix Factorization (NMF). Two diversification techniques were then applied: a Random sampling approach and a Greedy algorithm designed to maximize the total number of unique items shown across the entire user base.
Results: The Diversity-Fairness Trade-off
Our experimental results reveal a clear trade-off: there is no "free lunch" when it comes to aggregate diversity. As the variety of items shown across the platform increases, so does the disparity in how well individual users are served.
When applying the Greedy algorithm to KNN-based recommendations, aggregate diversity rose from a baseline of 1.6% to 60.5%. However, this gain came at a significant cost: Score Disparity increased from 0.01% to 3.6%, and Recommendation Disparity surged from 0.2% to 20.1%. As shown in Figure 1, there is a near-linear relationship between diversity gains and the widening gap between satisfied and unsatisfied users.
In some experimental configurations, the disparity becomes even more pronounced. Figure 3 illustrates cases where Recommendation Disparity can approach 1.0, indicating that the burden of diversification is borne almost entirely by a subset of the user population, effectively "taxing" their experience to benefit system-wide variety.
Conclusion and Future Directions
This study highlights that improving fairness for items (ensuring niche products are seen) can directly cause unfairness for users. By applying the Gini coefficient to recommendation utility, we provide a mathematically grounded framework for discussing satisfaction inequality in AI systems.
Future work should focus on integrating these fairness measures directly into the design of recommendation algorithms rather than treating them as post-processing metrics. Additionally, further research is needed to determine if specific protected demographic groups are disproportionately affected by the "diversity tax" identified here.
Summary updated