Introduction
Global, non-personalized recommendations, such as "Trending Topics" or "Top 10" lists, exert significant social influence as they are presented uniformly to an entire user base. While much research focuses on individual personalization, these global lists face a unique challenge: balancing user satisfaction with ideological neutrality. This paper explores the tension between User Fairness—representing the diverse preferences of the population—and Organisational Fairness—maintaining ideological balance. The authors demonstrate that optimizing for one often compromises the other, highlighting a critical trade-off in recommendation system design.
Methodology
The researchers modeled the recommendation task as a multi-winner election. In this framework, news articles serve as "candidates" and users as "voters."
Data and Scoring
The study utilized the Adressa dataset, comprising 865 Norwegian news articles and 63,260 users. User-article affinity was determined by "Active Time" (dwell time). To address sparsity, Non-negative Matrix Factorization (NMF) was employed to predict missing preference scores.
Voting Rules and Metrics
Six multi-winner voting rules were evaluated: SNTV, STV, k-Borda, Bloc voting, Chamberlin-Courant (CC), and Monroe. These rules aim to select a representative "committee" (the recommendation list) of size $\kappa = 10$.
- User Satisfaction: Measured how well the selected list aligned with predicted user preferences.
- Organisational Bias: Quantified using Pointwise Mutual Information (PMI) to identify ideological "seed words" in Norwegian media. Bias was calculated relative to a "reference bias" ($\rho = 1.423$), representing the population's actual consumption habits.
Results
The empirical evaluation reveals a clear conflict between maximizing satisfaction and minimizing bias. As shown in Table 1, different voting rules prioritize these objectives differently.
Single Transferable Vote (STV) achieved the highest user satisfaction (0.894) but exhibited significant negative bias (-0.117). Conversely, k-Borda was the most ideologically neutral (Bias = 0.033) but resulted in the lowest user satisfaction (0.800). Bloc voting performed poorly on the fairness front, producing the highest ideological skew (-0.240).
The results indicate that voting rules are inherently "content-blind." Because they optimize for preference scores without considering article substance, they can inadvertently amplify the ideological leanings of the most active user segments. In the Adressa dataset, this manifested as a consistent left-leaning bias across most methods.
Discussion and Conclusion
This research identifies a fundamental "Satisfaction-Bias Trade-off" in global recommendations. While social choice theory provides tools for fair representation, these tools do not automatically ensure ideological neutrality.
The study is limited by its lexicon-based bias detection and its focus on a single language (Norwegian). Future work should explore more sophisticated NLP techniques for bias analysis and, crucially, develop "cohesive mechanisms" that treat user satisfaction and ideological balance as joint constraints in a unified optimization process. Bridging this gap is essential for creating recommendation engines that are both representative and neutral.
Summary updated