cs.LGOct 29, 2026

Fair Recommender Systems for Intersecting Groups

Authors: Christine Ling, Juba Ziani

Organizations: Georgia Institute of Technology.

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Recommendation systems that optimize for the average user can fail to capture the preferences of diverse demographic groups. We consider a dynamic recommendation system, where arriving users can belong to different demographic groups. Importantly, these groups can intersect, and a user may belong to several groups. We first show that over time, any group-blind policy (which excludes all current-user group information) recommending K≥1K \geq 1 items incurs linear total regret under mild conditions. In turn, we introduce three group-aware algorithms based on the classical Upper Confidence Bound algorithm (UCB): i) Intersection-Fitting, which runs UCB on each intersection, but has a sublinear group- and intersection-level regret guarantee with an exponential dependence in the number of groups; and ii) two group-level heuristics, Inner-Outer UCB and Group-Item UCB which allocate different arms to different groups, and use a sleeping expert meta-learner to choose which active group's recommendation an intersection should inherit at any given time step. We provide experimental results on two datasets, Goodreads and BoardGameGeek, and highlight that our heuristics can be preferable to the intersection-fitting algorithm when intersection-level preferences are closely related to group-level preferences, but that intersection-fitting performs better when different intersections within a same group prefer different items.

Figures & tables

 Ij,t=v^j,t+Vmax⁡2ln⁡max⁡{2,t+1}Nj,tI_{j,t}=\hat{v}_{j,t}+V_{\max}\sqrt{\frac{2\ln\max\{2,t+1\}}{N_{j,t}}} 
Algorithm 1 Group-Blind top- KK UCB
 Ij,ιP,t=v^j,ιP+Vmax⁡2ln⁡max⁡{2,τιP(t)}Nj,ιPI_{j,\iota_{P},t}=\hat{v}_{j,\iota_{P}}+V_{\max}\sqrt{\frac{2\ln\max\{2,\tau_{\iota_{P}}(t)\}}{N_{j,\iota_{P}}}} 
Algorithm 2 Intersection-Fitting UCB
 Ij,k,t=v^j,k+Vmax⁡2ln⁡max⁡{2,t+1}Nj,kI_{j,k,t}=\hat{v}_{j,k}+V_{\max}\sqrt{\frac{2\ln\max\{2,t+1\}}{N_{j,k}}} 
Algorithm 3 Inner–Outer UCB
 Ij,k,t=v^j,k+Vmax⁡2ln⁡max⁡{2,t+1}Nj,kI_{j,k,t}=\hat{v}_{j,k}+V_{\max}\sqrt{\frac{2\ln\max\{2,t+1\}}{N_{j,k}}} 
Algorithm 4 Group–Item UCB
Figure 1: Goodreads primary catalog ( n=15n=15 ) at slate size K=1K=1 . At the group level, Intersection-Fitting and both pooling heuristics (Group–Item UCB and Inner–Outer UCB) achieve comparable low cumulative regret. At the intersection level, common demographics (solid lines) and rare demographics (dotted lines) are plotted together. Across both traffic tiers, the pooling heuristics achieve strictly lower cumulative regret than exact Intersection-Fitting.
Figure 2: Goodreads primary catalog ( n=15n=15 ) at slate size K=5K=5 . Group-Blind UCB incurs linear regret across all panels. Group–Item UCB achieves the lowest regret across groups and intersections, whereas Inner–Outer UCB’s cumulative regret exceeds Intersection-Fitting on common intersections.
Figure 3: BoardGameGeek primary catalog ( n=15n=15 ) at slate size K=1K=1 . Intersection-Fitting achieves negative group-level regret by recommending intersection-specific slates that outperform the group’s best fixed slate. Both pooling heuristics (Group–item UCB and Inner–Outer UCB) result in higher cumulative regret than Intersection-Fitting across the groups and intersections.
Figure 4: BoardGameGeek primary catalog ( n=15n=15 ) at slate size K=5K=5 . While the pooling heuristics also achieve negative group-level regret, Intersection-Fitting drives group regret to be more negative and achieves lower cumulative intersection regret across all displayed panels.
Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Goodreads primary catalog ( n=15n=15 ) at slate size K=3K=3 . On rare intersections (dotted lines), Group–item UCB and Inner–Outer UCB achieve lower regret than Intersection-Fitting. On common intersections (solid lines), Group–item UCB maintains the lowest regret, while Inner–Outer UCB achieves comparable cumulative regret to Intersection-Fitting.
Figure 6: BoardGameGeek primary catalog ( n=15n=15 ) at slate size K=3K=3 . Intersection-Fitting achieves the lowest cumulative regret at both the group and intersection levels. Group–Item UCB and Inner–Outer UCB have distinct regret trajectories that end strictly higher than Intersection-Fitting across the panels.
Figure 7: Goodreads extended catalog ( n=72n=72 ) at slate size K=1K=1 . At this slate size, Group–Item UCB and Inner–Outer UCB are algorithmically equivalent. Both pooled heuristics efficiently leverage the inherited group structure to achieve strictly lower cumulative regret than Intersection-Fitting across both group and intersection levels.
Figure 8: Goodreads extended catalog ( n=72n=72 ) at slate size K=3K=3 . Group–Item UCB achieves the lowest regret. Unlike the 1515 -item catalog where Inner–Outer UCB and Intersection-Fitting were comparable at K=3K=3 , the expanded 7272 -item catalog separates Intersection-Fitting’s budget enough that Inner–Outer UCB establishes an advantage on common intersections.
Figure 9: Goodreads extended catalog ( n=72n=72 ) at slate size K=5K=5 . This is a reversal of the dynamic observed in the 1515 -item catalog: the expanded action space penalizes Intersection-Fitting’s lack of data sharing more, allowing Inner–Outer UCB to maintain lower cumulative regret on common intersections. Group–Item UCB remains the most efficient policy.
Figure 10: BoardGameGeek extended catalog ( n=72n=72 ) at slate size K=1K=1 . Because the empirical intersections exhibit heterogeneous preferences that do not smoothly compose from their base themes, pooling data actively harms learning. Intersection-Fitting avoids this cross-group interference, driving group regret deeply negative and achieving the lowest cumulative intersection regret.
Figure 11: BoardGameGeek extended catalog ( n=72n=72 ) at slate size K=3K=3 . Intersection-Fitting consistently achieves the lowest cumulative regret across both evaluation levels. The skewed UCB-style indices caused by active-group updates prevent Inner–Outer UCB and Group–Item UCB from efficiently learning the optimal slates for atypical intersections.
Figure 12: BoardGameGeek extended catalog ( n=72n=72 ) at slate size K=5K=5 . Even at larger slate sizes within the expanded catalog, Intersection-Fitting remains the dominant policy. The structural penalty of pooling heterogeneous data persists, allowing exact, unbiased intersection-level learning to outpace both heuristics.
CatalogKKLevelAlgorithmAvg Final ∣Regret∣|\text{Regret}|Avg Std DevRelative Std
n=15n=151GroupGroup-Blind97,102.9863.00.8888%
n=15n=151GroupIntersection-Fit9,105.6203.32.2331%
n=15n=151GroupInner-Outer5,616.8166.02.9552%
n=15n=151GroupGroup-Item5,616.8166.02.9552%
n=15n=151IntersectionGroup-Blind728.434.54.7325%
n=15n=151IntersectionIntersection-Fit513.821.64.1955%
Appendix
Table 1: Robustness Analysis: Multi-Seed Variance in Goodreads ( T=200,000T=200{,}000 )
CatalogKKLevelAlgorithmAvg Final ∣Regret∣|\text{Regret}|Avg Std DevRelative Std
n=15n=151GroupGroup-Blind79,064.81,080.51.3666%
n=15n=151GroupIntersection-Fit72,146.8532.80.7385%
n=15n=151GroupInner-Outer36,677.2599.51.6345%
n=15n=151GroupGroup-Item36,677.2599.51.6345%
n=15n=151IntersectionGroup-Blind89,014.41,234.11.3864%
n=15n=151IntersectionIntersection-Fit36,947.7481.41.3031%
Appendix
Table 2: Robustness Analysis: Multi-Seed Variance in BoardGameGeek ( T=10,000,000T=10{,}000{,}000 )
More in this topic

Recommender Systems

Recommender systems predict and rank items for people using interaction histories, content, context, and feedback.

See all 545 papers

Explore similar work

CardsList
Jan 23, 2026

Explaining Group Recommendations via Counterfactuals

Group recommender systems help users make collective choices but often lack transparency, leaving group members uncertain about why items are suggested. Existing explanation methods focus on individuals, offering limited support for groups where multiple preferences interact. In this work, we introduce and formalize the problem of finding counterfactual explanations for group recommendations. We introduce a comprehensive set of item- and explanation-level metrics that capture different aspects, such as relevance, popularity, explanation utility, interpretability, and fairness. We design five algorithms that employ techniques such as search-space pruning, greedy selection, early stopping, and incremental refinement to reduce computational cost and the number of recommender calls while producing high-quality explanations. The algorithms consider different trade-offs; for example, some prioritize computational efficiency at the expense of explanation quality, while others focus on improving explanation quality at a higher computational cost. Experiments on real-world datasets analyze these trade-offs in terms of explanation size, computational cost, utility, and fairness.

Maria Stratigi, Nikos Bikakis, Kostas Stefanidis

Data Science Research Centre, Tampere University, Tampere, Finland · Hellenic Mediterranean University & Archimedes/Athena RC, Greece

Recommender SystemsExplainable Recommendation
AskPDF
Jul 1, 2025

Rethinking Group Recommender Systems in the Era of Generative AI: From One-Shot Recommendations to Conversational Group Decision Support

More than twenty-five years ago, early ideas emerged on how to design systems that provide recommendations to groups of users rather than individuals. Since then, a wide range of algorithmic approaches has been proposed to elicit individual preferences, aggregate them, and generate group recommendations. However, despite this substantial body of work, relatively few examples of real-world group recommender systems are in use. This observation invites reflection on the assumptions of the academic research, particularly regarding how inter-group communication unfolds and how recommendation-supported decisions should be made. In this essay, we suggest that some of these assumptions and corresponding system designs may not fully align with users' needs or expectations. We therefore outline a forward-looking perspective that explores how insights from group-decision making research and recent advances in Generative AI, including systems such as ChatGPT, could better inform the design of group recommender systems. As one possible direction, we consider settings in which human group members interact and exchange information through a chat based interface, while an AI-based agent supports the decision-making process. Such an approach may contribute to design more natural group and system interactions, and could help facilitate a broader adoption of group recommender systems in practice.

Dietmar Jannach, Amra Delić, Francesco Ricci +1

University of Klagenfurt, Austria · University of Sarajevo, Bosnia · Free University of Bozen-Bolzano, Italy

Recommender SystemsHuman-AI Decision Making
AskPDF
Jun 1, 2026

Rank-Constrained Deep Matrix Completion for Group Recommendation

The growing popularity of group activities has increased the need for methods that provide recommendations to groups of users given their individual preferences. Many existing group recommender systems rely on aggregating individual user preferences, but they often struggle with high-dimensional and highly sparse rating data commonly found in real-world scenarios. We propose Group Rank-Constrained Deep Matrix Completion (Group RC-DMC), a novel framework that extends RC-DMC by integrating group-level representation learning via a Set-Transformer aggregator, jointly leveraging low-rank structure and attention-based nonlinear modeling. Unlike most existing group recommender systems, Group RC-DMC unifies explicit low-rank regularization, linear encoder-decoder architectures, and attention-based nonlinear group modeling within a single framework, yielding accurate predictions at both the individual and group levels. Group RC-DMC addresses data sparsity through low-rank matrix completion, computing per-user latent representations from observed ratings only, and enforcing a rank constraint on the latent space using a nuclear-norm proximal step based on periodic singular value thresholding. The decoder is parametrized as a low-rank factorization, enabling efficient inference. Experimental results on the MovieLens and Goodbooks datasets demonstrate that Group RC-DMC achieves superior reconstruction accuracy, measured by lower group RMSE, while remaining computationally efficient and competitive in group-level performance in terms of precision, recall, and F1 score compared with weighted-before-factorization (WBF) and after-factorization (AF) baselines. The results highlight the model's ability to recover the underlying low-rank structure of user-item interactions and provide robust group recommendations across small, medium, and large user groups.

Mubaraka Sani Ibrahim, Lehel Csató, Isah Charles Saidu

Department of Computer Science, African University of Science and Technology, Abuja, Nigeria. · Faculty of Mathematics and Computer Science, Babes-Bolyai University, Cluj-Napoca, Romania. · Department of Computer Science, Baze University, Abuja, Nigeria.

Recommender SystemsMatrix Completion
AskPDF