Achieving true fairness in recommender systems requires moving beyond broad demographic categories to address disparities at a fine-grained level, ensuring equitable representation for all subgroups. This goal can be made feasible through advanced re-ranking methodologies like CONFIGRE.

Recommender systems are ubiquitous in today’s digital landscape, providing tailored suggestions to users in domains like e-commerce, entertainment, and education. However, as these systems influence user choices and provider visibility, ensuring fairness becomes a critical challenge.

In a recent study, with Elizabeth Gómez, David Contreras, and Maria Salamó, published in the proceedings of the ACM UMAP 2024 conference, we introduce CONFIGRE, a novel re-ranking algorithm designed to address equity across both coarse and fine-grained provider groups.

Understanding fairness in recommender systems

Equity in recommender systems often focuses on provider fairness, ensuring visibility for items from various groups proportional to their representation in user interactions. Traditionally, fairness has been applied to coarse-grained groups, such as continents or binary gender classifications, due to their manageability. However, this approach can obscure inequities within these groups, disadvantaging smaller, fine-grained subgroups, like specific countries within a continent.

The CONFIGRE Approach

CONFIGRE (COarse aNd FIne Grained RE-ranking) is a post-processing algorithm that tackles this limitation by redistributing visibility among providers at both coarse and fine-grained levels. Here’s how it works:

1. Group Representation calculation. CONFIGRE first computes the representation of each group (e.g., continents and countries) based on the training dataset.
2. Bucket creation. Items are grouped into buckets by their coarse (e.g., continent) and fine-grained (e.g., country) group attributes, sorted by their predicted relevance.
3. Re-ranking in phases.
   • Phase 1. Prioritizes underrepresented fine-grained groups, ensuring their representation matches their training set proportions.
   • Phase 2. Balances recommendations across coarse-grained groups while respecting fine-grained representation as much as possible.
   • Phase 3. Completes the recommendation list with the most relevant items if earlier phases leave gaps.

Experimental Validation

The effectiveness of CONFIGRE was tested on two datasets:

• MovieLens-1M. A movie recommendation dataset enriched with geographic metadata.
• DataSongs. A newly created dataset of song ratings with detailed geographic attributes.

Our main findings include:

• Reduction in disparities. CONFIGRE significantly reduced disparities in visibility for fine-grained groups. For example, in the DataSongs dataset, it reduced country-level visibility disparity by over 50% for some algorithms.
• Enhanced coverage. The number of represented countries increased across all algorithms, demonstrating CONFIGRE’s ability to improve diversity in recommendations.
• Minimal impact on effectiveness. While fairness adjustments often compromise recommendation quality, CONFIGRE achieved its goals with negligible loss in accuracy (measured by NDCG).

Conclusions

By addressing fairness at both coarse and fine-grained levels, CONFIGRE ensures that smaller providers within larger groups are not under-recommended. This has far-reaching implications, especially in domains like education and entertainment, where equitable exposure can significantly influence opportunities and success.

This study highlights the importance of considering multiple levels of granularity in fairness-aware recommender systems. Future research could explore applying CONFIGRE to other sensitive attributes, such as socio-economic status or linguistic diversity, and extending its utility to multi-stakeholder environments.