TLDR: This research paper empirically demonstrates that recommender systems can be fair at a group level but still highly unfair to individuals. It highlights that group and individual fairness measures are not interchangeable, and within-group disparities are often as significant as overall individual unfairness. The study advocates for evaluating both individual and within-group fairness alongside traditional group fairness to ensure truly equitable recommendation outcomes.
Fairness in artificial intelligence, particularly in recommender systems, is a topic of increasing importance. These systems, which suggest movies, jobs, or music, are now under scrutiny to ensure they don’t systematically disadvantage users. Traditionally, fairness has been categorized into two main types: group fairness and individual fairness.
Group fairness aims for equitable outcomes across different user groups – for example, ensuring similar recommendation quality for different age demographics or genders. Individual fairness, on the other hand, focuses on treating similar users or items equally, meaning that users with similar preferences should receive recommendations of comparable quality.
Despite the critical nature of both concepts, there has been a significant gap in our scientific understanding of how these two types of fairness relate to each other. Prior research often used different evaluation measures or objectives for each type, making a direct comparison challenging. This left a crucial question unanswered: how does improving one type of fairness impact the other?
Bridging the Fairness Gap
A recent research paper, “Stairway to Fairness: Connecting Group and Individual Fairness,” by Theresia Veronika Rampisela, Maria Maistro, Tuukka Ruotsalo, Falk Scholer, and Christina Lioma, addresses this very gap. The authors conducted a comprehensive study comparing evaluation measures applicable to both group and individual fairness, aiming to understand their intricate relationship.
The study utilized eight experimental runs across three diverse datasets: ML-1M (movie ratings), JobRec (job applications), and LFM-1B (music playcounts). To simulate modern recommender systems, they employed Large Language Model-based Recommenders (LLMRecs), specifically four open-source models: Llama-3.1-8B-Instruct, Qwen2.5-7B-Instruct, GLM-4-9B-chat, and Ministral-8B-Instruct-2410. Users were grouped based on various sensitive attributes like gender, age, occupation, academic degree, years of experience, study major, and country, including intersectional groupings (combinations of attributes).
Key Insights from the Research
The findings of this research offer profound insights for anyone involved in developing or deploying recommender systems:
1. Group Fairness Doesn’t Guarantee Individual Fairness: The most striking discovery was that recommendations deemed highly fair for groups could, at the same time, be very unfair for individuals. This is a novel empirical finding, demonstrating that achieving fairness at a group level does not automatically translate to fairness for each individual user within those groups.
2. No Reliable Proxy for Fairness Measures: The study found that individual fairness measures do not consistently provide equivalent rankings to group fairness measures. This means that one cannot simply use an individual fairness metric as a stand-in for a group fairness metric, or vice versa. Both types of fairness need to be explicitly evaluated to get a complete picture.
3. The Impact of Intersectional Grouping: As more sensitive attributes were combined to form user groups (e.g., considering both gender and age), the difficulty of achieving group fairness increased. This highlights the critical importance of evaluating “intersectional fairness,” which considers the unique experiences of individuals at the intersection of multiple social identities.
4. Hidden Unfairness Within Groups: The research revealed that within-group unfairness (disparities among individuals within the same group) was almost as high as overall individual unfairness and consistently worse than between-group unfairness. This implies that focusing solely on between-group fairness metrics can mask significant disparities and unfairness experienced by individual users within those groups.
Also Read:
- Streamlining Recommender Systems: Unpacking the Necessity of User Data
- Navigating the Complex World of Text Anonymization: A Comprehensive Review
Implications for Recommender System Design
These findings are crucial for practitioners aiming to improve the fairness of their systems. The paper strongly advocates for evaluating individual and within-group fairness alongside traditional group fairness. Relying solely on between-group fairness scores can create a false sense of security, as they may hide substantial disparities in recommendation effectiveness across individual users, even when accounting for within-group variations.
The research provides the first empirical evidence of the disjoint nature of group and individual fairness concepts in recommender systems. It suggests that future fairness mitigation strategies should consider incorporating terms for both between-group and within-group fairness in their optimization objectives.
You can read the full research paper here.
