TLDR: This paper presents the first comprehensive reproducibility study of product-side fairness in bundle recommendation (BR) systems. It investigates how popularity bias in historical data affects exposure at both bundle and item levels, evaluates the fairness of four state-of-the-art BR methods across three datasets, and analyzes the impact of user interaction tendencies. Key findings include that exposure patterns differ between bundles and items, fairness assessments vary by metric, and user behavior significantly influences fairness outcomes, highlighting the need for dual-layer fairness interventions in BR.
Recommender systems are everywhere, from online shopping to music playlists, helping us discover new things. While these systems are designed to show us what we might like, a growing concern is whether they are fair, especially to the products or items being recommended. This is known as product-side fairness, ensuring that products and their suppliers get equal or proportional visibility.
Traditionally, fairness in recommender systems has focused on individual items. However, a new challenge arises with ‘bundle recommendation’ (BR), where systems suggest collections of items, like a fashion outfit or a set of books. This introduces a unique complexity: recommendations are made at the bundle level, but user satisfaction and product visibility depend on both the bundle and the individual items within it. Existing fairness frameworks often don’t directly apply to this multi-layered scenario.
A Deep Dive into Bundle Recommendation Fairness
A recent study, titled “A Reproducibility Study of Product-side Fairness in Bundle Recommendation,” by Huy-Son Nguyen, Yuanna Liu, Masoud Mansoury, Mohammad Aliannejadi, Alan Hanjalic, and Maarten de Rijke, takes the first comprehensive look at product-side fairness in bundle recommendation. The researchers aimed to understand how fairly current BR methods allocate exposure to both bundles and the individual items they contain, and how biases might spread through these systems. You can read the full research paper here: Research Paper.
The study explored three key questions:
- How does popularity bias in historical user interactions lead to unfair exposure at both the bundle and item levels in BR methods?
- How fairly do BR methods expose bundles of varying popularity, and how does this affect the exposure of items within those bundles?
- How do users’ interaction preferences (e.g., preferring bundles vs. individual items) impact fairness outcomes in BR?
Methodology and Findings
To answer these questions, the team conducted an extensive empirical study using four advanced BR methods across three real-world datasets: Youshu (book lists), NetEase (music playlists), and iFashion (fashion outfits). They used six widely accepted fairness metrics to measure disparities at both bundle and item levels. A novel approach was also devised to investigate the impact of user preferences by grouping users based on their interaction tendencies.
The findings revealed several important patterns:
- Popularity Bias Amplification: The study found that popularity bias present in historical user interaction data often gets amplified in the recommendation results. This means that already popular bundles and items tend to receive even more exposure, while less popular ones become even less visible. Interestingly, the distribution of interactions for bundles and items doesn’t always align, especially in datasets like NetEase, where user preferences for bundles and individual items were less consistent.
- Fairness Trade-offs: When evaluating the fairness of BR methods, the researchers observed a common trade-off: methods that performed well in terms of accuracy (recommending relevant bundles) didn’t always achieve the best fairness scores. Furthermore, achieving fairness at the bundle level did not automatically translate to fairness at the item level. This highlights the complex nature of fairness in BR, where interventions might be needed at both layers.
- User Behavior Matters: The study categorized users into groups based on whether they tended to interact more with bundles, individual items, or equally with both. It was found that BR methods generally achieved fairer exposure distributions when serving users who primarily interacted with bundles. This could be because bundles often have less overlap, leading to greater diversity in exposure.
Also Read:
- Navigating the Complexities of Point of Interest Recommendation: Challenges and a Path Forward
- GRACE: Unpacking User Shopping Journeys for Better Recommendations
Conclusion and Future Directions
This pioneering study underscores the intricate nature of fairness in bundle recommendation. It emphasizes that simply focusing on bundle-level fairness is insufficient; item-level dynamics must also be considered. The varying assessments from different fairness metrics also reinforce the need for a multi-faceted evaluation approach.
The research provides actionable insights for building fairer bundle recommender systems and lays a crucial foundation for future work. The authors plan to expand their analysis beyond popularity, exploring other grouping strategies like item categories, supplier identities, or sensitive user and item attributes to uncover additional fairness issues and develop more holistic and inclusive fairness-aware BR methods.
