New AI Fixes Biased Recommendations for Everyone
Sick of generic suggestions? This AI makes recommendations fair for all groups.
The research focuses on recommendation systems that optimize for the average user and can fail to capture the preferences of diverse demographic groups. It considers a dynamic system where arriving users can belong to different demographic groups, these groups can intersect, and a user may belong to several groups. The article shows that, under mild conditions, any group-blind policy recommending K ≥ 1 items incurs linear total regret over time.
To address this, the article introduces three group-aware algorithms based on UCB. Intersection-Fitting runs UCB on each intersection and has a sublinear group- and intersection-level regret guarantee with an exponential dependence in the number of groups. Inner-Outer UCB and Group-Item UCB are group-level heuristics that 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.
The article reports experiments on Goodreads and BoardGameGeek and highlights that the heuristics can be preferable to Intersection-Fitting when intersection-level preferences are closely related to group-level preferences, but that Intersection-Fitting performs better when different intersections within the same group prefer different items.
- Standard recommendation systems often ignore group differences, leading to worse suggestions over time.
- New algorithms that consider intersecting groups (like age and gender) can provide fairer and more accurate recommendations.
- Tests on Goodreads and BoardGameGeek show these methods improve user satisfaction by tailoring suggestions to specific group preferences.
Why It Matters
Fairer recommendations mean you see content you actually like, saving time and improving your online experience.