Source-linked AI summary
Feedback Loop and Bias Amplification in Recommender Systems
Masoud Mansoury, Himan Abdollahpouri, Mykola Pechenizkiy, Bamshad Mobasher, Robin Burke
TL;DR
Recommender systems can amplify popularity bias as users interact with recommendations and feed selected items back into their profiles. The paper simulates this process offline across several algorithms and finds that amplification also affects diversity, taste representation, and user-group experience, with stronger effects for minority users.
Problem
Popularity-biased recommendations and their feedback with user profiles require analysis of how bias changes over repeated interactions, beyond a single-step view.
Method
The paper simulates repeated offline interactions by applying recommendation algorithms, modeling item acceptance and ratings, and adding selected interactions to user profiles.
Results
Different recommendation algorithms amplify existing popularity bias over iterations, alongside declining aggregate diversity, shifting user-preference profiles, and user-group homogenization.
Takeaways & Limitations
Bias amplification can greatly increase a recommender system's existing bias over time, emphasizing the importance of addressing popularity bias and increasing recommendation diversity.
Takeaways & Limitations
The study groups users using a predefined gender label, while alternative groupings based on profile or rating characteristics remain possible.
Abstract
from arXiv · showhide
Recommendation algorithms are known to suffer from popularity bias; a few popular items are recommended frequently while the majority of other items are ignored. These recommendations are then consumed by the users, their reaction will be logged and added to the system: what is generally known as a feedback loop. In this paper, we propose a method for simulating the users interaction with the recommenders in an offline setting and study the impact of feedback loop on the popularity bias amplification of several recommendation algorithms. We then show how this bias amplification leads to several other problems such as declining the aggregate diversity, shifting the representation of users' taste over time and also homogenization of the users experience. In particular, we show that the impact of feedback loop is generally stronger for the users who belong to the minority group.
1 INTRODUCTION
The paper frames recommender bias as both inherited from skewed interaction data and potentially intensified through feedback loops. It investigates this amplification offline, including its effects on diversity, taste representation, user homogenization, and minority groups.
- Collaborative Filtering uses historical user–item interactions to generate personalized recommendations.
- Popularity bias arises because a few items receive many ratings while most items receive little attention.
- Algorithms can propagate existing data bias and recommend popular items even to users uninterested in them.
- User reactions to biased recommendations are added to profiles, creating a feedback loop in which recommendations and profiles dynamically affect each other.
- The study simulates user interactions offline to examine popularity-bias amplification and resulting changes in diversity, taste representation, and user homogenization.
- The feedback loop generally has a stronger impact on users in the minority group.
2 FEEDBACK LOOP SIMULATION
The paper simulates repeated recommender interactions offline by generating recommendations, probabilistically selecting items, estimating ratings, and adding new interactions to user profiles.
- The offline simulation iteratively generates recommendation lists and updates user profiles with selected items based on acceptance probabilities.
- At each iteration, rating data are split into 80% training and 20% test sets.
- A recommendation model trained on the training data produces a list for every user.
- Acceptance probabilities assign higher selection chances to higher-ranked items in each recommendation list.
- The simulation separately estimates a selected item's potential rating using the user's profile statistics, the item's average rating, and Gaussian noise.
- New item–rating pairs are added only when the item is absent from the user's profile, and the process repeats to form the next iteration's data.
3 MODELING FEEDBACK
The paper models feedback-loop propagation by relating recommendation popularity to rating-data popularity and showing that added interactions increase popularity across iterations.
- PDt and PRt denote the average popularity of items in the rating data and recommendations at iteration t.
- θt represents the percentage increase in recommendation popularity relative to rating-data popularity at iteration t.
- Adding K interactions increases the next iteration's rating-data size to |Dt| + K.
- The average popularity of next-iteration rating data increases by K×θt/(|Dt|+K).
- Because recommendation popularity is proportional to rating-data popularity plus a positive increment, PRt+1 > PRt across iterations.
4 EXPERIMENTS
Experiments simulate feedback-loop interactions on MovieLens 1M across three recommendation algorithms. Over iterations, popularity bias increases alongside declining catalog coverage, shifting taste representations, preference homogenization, and stronger deviation for the minority group.
- 4.2 Algorithms: The evaluation compares UserKNN, BPR, and MostPopular over 20 simulated feedback-loop iterations.UserKNN uses explicit ratings, BPR uses binary data, and MostPopular recommends unseen popular items.
- 4.3.1 Popularity bias amplification: Average recommendation popularity rises across all algorithms, with BPR showing the steepest increase and therefore stronger bias propagation.The algorithms begin at different popularity levels because of their inherent designs.
- 4.3.1 Popularity bias amplification: Aggregate diversity, defined as catalog coverage, declines as recommendations concentrate on popular items, especially for BPR.Catalog coverage is the percentage of items appearing at least once across users’ recommendation lists.
- 4.3.2 Shifting users’ taste representation: Users’ profile-based taste representations increasingly deviate from their initial preferences over time, although this does not establish changes in intrinsic preferences.The study uses KLD between initial and later genre distributions; higher KLD indicates greater deviation.
- 4.3.3 Homogenization: KLD between male and female preference representations decreases dramatically for all algorithms, indicating strong homogenization of users’ preferences.The analysis measures the distance between separately extracted male and female genre distributions over iterations.
- 4.3.3 Homogenization: Female preference representations move toward the initial population representation while male representations become slightly more distant, and female deviation from initial preferences is significantly higher.The authors attribute the group asymmetry to males contributing most ratings and recommended items being more likely to reflect male-rated items.
5 DISCUSSION AND FUTURE WORK
The paper identifies several directions for extending its offline feedback-loop study, including alternative item-selection policies, user-grouping strategies, and popularity-bias mitigation algorithms. It also notes that repeated interactions with already-profiled items could be modeled differently in domains such as music.
- In music and similar domains, ratings for items already in a user’s profile could be updated across iterations instead of excluding those items.The current selection restriction prevents already-profiled items from being added again.
- User groups could be defined using profile size, rated-item popularity, or other statistical characteristics instead of predefined gender labels.
- Popularity-bias mitigation algorithms could be evaluated for whether they reduce feedback-loop amplification.The paper hypothesizes that reducing popularity bias per iteration would also reduce amplification over time.
- Alternative user-selection policies, including top-1 or random selection, could replace the ranking-position-based technique used in Equation 1.
6 CONCLUSION
Using offline simulation, the paper shows that iterative user–recommender interactions amplify existing bias and affect diversity and preference representation. The amplification is stronger for the minority group, highlighting the need to address popularity bias over time.
- Offline simulation shows that recommendation algorithms amplify existing bias across successive iterations of user interaction.
- Feedback-loop amplification is associated with declining aggregate diversity, shifting preference representations, and homogenization of user groups.
- The bias amplification observed for females, the minority group in this dataset, was stronger than for males.