Feedback Loop and Bias Amplification in Recommender Systems

Abstract

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. In particular, we show that the impact of feedback loop is generally stronger for the users who belong to the minority group. CCS CONCEPTS • Information systems → Recommender systems; • Humancentered computing → Interactive systems and tools.

Simulations for novel problems in recommendation: analyzing misinformation and data characteristics

“…When done properly, this would allow generating and studying feedback loops, like those created by reinforcement learning algorithms, but at a higher, more global level. As observed recently, and in agreement with our misinformation analysis [21], some algorithms are more prone to reproduce biases at each feedback loop [29].…”

Towards responsible media recommendation

“…Algorithms can also amplify already existing biases. For example, recommendation algorithms that are trained on the MovieLens data set 4 -a very popular data set in the research community-were found to strongly intensify the preexisting popularity bias they inherited from the data set [7,60].…”

“…The numbers clearly indicate that a small fraction of the movies that are recommended on the front page accounts for a large number of the recorded page impressions. Interestingly, while recommender systems are often considered as means to increase sales in the long tail, the concentration on the short head can in fact be increased by a recommender system [36,48,60]. A popularity bias can be amplified by a recommender system when it learns from the recorded data to recommend popular items more frequently than less popular items.…”

Non-technical losses detection in energy consumption focusing on energy recovery and explainability

“…In some cases, the problem is bias-related, and if there is a feedback loop (i.e. the model learns based on its previous predictions), the bias is aggravated in each new prediction made (Mehrabi et al 2019;Mansoury et al 2020). In other cases, the problem is related to the fact that the dataset evolves over time.…”