[Paper Review] Towards Fair Personalization by Avoiding Feedback Loops

Gökhan Çapan, Özge Bozal+2 more|arXiv (Cornell University)|Dec 20, 2020

Recommender Systems and Techniques22 references20 citations

TL;DR

This paper proposes a fair personalization framework that avoids feedback loops in recommender systems by explicitly modeling systematic exposure in preference inference. Using a Dirichlet-Luce model with Thompson sampling, it corrects biases from over- or under-presentation, enabling accurate preference estimation and discovery of initially censored top alternatives without overestimating promoted items.

ABSTRACT

Self-reinforcing feedback loops are both cause and effect of over and/or under-presentation of some content in interactive recommender systems. This leads to erroneous user preference estimates, namely, overestimation of over-presented content while violating the right to be presented of each alternative, contrary of which we define as a fair system. We consider two models that explicitly incorporate, or ignore the systematic and limited exposure to alternatives. By simulations, we demonstrate that ignoring the systematic presentations overestimates promoted options and underestimates censored alternatives. Simply conditioning on the limited exposure is a remedy for these biases.

Motivation & Objective

• To address fairness issues in interactive recommender systems caused by self-reinforcing feedback loops due to over- or under-presentation of content.
• To correct biased preference estimates arising from systematic exposure, such as overestimation of promoted or popular items and underestimation of rarely presented options.
• To ensure that all alternatives, including initially censored or less popular ones, have a fair chance of being discovered and recommended.
• To develop a personalization system that is robust to initial biases, promotes discovery of true user preferences, and avoids filter bubbles.

Proposed method

• The paper employs a Dirichlet-Luce model to perform Bayesian inference on user preferences, explicitly conditioning on the set of presented alternatives to correct exposure bias.
• It uses Thompson sampling to generate presentation sets based on posterior samples of preferences, enabling exploration of under-presented options and balancing exploitation and exploration.
• The system models user choices using a discrete choice model that satisfies Luce’s choice axiom, ensuring that choice probabilities are proportional to underlying preference strengths.
• Preference estimation is updated online using partial feedback from user choices within presented subsets, avoiding reliance on full-itemset feedback.
• The method contrasts with the naive Dirichlet-Multinomial model, which ignores presentation structure and thus fails to correct for exposure bias.
• Monte Carlo methods are used to approximate intractable quantities in the Dirichlet-Luce model, enabling practical implementation despite computational cost.

Experimental results

Research questions

• RQ1 How does ignoring systematic exposure in preference inference lead to biased estimates of user preferences in recommender systems?
• RQ2 Can explicitly modeling the presentation set in the inference process correct overestimation of over-presented items and underestimation of under-presented ones?
• RQ3 To what extent can a recommendation system discover and eventually present the user’s true favorite options when they are initially censored from presentations?
• RQ4 How does the system’s initial choice bias affect long-term preference estimation, and can it be corrected over time?
• RQ5 Does incorporating exposure structure into the model improve fairness, discovery, and robustness to unfair comparisons?

Key findings

• The Dirichlet-Luce model, which conditions on the presentation set, successfully corrects overestimation of promoted options—such as option 3 included in every presentation—whereas the Dirichlet-Multinomial model overestimates it.
• When initially censored options (e.g., options 4 and 5) are not presented, the Dirichlet-Luce model avoids imposing negative bias and eventually learns to present the true top alternatives, while the Dirichlet-Multinomial model fails to recover their true preference strength.
• The Dirichlet-Luce model corrects initial misconceptions caused by unfair comparisons, such as repeatedly presenting a good option alongside a better one, by gradually adjusting preference estimates as more data accumulates.
• Thompson sampling-based presentation mechanisms outperform greedy strategies by enabling exploration of under-presented alternatives, leading to better long-term discovery of user preferences.
• The model achieves robustness to both promotion bias and censorship bias, demonstrating that conditioning on exposure is essential for fair and accurate personalization.
• Despite higher computational cost due to intractable expectations, Monte Carlo approximation enables effective implementation of the Dirichlet-Luce model in practice.