REVEALING AND ADDRESSING BIAS IN RECOMMENDER SYSTEMS
Abstract
In the past decades, recommenders have achieved outstanding success in delivering personalized and accurate recommendations. Highly customized recommendations bring individuals great convenience, while helping content providers to connect with interested consumers accurately. However, they may also lead to undesirable outcomes, often through the introduction of bias in the training, deployment, and maintenance of recommender systems. For example, recommenders may impose unfair burdens on certain user groups, disadvantaging their prominence on job-based recommenders. And they may narrow down a user’s interest areas, raising concerns of echo chambers, fairness, and diversity. In this dissertation, we ground our work in identifying gaps in the literature for identifying and addressing bias in recommender systems, then introduce four approaches for mitigating biases from multiple perspectives, listed as below:
• First, we identify an inherent bias in many recommendation algorithms which optimize for the head (or popular portion) of the rating distribution, thus lead to large estimation errors for tail ratings. We conduct a data-driven investigation and theoretical analysis of the challenges posed by traditional latent factor models for estimating such tail ratings. With these challenges in mind, we propose a new multi-latent representation method designed specifically to estimate these tail ratings better, to reduce the bias associated with tail ratings in recommender systems.
• Second, we address another unexplored bias – the target customer distribution distortion. Traditional recommender systems typically aim to optimize an engagement metric without considering the overall distribution of target customers, thereby leading to serious distortion problems. Through a data-driven study, we reveal several distortions that arise from conventional recommenders. Toward overcoming these issues, we propose a target customer re-ranking algorithm to adjust the population distribution and composition in the Top-k target customers of an item while maintaining recommendation quality.
ii
• Third, we focus on mitigating the next unexplored bias – user’s taste distortion. We show how existing approaches assume a static view of user’s tastes, and so previously proposed calibrated recommenders result in poor modeling of the shift of a user’s evolution. Thus, we empirically identify the taste distortion problem through a data-driven study over multiple datasets. We propose a taste-enhanced calibrated recommender system designed with the shifts and trends of user’s taste preferences in mind, which results in improved taste distribution estimation and recommendation results.
• Last but not least, we study the distribution bias in a dynamic recommendation environment. Previous studies of such distribution-aware recommendation have focused exclusively on static scenarios, ignoring important challenges that manifest in real-world dynamics and leading to poor performance in practice. Hence, we present the first study of distribution bias in dynamic recommendations, and propose new methods to mitigate this bias even in the presence of feedback loops and other dynamics.
Citation
Zhao, Xing (2021). REVEALING AND ADDRESSING BIAS IN RECOMMENDER SYSTEMS. Doctoral dissertation, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /195114.