Evaluating, Understanding, and Mitigating Unfairness in Recommender Systems

Recommender systems are information filtering tools that discover potential matchings between users and items and benefit both parties. This benefit can be considered a social resource that should be equitably allocated across users and items, especially in critical domains such as education and employment. Biases and unfairness in recommendations raise both ethical and legal concerns. In this dissertation, we investigate the concept of unfairness in the context of recommender systems. In particular, we study appropriate unfairness evaluation metrics, examine the relation between bias in recommender models and inequality in the underlying population, as well as propose effective unfairness mitigation approaches.

We start with exploring the implication of fairness in recommendation and formulating unfairness evaluation metrics. We focus on the task of rating prediction. We identify the insufficiency of demographic parity for scenarios where the target variable is justifiably dependent on demographic features. Then we propose an alternative set of unfairness metrics that measured based on how much the average predicted ratings deviate from average true ratings. We also reduce these unfairness in matrix factorization (MF) models by explicitly adding them as penalty terms to learning objectives.

Next, we target a form of unfairness in matrix factorization models observed as disparate model performance across user groups. We identify four types of biases in the training data that contribute to higher subpopulation error. Then we propose personalized regularization learning (PRL), which learns personalized regularization parameters that directly address the data biases. PRL poses the hyperparameter search problem as a secondary learning task. It enables back-propagation to learn the personalized regularization parameters by leveraging the closed-form solutions of alternating least squares (ALS) to solve MF. Furthermore, the learned parameters are interpretable and provide insights into how fairness is improved.

Third, we conduct theoretical analysis on the long-term dynamics of inequality in the underlying population, in terms of the fitting between users and items. We view the task of recommendation as solving a set of classification problems through threshold policies. We mathematically formulate the transition dynamics of user-item fit in one step of recommendation. Then we prove that a system with the formulated dynamics always has at least one equilibrium, and we provide sufficient conditions for the equilibrium to be unique. We also show that, depending on the item category relationships and the recommendation policies, recommendations in one item category can reshape the user-item fit in another item category.

To summarize, in this research, we examine different fairness criteria in rating prediction and recommendation, study the dynamic of interactions between recommender systems and users, and propose mitigation methods to promote fairness and equality.