Learning recommender systems from biased user interactions

Recommender systems have been widely deployed to help users quickly find what they need from a collection of items. Predominant recommendation methods rely on supervised learning models to predict user ratings on items or the probabilities of users interacting with items. In addition, reinforcement learning models are crucial in improving long-term user engagement within recommender systems. In practice, both of these recommendation methods are commonly trained on logged user interactions and, therefore, subject to bias present in logged user interactions. This thesis concerns complex forms of bias in real-world user behaviors and aims to mitigate the effect of bias on reinforcement learning-based recommendation methods.
The first part of the thesis consists of two research chapters, each dedicated to tackling a specific form of bias: dynamic selection bias and multifactorial bias. To mitigate the effect of dynamic selection bias and multifactorial bias, we propose a bias propensity estimation method for each. By incorporating the results from the bias propensity estimation methods, the widely used inverse propensity scoring-based debiasing method can be extended to correct for the corresponding bias.
The second part of the thesis consists of two chapters that concern the effect of bias on reinforcement learning-based recommendation methods. Its first chapter focuses on mitigating the effect of bias on simulators, which enables the learning and evaluation of reinforcement learning-based recommendation methods. Its second chapter further explores different state encoders for reinforcement learning-based recommendation methods when learning and evaluating with the proposed debiased simulator.