New debiasing strategies in collaborative filtering recommender systems: modeling user conformity, multiple biases, and causality.

Recommender Systems are widely used to personalize the user experience in a diverse set of online applications ranging from e-commerce and education to social media and online entertainment. These State of the Art AI systems can suffer from several biases that may occur at different stages of the recommendation life-cycle. For instance, using biased data to train recommendation models may lead to several issues, such as the discrepancy between online and offline evaluation, decreasing the recommendation performance, and hurting the user experience. Bias can occur during the data collection stage where the data inherits the user-item interaction biases, such as selection and exposure bias. Bias can also occur in the training stage, where popular items tend to be recommended much more frequently given that they received more interactions to start with. The closed feedback loop nature of online recommender systems will further amplify the latter biases as well. In this dissertation, we study the bias in the context of Collaborative Filtering recommender system, and propose a new Popularity Correction Matrix Factorization (PCMF) that aims to improve the recommender system performance as well as decrease popularity bias and increase the diversity of items in the recommendation lists. PCMF mitigates popularity bias by disentangling relevance and conformity and by learning a user-personalized bias vector to capture the users\u27 individual conformity levels along a full spectrum of conformity bias. One shortcoming of the proposed PCMF debiasing approach, is its assumption that the recommender system is affected by only popularity bias. However in the real word, different types of bias do occur simultaneously and interact with one another. We therefore relax the latter assumption and propose a multi-pronged approach that can account for two biases simultaneously, namely popularity and exposure bias. our experimental results show that accounting for multiple biases does improve the results in terms of providing more accurate and less biased results. Finally, we propose a novel two-stage debiasing approach, inspired from the proximal causal inference framework. Unlike the existing causal IPS approach that corrects for observed confounders, our proposed approach corrects for both observed and potential unobserved confounders. The approach relies on a pair of negative control variables to adjust for the bias in the potential ratings. Our proposed approach outperforms state of the art causal approaches, proving that accounting for unobserved confounders can improve the recommendation system\u27s performance