Date on Master's Thesis/Doctoral Dissertation

5-2022

Document Type

Doctoral Dissertation

Degree Name

Ph. D.

Department

Computer Engineering and Computer Science

Degree Program

Computer Science and Engineering, PhD

Committee Chair

Nasraoui, Olfa

Committee Co-Chair (if applicable)

Altiparmak, Nihat

Committee Member

Altiparmak, Nihat

Committee Member

Frigui, Hichem

Committee Member

Park, Juw Won

Committee Member

Cashon, Cara

Author's Keywords

machine learning; recommender systems; artificial intelligence; fairness; AI ethics

Abstract

Artificial Intelligence (AI)-driven recommender systems have been gaining increasing ubiquity and influence in our daily lives, especially during time spent online on the World Wide Web or smart devices. The influence of recommender systems on who and what we can find and discover, our choices, and our behavior, has thus never been more concrete. AI can now predict and anticipate, with varying degrees of accuracy, the news article we will read, the music we will listen to, the movies we will watch, the transactions we will make, the restaurants we will eat in, the online courses we will be interested in, and the people we will connect with for various ends and purposes. For all these reasons, the automated predictions and recommendations made by AI can lead to influencing and changing human opinions, behavior, and decision making. When the AI predictions are biased, the influences can have unfair consequences on society, ranging from social polarization to the amplification of misinformation and hate speech. For instance, bias in recommender systems can affect the decision making and shift consumer behavior in an unfair way due to a phenomenon known as the feedback loop. The feedback loop is an inherent component of recommender systems because the latter are dynamic systems that involve continuous interactions with the users, whereby data collected to train a recommender system model is usually affected by the outputs of a previously trained model. This feedback loop is expected to affect the performance of the system. For instance, it can amplify initial bias in the data or model and can lead to other phenomena such as filter bubbles, polarization, and popularity bias. Up to now, it has been difficult to understand the dynamics of recommender system feedback loops, and equally challenging to evaluate the bias and filter bubbles emerging from recommender system models within such an iterative closed loop environment. In this dissertation, we study the feedback loop in the context of Collaborative Filtering (CF) recommender systems. CF systems comprise the leading family of recommender systems that rely mainly on mining the patterns of interaction between the users and items to train models that aim to predict future user interactions. Our research contributions target three aspects of recommendation, namely modeling, debiasing and evaluating feedback loops. Our research advances the state of the art in Fairness in Artificial Intelligence on several fronts: (1) We propose and validate a new theoretical model, based on Martingale differences, to model the recommender system feedback loop, and allow a better understanding of the dynamics of filter bubbles and user discovery. (2) We propose a Transformer-based deep learning architecture and algorithm to learn diverse representations for users and items in order to increase the diversity in the recommendations. Our evaluation experiments on real world datasets demonstrate that our transformer model recommends 14\% more diverse items and improves the novelty of the recommendation by more than 20\%. (3) We propose a new simulation and experimentation framework that allows studying and tracking the evolution of bias metrics in a feedback loop setting, for a variety of recommendation modeling algorithms. Our preliminary findings, using the new simulation framework show that recommender systems are deeply affected by the feedback loop, and that without an adequate debiasing or exploration strategy, this feedback loop limits the discovery of the user and increases the disparity in exposure between items that can be recommended. To help the research and practice community in studying recommender system fairness, all the tools developed to model, debias, and evaluate recommender systems are made available to the public as open source software libraries \footnote{https://github.com/samikhenissi/TheoretUserModeling}. (4) We propose a novel learnable dynamic debiasing strategy that learns an optimal rescaling parameter for the predicted rating and achieves a better trade-off between accuracy and debiasing. We focus on solving the popularity bias of the items and test our method using our proposed simulation framework and show the effectiveness of using a learnable debiasing degree to produce better results.

Download

Share

COinS