Graduate Thesis Or Dissertation

Abstract

• Recommender systems are one of the most pervasive applications of machine learning. They play a pivotal role in helping users find items tailored to their taste. Although these systems intend to assist people in their information needs, they can cause implicit or explicit discrimination against individuals or groups. There are several ways that different biases can creep into recommender systems. Reflection of societal and historical prejudices in datasets and during the data collection process, lack of sufficient data on minority groups, lack of suitable evaluation methods and model designs to detect these biases and lessen the unfairness caused by them are among the many reasons for unfairness in these systems. A system needs to defend against the biases in recommendation output to prevent harm and unfairness. However, integrating the goal of fairness with accuracy in recommender systems is challenging, primarily because of this goal's significant trade-offs with accuracy. Accuracy in recommender systems is the ability of that system to predict users' needs and interests accurately. On the other hand, fairness is a complicated concept with a variety of definitions. To use fairness as an objective, we need to define it based on the application area and the context of a problem. Additionally, we need to specify the fairness concerns of the different stakeholders involved in the recommender systems and the fairness priorities of a system. Any of these aspects might disagree with the goal of accuracy. For example, if fairness for content providers is more exposure to users, increasing it might cause a reduction in accuracy. Therefore, controlling the trade-off between accuracy and fairness becomes essential. Throughout this dissertation, several recommendation models and re-ranking approaches are presented that aim to address this problem using in- and post- processing methods. These approaches show promising results, but it is worth mentioning that they have intrinsic limitations and, therefore, shouldn't be considered ultimate solutions.

Creator

• Sonboli, Nasim

Date Issued

• 2022-04-13

Academic Affiliation

• Information Science

Advisor

• Burke, Robin

Committee Member

• Mobasher, Bamshad
• Fiesler, Casey
• Waggoner, Bo
• Ekstrand, Michael

Degree Grantor

• University of Colorado Boulder

Commencement Year

• 2022

Subject

• Information science
• Fairness-aware Recommender systems
• Data Mining
• Fairness
• Computer science
• Machine Learning

Publisher

• University of Colorado Boulder

Last Modified

• 2022-07-07

Resource Type

• Dissertation

Rights Statement

• In Copyright

Language

• English [eng]

Relationships

Items

Thumbnail	Title	Date Uploaded	Visibility	Actions
	Sonboli_colorado_0051E_17692.pdf	2022-07-07	Public	Download
	Thesis_Approval_Form.pdf	2022-07-07	Public	Download
	IRB_for_one_of_the_studies.pdf	2022-07-07	Public	Download