Sensitive Attribute Association Bias in Latent Factor Recommendation Algorithms: Theory and In Practice

Beattie, Alexandra

Sensitive Attribute Association Bias in Latent Factor Recommendation Algorithms: Theory and In Practice

dc.contributor.advisor	Hougen, Dean
dc.contributor.author	Beattie, Alexandra
dc.contributor.committeeMember	Nicholson, Charles
dc.contributor.committeeMember	Razzaghi, Talayeh
dc.contributor.committeeMember	Diochnos, Dimitrios
dc.contributor.committeeMember	Harris-Watson, Alexandra
dc.date.accessioned	2023-12-06T22:27:22Z
dc.date.available	2023-12-06T22:27:22Z
dc.date.issued	2023-12-15
dc.date.manuscript	2023-12-06
dc.description.abstract	This dissertation presents methods for evaluating and mitigating a relatively unexplored bias topic in recommendation systems, which we refer to as attribute association bias. Attribute association bias (AAB) can be introduced when leveraging latent factor recommendation models due to their ability to entangle model and implicit attributes into the trained latent space. This type of bias occurs when entity embeddings showcase significant levels of association with specific types of explicit or implicit entity attributes, thus having the potential to introduce representative harms for both consumer and provider stakeholders. We present a novel analysis method framework to help practitioners evaluate their latent factor recommendation models for AAB. This framework consists of three main techniques for gaining insight into sensitive AAB in the recommendation latent space: bias direction creation, bias evaluation metrics, and multi-group evaluation. Methods within our evaluation framework were inspired by techniques presented by the natural language processing research community for measuring gender bias in learned language representations. Additionally, we explore how this bias can be reinforced and produce feedback loops via retraining. Finally, we explore possible mitigation techniques for addressing said bias. Primarily, we demonstrate our methodology with two case studies that evaluate user gender association bias in latent factor recommendation. With our methods, we uncover the existence of user gender association bias and compare the various methods we propose to help guide practitioners in how best to use our techniques for their systems. In addition to exploring user gender, we experiment with measuring user age association bias as a means for evaluating non-binary AAB.	en_US
dc.identifier.uri	https://hdl.handle.net/11244/340008
dc.language	en_US	en_US
dc.subject	Computer Science	en_US
dc.subject	Recommender Systems	en_US
dc.subject	Responsible AI	en_US
dc.subject	Machine Learning Evaluation	en_US
dc.thesis.degree	Ph.D.	en_US
dc.title	Sensitive Attribute Association Bias in Latent Factor Recommendation Algorithms: Theory and In Practice	en_US
ou.group	Gallogly College of Engineering	en_US
shareok.orcid	0000-0003-3351-5951	en_US