Publication Date

Summer 2024

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Teng-Sheng Moh

Keywords

Artificial Intelligence, Machine Learning, Popularity Bias, Recommendation Systems

Abstract

The ways most people consume the media have become very much driven by some pre-set algorithms. It is increasingly important to examine the outcome of these artificial intelligence (AI) models and ensure that any potentially dangerous long-term effects are addressed before they have a significant negative impact in our society. Popularity bias is one of these potentially harmful impacts, which stemmed from the shift from human intelligence to AI, or machine intelligence/machine learning (ML), when one explores the media and receives recommendations (often without requesting). In ML, three key steps usually occur; i.e, pre-processing, in-processing, and post- processing steps. The most common approach to address the issue of popularity bias is at the start of the pre-processing stage. This approach deals with the exposure bias, which believes that popularity bias occurs due to the inherited bias in the popularity distribution of the base version of the given data set the model trains on. This method, however, provides only a temporary solution to the root problem. To break the cycle of exposure bias and help bring an end to the resulting popularity bias, in this paper we propose a novel post-processing algorithm called Closer-To-Average (CTA). Continuing and combining our previous work in the in-processing stage of the ML cycle, as the proposed countermeasures are used in the real world, will result in new, less biased datasets that future models can be trained on. This not only also provides users with more relevant recommendations, but more importantly, it helps break the continued cycle of popularity bias. Experimental results show that the CTA algorithm, when applied at the end of a modified version of the SASRec recommendation model, provided a 6% improvement in Blockbuster scores compared to the baseline models.

In addition, when applying the proposed solution in a real-time, web-based system for movie recommendations, it only resulted in an additional overhead time of 0.3 seconds, a negligible increase for the application. We believe that the proposed method may be widely applicable to many RS and data sets, and will contribute significantly towards a more genuine, less divisive society through unbiased recommendations.

Download

Available for download on Friday, August 01, 2025

Share

COinS