Toward Responsible Recommender Systems
Abstract
Recommender systems have become essential conduits: they can shape the media we consume, the jobs we seek, and even the friendships and professional contacts that form our social circles. With such a wide usage and impact, recommender systems can exert strong, but often unforeseen, and sometimes even detrimental influence on the social processes connected to culture, lifestyles, politics, education, ethics, economic well-being, and even social justice. Hence, in this dissertation
research, we aim to identify, analyze, and alleviate potential risks and harms on users, item providers, the platforms, and ultimately the society, and to lay the foundation for new responsible recommender systems. In particular, we make three unique contributions toward responsible recommender systems:
• First, we study how to counteract the exposure bias in user-item interaction data. To overcome the challenge that the user-item exposure information is hard to be estimated when aiming to produce unbiased recommendations, we develop a novel combinational joint learning framework to learn unbiased user-item relevance and unbiased user-item exposure information simultaneously. Then, we push the problem to an extreme where we aim to predict relevance for items with zero exposure in the interaction data. For this, we propose a neural network utilizing a randomized training mechanism and a Mixture-of-Experts Transformation structure. Experiments validate the effective performance by the proposed methods.
• Second, we study what bias the machine learning based recommendation algorithms can bring and how to alleviate these bias. We uncover the popularity-opportunity bias on items and the mainstream bias on users. We conduct extensive data-driven study to show the existence of these bias in fundamental recommendation algorithms. Then, we explore and propose potential solutions to relieve these two types of bias, which empirically demonstrate outstanding performance for debiasing.
• At last, we move our attention to the problem of how to measure and enhance fairness in recommendation results. We study the recommendation fairness in three different recommendation scenarios – the multi-dimension recommendation scenario, the personalized ranking recommendation scenario, and the cold-start recommendation scenario. With respect to different recommendation scenarios, we develop different algorithms to enhance the recommendation fairness. We also conduct extensive experiments to empirically show the effectiveness of the proposed solutions.
Citation
Zhu, Ziwei (2022). Toward Responsible Recommender Systems. Doctoral dissertation, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /197810.