Abstract
Personalization systems (PS) are applied in nearly every corner of the internet through recommendation, content supply, messaging, and so on. Not only do business owners depend on personalization for recommending the relevant items to the right users, but also consumers need personalization to find useful information without being overwhelmed in the info-times. Due to these reasons, enormous efforts have long been devoted in developing more powerful AI and machine learning techniques to improve the performance of PS.
In recent years, however, people start to realize that PS empowered by these techniques may lead to undesired effects and undermine the original good purposes of personalization. One of the main issues with PS is their lack of transparency, which means that users may not fully understand why they are being recommended certain items or content, which can lead to confusion and mistrust. Another issue is fairness. Due to the vast amounts of data used to train these systems, PS can inadvertently learn and perpetuate biases that exist in society. This can have serious consequences, such as perpetuating systemic inequality and discrimination. Addressing the issues of transparency and fairness in PS is essential for ensuring that they are trusted and effective tools to continue to benefit users, businesses, and society as a whole.
This dissertation focuses on improving the transparency and fairness of PS, which contributes to the development of more responsible PS that benefit all stakeholders. To enhance transparency, I propose to generate intuitive, textual explanations for personalized results. The explanations are expected to help users make more informed decisions and build trust in the system. For fairness, I investigate and propose algorithms to address the issues from different perspectives in PS. In general, the algorithms aim to generate personalized results without discrimination in serving users and business owners with different social constructs. Comprehensive and rigorous analysis and experiments demonstrates the approaches' effectiveness in various contexts and applications.
