Abstract

Recommender systems serve as an essential bridge for connecting users and items on a digital platform, which helps users find relevant items and helps items reach their potential users.
People exhibit both regular habits and curiosity, demonstrating repetition behaviors as well as exploration behaviors when engaging with platforms. The coexistence of repetition and exploration imposes challenges for performance evaluation and designing recommendation models. Thus, this thesis aims to provide insights and understand the repetition and exploration in various recommendation tasks. Considering the repetition and exploration in recommendation, this thesis covers two research aspects: (i) evaluate recommendation performance, and (ii) optimize and design recommendation methods.
In terms of recommendation evaluation, this thesis introduces a comprehensive set of evaluation metrics that take into account both user-side and item-side aspects, which allow for a more detailed understanding of recommendation performance. Additionally, the thesis presents guidelines for evaluating recommendation models in a scenario with both repetition and exploration behaviors. In the context of recommendation optimization and design, the thesis puts forward a range of training strategies and recommendation model designs based on the insights derived from the analysis of repetition and exploration.
To sum up, this thesis uncovers the key differences between repetition and exploration in recommendation, and highlights the importance of evaluating and optimizing recommendation models from the perspective of repetition and exploration.