| dc.description.abstract | Recommender systems have been existing accompanying by web development, driving personalized experience for billions of users. They play a vital role in the information retrieval process, overcome the information overload by facilitating the communication between business people and the public, and boost the business world. Powered by the advances of machine learning techniques, modern recommender systems enable tremendous automation on the data preprocessing, information distillations, and contextual inferences. It allows us to mine patterns and relationships from massive datasets and various data resources to make inferences. Moreover, the fast evolvement of deep learning techniques brings vast vitality and improvements dived in both academic research and industry applications. Despite the prominence achieved in the recent recommender systems, the automation they have been achieved is still limited in a narrow scope. On the one hand, beyond the static setting, real-world recommendation tasks are often imbued with high-velocity streaming data. On the other hand, with the increasing complexity of model structure and system architecture, the handcrafted design and tuning process is becoming increasingly complicated and time-consuming. With these challenges in mind, this dissertation aims to enable advanced automation in recommender systems. In particular, we discuss how to update factorization-based recommendation models adaptively and how to automatically design and tune recommendation models with automated machine learning techniques. Four main contributions are made via tackling the challenges:
(1) The first contribution of this research dissertation is the development of a tensor-based algorithm for streaming recommendation tasks.
(2) As deep learning techniques have shown their superiority in recommendation tasks and become dominant in both academia and industry applications, the second contribution is exploring and developing advanced deep learning algorithms to tackle the recommendation problem with the streaming dataset.
(3) To alleviate the burden of human efforts, we explore adopting automated machine learning in designing and tuning recommender systems. The third contribution of this dissertation is the development of a novel neural architecture search approaches for discovering useful features interactions and designing better models for the click-through rate prediction problem.
(4) Considering a large number of recommendation tasks in industrial applications and their similarities, in the last piece of work work, we focus on the hyperparameter tuning problem in the transfer-learning setting and develop a transferable framework for meta-level tuning of machine learning models. | en |
