Work

Data-centric approaches for advanced and efficient recommendation modeling

Type: dissertation
Abstract: Recommender systems have become one of the most successful applications of artificial intelligence's in daily life, significantly relieving people from overwhelming information load and assisting users in decision-making by providing personalized recommendations.Advanced recommendation modeling techniques have been deployed across various platforms, i.e., Tiktok in video recommendations, Linkedin in friend recommendations, Goodreads for book recommendations, and Taobao for item recommendations.Despite their widespread success, existing recommendation approaches face several limitations, particularly in terms of multi-faceted user modeling and scalability.For instance, these models often experience performance degradation when handling multifaceted user data.Additionally, many state-of-the-art recommendation models are trained on large-scale datasets, making the training process both computationally expensive and inefficient, especially when fine-tuning hyperparameters or architectures to optimize recommendation performance.To address these challenges, this dissertation adopts a data-centric perspective, proposing curated data manipulation and optimization strategies to enhance the performance of recommendation models across different scenarios.Firstly, we demonstrate that disentangling various facets of data (i.e., social data and user-item data) for independent modeling can significantly improve recommendation performance.Furthermore, we explore the potential to substantially reduce the size of i user-item interaction and content data in recommendation datasets while preserving critical information, thus markedly decreasing training costs.Different from modelcentric approaches that are often tailored to specific recommendation models, datacentric strategies generate enhanced datasets that offer benefits across a range of existing models.By addressing key challenges related to personalized modeling and computational efficiency from a perspective of data-centric, this dissertation lays the foundation for next-generation recommendation techniques.First and foremost, I would like to express my profound gratitude to my supervisors, Prof. Qing Li and Prof. Ke Tang, for their invaluable guidance, inspiration, encouragement, and unwavering support throughout my Ph.D. journey.Their patience, meticulous attention to detail, professionalism, and humility have profoundly shaped both my personal growth and career aspirations.They have cultivated an open and intellectually stimulating environment, which has encouraged me to explore meaningful and challenging research topics.I will always cherish the thoughtful conversations with them, during which they provided not only expert advice but also heartfelt encouragement during the uncertain early stages of my Ph.D. Without their constant support and mentorship, the completion of this thesis would not have been possible.I would also like to extend my sincere thanks to my mentors, Dr. Wenqi Fan and Dr. Shengcai Liu, for their expert guidance in research.Their insightful suggestions on topic selection, the formulation of rigorous ideas, and the refinement of academic writing have been instrumental in shaping my research skills.Dr. Wenqi Fan, in particular, introduced me to the field of data mining, provided me with the essential skills to conduct high-quality research, and consistently encouraged me to strive for excellence and aim high in my endeavors.I am equally grateful to Dr. Shengcai
Author: Wu, Jiahao
Language: English
Open Access status: green
OpenAlex ID: W7217997104
Cites: 95
Cited by:
Related to: 9