PhD Dissertation Defense by DONG Viet Hoang | Hybrid Learning for Recommendation: From Contrastive Alignment to Language-Model Fusion

ABOUT THE TALK

Recommender systems have largely relied on collaborative filtering, which maps users and items to compact latent vectors and ranks by their dot product. This makes large-scale ranking fast, but interaction data records only that a preference was expressed, never why, and is often sparse. User reviews and pre-trained language models can supply the missing semantics, yet existing methods either treat reviews as input features, underserving entities with few reviews, or place a language model in the inference path at a per-query cost that does not scale.

This dissertation brings semantic understanding into collaborative recommendation while preserving efficient retrieval. First, a review-centric contrastive alignment framework moves reviews from the input pipeline into the training objective. Second, a large language model extends this framework by constructing contrastive views offline. Third, a fusion framework jointly trains user and item ID-token embeddings with masked identifier modeling through a frozen language model and Bayesian personalized ranking.

A key novelty lies in determining where text should act: it shapes the learned representations during training but is never required at serving time. This improves robustness to sparse and missing reviews and reduces inference latency by orders of magnitude compared with LLM-based recommenders. The findings, including the caution that fluent generated text is not automatically valid contrastive supervision, extend to other sparse and heterogeneous domains with auxiliary text.

ABOUT THE SPEAKER

I received my B.Sc. in Computer Science (Honors Program) from the University of Engineering and Technology, Vietnam National University, Hanoi. My research interests include recommender systems, contrastive learning, and large language models for recommendation; in particular, I integrate user reviews and language models into collaborative filtering to build recommender systems that are both semantically informed and efficient. This work has resulted in publications at leading conferences and journals in data mining and information systems, and was selected as a top-ranked paper at WSDM 2025, where I presented it with the support of a SIGIR Travel Award. I have served as a volunteer for TheWebConf 2024, as a program committee member for conferences such as SIGIR, KDD, WSDM, CIKM, and TheWebConf, and as a reviewer for the journals IEEE TKDE and Tsinghua Science and Technology. Beyond research, I apply AI and data technologies to real-world challenges in Vietnam, work that was recognized with the Vietnam Digital Award 2025. In my leisure time, I lead the Singapore chapter of the Dong Hanh Foundation, a non-profit supporting underprivileged Vietnamese youth, and enjoy exploring new and interesting areas of knowledge.