Publications

Studies how LLM-derived semantic representations and collaborative ID representations can be jointly modeled for sequential recommendation.

Introduces geographic item tokens for local-life recommendation, connecting semantic embeddings with spatial and collaborative signals.

Builds a hierarchical real-world benchmark and studies human-aligned, LLM-as-a-judge evaluation.

Transfers text-enhanced preferences across domains by combining textual semantics, collaborative signals and contrastive learning.

Aligns ID, text and image signals through self-supervised learning for multimodal sequential recommendation.

Models time-attenuated sentiment signals extracted from reviews for sequential recommendation.