KDD 2027 · Jul 2026 · Under Review
Zhexun Chen, Haolin Song, Anwei Luan, Yuxun Zhang, Han Wang, Guangneng Hu
Proposes a semantic supervision framework that treats different permutations of the same semantic identifier set as equivalent views and aligns them via a contrastive objective. Improves NDCG@5 by up to 28% and representation uniformity across four benchmarks, with consistent gains on long-tail items and sparse users.
generative recommender · permutation-invariant · semantic equivalency · contrastive learning
AAAI 2027 · Jul 2026 · Under Review
Zhexun Chen, Guangneng Hu
Transfers semantic knowledge from LLM retrievers into generative recommenders through a lightweight adaptor and contrastive alignment, bypassing the expensive generative encoder. Achieves up to 29.2% improvement over the backbone with 6.6% additional training overhead.
generative recommendation · retrieval-augmented · LLM · contrastive learning
AAAI 2027 · Jul 2026 · Under Review
Yuxun Zhang, Zhexun Chen, Guangneng Hu
Models interest structure via time-aware hypergraphs, interest drift via transition flow fields, and multi-scale temporal dynamics through decomposed long/short-term fusion. Effective across five real-world benchmarks for continuous-time sequential recommendation.
sequential recommendation · hypergraph · interest dynamics · continuous-time