김혜영 박사과정, 박성민 박사 졸업생 WSDM 2027 Findings 국제 학술대회 논문 채택

04 Nov 2026

DIAL 연구실 소속 인공지능학과 김혜영 박사과정, 박성민 박사 졸업생, 이종욱(교신저자) 교수가 참여한 논문 “Intent-guided Semantic Regularization for LLM-enhanced Sequential Recommendation” 이 데이터마이닝 분야 최우수 국제 학술대회인 The 20th ACM International Conference on Web Search and Data Mining (WSDM 2027)에 최종 게재가 승인되었으며 2027년 2월에 발표될 예정입니다.

Abstract

Large Language Model (LLM)-enhanced Sequential Recommendation (SR) leverages rich textual semantics of items to alleviate the long-tail problem. A prominent paradigm, LLM-as-Extractor, initializes item embeddings in SR models with LLM-generated representations, offering both efficiency and performance gains. However, existing models exploit LLM semantics in a limited way. (i) They use LLM representations only as a starting point; as training proceeds on user behaviors, the original semantic structure gradually deteriorates. (ii) They capture semantics only at the level of individual items and fail to effectively account for user intents, i.e., the higher-level semantic patterns that underlie user interactions. Since such intents are shared across items and users, they can provide stable guidance under sparse interactions. To address these limitations, we propose Intent-guided semantic Regularization (IntReg), which leverages natural-language user intents as semantic guidance in training SR models. Specifically, IntReg first constructs a domain-specific intent pool. Each user sequence is then represented as a context-dependent mixture of intents, yielding a semantic distribution over candidate items. This distribution is kept frozen and used to regularize the collaborative prediction with sequence-adaptive strength, thereby preserving LLM semantics throughout behavior-driven SR training. Extensive experiments on four real-world datasets show that IntReg consistently outperforms existing LLM-as-Extractor models, achieving gains of up to 14.6% in N@10 overall and up to 22.4% in H@10 for tail items.