Yujie Jia, Guohao Sun, Xiu Susie Fang, Ge Zhang, Jinhu Lu · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3841015
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Sequential Recommendation Systems (SRS) aim to predict the next item based on their historical interaction sequence. To address the data sparsity issue that commonly existed in SRS, Cross-Domain Sequential Recommendation (CDSR) has emerged, which aims to improve recommendation performance in a sparse domain (usually called target domain) by transferring knowledge from an information-rich domain (usually called source domain). However, existing CDSR methods mainly rely on ID-based representations and provide limited semantics across domains. With the rapid development of large language models (LLMs), recent studies have begun to incorporate LLM-generated item or user embeddings with rich semantic information into CDSR models, substantially enriching the representation space and improving modeling expressiveness. However, these LLM-enhanced methods typically treat all source domains equally and overlook the varying semantic relevance of different domains, which will lead to low-relevance domains introducing noise and consequently degrade recommendation performance. To tackle these issues, we propose LLMAW4CDSR, a Large Language Model-guided Adaptive source-domain Weighting framework. The core idea is to first leverage the semantic reasoning abilities of LLMs to produce semantically enriched user and item representations, and then use them as semantic guidance for more effective cross-domain knowledge transfer. Building on this guidance, we design an source-domain weighting mechanism that estimates the contribution of each source domain to the target domain. By integrating LLM-driven semantic alignment with adaptive domain weighting, LLMAW4CDSR highlights relevant domains while suppressing noisy ones, enabling fine-grained cross-domain knowledge transfer. Comprehensive experiments on four cross-domain datasets demonstrate that LLMAW4CDSR consistently outperforms both state-of-the-art CDSR methods and LLM-based baselines.
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
No comments yet — start the discussion below.