Sijin Lu, Jun Wu · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3839913
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).
Large Language Models (LLMs) have shown promise as embedding generators for sequential recommendation. Effective adaptation typically involves two stages: collaborative signal injection and discriminative representation learning. However, optimization objective and item semantic misalignments arise between these two stages, resulting in catastrophic forgetting of collaborative relationships learned in the first stage. To address these issues, we propose a novel LLM-based embedding method called GCTP, which obtains both collaborative and discriminative embeddings. Specifically, we maintain a generative training objective throughout the training process, thereby alleviating potential optimization conflicts. Moreover, we preserve local semantics to ensure semantic consistency during training, and capture global semantics with minimal modifications when extracting item embeddings. Experiments on real-world datasets show that GCTP significantly outperforms state-of-the-art LLM-based embedding methods.
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.