JevRec uses variable-length item codes for generative recommendation
Original title: JevRec: Linguistic Encoding with Jev-like Decision Modeling for LLM-based Generative Recommendation
On three public datasets it beat compared baselines and used less train and inference time than TIGER.
JevRec encodes recommendation items as variable-length semantic identifiers, giving popular items shorter codes and long-tail items richer descriptions from shared codebooks. Separate heads predict sequence length and generate content while sharing the interaction-history representation, avoiding a conflict under one output head. Experiments on three public datasets show it outperforms all compared baselines on the evaluated metrics, with lower training and inference time than TIGER and less gradient conflict between length and content prediction.
- Popular items get shorter SIDs; long-tail items keep richer descriptions.
- A dedicated head predicts length; a second head generates content.
- Further analysis reports lower training and inference time than TIGER.
- The paper also reports reduced gradient conflict between the two heads.
From the original article
1 min read at arxiv.orgGenerative recommendation (GR) represents items as semantic identifiers (SIDs) and autoregressively generates the next item's SID from interaction history. However, most methods assign the same SID length to all items despite highly unequal interaction frequencies.
By Qiyong Zhong · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.