You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Owing to their powerful semantic reasoning capabilities, Large LanguageModels (LLMs) have been effectively utilized as recommenders, achievingimpressive performance. However, the high inference latency of LLMssignificantly restricts their practical deployment. To address this issue, thiswork investigates knowledge distillation from cumbersome LLM-basedrecommendation models to lightweight conventional sequential models. Itencounters three challenges: 1) the teacher's knowledge may not always bereliable; 2) the capacity gap between the teacher and student makes itdifficult for the student to assimilate the teacher's knowledge; 3) divergencein semantic space poses a challenge to distill the knowledge from embeddings.To tackle these challenges, this work proposes a novel distillation strategy,DLLM2Rec, specifically tailored for knowledge distillation from LLM-basedrecommendation models to conventional sequential models. DLLM2Rec comprises: 1)Importance-aware ranking distillation, which filters reliable andstudent-friendly knowledge by weighting instances according to teacherconfidence and student-teacher consistency; 2) Collaborative embeddingdistillation integrates knowledge from teacher embeddings with collaborativesignals mined from the data. Extensive experiments demonstrate theeffectiveness of the proposed DLLM2Rec, boosting three typical sequentialmodels with an average improvement of 47.97%, even enabling them to surpassLLM-based recommenders in some cases.
AkihikoWatanabe
changed the title
あ
Distillation Matters: Empowering Sequential Recommenders to Match the
Performance of Large Language Model, Yu Cui+, N/A, arXiv'24
May 3, 2024
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)
The text was updated successfully, but these errors were encountered: