Research map / Large language models / Language, safety and interpretability

LLMEmb: Large Language Model Can Be a Good Embedding Generator for Sequential Recommendation

AAAI 2025 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks · 31 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LLMEmb, which fine-tunes LLaMA-7B with attribute-dropout contrastive learning and then trains a frozen-embedding adapter with collaborative alignment, producing item embeddings that boost GRU4Rec, Bert4Rec, and SASRec on long-tail sequential recommendation.

Problem

Sequential recommenders perform poorly on long-tail (low-popularity) items because item embeddings are skewed; existing LLM-enhanced SRS methods suffer a semantic gap (general LLM not tailored to recommendation) and semantic loss (aggressive PCA and continued training of LLM embeddings).

Approach

Two-stage LLM embedding generator: supervised contrastive fine-tuning (LoRA on LLaMA-7B) with attribute-level random dropping as positive augmentation, followed by recommendation adaptation training that keeps LLM embeddings frozen, uses a two-layer adapter after PCA to 1536 dims, and aligns adapter outputs with a collaborative SRS embedding layer via in-batch contrastive loss.

Key results

• LLMEmb with SASRec on Yelp reaches overall H@10 0.6647 and N@10 0.4113, versus 0.6415 for the best baseline (LLM2X).
• On the Yelp tail group with SASRec, LLMEmb reaches H@10 0.2951 vs 0.1760 for LLM2X, a large gain on long-tail items.
• Ablation on Yelp (SASRec) shows removing SCFT drops tail H@10 from 0.2951 to 0.2474 and removing the freeze drops it to 0.1710.

Measured effects reported in the paper

Surprising findings

• Fine-tuning the LLM embeddings during RAT (w/o freeze) severely hurts performance, with overall H@10 falling to 0.6257.
• LLMEmb slightly underperforms LLM2X on the most popular items (40+ group), indicating a seesaw between popular and long-tail items.

Limitations

• Tail-group gains come with a slight drop on the most popular items, a seesaw effect.
• Relies on a 7B LLM for embedding generation, which is costly to run over item sets.
• Evaluation uses 100 sampled negatives and only three datasets.

Experimental setup

Datasets: Yelp Amazon Beauty Amazon Fashion

Benchmarks: Yelp Amazon Beauty Amazon Fashion

Metrics: H@10 N@10

Baselines: MELT LLM2X SAID TSLRec

Builds on: SASRec LLaMA LoRA E4SRec LLM2X SAID

Closest papers by meaning

• LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.93)
• Can LLMs Outshine Conventional Recommenders? A Comparative Evaluation (NeurIPS 2025 · similarity 0.91)
• SLMRec: Distilling Large Language Models into Small for Sequential Recommendation (ICLR 2025 · similarity 0.90)
• Improving LLMs for Recommendation with Out-Of-Vocabulary Tokens (ICML 2025 · similarity 0.89)
• LLM4RSR: Large Language Models as Data Correctors for Robust Sequential Recommendation (AAAI 2025 · similarity 0.89)
• SRA-CL: Semantic Retrieval Augmented Contrastive Learning for Sequential Recommendation (NeurIPS 2025 · similarity 0.89)