Data-efficient Fine-tuning for LLM-based Recommendation (SIGIR 2024)
This is a catalog entry from a curated reading list (Recommender Engines section) covering the SIGIR 2024 paper "Data-efficient Fine-tuning for LLM-based Recommendation."
- Source: ACM Digital Library
- Venue: SIGIR 2024
- Resource type: Academic paper
- The paper targets data-efficient fine-tuning of LLMs for recommendation, addressing the cost of training on large-scale user interaction data.
- It sits at the intersection of recommendation and large-scale search/personalization systems, where efficiency, scalability, and user intent understanding remain core challenges.
- The post frames the work against the evolution of recommenders: from matrix factorization and deep CTR models, to sequential Transformers, to LLM instruction-following and generative recommendation (Gen-Rec).
- LLMs bring semantic priors and cold-start capability, but online inference cost and hallucination risk demand careful system-level design.
- Datasets: MS MARCO, BEIR, Natural Questions, domain corpora, public recommendation benchmarks
- Metrics: nDCG@10, MRR, Recall@k, Hit@k, human preference, task success rate, latency and token cost
- Baselines: BM25, dense retrieval, cross-encoder reranking, retrieval-free LLMs, commercial search APIs
- 360Brew: A Decoder-only Foundation Model for Personalized Ranking
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers
- Augmenting Netflix Search with In-Session Adapted Recommendations
- Bridging Language and Items for Retrieval and Recommendation
- DiffKG: Knowledge Graph Diffusion Model for Recommendation (WSDM 2024)
- EAGER-LLM: Enhancing LLMs as Recommenders through Exogenous...
- RAG: Retrieval-Augmented Generation
- LTR: Learning to Rank
- nDCG: Normalized Discounted Cumulative Gain, a ranking-quality metric
- Agentic Search: Modeling search as sequential decision-making with tool calls
- Gen-IR: Generative Information Retrieval
Key points
Context from the catalog post
The entry decomposes typical method components in this space — representation learning, retrievers, rerankers, planners, generators, and feedback mechanisms — and describes a common pipeline: problem formulation → model/system design → training/construction → inference pipeline. Learning strategies mentioned include supervised fine-tuning, contrastive learning, distillation, reinforcement learning (including process rewards), and bootstrapped data synthesis.
Evaluation practices referenced
Engineering checklist highlighted by the post
| Area | Concern | Suggestion | |------|---------|------------| | Data | PII, versioning | Partitioned indexes, anonymization, rollback-able embeddings | | Latency | p99 budget, retrieval depth | Cascades + early stopping, query caching, async reranking | | Quality | Offline vs. online gains | Interleaving experiments, human audits, citation checks | | Safety | Open-retrieval poisoning/bias | Source whitelisting, adversarial detection, output filtering | | Cost | Token/GPU spend per query | Small-model routing, distillation, hybrid sparse+dense |
Related entries in the same collection
Caveat
The post is largely a structured overview based on the abstract and public metadata. Specific numerical results are not reproduced; consult the original PDF via the DOI link before citing quantitative conclusions.