Data-efficient Fine-tuning for LLM-based Recommendation (SIGIR 2024)

Overview

This entry covers the SIGIR 2024 paper "Data-efficient Fine-tuning for LLM-based Recommendation".

| Field | Content | |-------|---------| | Title | Data-efficient Fine-tuning for LLM-based Recommendation, SIGIR 2024 | | Venue | SIGIR 2024 (ACM SIGIR Conference on Research and Development in Information Retrieval) | | Link | https://dl.acm.org/doi/abs/10.1145/3626772.3657807 | | Type | Academic paper | | Section | Recommender Engines |

Background and Motivation

LLM-based recommendation has emerged as a promising direction: large language models bring semantic priors, world knowledge, and natural-language interaction capabilities to recommender systems, helping with user intent understanding and cold-start situations. However, fine-tuning LLMs for recommendation typically demands large amounts of user-interaction data, which is expensive to collect, clean, and annotate — and often unavailable for new platforms or long-tail domains.

This paper targets precisely that bottleneck: how to fine-tune LLMs for recommendation in a data-efficient manner, adapting the model to user preferences and item semantics without requiring massive supervised interaction datasets.

Why It Matters

• Cost: Reducing fine-tuning data requirements directly lowers GPU and annotation costs for industrial deployment.
• Cold start: Data-efficient adaptation is critical for new users, new items, and new domains where interaction logs are sparse.
• Knowledge preservation: Naive fine-tuning can erode the general language abilities that make LLMs valuable; efficient adaptation strategies aim to align recommendation behavior while retaining this knowledge.

Context in the Field

Recommendation has evolved from matrix factorization and deep CTR models to sequential Transformers and now instruction-following / generative recommendation (Gen-Rec) with LLMs. The core tension remains: user behavior is sparse, item catalogs are huge, and online inference costs and hallucination risks demand careful system design. Work on data-efficient fine-tuning sits at the intersection of LLM adaptation (e.g., parameter-efficient fine-tuning), synthetic/instruction data generation, and recommendation-specific objectives.

Notes for Readers

• Consult the original PDF via the ACM DOI link for the exact method, datasets, baselines, and quantitative results.
• Recommended metrics when evaluating this line of work: Recall@k, NDCG@k, Hit@k, plus training-data scale ablations and token/GPU cost.
• Readers may also find related entries useful, e.g., 360Brew (decoder-only foundation model for personalized ranking), "Actions Speak Louder than Words" (trillion-parameter sequential transducer), and "Bridging Language and Items for Retrieval and Recommendation."

Reference

• Original paper: *Data-efficient Fine-tuning for LLM-based Recommendation*, SIGIR 2024. https://dl.acm.org/doi/abs/10.1145/3626772.3657807