Research map / Large language models / Retrieval and knowledge

Towards S²-Challenges Underlying LLM-Based Augmentation for Personalized News Recommendation

AAAI 2025 · Large language models › Retrieval and knowledge · approach: Knowledge-augmented question answering

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 S2LENR, which augments MIND news data with GPT-4-generated news and adds a structure-aware prototype refinement module and a contrastive semantic-aware denoising module to improve news recommendation.

Problem

LLM-based data augmentation for personalized news recommendation produces user-news bipartite graphs that are isolated cliques (lacking collaborative signals) and injects noisy open-world knowledge into user behavior.

Approach

Builds a parametric user-prototype bipartite by clustering generated news into learnable prototypes to inject pseudo collaborative signals, and uses a gated contrastive objective to separate useful signal from noise in user representations.

Key results

• On MIND-Large, S2LENR reaches 71.13 AUC vs 69.98 for the best baseline HDNR and 69.47 for LENR.
• On MIND-Small, S2LENR reaches 69.17 AUC, about 1.4 points above HDNR (68.23) and 2.1 above LENR (67.04).
• Ablation on MIND-Small: removing the structure-aware refinement module drops AUC from 69.17 to 68.20, more than removing the denoising module (68.49).

Measured effects reported in the paper

Surprising findings

• Replacing LLM augmentation with traditional Mixup or Concat augmentation yields lower AUC than LLM-based augmentation, but LLM augmentation alone (LENR) gives only modest gains over NRMS.

Limitations

• Prototype count K is a fixed hyperparameter and performance drops when K becomes large, so the collaborative signal fades.
• LLM-generated news may introduce irrelevant open-world noise, which the denoising module only partially addresses.
• Temporal information is ignored in generation because LLMs handle it poorly.

Experimental setup

Datasets: MIND-Large MIND-Small

Benchmarks: MIND-Large MIND-Small

Metrics: AUC MRR nDCG@5 nDCG@10

Baselines: EBNR DKN NPA NAML LSTUR NRMS GNewsRec GERL User-as-Graph HDNR LENR

Builds on: NRMS GERNE GERL contrastive learning

Closest papers by meaning

• Bridging the User-side Knowledge Gap in Knowledge-aware Recommendations with Large Language Models (AAAI 2025 · similarity 0.86)
• LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.86)
• Can LLMs Outshine Conventional Recommenders? A Comparative Evaluation (NeurIPS 2025 · similarity 0.86)
• Improving LLMs for Recommendation with Out-Of-Vocabulary Tokens (ICML 2025 · similarity 0.86)
• CoRA: Collaborative Information Perception by Large Language Model’s Weights for Recommendation (AAAI 2025 · similarity 0.86)
• LLMEmb: Large Language Model Can Be a Good Embedding Generator for Sequential Recommendation (AAAI 2025 · similarity 0.85)