Reproduce “Fusion and Alignment Enhancement with Large Language Models for Tail-item Sequential Recommendation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhifu Wei, Yizhou Dang, Guibing Guo, Chuang Zhao, Zhu Sun
The FAERec framework enhances sequential recommendation performance, particularly for tail items, by fusing collaborative and semantic signals from Large Language Models and explicitly aligning their embedding spaces. It recorded relative improvements up to 29.62% in tail H@10 on Amazon Beauty and average H@10 gains up to 1500.2% for long-tail items across diverse backbones, while also increasing recommendation diversity.
Key insights
- Explicitly fusing collaborative and semantic information, rather than merely using LLM embeddings as auxiliary signals, is crucial for improving item representation quality, especially for tail items.
- Active alignment of heterogeneous embedding spaces (ID and LLM) at both item and feature levels is essential to resolve structural inconsistencies and prevent conflicting signals that degrade recommendation accuracy.
- A progressive alignment strategy combined with popularity-based grouping effectively manages optimization complexity and mitigates popularity bias, leading to more robust and accurate recommendations for both popular and long-tail items.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.