Reproduce “AgenticTagger: Structured Item Representation for Recommendation with LLM Agents” 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

Zhouhang Xie, Bo Peng, Zhankui He, Ziqi Chen, Alice Han, Isabella Ye, Benjamin Coleman, Noveen Sachdeva and 6 more

AgenticTagger, a multi-LLM framework, generates interpretable, low-cardinality, and hierarchically structured natural language descriptors for items in recommender systems. This approach enhanced generative recommendation performance by up to 11.6% NDCG@10 on public benchmarks and improved ranking accuracy on a private news-feed dataset, also enabling more controllable and interpretable recommendations.

Key insights

  • Leveraging LLMs to generate structured, low-cardinality natural language features effectively overcomes the limitations of free-form text generation, making them compatible with existing recommender system architectures.
  • A multi-agent, iterative self-refinement approach is critical for scaling LLM-based vocabulary construction to large item corpora, enabling parallel processing and the learning of coherent, hierarchical descriptor sets.
  • Constraining LLM output to a pre-defined, structured vocabulary prevents vocabulary explosion and ensures the creation of usable, interpretable, and generalizable features for downstream recommendation tasks.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.