Reproduce “Preserving Privacy and Utility in LLM-Based Product Recommendations” 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

Tina Khezresmaeilzadeh, Jiang Zhang, Dimitrios Andreadis, Konstantinos Psounis

Researchers at USC developed a hybrid framework for LLM-based product recommendations that processes sensitive user data locally while leveraging cloud LLMs for non-sensitive information. This framework achieved utility scores (Categorical HR@10 of 0.6061 and Semantic HR@10 of 0.3117) that closely approach a non-private baseline, effectively minimizing privacy leakage during recommendation inference.

Key insights

  • Context-aware, BERT-based obfuscation is crucial for effectively filtering sensitive information while preserving more non-sensitive context than simple categorical filtering.
  • Recovering recommendations for sensitive data locally using a lightweight LLM (Llama 3.2 1B) effectively restores overall recommendation utility, closely approaching non-private baselines.
  • The hybrid architecture provides a practical balance, leveraging powerful cloud LLMs for general recommendations while confining sensitive user data processing to the client device for privacy.

Summary by alphaXiv.

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