Logits-Level Balanced Machine Unlearning for LLM-Based Recommendation System.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Recommendation systems are crucial for digital platforms, with Large Language Models (LLMs) enhancing personalization and accuracy.
• LLM-based recommendation (LLMRec) systems face data governance challenges, including privacy, outdated, poisoned, and copyrighted data, necessitating effective data removal.
• Existing methods struggle to accurately delete data and its impact from complex LLMRec systems.

Purpose of the Study:

• To propose and evaluate an LLMRec unlearning system capable of precise data removal while maintaining system performance.
• To address the challenges of data governance in LLMRec systems through a novel unlearning mechanism.

Main Methods:

• Developed an adapter-driven logits modification system for LLMRec unlearning.
• Utilized adapters to reduce training costs during the unlearning process.
• Employed knowledge distillation (KD)-powered logits modification to ensure continued effective reasoning and recommendation quality post-unlearning.
• Incorporated an adapter-based general knowledge retention module to preserve core LLM capabilities.

Main Results:

• The proposed system demonstrates effective and efficient unlearning in LLMRec systems.
• The adapter-driven approach successfully removes target data while preserving recommendation accuracy.
• Knowledge distillation and general knowledge retention modules maintain model reasoning and foundational language abilities.

Conclusions:

• The adapter-driven logits modification system offers a precise and effective solution for unlearning in LLMRec systems.
• The method balances data removal requirements with the need to maintain recommendation performance and general language understanding.
• This work provides a significant advancement in addressing data governance challenges within LLM-based recommendation frameworks.