Researchers have developed CUNO, a new framework for graph unlearning designed to mitigate catastrophic unlearning, a phenomenon where model utility sharply declines with large amounts of deleted data. CUNO addresses this by progressively removing data samples based on their estimated unlearning difficulty and employing a negative preference optimization objective to guide the model away from its original behavior without sacrificing retained performance. Experiments show CUNO significantly improves utility retention, maintaining over half the original performance even at 50% data deletion, outperforming existing methods. AI
IMPACT This research could improve the ability to selectively remove data from trained models, enhancing privacy and data management in graph-based machine learning applications.
RANK_REASON The cluster contains a research paper detailing a new method for graph unlearning. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- CUNO
- Graph Unlearning with Efficient Partial Retraining
- machine learning
- Negative Preference Optimization
- Preference Optimization
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
