A new paper proposes that per-point data valuation scores, such as Data Shapley, are not ideal for selecting training data when the goal is to identify the most valuable points. The authors argue that a point's value is context-dependent, diminishing if it's a near-duplicate of already selected data. They suggest that adaptive coverage, which considers what data has already been selected, is a more effective strategy than fixed top-k scoring, especially when data selection is critical and budget is limited. AI
IMPACT Suggests alternative methods for selecting training data that may improve model performance by avoiding redundant data points.
RANK_REASON Academic paper on data valuation methods for machine learning. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
