FLOWGEM method tackles non-monotone missing data with Wasserstein gradient flows

Researchers have introduced FLOWGEM, a novel iterative method designed to generate complete datasets from data containing Missing at Random (MAR) values. This approach aims to recover the correct data distribution by minimizing the Kullback-Leibler divergence between observed and generated data distributions across various missingness patterns. FLOWGEM utilizes a particle evolution of Wasserstein Gradient Flow, with velocity fields approximated by local linear estimators, demonstrating state-of-the-art performance in simulations and real-world benchmarks, particularly for non-monotone MAR mechanisms. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Presents a principled and practical method for improving data imputation, potentially enhancing the reliability of downstream machine learning analyses.

RANK_REASON This is a research paper detailing a new method for handling missing data in machine learning.

Read on arXiv cs.LG →

• FLOWGEM
• arXiv
• Kullback-Leibler divergence
• Wasserstein Gradient Flow

• paper
• other