Researchers have developed a new method for robust covariance estimation, particularly useful for data that is partially observed and contains entry-wise contamination. The proposed approach extends inference for the Student-t* model to incomplete observations using a stochastic approximation EM (SAEM) algorithm. This method combines stochastic updates of coordinate-wise scales with Rao-Blackwellized Gaussian conditional moments, avoiding the need to explicitly sample latent Gaussian variables. Experiments indicate that this SAEM algorithm outperforms existing estimators, including Gaussian, Student-t, and imputation-based methods, in terms of Fisher-Rao error under moderate to high outlier rates with diffuse contamination. AI
IMPACT Introduces a novel statistical method that could improve the reliability of machine learning models dealing with noisy and incomplete datasets.
RANK_REASON Academic paper detailing a new statistical algorithm and its experimental validation. [lever_c_demoted from research: ic=1 ai=0.7]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
