New research explores uncertainty-dependent missing labels for classification

A new research paper explores a semi-supervised classification method where the probability of a label being missing is dependent on the observed features and the classifier's uncertainty. This approach treats missingness not just as lost information but as a signal generated by a mechanism linked to the classifier. The paper develops a likelihood-based information theory for these uncertainty-dependent missing labels, deriving a Fisher information decomposition and covariance partitions under various model specifications. Calculations involving Gaussian mixture models and a medical diagnosis example illustrate how this method can enhance estimation and classification efficiency within a fixed labeling budget. AI

IMPACT This research could lead to more efficient semi-supervised learning methods by leveraging label missingness as an informative signal.

RANK_REASON The cluster contains a single academic paper submission to arXiv. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• arXiv
• CORE Recommender
• Fisher information
• Gaussian mixture model
• Geoffrey McLachlan
• Godambe--Eicker--Huber--White
• Hugging Face
• Learning from Uncertainty-dependent Missing Labels for Semi-supervised Classification

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →