AI Just Got Smarter at Spotting Hard Drive Failures

A new study introduces a semi-supervised classification method for two-component Weibull mixture models, where features are observed for all data but some class labels are missing. Missingness is modeled as a function of classification uncertainty, so the missing-label indicators themselves can carry information about the classifier. The authors characterize the possible decision boundaries, derive Fisher information after accounting for nuisance parameters, and present asymptotic relative efficiency results for one- and two-boundary cases. Numerical studies and a semi-synthetic analysis based on hard-drive failure data show that modeling feature-dependent label missingness may reduce expected error and improve decision-boundary estimation.