Favourable Missingness in Semi-Supervised Classification for Exponential Mixture Models

What the math gives to ML

The paper isolates a useful missing-not-at-random phenomenon: the indicator that a label is absent can itself constrain the class posterior when missingness depends on classification uncertainty. This suggests treating label availability as an auxiliary probabilistic observation rather than discarding unlabeled examples or assuming missingness is ignorable. In a neural classifier, a joint likelihood can train both from observed labels and from the pattern of missingness, provided the missingness mechanism is explicitly parameterized and regularized to avoid degenerate solutions. The most direct experiment is a semi-supervised classifier whose unlabeled loss is the likelihood of its own uncertainty-dependent missingness model.