Probabilistic Decoupling of Labels in Classification
This addresses the challenge of handling ambiguous or incomplete labels in classification for machine learning practitioners, though it appears incremental as it builds on existing probabilistic frameworks.
The paper tackles the problem of non-standard classification tasks like semi-supervised and noisy-label learning by developing a probabilistic approach to decouple labels from underlying classes, resulting in a unified method for inferring class distributions.
In this paper we develop a principled, probabilistic, unified approach to non-standard classification tasks, such as semi-supervised, positive-unlabelled, multi-positive-unlabelled and noisy-label learning. We train a classifier on the given labels to predict the label-distribution. We then infer the underlying class-distributions by variationally optimizing a model of label-class transitions.