Adapting Supervised Classification Algorithms to Arbitrary Weak Label Scenarios

Abstract

In many real-world problems, labels are often weak, meaning that each instance is labelled as belonging to one of several candidate categories, at most one of them being true. Recent theoretical contributions have shown that it is possible to construct proper losses or classification calibrated losses for weakly labelled classification scenarios by means of a linear transformation of conventional proper or classification calibrated losses, respectively. However, how to translate these theoretical results into practice has not been explored yet. This paper discusses both the algorithmic design and the potential advantages of this approach, analyzing consistency and convexity issues arising in practical settings, and evaluating the behavior of such transformations under different types of weak labels.