Online learning with binary feedback for multi-class problems

Evan Lucas, Steven Whitaker, Timothy Havens

Abstract: Online learning methods often focus on data selection, as human labeling is a noted bottleneck in resources needed to train a model. This work chooses to focus on reducing the human effort necessary for providing labels by attempting to usefully utilize a binary feedback method where the human indicates whether the prediction is correct or incorrect. More specifically, this work investigates methods for using and labeling training data in the absence of complete information. Various methods to generate labels in response to human feedback are proposed and then compared. These methods are tested on a variety of common classification tasks and results showing their usefulness are presented. Although the maximum accuracy achieved is not as high, the methods presented allow a model to learn faster, in terms of number of human interactions required.