General Algorithm for Learning from Grouped Uncoupled Data and Pairwise Comparison Data

Abstract

Uncoupled regression is the problem of learning a regression model from uncoupled data that consists of a set of input values (unlabeled data) and a set of output values where the correspondence between the input and output is unknown. A recent study showed that a method using both uncoupled data and pairwise comparison data can learn the optimal model under some assumptions. However, this method cannot use grouped information and may require the implementation (almost) from scratch for some models. In this study, we extend the above existing method for handling group information and derive a general algorithm that can learn a model using the standard regression method by approximating the loss function. The effectiveness of the proposed method is confirmed by synthetic and benchmark data experiments.