A theoretical framework for surrogate supervision multiview learning

We consider the problem of classification under the multi-view learning setting referred to as surrogate supervision multi-view learning (SSML). In this setting, training data is provided for two parts of the feature vector (views) in the following format (i) labeled first view examples and (ii) unlabeled first and second view examples. The goal in this setting is to obtain a classifier for the unlabeled view. For example, consider the task of training a classifier for a new sensor using a set of labeled measurements taken with a legacy sensor and a set of unlabeled measurements taken with both sensors. This paper provides theoretical performance bounds on the classification accuracy for classifiers trained in the SSML setting. The bounds rely on a characteristic we introduce termed classifier correlation distance, which characterizes the ability to approximate a classifier on one view by a classifier on the other. We show that when a classifier on one view can be well approximated with a classifier on the other view and vice versa, the classification accuracy of a classifier in the SSML setting approaches that of a classifier trained with labeled examples on the desired view.