Performance analysis of surrogate supervision multi-view learning linear classifiers in Gaussian data

Multi-view learning is a classification setting in which feature vectors consist of multiple views. The goal in this setting is to find a classifier for some or all of the views. We consider a limiting case of multi-view learning termed surrogate supervision multi-view learning (SSML). In the SSML setting, training data consists of two types: unlabeled two-view data examples and labeled single view examples. The goal in this setting is to find a classifier for the view for which no labels are available. In this paper, we analyze the case in which the data is Gaussian distributed and the classifiers on each view are linear. For this setting, we provide a theoretical analysis for the performance mismatch between the error associated with a classifier trained in the SSML setting and a classifier trained in the direct supervision setting.