On surrogate supervision multiview learning

Abstract

In semi-supervised multi-view learning, the input vector is partitioned into two views and a classifier based on each view is sought after. In such settings, often examples which include the two views and a label are available [1]. In this paper, we are interested in the setting where a classifier for examples from one view is sought after although no labeled examples are provided for that view. Specifically, we consider the setting where labeled examples are provided only for the other view along with additional unlabeled examples of the two views jointly. To solve this problem, we present the Classification-Constrained Canonical Correlation Analysis (C4A) algorithm. We apply our algorithm to an audiovisual classification task. In comparison to two alternatives, the proposed method demonstrates superior performance.