Two-View Online Learning

Abstract

We propose a two-view online learning algorithm that utilizes two different views of the same data to achieve something that is greater than the sum of its parts. Our algorithm is an extension of the single-view Passive Aggressive (PA) algorithm, where we minimize the changes in the two view weights and disagreements between the two classifiers. The final classifier is an equally weighted sum of the individual classifiers. As a result, disagreements between the two views are tolerated as long as the final combined classifier output is not compromised. Our approach thus allows the stronger voice (view) to dominate whenever the two views disagree. This additional allowance of diversity between the two views is what gives our approach the edge, as espoused by classical ensemble learning theory. Our algorithm is evaluated and compared to the original PA algorithm on three datasets. The experimental results show that it consistently outperforms the PA algorithm on individual views and concatenated view by up to 3%.