Multiple view semi-supervised discriminant analysis

Abstract

Beyond conventional semi-supervised dimensionality reduction methods which data are represented in a single vector or graph space, multiple view semi-supervised ones are to learn a hidden consensus pattern from multiple representations of multiple view data together with some domain knowledge. Under multiple view settings, we propose a new Multiple view Semi-supervised Discriminant Analysis (MSDA). Specifically, the labeled data are used to infer the discriminant structure in each view. Simultaneously, all the data, including the labeled and the unlabeled instances, are used to discover the intrinsic geometrical structure in each view. Thus, we can learn an optimal pattern from the multiple patterns of multiple representations with serial combination after getting the projection of each view. Experiments carried out on real-world data sets by MSDA show a clear improvement over the results of representative dimensionality reduction algorithms.