Direct Multi-label Linear Discriminant Analysis

Abstract

Multi-label problems arise in different domains such as digital media analysis and description, text categorization, multi-topic web page categorization, image and video annotation etc. Such a situation arises when the data are associated with multiple labels simultaneously. Similar to single label problems, multi label problems also suffer from high dimensionality as multi label data often happens to have large number of features. In this paper, the Direct Multi-label Linear Discriminant Analysis method is proposed for dimensionality reduction of multilabel data. In particular we extend Multi-label Discriminant Analysis (MLDA) and modify the between-class scatter matrix in order to improve classification accuracy. The problem that Direct MLDA overcomes is the limitation of the produced projections that in MLDA are defined as K - 1 for a K class problem. Experimental results on video based human activity recognition for digital media analysis and description as well as on other challenging problems indicate the superiority of the proposed method. © Springer-Verlag Berlin Heidelberg 2013.