Subclass Marginal Fisher Analysis

Subspace learning techniques have been extensively used for dimensionality reduction (DR) in many pattern classification problem domains. Recently, Discriminant Analysis (DA) methods, which use subclass information for the discrimination between the data classes, have attracted much attention. As DA methods are strongly dependent on the underlying distribution of the data, techniques whose functionality is based on neighbourhood information among the data samples have emerged. For instance, based on the Graph Embedding (GE) framework, which is a platform for developing novel DR methods, Marginal Fisher Analysis (MFA) has been proposed. Although MFA surpasses the above distribution limitations, it fails to model potential subclass structure that might lie within the several classes of the data. In this paper, motivated by the need to alleviate the above shortcomings, we propose a novel DR technique, called Subclass Marginal Fisher Analysis (SMFA), which combines the strength of subclass DA methods with the versatility of MFA. The new method is built by extending the GE framework so as to include subclass information. Through a series of experiments on various real-world datasets, it is shown that SMFA outperforms in most of the cases the state-of-the-art demonstrating the potential of exploiting subclass neighbourhood information in the DR process.