Optimal Subclass Discovery for Discriminant Analysis

Discriminant Analysis (DA) has had a big influence in many scientific disciplines. Unfortunately, DA algorithms need to make assumptions on the type of data available and, therefore, are not applicable everywhere. For example, when the data of each class can be represented by a single Gaussian and these share a common covariance matrix, Linear Discriminant Analysis (LDA) is a good option. In other cases, other DA approaches may be preferred. And, unfortunately, there still exist applications where no DA algorithm will correctly represent reality and, therefore, unsupervised techniques, such as Principal Components Analysis (PCA), may perform better. This paper first presents a theoretical study to define when and (most importantly) why DA techniques fail (Section 2). This is then used to create a new DA algorithm that can adapt to the training data available (Sections 2 and 3). The first main component of our solution is to design a method to automatically discover the optimal set of subclasses in each class. We will show that when this is achieved, optimal results can be obtained. The second main component of our algorithm is given by our theoretical study which defines a way to rapidly select the optimal number of subclasses. We present experimental results on two applications (object categorization and face recognition) and show that our method is always comparable or superior to LDA, Direct LDA (DLDA), Nonparametric DA (NDA) and PCA.