Feature Selection Based on Sequential Orthogonal Search Strategy

This thesis introduces three new feature selection methods based on sequential orthogonal search strategy that addresses three different contexts of feature selection problem being considered. The first method is a supervised feature selection called the maximum relevance–minimum multicollinearity (MRmMC), which can overcome some shortcomings associated with existing methods that apply the same form of feature selection criterion, especially those that are based on mutual information. In the proposed method, relevant features are measured by correlation characteristics based on conditional variance while redundancy elimination is achieved according to multiple correlation assessment using an orthogonal projection scheme. The second method is an unsupervised feature selection based on Locality Preserving Projection (LPP), which is incorporated in a sequential orthogonal search (SOS) strategy. Locality preserving criterion has been proved a successful measure to evaluate feature importance in many feature selection methods but most of which ignore feature correlation and this means these methods ignore redundant features. This problem has motivated the introduction of the second method that evaluates feature importance jointly rather than individually. In the method, the first LPP component which contains the information of local largest structure (LLS) is utilized as a reference variable to guide the search for significant features. This method is referred to as sequential orthogonal search for local largest structure (SOS-LLS). The third method is also an unsupervised feature selection with essentially the same SOS strategy but it is specifically designed to be robust on noisy data. As limited work has been reported concerning feature selection in the presence of attribute noise, the third method is thus attempts to make an effort towards this scarcity by further exploring the second proposed method. The third method is designed to deal with attribute noise in the search for significant features, and kernel pre-images (KPI) based on kernel PCA are used in the third method to replace the role of the first LPP component as the reference variable used in the second method. This feature selection scheme is referred to as sequential orthogonal search for kernel pre-images (SOS-KPI) method. The performance of these three feature selection methods are demonstrated based on some comprehensive analysis on public real datasets of different characteristics and comparative studies with a number of state-of-the-art methods. Results show that each of the proposed methods has the capacity to select more efficient feature subsets than the other feature selection methods in the comparative studies.