home        research        publications        people        open positions        contact
	Joint Feature and Classifier Selection The accuracy of any classifier is affected the three major components of the system: 1) classification algorithm: algorithm selection and tuning 2) inputs: how the features are extracted and/or selected 3) outputs: number and type of outputs the system is classifying While classifier selection and feature selection/extraction are well researched problems, there has been little research on output selection. The common method of solving a multi-class classification problem is to use a single classifier with multiple outputs. However, it is also possible to divide the problem into several smaller sub-classification problems. There are two major advantages of this approach: 1) a smaller classifier with fewer outputs may be easier to train 2) more specific feature sets can be used to divide the different sets of classes For example, in human motion applications, actions such as punching or reaching are likely best described by arm-based features, whereas kicking and walking are described mainly by leg-based features. A single multi-class classifier would require all these features, whereas a hierarchical classifier could use a more specific feature set for each action. There are also some challenges, however. 1) it is unclear which divisions will be beneficial (this is also likely classifier specific) 2) the best features for each sub-classifier will change as the outputs change This work uses genetic algorithm to simultaneously design a hierarchical tree-based classifier, and perform feature selection for the individual sub-classifiers. The first part of the genome specifies which features are used in each classifier, and the second part of the genome specifies the tree structure by specifying the output for each class at each level of the tree. The hierarchical classifier is able to outperform a flat KNN on artificial data sets, and has a performance that is comparable to the standard one-vs.-rest and one-vs.-one binary classifier extensions for SVN, but with fewer classifiers and a much shorter test time for new points.
	Researchers: Cecille Freeman, Otman Basir, Dana Kulić Key Publications: C. Freeman, D. Kulić and O. Basir, Feature-selected tree-based classification, IEEE Transactions on Systems, Man and Cybernetics, Part B: Cybernetics, 2013, In Press. link C. Freeman, D. Kulić and O. Basir, Joint feature selection and hierarchical classifier design, IEEE International Conference on Systems, Man, and Cybernetics, pp. 1728 - 1734, 2011. pdf