Eager Hierarchical Feature Selection

Abstract

This chapter discusses four different eager hierarchical feature selection methods, i.e. Tree-based Feature Selection (TSEL) (Jeong and Myaeng, Proceedings of the international joint conference on natural language processing, Nagoya, Japan, 2013, [1]) , Bottom-up Hill Climbing Feature Selection (HC) (Wang et al, Proceedings of the 26th Australasian computer science conference, Darlinghurst, Australia, 2003, [5]) , Greedy Top-down Feature Selection (GTD) (Lu et al, Proceedings of the international conference conference on collaborative computing, Austin, USA, 2013, [2]) and Hierarchy-based Feature Selection (SHSEL) (Ristoski and Paulheim, Proceedings of the international conference on discovery science (DS 2014), 2014, [3]) . All of those four hierarchical feature selection methods are also categorised as filter methods. Those methods aim to alleviate the feature redundancy by considering the hierarchical structure between features and the predictive power of features (e.g. information gain). Unlike the lazy hierarchical feature selection methods discussed in last chapter, those eager hierarchical feature selection methods only consider the relevance value of those features calculated by the training dataset and the hierarchical information, without considering the actual value of features for individual testing instance.