Cross classification approach for feature selection

In this paper we present a novel method for objective function specification and feature selection method by combining unsupervised learning with supervised cross validation. A one dimensional Kohonen-SOM (Self-Organizing Map) and k-mean clustering algorithm are used to perform a clustering of sample data for a chosen subset of input features and a given number of clusters. The resulting object clusters are compared with the predefined original target classes and a matching factor is calculated. This factor is used as an objective function for heuristic (forward/backward) sequential feature selection. Additionally, a novel method for feature selection, called cross feature selection, is introduced. This method can be applied only if there exist more than two target classes, because it uses the grouping of target classes into larger hyper-classes. For each group of target classes (hyper classes) and for individual target classes it is possible to use this matching factor to test the significance of every subset of input variables due to its target predictability.