Feature Selection in Cancer Classification from mRNA Data Based on Localized Dimension Reduction

In response to the ICMLA 2009 "Functional Clustering of Gene Expression Profiles in Human Cancers Challenge", we present a new dimension reduction approach that ranks features based on their localized discriminative power. The proposed method is based on a localized dimension reduction penalty added to the objective function for training a hyper basis function (hyper BF or generalized RBF) neural network. The localized dimension reduction in a hyper BF network is motivated by the fact that many patterns are recognized to belong to one class due to the co-occurrence of specific values along a specific set of dimensions. Furthermore, one class is very likely to be composed of multiple sub-models that are different in their characteristics. The proposed method is applied at multiple iterations. Features are ranked each iteration and a smaller subset is passed to the next iteration. A support vector machine (SVM) is used to evaluate the remaining dimensions in a ten-fold cross validation setting. Experimental results show the proposed method effectively reduces the number of dimensions from 54,613 to 65 while increasing the cross-validation classification accuracy from 92.3% to 96.4%. The classification accuracy of the final model on unseen data was reported by the challenge coordinators to be 94% accurate.