Applying Evolutionary Hypernetworks for Multiclass Molecular Classification of Cancer

This paper presents a pattern recognition method for multiclass cancer molecular classification using evolutionary hypernetworks. A multiclass classification issue is decomposed into a set of binary classification issues by One-Versus-All (OVA) approach. The signal-to-noise ratio method is employed for informative genes selection from the DNA microarray. A series of binary classifiers are evolved and used to build a final ensemble classifier for multiclass classification through an evolutionary learning procedure of the hypernetwork. The test sample is classified by using the ensemble classifier. Experimental results show that the Leave One Out Cross Validation (LOOCV) accuracy of the acute leukemia dataset, the small, round blue cell tumor dataset, and the GCM dataset is 98.61%, 100% and 85.35%, respectively. The evolutionary hypernetworks is fit to find cancer-related genes and has a good readability of the learned results.