Classification and Clustering Using Intelligent Techniques: Application to Microarray Cancer Data

Analysis and interpretation of DNA Microarray data is a fundamental task in bioinformatics. Feature Extraction plays a critical role in better performance of the classifier. We address the dimension reduction of DNA features in which relevant features are extracted among thousands of irrelevant ones through dimensionality reduction. This enhances the speed and accuracy of the classifiers. Principal Component Analysis (PCA) is a technique used for feature extraction which helps to retrieve intrinsic information from high dimensional data in eigen spaces to solve the curse of dimensionality problem. The curse of dimensionality means n >> m, where n is a large number of features and m is a small number of samples (may be too less). Neural Networks (NN) and Support Vector Machine (SVM) are implemented and their performances are measured in terms of predictive accuracy, specificity, and sensitivity. First, we implement PCA for significant feature extraction and then FFNN trained using Backpropagation (BP) and SVM are implemented on the reduced feature set. Next, we propose a Multiobjective Genetic Algorithm-based fuzzy clustering technique using real coded encoding of cluster centers for clustering and classification. This technique is implemented on microarray cancer data to select training data using multiobjective genetic algorithm with non-dominated sorting (MOGA-NSGA-II). The two objective functions for this multiobjective techniques are optimization of cluster compactness as well as separation simultaneously. This approach identifies the solution i.e. the individual chromosome which gives the optimal value of the compactness and separation. Then we find high confidence points for these non-dominated set using a fuzzy voting technique. Support Vector Machine (SVM) classifier is further trained by the selected training points which have high confidence value. Then remaining points are classified by trained SVM classifier. Finally, the four clustering label vectors through majority voting ensemble are combined, i.e., each point is assigned a class label that obtains the maximum number of votes among the four clustering solutions. The performance of the proposed MOGA-SVM, classification and clustering method has been compared to MOGA-BP, SVM, BP. The performance are measured in terms of Silhoutte Index, ARI Index respectively. The experiment were carried on three public domain cancer data sets, viz., Ovarian, Colon and Leukemia cancer data to establish its superiority.