Hybrid ANOVA and LASSO Methods for Feature Selection and Linear Support Vector, Multilayer Perceptron and Random Forest Classifiers Based on Spark Environment for Microarray Data Classification

Abstract

Microarray dataset frequently contains a countless number of insignificant and irrelevant genes that might lead to loss of valuable data. The classes with both high importance and high significance gene sets are commonly preferred for selecting the genes, which determines the sample classification into their particular classes. This property has obtained a lot of importance among the specialists and experts in microarray dataset classification. The trained classifier model is tested for cancer datasets and Huntington disease data (HD) which consists of Prostate cancer (Singh) dataset comprising 102 samples, 52 of which are tumors and 50 are normal with 12625 genes. The lung cancer (Gordon) dataset comprises 181 samples, 150 of which are normal and 31 are tumors with 12533 genes. The breast cancer (Chin) dataset comprises 118 samples, 43 of which are normal and 75 are tumors with 22215 genes. The breast cancer (Chowdary) dataset comprises 104 samples, 62 of which are normal and 42 are tumors with 22283 genes. Finally, the Huntington disease (Borovecki) dataset comprises 31 samples, 14 of which are normal and 17 are with Huntington’s disease with 22283 genes. This paper uses Multilayer Perceptron Classifier (MLP), Random Forest (RF) and Linear Support Vector classifier (LSVC) classification algorithms with six different feature selection methods named as Principal Component Analysis (PCA), Extra Tree Classifier (ETC), Analysis of Variance (ANOVA), Least Absolute Shrinkage and Selection Operator (LASSO), Chi-Square and Random Forest Regressor (RFR). Further, the paper presents a comparative analysis on the obtained classification accuracy and time consumed among the models in Spark environment and in conventional system. Performance parameters such as accuracy and time consumed are applied in this comparative analysis to analyze the behavior of the classifiers in the two environments. Th results indicate that the models in spark environment was extremely effective for processing large-dimension data, which cannot be processed with conventional implementation related to a some algorithms. After that, a proposed hybrid model containing embedded approach (LASSO) and the Filter (ANOVA) approach was used to select the optimized features form the high dimensional dataset. With the reduced dimension of features, classification is performed on the reduced data set to classify the samples into normal or abnormal and applied in spark in hadoop cluster (distributed manner). The proposed model achieved accuracy of 100% in case of Borovecki dataset when using all classifiers, 100% in case of Singh, Chowdary and Gordon datasets when classified with RF and LSVC classifiers. Also, accuracy was 96% in case of Chin dataset when using RF classifier with optimal genes with respect to accuracy and time consumed.