Markov blanket: efficient strategy for feature subset selection method for high dimensionality microarray cancer datasets

Abstract

Currently, feature subset selection methods are very important, especially in areas of application for which datasets with tens or hundreds of thousands of variables (genes) are available. Feature subset selection methods help us select a small number of variables out of thousands of genes in microarray datasets for a more accurate and balanced classification. Efficient gene selection can be considered as an easy computational hold of the subsequent classification task, and can give subset of gene set without the loss of classification performance. In classifying microarray data, the main objective of gene selection is to search for the genes while keeping the maximum amount of relevant information about the class and minimize classification errors. In this paper, explain the importance of feature subset selection methods in machine learning and data mining fields. Consequently, the analysis of microarray expression was used to check whether global biological differences underlie common pathological features in different types of cancer datasets and identify genes that might anticipate the clinical behavior of this disease. Using the feature subset selection model for gene expression contains large amounts of raw data that needs analyzing to obtain useful information for specific biological and medical applications. One way of finding relevant (and removing redundant ) genes is by using the Bayesian network based on the Markov blanket [1]. We present and compare the performance of the different approaches to feature (genes) subset selection methods based on Wrapper and Markov Blanket models for the five-microarray cancer datasets. The first way depends on the Memetic algorithms (MAs) used for the feature selection method. The second way uses MRMR (Minimum Redundant Maximum Relevant) for feature subset selection hybridized by genetic search optimization techniques and afterwards compares the Markov blanket model’s performance with the most common classical classification algorithms for the selected set of features. For the memetic algorithm, we present a comparison between two embedded approaches for feature subset selection which are the wrapper filter for feature selection algorithm (WFFSA) and Markov Blanket Embedded Genetic Algorithm (MBEGA). The memetic algorithm depends on genetic operators (crossover, mutation) and the dedicated local search procedure. For comparisons, we depend on two evaluations techniques for learning and testing data which are 10-Kfold cross validation and 30-Bootstraping. The results of the memetic algorithm clearly show MBEGA often outperforms WFFSA methods by yielding more significant differentiation among different microarray cancer datasets. In the second part of this paper, we focus mainly on MRMR for feature subset selection methods and the Bayesian network based on Markov blanket (MB) model that are useful for building a good predictor and defying the curse of dimensionality to improve prediction performance. These methods cover a wide range of concerns: providing a better definition of the objective function, feature construction, feature ranking, efficient search methods, and feature validity assessment methods as well as defining the relationships among attributes to make predictions. We present performance measures for some common (or classical) learning classification algorithms (Naive Bayes, Support vector machine [LiBSVM], K-nearest neighbor, and AdBoostM Ensampling) before and after using the MRMR method. We compare the Bayesian network classification algorithm based on the Markov Blanket model’s performance measure with the performance of these common classification algorithms. The result of performance measures for classification algorithm based on the Bayesian network of the Markov blanket model get higher accuracy rates than other types of classical classification algorithms for the cancer Microarray datasets. Bayesian networks clearly depend on relationships among attributes to make predictions. The Bayesian network based on the Markov blanket (MB) classification method of classifying variables provides all necessary information for predicting its value. In this paper, we recommend the Bayesian network based on the Markov blanket for learning and classification processing, which is highly effective and efficient on feature subset selection measures.