Feature selection using Markov clustering and maximum spanning tree in high dimensional data

Feature selection is the most important preprocessing step for classification of high dimensional data. It reduces the load of computational cost and prediction time on classification algorithm by selecting only the salient features from the data set for learning. The main challenges while applying feature selection on high dimensional data (HDD) are: handling the relevancy, redundancy and correlation between features. The proposed algorithm works with the three main steps to overcome these issues. It focuses on filtering strategy for its effectiveness in handling the data sets with large size and high dimensions. Initially to measure the relevancy of features with respect to class, fisher score is calculated for each feature independently. Next, only relevant features are passed to the clustering algorithm to check the redundancy of features. Finally the correlation between features is calculated using maximum spanning tree and the most appropriate features are filtered out. The classification accuracy of the presented approach is validated by using C4.5, IB1 and Naive Bayes classifier. The proposed algorithm gives high classification accuracy when compared against the accuracies given by three different classifiers on the datasets containing features extracted from fisher score method and dataset containing all the features or full-featured dataset.