Feature Selection Based on Top Frequency

High-dimensional data has become a challenging problem for feature selection. This paper proposes a top frequency-based feature selection method that yields subsets of features and give better results. First, we use a given high-dimensional dataset to form a random forest model. Then, we go through the forest model to achieve features and the number of occurrences that these achieved features are used as nodes in all decision trees in the forest. Finally, these achieved features are sorted in the descending order of their frequency. We only choose the top K features with the highest frequency from the sorted feature list to build a new model for the classification task. Fifteen high-dimensional datasets were conducted in experiments to evaluate the proposed feature selection model. The experimental results have shown that three classifiers built with small subsets of obtained features outperformed those built with all features.