A Method to Clustering the Feature Ranking on Data Classification Using an Ensemble Feature Selection

Abstract

The aim of this paper is to improve the predictive performance of the classification process by means of building multiple data classification models based on the output from feature selection methods that use ensemble strategy to find the optimal set of features. Currently, the data volume has grown at an extreme rate causing a variety of problems. The big data situation has made automatic analysis tasks such as data classification facing low performance and high computational time problems when dealing with big data that are huge in both volume and dimensions. In this research work, we tackle the big data problem in the high dimensionality aspect. We propose an ensemble method to reduce data dimension by means of feature clustering to rank the potential features and also return suitable subset of features for further classifying the training data. The two paradigms of feature selection based on ensemble strategy are proposed and evaluated. Experimental results confirm the efficacy of our proposed feature ensemble method.

The Silhouette Width Criterion for Clustering and Association Mining to Select Image Features

“…This research, thus, aims at extending the previous work of Nuntawut et al [7], [8] by proposing a silhouette width criterion for automatic setting of initial cluster numbers. We also add confidence criteria into feature selection based on association rule mining technique to increase performance.…”

“…But this feature selection algorithm does not work automatically because human is the one who select the features one by one based on the feature scores reported from the algorithm. Therefore, Nuntawut et al [8] improved the algorithm by proposing clustering technique to cluster the feature scores to assist users on finding an appropriate groups of features. The clustering process is supposed to be automatic in the sense that the number of clusters should be judged by the process itself.…”