Feature subset selection using mutual standard deviation in sentiment mining

The complexity of an optimal feature subset selection for sentiment classification problem grows exponentially based on the number of features. The paper aims to enhance the selection of feature subset in high-dimensional feature space based on filter approach. Filter-based methods evaluate features relevance by considering only at the properties of the data. The score of the feature is calculated, then low-scoring features will be removed. In most feature selection methods, the selection of a feature subset is based on the feature space while the model hypothesis space is ignored. To assess the relevance of features, a feature ranking by distance measure is conducted to minimize intra-class and maximize inter-class distance. A filter method based on distribution and dispersion of features on feature space namely mutual standard deviation is proposed. Wide ranges of comparative experiments are performed on two widely used datasets, namely, movie, and book review dataset, in sentiment analysis. The results showed that proposed feature ranking method outperforms the other baseline methods regarding accuracy.