Integration of feature subset selection methods for sentiment analysis

Abstract

Feature selection is one of the main challenges in sentiment analysis to find an optimal feature subset from a real-world domain. The complexity of an optimal feature subset selection grows exponentially based on the number of features for analysing and organizing data in high-dimensional spaces that lead to the high-dimensional problems. To overcome the problem, this study attempted to enhance the feature subset selection in high-dimensional data by removing irrelevant and redundant features using filter and wrapper approaches. Initially, a filter method based on dispersion of samples on feature space known as mutual standard deviation method was developed to minimize intra-class and maximize inter-class distances. The filter-based methods have some advantages such as they are easily scaled to high-dimensional datasets and are computationally simple and fast. Besides, they only depend on feature selection space and ignore the hypothesis model space. Hence, the next step of this study developed a new feature ranking approach by integrating various filter methods. The ordinal-based and frequency-based integration of different filter methods were developed. Finally, a hybrid harmony search based on search strategy was developed and used to enhance the feature subset selection to overcome the problem of ignoring the dependency of feature selection on the classifier. Therefore, a search strategy on feature space using integration of filter and wrapper approaches was introduced to find a semantic relationship among the model selections and subsets of the search features. Comparative experiments were performed on five sentiment datasets, namely movie, music, book, electronics, and kitchen review dataset. A sizeable performance improvement was noted whereby the proposed integration-based feature subset selection method yielded a result of 98.32% accuracy in sentiment classification using POS-based features on movie reviews. Finally, a statistical test conducted based on the accuracy showed significant differences between the proposed methods and the baseline methods in almost all the comparisons in k-fold cross-validation. The findings of the study have shown the effectiveness of the mutual standard deviation and integration-based feature subset selection methods have outperformed the other baseline methods in terms of accuracy.