Unionization method for changing opinion in sentiment classification using machine learning

Abstract

Sentiment classification aims to determine whether an opinionated text expresses a positive, negative or neutral opinion. Most existing sentiment classification approaches have focused on supervised text classification techniques. One critical problem of sentiment classification is that a text collection may contain tens or hundreds of thousands of features, i.e. high dimensionality, which can be solved by dimension reduction approach. Nonetheless, although feature selection as a dimension reduction method can reduce feature space to provide a reduced feature subset, the size of the subset commonly requires further reduction. In this research, a novel dimension reduction approach called feature unionization is proposed to construct a more reduced feature subset. This approach works based on the combination of several features to create a more informative single feature. Another challenge of sentiment classification is the handling of concept drift problem in the learning step. Users’ opinions are changed due to evolution of target entities over time. However, the existing sentiment classification approaches do not consider the evolution of users’ opinions. They assume that instances are independent, identically distributed and generated from a stationary distribution, even though they are generated from a stream distribution. In this study, a stream sentiment classification method is proposed to deal with changing opinion and imbalanced data distribution using ensemble learning and instance selection methods. In relation to the concept drift problem, another important issue is the handling of feature drift in the sentiment classification. To handle feature drift, relevant features need to be detected to update classifiers. Since proposed feature unionization method is very effective to construct more relevant features, it is further used to handle feature drift. Thus, a method to deal with concept and feature drifts for stream sentiment classification was proposed. The effectiveness of the feature unionization method was compared with the feature selection method over fourteen publicly available datasets in sentiment classification domain using three typical classifiers. The experimental results showed the proposed approach is more effective than current feature selection approaches. In addition, the experimental results showed the effectiveness of the proposed stream sentiment classification method in comparison to static sentiment classification. The experiments conducted on four datasets, have successfully shown that the proposed algorithm achieved better results and proving the effectiveness of the proposed method.