Comparative study on feature selection and ensemble methods for sentiment analysis classification

Zahir Mohammad Adnan Younis
زاهر "محمد عدنان" يونس
Journal Title
Journal ISSN
Volume Title
AL-quds University
People use the Web and social media to express their opinions and comments on various topics and posts generating huge amounts of data. Hence, comes the necessity to analyze this large amount of text regarding a certain subject and figuring out what people think of it. The interest and necessity of this analysis is continuously rising in many fields, such as politics, marketing, entertainment, sports, etc., to figure out people opinions, thinking, interests, preferences, and trends. Consequently, analysis, classification and clustering of this huge amount of text data regarding certain subjects became an interest of a vast number of researchers and beneficiaries. This analysis of text data content is known as sentiment analysis. Sentiment Analysis (SA) is a text-mining field that computationally treats and analyses these sentiments (opinions, thinks, subjectivity, interests, preferences, etc.,.) of available text. SA aims to classify expressions in a text as positive, negative or neutral opinion towards the subject of interest. The main objective of this research is to carry out a comparative study on the accuracy and performance of feature selection and ensemble methods for SA classification. The comparison was carried out using different combinations of classification algorithms for classifying text to being either positive or negative. During the comparison of the algorithms and methods, the results showed that better accuracy can be achieved based on the used feature selection method (i.e., statistical, wrapper, or embedded). Additionally, it showed which feature selection method outperforms and is more suitable than other methods for the type of data and classification algorithms. Furthermore, when using combined ensemble methods (Bagging, Boosting, Stacking and Vote) performed better than using a single classifier by means of accuracy. Moreover, merging feature subsets selected by embedded method improved classification accuracy. Finally, tuning the parameters of feature selection methods improved the classification accuracy and reduced the time needed to select feature subsets. Particularly, the results showed that accuracy depends on the feature selection method, ensemble methods, number of selected features, type of classifier, and tuning parameters of the algorithms used. A high accuracy of up to 99.85% was achieved by merging features of two embedded methods when using stacking ensemble method. Also, a high accuracy of 99.5% was achieved by tuning parameters in stacking method, and it reached 99.95% and iv 100% by tuning parameters in SVMAttributeEval method using statistical and machine learning approaches, respectively. Furthermore, tuning algorithms' parameters reduced the time needed to select feature subsets.