Improving Sentiment Analysis of Microblogs through Bagging of Ensemble Classifiers

The growth of social media and micro-blogs has greatly shifted the dynamics of businesses and the way advertising is carried out. Micro-blogs have transformed the consumer from being mere shoppers to advertiser and reviewers. Micro-blog opinions have become the reflection of society’s opinions, attitudes, and preferences at large hence the greater need to not only access data stemming from microblogs, but to be able to analyze the data and make predictions based on it, whether a product is seen in a positive light or negatively. This fierce battle for consumers ‘attention has resulted in many corporations investing in data analysis to capture the market; consumers nowadays heavily rely on the opinions and reviews shared across microblogs in order to make a decision on products and services on offer. Thus, the need for organizations to be able to classify these reviews quickly and as proficiently as possible. However, the task of combing through millions of reviews to determine the sentiment of the feedback is humanly tasking henceforth a number of machine learning techniques to detect and perform binary classification – positive and negative- on reviews have already been proposed. However, the nature of the reviews of micro-blogs has resulted in classification increasingly becoming more complex with the usage of emoticons, slang and short phrase which we have dubbed as “social media language”. Classifying such complex reviews or blog posts using simplistic single classifiers no longer suffices hence in this paper, we proposed an ensemble classifier-based approach to detect polarity of reviews. The proposed ensemble classifier uses 7 classifiers- Random Forest (RF), Support Vector Machine (SVM), Logistic Regression (LR), Naïve Bayes, K- Nearest Neighbors (KNN), Xgboost and Adaboost classifiers. The proposed technique is assessed on Pang et al.’s polarity dataset v1.0, Bo Pang and Lillian Lee’s 2004 ACL polarity dataset v2.0 and ACL’s IMDb dataset. The evaluation results show that the proposed classifier provides better classification accuracy on both datasets than simple classifiers. Keywords: Ensemble, Bagging, Sentiment Analysis, F1-Score, Accuracy, Classification.