Imbalanced Sentiment Polarity Detection Using Emoji-Based Features and Bagging Ensemble

The research on the sentiment analysis of social media content is remarkably growing for constructing resources and investigating new ideas and techniques to address various challenges. This paper explores a new approach for sentiment polarity detection in Arabic text using non-verbal emoji-based features while addressing the class imbalance problem. The proposed method is based on Bootstrap Aggregating (Bag-ging) algorithm and Synthetic Minority Oversampling Technique (SMOTE) to build and combine multiple models from the training dataset. Three different classifiers are evaluated as single and ensemble classifiers: naive Bayes, k-NN and decision trees. The performance is evaluated and compared on three datasets with a varying imbalance ratio ranging from two to more than seven. The experimental results show that the proposed approach performs better than other approaches in most of the considered cases.