Improving Sentiment Analysis on Imbalanced Airlines Twitter Data Using DSMOTE Technique

Abstract

Recently, social media and microblogging have gained popularity and traffic. Customer tweets to US airlines take time to analyse. A sentiment analysis model for unbalanced datasets fixes this with the help of the SMOTE method. This paper uses a new over-sampling technique to synthesise more samples near easily misclassified cases, unlike standard SMOTE, which treats all minority group samples equally. We target misclassified minority class classification to improve accuracy. The model steps through tweet sentiment classification. First, remove tweets with special characters, URLs, and stop words. It cleans tweets and extracts features to create numerical feature vectors. The Bag of Words (BoW) model uses all unique tweet terms to develop a lexicon—the presence or absence of these words numbers each tweet. We use Random Forest (RF) and Recurrent Neural Network (RNN) classification models after transforming tweets into feature vectors. A Random Forest ensemble learning system classifies using many decision trees. RNNs process sequential text using internal memory states. RF and RNN models use tweet feature vectors. Models learn feature-sentiment label patterns. They can label new tweets positive, negative, or neutral. These classification models let the installed system classify tweets by sentiment, providing valuable sentiment analysis insights. These models accurately classify tweets as positive, negative, or neutral. The density-based SMOTE results show our model’s efficiency. TFIDF vectoriser Random Forest has 81% accuracy and 70% F1 score. These measures show the model can classify sentiment in imbalanced datasets, making it useful for sentiment analysis.