A Comparative Study of Different Machine Learning Techniques to Analyze Sentiments

Abstract

The advent of social networks has transformed communication patterns, offering newlinevaluable data, information, and content that can be harnessed to analyze user newlineopinions. This study focuses on sentiment analysis, a prominent natural language newlineprocessing (NLP) task, aimed at determining user sentiments and evaluations towards newlineproducts, entities, or services they review. Various methods, for example word newlineembeddings, sentiment lexicons, and annotated data, have been employed in newlinesentiment analysis. The research presents a comparative evaluation of machine newlinelearning approaches, encompassing three main steps. In the first step, traditional newlinefeature extraction techniques were assessed against machine learning methods. newlineWord2Vec emerged as the most effective, outperforming TF-IDF and Doc2Vec, newlineowing to its dense vector representations that captured semantic meanings adeptly. newlineLogistic Regression achieved the highest F1-score with Word2Vec, showcasing its newlineability to capture word relationships and achieve superior classification performance. newlineSVM achieved highest F1-score for Word-2-Vec and lowest in Doc-2-Vec. Random newlineForest excelled with TF-IDF, achieving the highest F1-score among the algorithms, newlinewhile XGBoost achieved highest F1-score for Word-2-Vec and lowest in Doc-2-Vec. newlineThe second step incorporated a convolutional neural network (CNN) approach to newlineassess the overall accuracy of the proposed methodologies, which yielded an 87% newlineaccuracy. The suggested methodology outperformed the most recent methods by 1- newline2% in terms of overall accuracy. In the third step, domain-independent sentiment newlineanalysis was performed using a deep learning methodology. An attention-based newlineemotion-embedding BiLSTM-GRU Network for sentiment analysis was introduced newlineand compared against four baseline models. The proposed network demonstrated newlinesuperior performance, outperforming state-of-the-art models, with an impressive newlineaccuracy of 93%. In conclusion, this research provides valuable insights into newlinesentiment analysis, offering a comprehensive evaluation of machine