An Ensembled Based Machine Learning Technique of Sentiment Analysis

Abstract

User evaluations on social networking sites such as Twitter, YouTube, and Facebook have grown rapidly due to their widespread use in sentiment analysis, providing valuable insight for both governmental and non-governmental organizations. Analyzing these evaluations not only helps improve the effectiveness of products and services, but also aids in the developing operational and management strategies. Although various analysis models have been proposed, challenges remain in processing, classifying, and accessing user evaluations, such as dealing with complex sentences,  requiring more than sentiment words and achieving adequate accuracy and performance based on limited labeled data. This study primarily examines the performance of three commonly used machine learning algorithms proposes an Ensemble method, which combines Naive Bayes, Support Vector Machines and K-nearest Neihbor algorithms. The proposed method was tested on a Twitter dataset.The Ensemble method creates a classification model by applying the three classification algorithms: Naive Bayes, Support Vector Machines and K-nearest Neighbor, to for the prediction of unknown example and assigns the predicted class receiving the most votes. According to the results, the Ensemble method has an accuracy of 83.28% at 60/40 test split, 83.27% at 70/30 test split, 83.50% at 80/20 test split, and 86.12% at 90/10 test split, and  F-measure of 84.15% at 60/40 test split, 83.72% at 70/30 test split, 84.41% at 80/20 test split, and 86.58% at 90/10 test split.  In terms of individual performance, k-nearest Neighbor has better accuracy and F-measure than Support Vector Machine and Naive Bayes while the Ensemble method proves to be the most efficient in terms of accuracy and F-measure.