Analysis of the Impact of Vectorization Methods on Machine Learning-Based Sentiment Analysis of Tweets Regarding Readiness for Offline Learning

Abstract

Twitter users use social media to express emotions about something, whether it is criticism or praise. Analyzing the opinions or sentiments in the tweets that Twitter users send can identify their emotions for a particular topic. This study aims to determine the impact of vectorization methods on public sentiment analysis regarding the readiness for offline learning in Indonesia during the Covid-19 pandemic. The authors labeled sentiment using two different approaches: manually and automatically using the NLP TextBlob library. We compared the vectorization method used by employing count vectorization, TF-IDF, and a combination of both. The feature vectors were then classified using three classification methods: naïve Bayes, logistic regression, and k-nearest neighbor, for both manual and automatic labeling. To assess the performance of sentiment analysis models, we used accuracy, precision, recall, and F1-score for performance metrics. The best results showed that the Logistic regression classifier with the feature extraction technique that combines count vectorization and TF-IDF provided the best performance for both data with manual and automatic labeling.