Improving Sentiment Classification Accuracy using Engagement and Temporal Features from Social Media Data

Sentiment analysis has become an important instrument to gain insight on user attitudes, brands and trends in online behavior and digital platforms. This paper addresses this aspect by conducting a recognizable study of an approach encompassing all the available methodologies on classifying sentiment in three categories, i.e. positive, negative and neutral as well as traditional Machine Learning (ML) models, ensemble approaches and a stacking-based combination approach. Different independent classifier models such as Support Vector Machine (SVM), Logistic Regression (LR), Naive Bayes (NB) mode have been tested with Term Frequency-Inverse Document Frequency (TF-IDF) features on a sentiment dataset that has undergone preprocessing. Hard voting and extended voting using multiple learners are applied to make a more significant improvement in performance. It shows the biggest benefit in a stacking ensemble model, where LR, NB and SVM are implemented as base learners, and Random Forest (RF) serves as a meta-learner. The findings confirm the usefulness of using hybrid ensemble models in sentiment classification particularly when sparse and moderate sized data sets are involved.