A Stacked Ensemble Approach for Robust Sentiment Classification on Twitter Data

This study focuses on how a stacked ensemble can be employed for binary sentiment classification on short and noisy Twitter data. The model integrates Random Forest and Support Vector Machine (SVM) as the base models, with Logistic Regression as the final decision maker. Tweets were first pre-processed and then modified into numerical features using TF-IDF so that the algorithms could process them. The model showed only moderate performance, which reflects how difficult it is to work with noisy and high-dimensional text. More importantly, this work points out the shortcomings that arise when utilizing and combining different types of models in an ensemble. Limitation such as calibration compatibility amongst models, conflicts in decision boundaries, and the limits of TF-IDF in capturing meaning are highlighted. These findings indicate that even well-known techniques can struggle when applied to real-world sentiment analysis. Overall, this study provides a clear and reproducible baseline and encourages future work to use richer text features, better-calibrated ensembles, and more diverse models to improve sentiment classification.