Enhanced Sentiment Classification using a Multi-layered Stacked Ensemble Architecture
Marriott international, Dallas, Texas, USA, Nitin Kumar, Vipin Kataria · International Journal of Innovative Research in Science Engineering and Technology · 2025
Sentiment analysis has emerged as a crucial tool for understanding public opinion, consumer feedback, and social media discourse in today's data-driven business landscape. This paper presents a comprehensive study on sentiment analysis using various machine learning approaches, addressing the growing need for accurate and efficient sentiment classification systems. This paper presents a novel multi-layered stacked ensemble architecture for Twitter sentiment analysis that significantly outperforms existing methods. We integrate complementary feature extraction techniques—TF-IDF vectorization and BERT embeddings—with diverse classification algorithms to create a robust system for categorizing tweets into positive, negative, neutral, and irrelevant sentiment classes. Our architecture systematically leverages the strengths of traditional machine learning models (XGBoost, Random Forest) and deep learning approaches (BERT, LSTM) through a cascading structure of meta-feature generation and specialized classifier layers. Experimental results demonstrate the superior performance of our proposed model, achieving 98.23% accuracy, 98.12% precision, 98.89% recall, and 98.25% F1-score, representing substantial improvements over strong baseline models including ExtraTreesClassifier (0.63% higher accuracy) and other state-of-the-art approaches. The model particularly excels at maintaining exceptional balance between precision and recall while effectively handling challenging cases such as sarcasm and ambiguous sentiment expressions. Our findings confirm that the dual feature representation approach and synergistic model combination create a more robust sentiment analysis system than any individual approach, providing valuable insights for future research in social media text analysis.