A Comparative Analysis of Sentiment Classification Based on Two-Level Ensemble Models

Sentiment analysis is a method to classify text into positive, negative, and neutral labels. A machine learning model is needed to perform sentiment analysis accurately. A machine learning model called the two-level ensemble model performs better than individual ensemble models. This paper aims to find the best performance of two-level ensemble models with different compositions. The models consist of traditional models and ensemble models. By combining traditional models and ensemble models, six ensemble models can be formed: an ensemble of traditional models, a combination of tree-based models, a combination of two-level bagging-based and boosting-based ensemble models, a two-level bagging-based, a boosting-based ensemble model, and a complete two-level ensemble model. Accuracy, precision, recall, and F1-Score are considered metrics for the evaluation process. The experimental results show that the complete two-level ensemble model has the best overall performance, with an average accuracy of 83.5%. In contrast, the two-level bagging-based ensemble performs best for specific labels in 50% of cases. This is due to the fact that combining all models allows one model’s strengths to compensate for another model’s flaws. In our situation, the strength of the bagging-based models in the training set and the tree-based models in the testing set is absorbed by the complete two-level ensemble model.