Hybrid Feature Selection and Ensemble Classifier with Optimization for Sentiment Classification
C Vanlalnunpuia, Lalhmingliana · 2023
This paper presents a Hybrid Feature Selection Technique and an Ensemble Classifier with Optimization approach for Sentiment Classification in IMDB dataset. Firstly, Recursive Feature Elimination with Cross-Validation (RFECV) is employed using two classifiers: Random Forest and Logistic Regression. RFECV allows us to identify the most informative features by iteratively eliminating less relevant ones. Subsequently, the selected features from both classifiers are merged using Union set operation and further filtered based on mutual information scores to produce final feature set. The final feature set is utilized to train an ensemble model using a Voting Classifier which combines the strengths of Random Forest and Logistic Regression. The ensemble model is optimized using Bayesian Optimization, which automatically tunes hyperparameters to achieve the best performance. Experimental results showed that the Ensemble Model trained on the selected features achieves significant improvements in classification accuracy. The classification report reveals high precision, recall, and F1-score values across sentiment classes.