Optimizing social issue sentiment analysis with hybrid Chi-square and bayesian-optimized binary coordinate ascent

Abstract

Feature selection aims to reduce the dimensionality of the feature space and prevent overfitting. However, when striving to produce accurate models for sentiment classification, feature selection introduces several challenges, particularly concerning textual content. Consequently, many researchers are exploring hybrid feature selection methods to customize the selection process and develop more advanced automated techniques, recognizing that the performance of these methods depends on hyperparameters. Integrating Bayesian Optimization into binary coordinate ascent (BCA) enhances the search for optimal solutions and improves classification performance in sentiment analysis, explicitly focusing on classifying abortion sentiment using Naïve Bayes. The effectiveness of combining Chi2 feature selection with the hybridized BCA and Bayesian Optimization approach is tested across multiple n-gram configurations. Results demonstrate significant improvements in accuracy and recall compared to Chi2 and BCA hybrid methods. For instance, the Bayesian Optimization-enhanced approach achieved up to 93.80% accuracy (1-gram) and 100% recall (4-gram), outperforming the baseline method. The study highlights trade-offs between computational efficiency and performance, noting that while the Chi2 and BCA hybrid method has lower training time complexity, the Bayesian Optimization-enhanced method excels in accuracy and recall during testing. The findings suggest that integrating Bayesian Optimization into feature selection improves sentiment classification performance and recommend further exploration of this approach with other classification algorithms, especially for social issues like abortion sentiment analysis.