Support Vector Machine Algorithm Optimization for Sentiment Analysis using Bayesian Optimization

Abstract

This study examines the effect of Bayesian Optimization in improving the performance, computational efficiency, and sustainability of Aspect-Based Sentiment Analysis models using Support Vector Machine (SVM). A dataset consisting of 988 customer reviews about Borobudur Temple, classified into six dimensions: Attractiveness, Facilities, Accessibility, Visual Image, Price, and Human Resources is used to compare two scenarios, namely Baseline SVM and SVM enhanced with Bayesian Optimization (BO). Important metrics used include accuracy, computational duration, energy usage, and carbon emissions. The results show that BO significantly improves accuracy, especially on difficult aspects such as Facilities (from 0.7294 to 0.8682) and Price (from 0.8047 to 0.9576). The most complicated aspect, namely visual image due to the very minimal number of datasets (unbalanced), achieved an increase in accuracy from 0.6729 to 0.72. In addition, BO reduces training time, especially for resource-intensive tasks such as the visual image aspect, reducing training time from 13.04 seconds to 9.4 seconds. Substantial reductions in energy consumption and CO₂ emissions are seen in line with sustainable machine learning principles. The hyperparameter adaptability of SVM, with linear kernels performing well in simpler tasks, while polynomial and sigmoid kernels improve performance for more complex parts. BO substantially alleviates the limitations of Baseline SVM, offering a robust, efficient, and environmentally friendly solution for ABSA. Future research can explore more enhancements for complex tasks to improve performance and efficiency.