Enhancing Support Vector Machines Accuracy Through Firefly Algorithm-Driven Feature Optimization for Forest-Fire Sentiment Classification

Abstract

Forest fires are a growing global environmental issue that significantly impact ecosystems, human health, and exacerbate climate change. Data from 2023 indicate that approximately 11.91 million hectares of forest were lost due to fires, highlighting the urgent need for technological approaches in addressing this issue. One relevant approach is public sentiment analysis based on the social media platform X, which enables the rapid and real-time capture of public perception. In text-based sentiment analysis, key challenges include high-dimensional feature space and class imbalance, which can degrade the performance of machine learning algorithms. Therefore, this study applies feature selection methods based on swarm intelligence, namely Particle Swarm Optimization (PSO) and Firefly Algorithm (FA), to enhance classification efficiency and accuracy. The models evaluated include Support Vector Machine (SVM), Naïve Bayes (NB), and K-Nearest Neighbors (KNN), using the Knowledge Discovery in Databases (KDD) approach, which involves selection, preprocessing, transformation with Term Frequency-Inverse Document Frequency (TF-IDF) and Synthetic Minority Oversampling Technique (SMOTE), and data splitting with an 80:20 ratio. Model performance was evaluated using four metrics: accuracy, precision, recall, and F1-score. The models achieved strong performance on the balanced training data, indicating effective learning after feature selection, with the fastest execution time recorded by FA optimization at 0.0071 seconds. To ensure a fair assessment of generalization, the main conclusions of this study are based on the testing results. On the testing data, SVM with FA achieved the highest accuracy of 96.36% with an execution time of 0.0064 seconds. Overall, swarm intelligence-based feature selection (PSO and FA) enhances the efficiency of conventional classifiers by reducing high-dimensional feature representations and execution time while maintaining strong predictive performance for forest-fire sentiment classification.