<![CDATA[The high dimensionality and noisy characteristics of Indonesian social media text present significant challenges for sentiment classification models. Redundant and irrelevant features may reduce classification efficiency and negatively affect model generalization performance. This study evaluates a hybrid wrapper-based feature selection approach that integrates Genetic Algorithm (GA) and Support Vector Machine (SVM) to optimize TF-IDF feature representations before classification using Long Short-Term Memory (LSTM). The experiments were conducted on 1,918 Indonesian Twitter comments related to SARS-CoV-2 sentiment, consisting of 1,044 negative and 874 positive labels. The proposed GA-SVM mechanism reduced the feature space from 35,343 to 17,931 selected features. Two evaluation scenarios were employed in this study. Under the hold-out train-test split evaluation, the GA-SVM+LSTM model achieved the best accuracy of 91.41% using a learning rate of 0.0001. Meanwhile, the 10-fold cross-validation evaluation produced an average accuracy of 89.41%, indicating stable generalization performance across different data partitions. The experimental results also show that the proposed feature selection approach improved computational efficiency by reducing training time from 263.64 seconds to 173.54 seconds. Overall, the findings indicate that hybrid GA-SVM feature selection can effectively improve TF-IDF-based sentiment classification performance for Indonesian social media text.]]>
