Comparative Performance Analysis of SVM, LSTM, and IndoBERT for Sentiment Analysis of User Reviews

Abstract

The Maganghub Kemnaker National Internship Program, launched on October 1, 2025, attracted significant public attention while also generating both criticism and praise on the social media platform X. This study compares the performance of three classification algorithms—Support Vector Machine (SVM), Long Short-Term Memory (LSTM), and IndoBERT—for sentiment analysis of public opinions regarding the program. The dataset consists of 6,825 Indonesian-language tweets collected between October and December 2025 after duplicate removal. The data were preprocessed and labeled as positive, neutral, or negative using two approaches: automated labeling with a Hugging Face model and manual annotation. Each algorithm was evaluated under five hyperparameter tuning scenarios using an 80:20 train–test split and assessed with accuracy, precision, recall, and F1-score. The results show that IndoBERT achieved the highest performance, with an accuracy of 85.27% using the Hugging Face–labeled dataset, outperforming SVM (79.05%) and LSTM (78.68%). IndoBERT's superior performance is attributed to its Transformer-based architecture and multi-head self-attention mechanism, which effectively capture bidirectional contextual information, including sarcastic expressions commonly found in social media posts. The findings were further implemented in a web-based dashboard to support interactive public opinion monitoring for policymakers.