<![CDATA[Sentiment analysis of public policy on social media faces significant challenges due to linguistic heterogeneity across platforms and limitations of single models in capturing the diversity of opinion expressions. Previous studies tend to employ single-platform and single-model approaches that potentially generate representational bias and accuracy degradation of up to 12–15% when applied to different platform contexts. This study aims to develop a soft voting-based ensemble model that integrates Support Vector Machine (SVM) and IndoBERT to analyze public sentiment toward the Free Nutritious Meal (MBG) Program across multiple platforms, and to evaluate the effectiveness of the ensemble approach compared to single models in addressing variations in linguistic characteristics of digital platforms. The research dataset consists of 7,500 comments from X, TikTok, and YouTube collected from January 6 to September 28, 2025, processed through informal Indonesian language preprocessing, lexicon-based labeling, and stratified split division. Results demonstrate that SVM performs optimally on TikTok (accuracy 98.1%, macro F1 98.0%) but weakly on the neutral class in X (F1 51.0%), while IndoBERT excels in handling pragmatic ambiguity in X (neutral F1 74.0%) despite slightly declining on TikTok (macro F1 93.0%). The ensemble model produces the most balanced performance with accuracies of 92.53% (YouTube), 95.73% (TikTok), 92.53% (X), and macro F1 scores of 85.07%, 94.33%, 84.92% respectively. The contributions of this research include the development of a multi-platform sentiment analysis approach that addresses single-platform bias, improved classification generalization capability across heterogeneous digital ecosystems, and provision of evidence-based evaluation instruments for improving government policy implementation and communication.]]>
