<![CDATA[Indonesia had 235.26 million internet users and 81.72% internet penetration in 2026, while hate speech threatens social cohesion. Based on the literature reviewed in this study, no prior study has been found that systematically conducted a three-way benchmark of MNB, CNB, and SVM with preprocessing ablation, McNemar testing, and Wilson CI error analysis in Indonesian hate speech detection, using a dataset of 13,169 tweets, with per-algorithm imbalanced-class handling (SMOTE for MNB/CNB; CSL and Lexicon-Based Coefficient Override for SVM), and ablation of six preprocessing configurations. MNB achieves F1 Macro 72.10% and highest Abusive Recall (84.30%), while SVM achieves Abusive Recall 80.20% with inference latency 0.10 ms on CPU; a direct comparison with IndoBERT on GPU T4 is not fully equivalent across hardware, though an additional CPU-only measurement shows IndoBERT at ~102.4 ms (~1,020× ratio). McNemar tests confirm SVM differs significantly from MNB/CNB (p<0.0001), CNB vs MNB non-significant (p=0.0704), indicating a negative SMOTE×CNB interaction specific to this study’s dataset and configuration. SVM is selected as the deployment model based on Utility Score (81.39) under the weighting scenarios tested, and may serve as a first-pass screening aid (not a substitute for human verification) for platforms operating under Indonesian ITE Law No. 1/2024.]]>
Copyrights © 2026
