<![CDATA[Implicit sexual harassment on Indonesian social media frequently evades conventional keyword-based moderation due to linguistic complexity, such as slang and sarcasm. This study evaluates the trade-off between the effectiveness and efficiency of Random Forest and IndoBERT for detecting such content. Using a purposive sampling strategy to select high-risk Instagram and X (Twitter) accounts, we developed a manually labeled "Gold Standard" dataset, validated through inter-annotator agreement, to ensure high-quality ground truth. Experiments were conducted on an NVIDIA RTX 5050 environment to simulate real-world constraints, measuring F1-Score, Recall, and inference latency. The results demonstrate that IndoBERT significantly outperforms Random Forest, particularly on the imbalanced Instagram dataset, achieving an F1-Score of 0.7667 compared to Random Forest’s 0.5924. Statistical verification using the Wilcoxon Signed-Rank Test confirms this superiority. While Random Forest is computationally lighter, requiring only 0.346 ± 0.016 seconds per training fold, it failed to detect subtle harassment, evidenced by a critically low Recall of 51.67% on the imbalanced Instagram dataset. Conversely, although IndoBERT requires a 345x higher training cost (119.28 ± 1.60 seconds), it achieved a rapid inference latency of 0.0239 ± 0.0008 seconds on GPU, proving its viability for real-time systems. Ultimately, this study's primary contribution lies in proving that IndoBERT significantly outperforms traditional Random Forest algorithms, thereby offering a major methodological step forward for Natural Language Processing in Indonesian. This ensures the reliable generalization of context-aware automated content moderation, delivering significant practical impacts for the field of Informatics and real-world social media safety.]]>
