<![CDATA[Spam detection for Indonesian short messages such as SMS and email remains challenging due to lexical variation, character obfuscation, and class imbalance. This study provides a systematic evaluation to determine the most balanced configuration between accuracy and efficiency for Indonesian spam filtering. We compare two pretrained backbones (IndoBERT and XLM RoBERTa), along with representation strategies (truncation versus chunking), summarization schemes (pooling), and feature fusion approaches. The system follows a feature based design with an emphasis on simplicity, and is assessed using F1 Macro, spam class recall, AUPRC (Area Under the Precision Recall Curve), and efficiency metrics in terms of embedding build time and training latency. Results indicate that IndoBERT achieves superior binary classification performance with high efficiency, while XLM RoBERTa slightly outperforms on AUPRC, making it more suitable for risk ranking scenarios. Truncation combined with mean pooling consistently yields stable results. Although late fusion only provides marginal improvements, it remains relevant as it highlights the potential of domain specific signals to enhance robustness under heavy obfuscation. The final recommendation for production is IndoBERT with truncation, mean pooling, and embedding only. Limitations include the focus on short messages and the lack of evaluation under extreme obfuscation. Future work should explore character level augmentation, cross domain evaluation, and cost sensitive threshold tuning.]]>
Copyrights © 2025
