DistilBERT with Temporal Features Improves SMS Spam Detection
November 13, 2026
A deployed DistilBERT model using inter-message temporal features achieves 0.8238 accuracy and 0.9128 recall on a 20,000-message production evaluation set. This outperforms a text-only DeBERTa-v3-base classifier which reached an AUC-ROC of 0.9011 on a separate test set.
HOW THIS AFFECTS YOU
●
builderAdding time-based features to your NLP models can help capture behavioral spam patterns.