Hadhari (حذارِ) - Arabic Spam Detection
Hadhari is a machine learning model designed to detect spam messages in Arabic text, targeting unsolicited advertisements.
Model Details
- Algorithm: Linear Support Vector Classification (
LinearSVC) - Feature Extraction: TF-IDF Vectorization (
max_features=2000,ngram_range=(1,3))
Dataset
The model was trained on a human-in-the-loop (HITL) verified dataset collected from a live WhatsApp bot.
- Size: ~1000 unique messages
- Distribution: Balanced between genuine conversations and student-targeted spam.
- Characteristics: Contains real-world Saudi colloquial Arabic, slang, emojis, and common local spam patterns.
Example Data
| Class | Message Example |
|---|---|
| Spam (1) | حل واجـــبات المواد بحــــــــــوثات علمــــــيةمشـــــــــــاريع تـخــــــرج* https://wa.me/... |
| Spam (1) | أعذار طبية(إجـازة مـرضـية) مـعتـمد فـي تطبيق صـحـتي (تقـبل لـجميع جهات العمل) |
Performance Metrics
- Overall Accuracy: 97.0%
- Spam Precision: 99.0%
Usage
Live API endpoint and documentation available at: Hadhari API Space