Yoge-2004/expense-intelligence-model

Expense Intelligence Model

Production-ready financial transaction classification and auxiliary intelligence suite for the Expense Tracker System.

Model Overview

• Architecture: Dual-Gram TF-IDF Vectorizer (Word 1-2 grams + Char-WB 3-5 grams, 200,000 max features) coupled with Calibrated Multiclass SAGA Logistic Regression.
• Inference Speed: Under 5ms per transaction on standard CPU.
• Accuracy: 99.30% overall test accuracy (99.32% Macro-F1) across 240,000+ holdout transactions.
• India Holdout: 98.85% accuracy (98.96% Macro-F1) on Indian banking and UPI transactions.

Supported Canonical Categories

1. food_dining
2. transportation
3. shopping_retail
4. entertainment_recreation
5. healthcare_medical
6. utilities_services
7. financial_services
8. income
9. government_legal
10. charity_donations

Auxiliary Artifacts Included

• models/category-tfidf/: Primary category classifier
• models/merchant-similarity/: Normalized merchant similarity search index
• models/duplicate-similarity/: Near-duplicate transaction detection model
• reports/: Complete validation and test metrics across countries (India, USA, UK, Canada, Australia)
• manifest.json: Full training provenance, environment, and quality gate scores