Tax-Deductible Transaction Classifier

This project classifies bank transactions as either tax-deductible or non-deductible using a hybrid system that combines rule-based logic and a machine learning model.

---

Live Demo: https://taxclassifier.streamlit.app/

---

Project Structure

├── app.py                # Streamlit web app interface
├── classifier.py         # ML prediction logic
├── data_loader.py        # Preprocessing and CSV loading
├── rules.py              # Rule matching function
├── rules_config.py       # Regex-based tax rules
├── train_model.py        # Training pipeline for ML model
├── requirements.txt      # Python dependencies

---

How It Works

Hybrid Classification Logic

1. Rule-Based Classifier: Uses regex patterns (e.g., business travel, meals, equipment) to assign labels and explanations.
2. ML Classifier: Falls back to a trained logistic regression model when no rule matches, using TF-IDF on text and one-hot encoding on merchant name.

The model outputs a clear explanation for every prediction.

---

Input Format

The input CSV must contain:

• date – e.g., 2024-05-10
• amount – e.g., 125.75
• merchant – e.g., Delta Airlines
• description – e.g., Flight to NYC for business conference

---

Output Format

The output is available in both CSV and JSON formats. Each record contains:

• date
• merchant
• description
• deductible – true or false
• reason – explanation (e.g., "Business travel", or "ML (p=0.84)")

---

How to Run

1. Install dependencies

pip install -r requirements.txt

2. (Optional) Retrain model

python train_model.py --input data/sample_transactions-2.csv --output models/tax_deductible_clf.joblib

3. Launch web app

streamlit run app.py

You can upload your own transaction CSV or load the sample.

---

Features

• ✅ Explainable rule-based deductions
• ✅ ML fallback with threshold tuning
• ✅ Streamlit UI with file upload and downloads
• ✅ JSON + CSV export support
• ✅ Modular and extensible codebase

---

Assumptions

• Transactions that don’t match any rule default to ML classification.
• The threshold for ML confidence is set to 0.5 by default (user-adjustable).
• Rule matching is prioritized by order — first match wins.

---

Sample JSON Output

[
  {
    "date": "2025-07-25",
    "merchant": "Apple",
    "description": "MacBook purchase for work",
    "deductible": true,
    "reason": "Business equipment purchase"
  }
]

---

Author

Arnav Gupta
AI/ML Internship Candidate
arnavgupta.info

---

source .venv/bin/activate