Intelligent Expense Categorization System (QuickBooks‑Style)

Project Duration: Feb 2026 Domain: Machine Learning, NLP, FinTech Automation Tech Stack: Python, Scikit‑learn, XGBoost, Pandas, NumPy, Flask, Jupyter Notebook

---

📌 Project Overview

Modern financial platforms like QuickBooks automatically categorize expenses (e.g., Travel, Food, Utilities, Office Supplies) from raw transaction data. This project replicates that real‑world system using Natural Language Processing (NLP) and structured numerical features, building a production‑ready ML pipeline for intelligent expense categorization.

The system takes raw transaction descriptions and metadata as input, predicts the most likely expense category, and assigns a confidence score. Low‑confidence predictions are flagged for human review, ensuring reliability and user trust — a core requirement in FinTech products.

---

🎯 Key Objectives

• Automate expense categorization at scale
• Combine text‑based NLP features with numerical transaction signals
• Handle uncertainty using confidence‑based decision logic
• Design the solution with production deployment in mind

---

🧠 Problem Statement

Manual expense categorization is:

• Time‑consuming
• Error‑prone
• Not scalable for large transaction volumes

The goal is to build a machine learning system that:

1. Accurately classifies expenses into predefined categories
2. Explains prediction confidence
3. Falls back to human intervention when the model is uncertain

---

📊 Dataset Description

The project uses a synthetic but realistic QuickBooks-style expense dataset designed to closely mirror real-world financial transaction data used in accounting platforms.

Dataset File: quickbooks_synthetic_expenses_7k_diverse.csv

🔢 Dataset Size

• ~7,000 transaction records
• Multi-class expense categorization problem

🧾 Columns Overview

Text Feature

• description – Raw transaction description containing merchant names, payment context, and notes (primary NLP signal)

Numerical / Structured Features

• amount – Transaction amount
• payment_mode – Mode of payment (card, UPI, cash, bank transfer, etc.)
• merchant_type – Merchant category signal
• is_recurring – Indicates recurring transactions
• day_of_week – Temporal pattern feature
• month – Seasonal spending signal

Target Variable

• expense_category – Ground truth expense class (e.g., Travel, Food, Utilities, Office Supplies, Entertainment, Healthcare, etc.)

🎯 Why This Dataset Works Well

• Mimics noisy, short-text financial descriptions
• Contains overlapping categories (realistic ambiguity)
• Supports confidence-based decision making
• Suitable for NLP + structured feature fusion

The dataset is explored, cleaned, and modeled end-to-end inside the Jupyter Notebook.

---

🔧 Feature Engineering

1️⃣ Text Processing (NLP)

• Lowercasing & normalization
• Stopword removal
• TF‑IDF vectorization to convert transaction descriptions into numerical vectors

2️⃣ Numerical Feature Handling

• Scaling and normalization
• Handling missing values
• Feature alignment between training and inference

3️⃣ Feature Fusion

• Combined TF‑IDF vectors with numerical features into a single feature space

---

🤖 Model Architecture

• Model Used: XGBoost (Multi‑Class Classifier)

• Why XGBoost?

  • Handles mixed feature types well
  • Strong performance on structured + sparse data
  • Robust to noise and feature interactions

Training Pipeline

1. Train‑validation split
2. Hyperparameter tuning
3. Multi‑class classification training
4. Performance evaluation

---

📈 Model Evaluation

Evaluation metrics used:

• Accuracy
• Precision / Recall / F1‑Score (per class)
• Confusion Matrix

Model performance analysis is documented step‑by‑step in the Jupyter Notebook.

---

✅ Confidence Scoring & Human‑in‑the‑Loop

To make the system production‑ready:

• Model outputs class probabilities
• Maximum probability is treated as confidence score

Decision Logic

• High confidence → Auto‑categorize
• Low confidence → Flag for manual review

This approach:

• Reduces incorrect auto‑classifications
• Improves long‑term trust
• Enables continuous feedback and retraining

---

🚀 Inference Service (Production‑Ready)

A Flask‑based inference service is implemented to:

• Accept transaction input via API

• Apply the same preprocessing pipeline

• Return:

  • Predicted category
  • Confidence score
  • Review flag (True / False)

This mirrors real‑world ML deployment patterns used in FinTech companies.

---

📁 Project Structure

project-root/
│
├── notebook/
│   └── Expense_Categorization.ipynb
│
├── app.py                 # Flask inference service
├── model.pkl              # Trained XGBoost model
├── vectorizer.pkl         # TF-IDF vectorizer
├── requirements.txt
├── README.md
└── data/
    └── transactions.csv

---

▶️ How to Run the Project

1️⃣ Install Dependencies

pip install -r requirements.txt

2️⃣ Run Jupyter Notebook

jupyter notebook

Open Expense_Categorization.ipynb to explore training and evaluation.

3️⃣ Start Inference API

python app.py

---

🧪 Sample Prediction Output

{
  "predicted_category": "Travel",
  "confidence": 0.87,
  "manual_review_required": false
}

---

💡 Real‑World Impact

This project closely aligns with:

• QuickBooks expense automation
• Invoice & transaction intelligence systems
• Enterprise‑grade ML pipelines

It demonstrates skills in:

• End‑to‑end ML system design
• NLP + structured data modeling
• Deployment‑ready ML thinking

---

🔮 Future Improvements

• Deep learning models (BERT‑based text encoders)
• Online learning with feedback loops
• Explainability using SHAP values
• Multi‑language transaction support

---

👤 Author

Utsav Kashyap Data Scientist Machine Learning | Backend | NLP

---

⭐ If you find this project useful, feel free to star the repository!