💳 Fraud Detection with Machine Learning on Google Cloud

📌 Overview

This project was completed as part of a Google Cloud Skills Boost lab.
The objective was to detect fraudulent financial transactions using machine learning on Google Cloud Platform (GCP) with both supervised and unsupervised approaches.

The dataset (≈228k transactions, originally from Kaggle) contains transaction type, amount, origin/destination balances, and a fraud label (isFraud).
The challenge is highly imbalanced (fraud <1%), typical of real-world fraud detection.

---

⚙️ Workflow

1. Data Engineering & Exploration
   • Ingested raw CSV into BigQuery via Cloud Storage.
   • Explored data distributions (fraud by type, transaction imbalances).
   • Identified that fraud mainly occurs in TRANSFER and CASH_OUT.
2. Feature Engineering
   • Created new features (origZeroFlag, destZeroFlag, amountError) to capture abnormal balance behaviors.
   • Applied undersampling to reduce class imbalance.
   • Partitioned into train, validate, and test datasets (~20% test holdout).
3. Modeling
   • Unsupervised: Trained k-means clustering for anomaly detection, profiled clusters with high fraud density.
   • Supervised:
     • Logistic Regression (baseline)
     • Boosted Tree Classifier (gradient boosting) → selected champion model.
4. Evaluation
   • Compared models using precision, recall, F1 score, accuracy, ROC/AUC.
   • Boosted Trees outperformed logistic regression with higher AUC and better fraud recall.
   • Fraud detection event rate increased by >95% compared to raw test data.
5. Prediction
   • Deployed champion model in BigQuery ML.
   • Scored test dataset, predicting fraudulent transactions with probability thresholds.

---

📊 Tech Stack

• Cloud Platform: Google Cloud Platform (BigQuery, Cloud Storage, Vertex AI)
• Languages: SQL, Python
• ML Methods: Logistic Regression · Boosted Trees · k-means clustering
• Concepts: Class imbalance handling · Feature engineering · Fraud/Anomaly detection

---

📊 Model Training & Evaluation

During this project, I trained supervised models in BigQuery ML to detect fraudulent transactions.
The evaluation metrics and curves below illustrate model performance:

Evaluation of logistic regression classifier in BigQuery ML

Key Metrics

• Accuracy: 98.2%
• ROC AUC: 0.93 (strong discriminatory power)
• Precision: 97.5% (very low false positives)
• Recall: 39.3% (fraud cases are rare, challenging to capture all)
• F1 Score: 0.56 (balanced trade-off between precision & recall)

Visual Insights

• Precision-Recall by Threshold: shows trade-off when adjusting probability cutoffs.
• Precision-Recall Curve: confirms high precision, recall drops due to class imbalance.

• ROC Curve: area under curve = 0.93, indicating robust fraud detection capability.

📂 Code Snippets

• Data Preparation and feature engineering
• Train an unsupervised model to detect anomalies

---

📝 Skills Gained

• Cloud-native ML workflows with BigQuery ML
• Handling imbalanced datasets in fraud detection
• Unsupervised anomaly detection vs supervised classification
• Model benchmarking with AUC, precision/recall, and confusion matrices