Dataset
This dataset, card_transdata.csv, is a public Kaggle dataset of 1,000,000 card transactions with 8 features (such as distance from home, distance from last transaction, ratio to median purchase price, and whether a chip or PIN was used). It contains 87,403 fraudulent and 912,597 legitimate transactions.
Python Code
AWS credentials are read from environment variables and are not included in this repository.
Purpose
This project aims to build a fraud detection pipeline that runs on cloud infrastructure at the scale of a million transactions, using Spark for processing and machine learning and Athena for SQL queries.
Architecture
- Stored the raw data in Amazon S3.
- Processed it with PySpark on a multi-node AWS EMR cluster.
- Engineered a distance ratio feature and assembled features with VectorAssembler.
- Wrote the transformed data back to S3 and created Athena tables with boto3 for SQL analysis.
- Trained a Spark MLlib Random Forest with 5-fold cross-validation, with and without class weighting to handle the 9% fraud rate.
The second diagram shows a proposed extension with scheduled runs; it was designed but not built.
Results
| Model | AUC | Accuracy | Fraud Precision | Fraud Recall | Fraud F1 |
|---|---|---|---|---|---|
| Random Forest | 0.990 | 0.987 | 1.00 | 0.86 | 0.92 |
| Random Forest, class-weighted | 0.995 | 0.981 | 0.83 | 0.98 | 0.90 |
Class weighting raised fraud recall from 86% to 98%, catching far more fraud at the cost of more false alarms. For fraud detection, missing fraud is usually the costlier error, so the weighted model is the better fit.
Notes
These metrics were computed on the full dataset rather than a separate held-out test set, so they likely overstate real-world performance. A held-out split is the next improvement.
Course: CSC 555 Mining Big Data, DePaul University (Fall 2024).