/Credit-Card-Fraud-Detection

Primary LanguageJupyter Notebook

Credit Card Fraud Detection

Introduction

This is the group project of COMP7404B. Original dataset can be found at Kaggle. It contains data about credit card transactions that occurred during a period of two days, with 492 frauds out of 284, 807 transactions. All variables in the dataset are numerical. The data has been transformed using PCA transformation(s) due to privacy reasons.

We use Autoencoder Neural Network + SVM for anomaly detection in credit card transaction data and evaluate its performences with only using logistic regression and svm.

File Structure

.
├── autoencoder.ipynb
├── constrast.ipynb
├── dataset
│   └── creditcard.csv
├── docs
│   ├── autoencoder.html
│   ├── constrast.html
│   ├── Data analysis for credit card fraud detection.pptx
│   └── fraud_detection.html
├── fraud_detection.ipynb
├── model.h5
├── original.png
└── README.md
  • fraud_detection.ipynb: is the main file with Autoencoder + SVM
  • autoencoder.ipynb: try to use Antoencoder to do classification
  • constrast.ipynb: logitsic regression and SVM

To view the result, just download docs/ folder is enough.

Results

Method Precision Recall F-1 Score
AE + SVM 0.84 0.75 0.79
Logistic Regression 0.85 0.61 0.71
SVM 0 0 NA

  1. SVM does not work for this imbalanced dataset.

  2. AE + SVM has nearly the same precision with logistic regression. But the recall of AE + SVM is higher, thus results in a better F-1 score.

  3. AE + SVM performs best among 3 methods.

Autoencoder is able to effectively extract sample features and can be applied to dealing with imbalanced dataset.