Fraud detection solution using AI and machine learning

The goal of this project is to develop a robust fraud detection system that can accurately identify fraudulent transactions . The system leverages machine learning and deep learning techniques to achieve high performance, especially in the context of imbalanced datasets typical in fraud detection scenarios.

Objectives

• To address the challenges posed by the significant class imbalance between fraudulent and non-fraudulent transactions through advanced data preprocessing, model selection, and threshold tuning techniques.

• To experiment with and optimize various machine learning models, including Random Forest, Gradient Boosting, and Deep Learning, to achieve the highest possible predictive performance (ROC AUC).

• To experiment with different classification thresholds to find the optimal balance between precision (minimizing false positives) and recall (maximizing fraud detection).

Dataset

Source: Kaggle - IEEE-CIS Fraud Detection Dataset

Files Used:

• train_transaction.csv: Transactional data including features like TransactionDT, TransactionAmt, and ProductCD.
• train_identity.csv: Identity data associated with transactions, including features like DeviceType and DeviceInfo.
• test_transaction.csv and test_identity.csv: Corresponding test datasets.

Target Variable: isFraud (binary indicator: 1 for fraud, 0 for non-fraud)

Data Exploration

Label Distribution

• The distribution is highly imbalanced, with a significantly larger number of non-fraud transactions (represented by "0" on the x-axis) compared to fraud transactions (represented by "1"). This imbalance is a common challenge in fraud detection datasets.

• The bar corresponding to non-fraud transactions is much taller, visually emphasizing its dominance in the dataset.

• F1-score are more suitable for evaluating the model's performance on the minority class.

Model Performance

Logistic Regression

Class	Precision	Recall	F1-Score	Support
0	0.97	1.00	0.99	113866
1	0.85	0.29	0.44	4242
Accuracy			0.97	118108
Macro Avg	0.91	0.65	0.71	118108
Weighted Avg	0.97	0.97	0.97	118108

Logistic Regression ROC AUC: 0.8742852391082819

Random Forest

Class	Precision	Recall	F1-Score	Support
0	0.98	1.00	0.99	113866
1	0.95	0.41	0.57	4242
Accuracy			0.98	118108
Macro Avg	0.97	0.70	0.78	118108
Weighted Avg	0.98	0.98	0.97	118108

Random Forest ROC AUC: 0.9309374755522329

xgboost

Class	Precision	Recall	F1-Score	Support
0	0.98	1.00	0.99	113866
1	0.77	0.46	0.57	4242
Accuracy			0.98	118108
Macro Avg	0.88	0.73	0.78	118108
Weighted Avg	0.97	0.98	0.97	118108

XGBoost ROC AUC: 0.9077410697552437

Deep Learning Model

Class	Precision	Recall	F1-Score	Support
0	0.98	1.00	0.99	113866
1	0.84	0.52	0.64	4242
Accuracy			0.98	118108
Macro Avg	0.91	0.76	0.82	118108
Weighted Avg	0.98	0.98	0.98	118108

Deep Learning Model ROC AUC: 0.9246724613055639

Conclusion

The deep learning model achieved an AUC of 0.92 and F1-score of 0.82 , demonstrating strong performance in detecting fraudulent transactions.