Fraud Detection from Customer Transactions

This project focuses on detecting fraudulent transactions in e-commerce data provided by Vesta Corporation. The goal is to build a robust machine learning model to classify transactions as fraudulent or non-fraudulent accurately.

Key Challenges

• Highly Imbalanced Dataset: The dataset is severely imbalanced, with only a tiny fraction of fraudulent transactions (3% of dataset). This requires special handling to ensure the model does not become biased towards the majority class.
• High Dimensionality: The dataset contains 393 features, many of which are masked or have missing values, adding complexity to the feature engineering process.
• Temporal Nature: The data is time-sensitive, requiring time-based splits for training and evaluation to avoid data leakage.

Project Structure

Data Exploration and Preprocessing

1. Handling missing values by imputing numerical columns with their mean and categorical columns with 'Unknown'.
2. Feature engineering to create time-based features, aggregated features, and interaction features.
3. Standardization and dimensionality reduction using PCA.
4. Hierarchical clustering to group similar features and reduce redundancy.

Feature Selection

1. Baseline Model: A LightGBM model is trained as a baseline, with feature importance analysis to identify key predictors.
2. Feature Selection: Correlation matrix analysis and Lasso regression are used to select the most relevant features.
3. Undersampling: The majority class is undersampled to balance the dataset, preserving the temporal order of transactions.

Modeling and Evaluation

1. CatBoost with Optuna: Hyperparameter tuning is performed using Optuna to optimize the CatBoost model.
2. The model is evaluated using time-based cross-validation to ensure robustness.
3. The final model is evaluated using ROC-AUC and Precision-Recall AUC scores.
4. A confusion matrix is generated to visualize the model's performance in classifying fraudulent and non-fraudulent transactions.
5. Feature importance is analyzed to understand the key drivers of the model's predictions.

Results

	Precision	Recall	F1-Score	Support
0	0.95	1.00	0.97	37254
1	0.93	0.51	0.66	4072
Accuracy			0.95	41326
Macro Avg.	0.94	0.75	0.82	41326
Weighted Avg.	0.95	0.95	0.94	41326

• The final CatBoost model achieved an ROC-AUC score of 0.935 on the test set, indicating strong performance distinguishing between fraudulent and non-fraudulent transactions.

Key Takeaways

• Handling Imbalanced Data: Undersampling the majority class helped improve the model's ability to detect fraudulent transactions without introducing significant bias.
• Feature Engineering: Time-based interaction features were crucial in improving model performance.
• Hyperparameter Tuning: Optuna was instrumental in finding the optimal hyperparameters for the CatBoost model, leading to improved performance.

👩‍💻 Author

📌 Developed by Sevilay Munire Girgin
📧 Contact Me
🌐 Portfolio