Project Overview:

• This Project focuses on building a Fraud Detection using highly unbalanced Dataset from Kaggle using different machine learning models and choose the one with the highest F1 score for positive Class(fraud)
• Using Logistic Regression ,RandomForest and Neural Networks with preprocessing on Training Dataset and hyperparameters Tunning

Key Challenges:

• Extreme Class imbalanced: Dataset has Few Fraud Transaction (1<<%)

• Features Anonymous: High dimension Features and can't use domain knowledge for feature engineering or selection

• Model Overfitting: Try to ensure that the model generalize well on unseen data

Dependencies:

• Install Miniconda/Conda

  • Download and install Miniconda (lightweight) or Anaconda (full bundle)

• Clone this repository

      git clone https://github.com/SherineTarek224/Credit_Card_Fraud_Detection.git
      cd Fraud_Detection
  

• conda create environment

      conda activate Fraud 
  

• Verify Installation

        conda list
  

Datasets:

• You can download the dataset from Google Drive. Place the extracted files inside the Dataset/ folder.

Repo Structure:

Fraud_Detection/

 ├── Dataset/
 │    ├── train.zip
 │    ├── val.zip
 │    ├── test.zip
 │    └── trainval.zip
 ├── Imagegs/
     ├── SMOTE(0.05)
     ├── Under_Sampling(0.05)
 ├── .gitignore
 ├── data_helper.py
 ├── evaluation.py
 ├── Models_Comparison.py
 ├── Test_chosen_model.py
 ├── train_model.py
 ├── fraud_eda.ipynb
 ├── environment.yml
 └── README.md

• README.md:Contains information about this project and important details
• environment.yml:contain configuration of conda environment and library names and its version
• EDA:Data analysis using jupyter notebook to explore features of dataset
• data_helper:load_data ,detect_outliers,choose the preprocessor(MinMax,Std,Robust)and balancing_data using(SMOTE,Under_Sampling)
• train_model:raining different models and evaluate them
• evaluation:evaluate models with different metrics(Recall,Percision,Recall ,AUC)
• Models_Comparison:Compare models according different metrics and export heatmap comparison
• Test_Chosen_Models:after save models.pkl now can choose which one of them to test them on unseen data

Models_Implemented:

• [1] Logistic Regression:

  • A simple and easy model.
  • Tuned using L2 regularization.
  • Hyperparameter tuning performed using Grid Search.
  • Handles class imbalance via weighted class adjustments.

• [2] Random_Forest:

  • An ensemble-based model.

  • Tuned using:

    • Number of trees (n_estimators)

    • Maximum tree depth (max_depth)

  • Class weights to handle imbalance.

  • RandomizedSearchCV was used to Tune hyperparameters.

• [2] Neural_Nets:

  • A Simple Neural Nets Structure
  • 2 hidden_layer_sizes
  • Used activation Function for none_linear transformation
  • Randomized_SearchCV was used to Tune hyperparameters

Results:

• Models_Performance Comparison (SMOTE(0.05) and optimal_threshold and its f1_score calculated on Validation_Data_only)

• SMOTE(0.05)

  • Validation_DATA
    

  • TEST_DATA

  

• UnderSampling(0.05)

  • Validation_DATA
    

  • TEST_DATA