Credit Card Fraud Detection

• This repository is for the Kaggle's Credit Card Fraud Detection to identify fraudulent credit card transactions.
• The datasets contain transactions made by credit cards in September 2013 by European cardholders.
• We did Exploratory Data Analysis, also did Data Visualization and Scaled the imbalance data because there was less number of fraudulent data than the normal one.
• For handling imbalance data we evaluated the final performance of the classification models with SMOTE and random undersampling method.
• In the model building, we used Random Forest Classifier, Logistic Regression, Decision Tree Classifier, Support Vector Machine (SVM), K -Nearest Neighbors, Decision Tree and Gaussian Naive Bayes.
• Also, we made the Classification Report, defined Confusion Matrix and plotted ROC Curve.

Resources We Used

• Jupyter Notebook.
• We used pandas, numpy, matplotlib, seaborn, imblearn (for imbalance data) and sklearn packages.
• Dataset link: https://www.kaggle.com/mlg-ulb/creditcardfraud

Exploratory Data Analysis & Data Visualization

• We analyzed this dataset to better understand.
• Did Data Visualization in different ways with matplotlib and seaborn libraries.
• Also scaled this dataset so it can be usable for our model.

Model Building

• First, we defined the features of our scaled data then split it into test and train set with 20% test set.
• Then we used two methods to resample our data, for Oversampling we used SMOTE and for Undersampling we used Random Undersampling method.
• After scaling here we used different Models to evaluate and evaluated on the basis of recall and precision scores. (Added Accuracy and F1-Score just to show)
• With this, we made the Classification Report and defined the Confusion Matrix for evaluation.

SMOTE Method

It is a type of oversampling but in this we will make the synthetic example of minority data and will give as a balanced data.

• In this method our Logistic Regression model has the highest Recall Score (0.928) compared to other models with poor Precision Score (0.05) from others.
• In terms of maintaining good Recall Score (0.84), Precision Score (0.91) the Random Forest Model performs better than Logistic Regression.

Model Used	Recall	Precision	Accuracy	F1-Score
Random Forest Classifier	0.84	0.91	0.99	0.87
Decision Tree Classifier	0.76	0.45	0.99	0.57
Logistic Regression Classifier	0.92	0.05	0.97	0.11
Support Vector Classifier (SVC)	0.90	0.09	0.98	0.16
K-Nearest Neighbors Classifier	0.86	0.48	0.99	0.61
Gaussian Naive Bayes	0.86	0.06	0.97	0.11

Undersampling Method

It means taking the less number of majority class, in our case taking less number of normal transactions so that our new data will be balanced.

• Random Forest performs well with excellent Recall (0.92), Precision Score (0.08) respectively.
• Followed by Logistic Regression with excellent Recall (0.92), Precision Score (0.05).

Model Used	Recall	Precision	Accuracy	F1-Score
Random Forest Classifier	0.92	0.08	0.98	0.14
Logistic Regression Classifier	0.92	0.05	0.97	0.10
K-Nearest Neighbors Classifier	0.90	0.07	0.97	0.13
Support Vector Classifier (SVC)	0.90	0.06	0.97	0.11
Decision Tree Classifier	0.90	0.01	0.91	0.03
Gaussian Naive Bayes	0.86	0.04	0.96	0.08