Detecting Credit Card Fraud As a Data Scientist

Another post starts with you beautiful people!
Hope you have learnt something from my previous post about machine learning classification real world problem
Today we will continue our machine learning hands on journey and we will work on an interesting Credit Card Fraud Detection problem.
The goal of this exercise is to anonymize credit card transactions labeled as fraudulent or genuine.
For your own practice you can download the dataset from here- Download the dataset!

About the dataset: The datasets contains transactions made by credit cards in September 2013 by european cardholders. This dataset presents transactions that occurred in two days, where we have 492 frauds out of 284,807 transactions. The dataset is highly unbalanced, the positive class (frauds) account for 0.172% of all transactions.

Let's start our analysis with loading the dataset first:-

As per the official documentation- features V1, V2, ... V28 are the principal components obtained with PCA. Feature 'Time' contains the seconds elapsed between each transaction and the first transaction in the dataset, feature 'Amount' is the transaction Amount and feature 'Class' is the response variable which takes value 1 in case of fraud and 0 otherwise.

Let's check our target variable- Class; how much it is balanced?

From the plot it's pretty clear that our target variable is highly unbalanced!

Now to handle such type of highly unbalanced classification problem, we should first test the data with resampling and sampling so that we can compare the result.
But before going for the resampling approach we need to normalize the 'Amount' feature. For normalization we will use a package from sklearn library- StandardScaler-

This operation will normalize the amount as below-

Now we will use traditional UNDER-sampling and we will create a 50/50 ratio for this. This will be done by randomly selecting "x" amount of sample from the majority class, being "x" the total number of records with the minority class-

Let's check number of data points in the minority class-

Next, pick the indices of the normal classes-

Now, randomly select "x" number out of the indices we picked above-

Let's append the 2 indices and prepare the under sample data-

Next, we will split the whole data and under sample data into train and test sets-

Let's check the no. of transactions of both datasets-

Now, our next step is to capture the most fraudulent transactions and since we are dealing with highly unbalanced data instead of using Accuracy and Precision metric we will use Recall here-

Now find out best model using k-fold score-

Let's predict the model for under sample data-

And compute the confusion matrix-

And the result is-

Now apply the model we fitted and test it on the whole data; it will give us-

It's a very decent recall accuracy when applying it to a much larger and skewed dataset!

Now to check if the model is also predicting as a whole correctly and not making many errors, we will use AUC and ROC curve-

Once we plot the above ROC curve, we get the AUC as 0.95.

The above approach we should try for the skewed data also and then build the final model with the whole training dataset and predict the classes in the test. I leave it to you for your own. Remember this step will ensure you that by undersampling the data, our algorithm does a much better job at detecting fraud.

So friends what should you can try more after the above approach? I give you a way- change the classification threshfold, investigate Precision-Recall curve, test the SVM and decision trees and share your experience with me.

Meanwhile Friends! Go chase your dreams, have an awesome day, make every second count and see you later in my next post.