Credit Card Fraud Detection Project – Machine Learning Project

Welcome to our credit card fraud detection project. Today, we’ll use Python and machine learning to detect fraud in a dataset of credit card transactions. Although we have shared the code for every step, it would be best to understand how each step works and then implement it.

Let’s begin!

How does Credit Card Fraud work?

Throughout my experience with the intricate issue of credit card fraud, I’ve observed the complexity and sophistication of these fraudulent schemes. Exploring the intricacies of how credit card fraud operates was crucial in developing the machine learning-based fraud detection project I led. Here’s a step-by-step breakdown, informed by real-life scenarios and the deployment of a credit card fraud detection project:

• Information Theft: The initial step involves the theft of credit card information, which can occur through various means, such as data breaches, phishing scams, or skimming devices. In one case I studied, fraudsters used a skimming device at a retail outlet to clone credit card details of unsuspecting customers. 
• Test Transactions: Once the information is stolen, fraudsters often conduct small transactions to test the validity of the credit card details. This step is crucial; a successful small transaction indicates the card is active and hasn’t been flagged for fraud. 
• Large Unauthorized Purchases: Following successful tests, the fraudsters proceed to make large purchases or withdraw funds, exploiting the credit card before any suspicious activity is detected by the cardholder or the bank. 
• Maximizing the Fraud Window: Fraudsters aim to maximize their window of opportunity by conducting as many transactions as possible before the fraud is detected. This involves possibly using the stolen information across multiple platforms or services. 
• Detection and Reporting: The final step in the fraud cycle involves detection by either the cardholder or the fraud detection systems employed by financial institutions. In our fraud detection using machine learning project, we developed algorithms that continuously learn from transaction patterns, enabling early detection of fraud by identifying anomalies in transaction behaviors. 

This sequential understanding of credit card fraud has been pivotal in our approach to building a machine learning-based credit card fraud detection system. 

Credit Card Fraud Detection Project With Steps

In our credit card fraud detection project, we’ll use Python, one of the most popular programming languages available. Our solution would detect if someone bypasses the security walls of our system and makes an illegitimate transaction.  

The dataset has credit card transactions, and its features are the result of PCA analysis. It has ‘Amount’, ‘Time’, and ‘Class’ features where ‘Amount’ shows the monetary value of every transaction, ‘Time’ shows the seconds elapsed between the first and the respective transaction, and ‘Class’ shows whether a  transaction is legit or not. 

In ‘Class’, value 1 represents a fraud transaction, and value 0 represents a valid transaction. 

You can get the dataset and the entire source code here.

Enrol for the Machine Learning Course from the World’s top Universities. Earn Masters, Executive PGP, or Advanced Certificate Programs to fast-track your career.

Step 1: Import Packages

We’ll start our credit card fraud detection project by installing the required packages. Create a ‘main.py’ file and import these packages:

import numpy as np

import pandas as pd

import sklearn

from scipy.stats import norm

from scipy.stats import multivariate_normal

from sklearn.preprocessing import MinMaxScaler

import matplotlib.pyplot as plt

import seaborn as sns

Step 2: Look for Errors

Before we use the dataset, we should look for any errors and missing values in it. The presence of missing values can cause your model to give faulty results, rendering it inefficient and ineffective. Hence, we’ll read the dataset and look for any missing values:

df = pd.read_csv(‘creditcardfraud/creditcard.csv’)

# missing values

print(“missing values:”, df.isnull().values.any())

We found no missing values in this dataset, so we can proceed to the next step. 

Join the Artificial Intelligence Course online from the World’s top Universities – Masters, Executive Post Graduate Programs, and Advanced Certificate Program in ML & AI to fast-track your career.

Step 3: Visualization

In this step of our credit card fraud detection project, we’ll visualize our data. Visualization helps in understanding what our data shows and reveals any patterns which we might have missed. Let’s create a plot of our dataset:

# plot normal and fraud

count_classes = pd.value_counts(df[‘Class’], sort=True)

count_classes.plot(kind=’bar’, rot=0)

plt.title(“Distributed Transactions”)

plt.xticks(range(2), [‘Normal’, ‘Fraud’])

plt.xlabel(“Class”)

plt.ylabel(“Frequency”)

plt.show()

In our plot, we found that the data is highly imbalanced. This means we can’t use supervised learning algorithms as it will result in overfitting. Furthermore, we haven’t figured out what would be the best method to solve our problem, so we’ll perform more visualisation. Use the following to plot the heatmap: 

# heatmap

sns.heatmap(df.corr(), vmin=-1)

plt.show()

Now, we’ll create data distribution graphs to help us understand where our data came from: 

fig, axs = plt.subplots(6, 5, squeeze=False)

for i, ax in enumerate(axs.flatten()):

   ax.set_facecolor(‘xkcd:charcoal’)

   ax.set_title(df.columns[i])

   sns.distplot(df.iloc[:, i], ax=ax, fit=norm,

                color=”#DC143C”, fit_kws={“color”: “#4e8ef5”})

   ax.set_xlabel(”)

fig.tight_layout(h_pad=-1.5, w_pad=-1.5)

plt.show()

With data distribution graphs, we found that nearly every feature comes from Gaussian distribution except ‘Time’. 

So we’ll use multivariate Gaussian distribution to detect fraud. As only the ‘Time’ feature comes from the bimodal distribution (and note gaussian distribution), we’ll discard it. Moreover, our visualisation revealed that the ‘Time’ feature doesn’t have any extreme values like the others, which is another reason why we’ll discard it. 

Add the following code to drop the features we discussed and scale others: 

classes = df[‘Class’]

df.drop([‘Time’, ‘Class’, ‘Amount’], axis=1, inplace=True)

cols = df.columns.difference([‘Class’])

MMscaller = MinMaxScaler()

df = MMscaller.fit_transform(df)

df = pd.DataFrame(data=df, columns=cols)

df = pd.concat([df, classes], axis=1)

Step 4: Splitting the Dataset

Create a ‘functions.py’ file. Here, we’ll add functions to implement the different stages of our algorithm. However, before we add those functions, let’s split our dataset into two sets, the validation set and the test set. 

import pandas as pd

import numpy as np

def train_validation_splits(df):

   # Fraud Transactions

   fraud = df[df[‘Class’] == 1]

   # Normal Transactions

   normal = df[df[‘Class’] == 0]

   print(‘normal:’, normal.shape[0])

   print(‘fraud:’, fraud.shape[0])

   normal_test_start = int(normal.shape[0] * .2)

   fraud_test_start = int(fraud.shape[0] * .5)

   normal_train_start = normal_test_start * 2

   val_normal = normal[:normal_test_start]

   val_fraud = fraud[:fraud_test_start]

   validation_set = pd.concat([val_normal, val_fraud], axis=0)

   test_normal = normal[normal_test_start:normal_train_start]

   test_fraud = fraud[fraud_test_start:fraud.shape[0]]

   test_set = pd.concat([test_normal, test_fraud], axis=0)

   Xval = validation_set.iloc[:, :-1]

   Yval = validation_set.iloc[:, -1]

   Xtest = test_set.iloc[:, :-1]

   Ytest = test_set.iloc[:, -1]

   train_set = normal[normal_train_start:normal.shape[0]]

   Xtrain = train_set.iloc[:, :-1]

   return Xtrain.to_numpy(), Xtest.to_numpy(), Xval.to_numpy(), Ytest.to_numpy(), Yval.to_numpy()

Step 5: Calculate Mean and Covariance Matrix

The following function will helps us calculate the mean and the covariance matrix:

def estimate_gaussian_params(X):

   “””

   Calculates the mean and the covariance for each feature.

   Arguments:

   X: dataset

   “””

   mu = np.mean(X, axis=0)

   sigma = np.cov(X.T)

   return mu, sigma

FYI: Free nlp course!

Step 6: Add the Final Touches

In our ‘main.py’ file, we’ll import and call the functions we implemented in the previous step for every set:

(Xtrain, Xtest, Xval, Ytest, Yval) = train_validation_splits(df)

(mu, sigma) = estimate_gaussian_params(Xtrain)

# calculate gaussian pdf

p = multivariate_normal.pdf(Xtrain, mu, sigma)

pval = multivariate_normal.pdf(Xval, mu, sigma)

ptest = multivariate_normal.pdf(Xtest, mu, sigma)

Now we have to refer to the epsilon (or the threshold). Usually, it’s best to initialise the threshold with the pdf’s minimum value and increase with every step until you reach the maximum pdf while saving every epsilon value in a vector.

After we create our required vector, we make a ‘for’ loop and iterate over the same. We compare the threshold with the pdf’s values that generate our predictions in every iteration. 

We also calculate the F1 score according to our ground truth values and the predictions. If the found F1 score is higher than the previous one, we override a ‘best threshold’ variable. 

Keep in mind that we can’t use ‘accuracy’ as a metric in our credit card fraud detection project. That’s because it would reflect all the transactions as normal with 99% accuracy, rendering our algorithm useless. 

We’ll implement all of the processes we discussed above in our ‘functions.py’ file:

def metrics(y, predictions):

   fp = np.sum(np.all([predictions == 1, y == 0], axis=0))

   tp = np.sum(np.all([predictions == 1, y == 1], axis=0))

   fn = np.sum(np.all([predictions == 0, y == 1], axis=0))

   precision = (tp / (tp + fp)) if (tp + fp) > 0 else 0

   recall = (tp / (tp + fn)) if (tp + fn) > 0 else 0

   F1 = (2 * precision * recall) / (precision +

                                    recall) if (precision + recall) > 0 else 0

   return precision, recall, F1

def selectThreshold(yval, pval):

   e_values = pval

   bestF1 = 0

   bestEpsilon = 0

   for epsilon in e_values:

       predictions = pval < epsilon

       (precision, recall, F1) = metrics(yval, predictions)

       if F1 > bestF1:

           bestF1 = F1

           bestEpsilon = epsilon

   return bestEpsilon, bestF1

In the end, we’ll import the functions in the ‘main.py’ file and call them to return the F1 score and the threshold. It will allow us to evaluate our model on the test set:

(epsilon, F1) = selectThreshold(Yval, pval)

print(“Best epsilon found:”, epsilon)

print(“Best F1 on cross validation set:”, F1)

(test_precision, test_recall, test_F1) = metrics(Ytest, ptest < epsilon)

print(“Outliers found:”, np.sum(ptest < epsilon))

print(“Test set Precision:”, test_precision)

print(“Test set Recall:”, test_recall)

print(“Test set F1 score:”, test_F1)

Here are the results of all this effort:

Best epsilon found: 5e-324

Best F1 on cross validation set: 0.7852998065764023

Outliers found: 210

Test set Precision: 0.9095238095238095

Test set Recall: 0.7764227642276422

Test set F1 score: 0.837719298245614

Challenges in Credit Card Fraud Detection 

In managing a credit card fraud detection project, I encountered several significant challenges: 

• Imbalanced Data: The primary challenge was the imbalanced nature of datasets, with legitimate transactions significantly outnumbering fraudulent ones. This imbalance often led to poor model performance in accurately detecting fraud. 
• Evolving Fraud Tactics: Fraudsters continually innovate, necessitating our models to be adaptable and capable of identifying new fraudulent patterns swiftly. 
• Data Privacy and Security: Ensuring the privacy and security of sensitive data was paramount. We employed advanced encryption and anonymization methods to protect user information in compliance with strict legal standards. 
• High False Positive Rates: A prevalent issue was the high rate of false positives, where legitimate transactions were mistakenly flagged as fraudulent. Reducing these false alarms without compromising the detection of actual fraud was a delicate balance to achieve. 
• Feature Engineering: Extracting and selecting the most predictive features from vast transaction data posed a significant challenge. Effective feature engineering was critical to improving model accuracy and efficiency. 

Overcoming these hurdles required innovative approaches, meticulous data management, and ongoing refinement of machine learning models. 

Conclusion

As an expert in machine learning, I’ve witnessed firsthand the evolving landscape of digital fraud and advanced analytics critical role in preempting and mitigating fraudulent activities. Through this article, I have shared insights and experiences that could serve as a beacon for professionals aspiring to venture into fraud detection. By navigating through the complexities of this project, I’ve gathered a wealth of knowledge that underscores the importance of innovative solutions in combating fraud.  

If you have any questions or suggestions regarding this project, let us know by dropping a comment below. We’d love to hear from you. 

With all the learnt skills you can get active on other competitive platforms as well to test your skills and get even more hands-on. If you are interested to learn more about the course, check out the page of the Execitive PG Program in Machine Learning & AI and talk to our career counsellor for more information.