Improving Credit Card Fraud Detection Using Ensemble Machine Learning Techniques

The Digital Detective and the Needle in the Haystack

Imagine you are walking through a massive, bustling city where millions of people are buying coffee, paying rent, and ordering pizza every single second. In this city, there are pickpockets trying to steal wallets, but they are incredibly rare—maybe only 2 or 3 thieves for every 100 honest shoppers. If you hired a security guard whose only job was to spot these thieves, they might get really good at saying "No one is stealing!" because, statistically, they are right almost all the time. But that's a bad job for a security guard! If they miss even one thief, the consequences are huge. This is the daily nightmare of online credit card companies: they have to find the tiny, sneaky fraudsters hiding inside a mountain of perfectly normal transactions.

To solve this, scientists use something called Machine Learning. Think of this as a super-smart robot that learns by reading history. It looks at thousands of past transactions to figure out what "normal" spending looks like and what "suspicious" spending looks like. However, there's a catch: because there are so few fraud cases, the robot gets confused. It's like trying to learn what a lion looks like by only showing the robot a picture of a lion once, while showing it a million pictures of cats. The robot will just think, "Oh, everything is a cat!" To fix this, researchers use a clever trick called Generative Adversarial Networks (GANs). You can think of a GAN as a master forger and a master detective working together. The forger tries to create fake "lion" pictures that look so real they fool the detective, while the detective gets better at spotting the fakes. By the time they are done, the detective has seen enough "lions" to finally recognize the real ones. This paper explores how to use these digital forgers and detectives to catch credit card thieves faster and smarter.

---

The Paper's Big Idea: Teaching Robots to Spot the Needle

In this study, a team of researchers from Cadi Ayyad University in Morocco decided to build a better digital security guard. They wanted to see if they could combine the "forger" trick (GANs) with a team of different machine learning robots (called Ensemble Learning) to catch credit card fraud in real-time. They didn't just want a robot that was accurate; they wanted one that was fast enough to stop a thief before the transaction even finished.

The researchers started with a real-world dataset from a website called AirportRental. It contained 76,900 transactions, but here was the problem: only 1,922 of them (about 2.5%) were actually fraudulent. It was a massive haystack with very few needles.

First, they tried to teach their robots on this messy, unbalanced data. As expected, the robots prioritized the majority class. They were so good at saying "This is a normal transaction" that they missed most of the fraud. But then, the researchers brought in the GANs. They used the forger-detective team to create synthetic (fake but realistic) examples of fraudulent transactions. They mixed these new examples with the real ones until the dataset was perfectly balanced: 50% normal transactions and 50% fraudulent ones. Suddenly, the robots had a much better study guide.

Next, they tested a lineup of different machine learning models to see which one was the best detective. They tried:

• Logistic Regression and Naïve Bayes (simple, fast thinkers).
• KNN and SVM (slower, more complex thinkers).
• Random Forest and XGBoost (powerful teams of decision-makers working together).

They measured two things: How good were they at catching the fraud? (This is called Recall). And how fast were they? (This is Execution Time).

The Results: The Speedy Team Wins

The experiments showed that the "forger" trick worked wonders. When the robots learned from the balanced data created by the GANs, their ability to spot fraud skyrocketed. On average, their Recall (the ability to find the bad guys) jumped by about 25 percentage points. This means they stopped missing the thieves.

Among all the models tested, one team stood out as the clear champion: XGBoost.

Here is why XGBoost was the star of the show:

1. It was the most accurate: On the balanced data, XGBoost achieved a 95% Precision (it was right almost every time it raised an alarm) and a 93% Recall (it caught almost all the fraud). It also had a 99% AUC-ROC, which is a fancy way of saying it was nearly perfect at distinguishing between good and bad transactions.
2. It was the fastest under pressure: This is the most exciting part. The researchers didn't just test the robots one by one; they simulated a busy airport rental counter where hundreds of transactions happened at the same time.
   • When handling just 1 transaction, the simple models like Naïve Bayes were the fastest, taking only 0.09 milliseconds.
   • But as the crowd grew to 30 transactions at once, the simple models and the complex ones (like KNN and SVM) started to slow down dramatically. KNN's time jumped to 18.45 ms, and SVM took 21.10 ms. Their performance "degraded" (got worse) by over 800%.
   • XGBoost, however, stayed cool. Even with 30 transactions, it only took 2.90 milliseconds. Its performance only got about 480% slower, which is the best stability of the group.

The paper explicitly rules out the idea that the fastest simple models (like Naïve Bayes) are the best choice for real-world systems. While they are quick for a single task, they can't handle the rush hour of a busy day. The paper also argues against relying on old, unbalanced data, showing that without the GAN balancing, even the best models struggle to find the fraud.

The Verdict

The study concludes that the best way to catch credit card fraud is to use a GAN to create a balanced training set and then let an Ensemble model like XGBoost do the detecting. This combination gives you the best of both worlds: a robot that is incredibly good at spotting the tiny, rare fraudsters, and a robot that is fast enough to stop them before the money leaves your account.

The authors suggest that this approach is a robust, scalable solution for real-world financial systems. They didn't just find a theoretical win; they simulated the real-time pressure of a busy transaction system and proved that XGBoost can handle the heat. While they didn't claim this is the final solution to all fraud (since thieves are always changing their tricks), they showed that this specific combination of tools is a very strong, reliable, and practical step forward for keeping our digital wallets safe.