It usually starts with a ping.
A text from your bank. A flagged charge. A canceled dinner reservation because your card was inexplicably declined.
Credit card fraud has become such a common part of modern life that many accept it as inevitable. But a team of graduate researchers at the Ambassador Crawford College of Business and Entrepreneurship at Kent State University—each bringing years of prior professional experience in finance, payments, machine learning, and information technology—are determined to challenge that assumption, one algorithm at a time.
“We’ve all felt the frustration,” said Kelechi Amamba, one of the project researchers. “You know your card should work, but something gets flagged, and suddenly you’re dealing with declined transactions or fraud claims. Our research is about solving that—not just catching fraud but making the entire process more intuitive and less disruptive.”
Amamba, a business analyst and machine learning expert with more than 15 years of experience across AI startups and multinational firms, worked alongside Olufemi S. Oloniluyi, an IT and financial services analyst with more than a decade of experience, and Olayinka H. Sikiru, a payment systems specialist with a background in global fintech companies. Drawing on their prior careers in finance and technology, the three launched an independent research project to explore how machine learning could improve fraud detection.
Together, they spent the past year immersed in more than half a million credit card transactions—running dozens of simulations on the Ohio Supercomputer Center’s (OSC) high performance systems to train and test their machine learning models. Their goal: build a fraud detection system that works faster, more accurately, and with fewer mistakes than those currently in use.
“Fraud detection is tricky,” Sikiru said. “It’s not just about catching the bad guys. If your model is too strict, you end up blocking legitimate transactions. If it’s too lenient, you let real fraud slip through.”
The tension between security and customer experience is a problem for everyone—banks, merchants, and especially consumers. American consumers reported losing $12.5 billion to scams and fraud in 2024—an increase from $10 billion in 2023. For every dollar lost to fraud, financial institutions can spend quadruple that amount investigating the case, handling customer support, and managing fallout.
Industry research shows that false positives—legitimate transactions wrongly flagged as fraud—cost banks even more than actual fraud, thanks to lost customer trust and abandoned cards.
“There’s this huge ripple effect,” Sikiru said. “People stop using their cards. They switch banks. Institutions lose customers and it all comes back to whether or not your fraud detection system can make the right call, in real time.”
To tackle that problem, Amamba, Oloniluyi, and Sikiru started examining what most fraud detection systems get wrong: the data.
“Many models use a large number of features—things like merchant name, transaction category, time of purchase—but they don’t always consider how those features interact or overlap,” Oloniluyi said. “We wanted to make the data cleaner, more precise, and more meaningful.”
The researchers developed a hybrid feature selection process that combined statistical tools with real-world domain knowledge. For example, instead of treating city, state, and ZIP code as separate features, they used latitude and longitude to create a more streamlined geospatial indicator. That slight change made a substantial difference in how well their models could identify unusual patterns.
“We didn’t just trust what the algorithm said was important,” Amamba explained. “We used our knowledge to fine-tune it, asking what each data point really tells us about behavior.”
They created three versions of their dataset: a small feature set with five inputs, a medium set with eight, and a full set with 11. Then they tested each set across 24 machine learning models.
To support their research, the team connected with OSC through Philip Thomas in Kent State’s Division of Information Services, who is part of OSC’s Campus Champions program. His role is to advocate for high performance computing (HPC) resources at the institution, onboard new researchers, and help them get access to computing time.
With support from OSC, the team was able to run more than 50 experimental simulations using MATLAB’s Classification Learner tool. They trained each model on historical data, adjusting the size of their testing and validation groups, and compared outcomes using metrics like precision, recall, and false positive rate.
“Working with OSC allowed us to move fast,” Amamba said. “We could test and retest different configurations and feature sets without waiting days for results. That made it possible to do much more thorough research.”
Beyond computing speed, the ability to parallelize jobs and automate testing conditions helped the team find the optimal balance between performance and practicality, especially when designing a model that could someday operate in real time.
“We found that simpler models, like Gaussian Naive Bayes, performed really well on speed,” Sikiru said. “They were fast and accurate, which is exactly what you need for real-time fraud detection.”
Meanwhile, more complex models like Cubic Support Vector Machines (SVMs) excelled at precision, correctly flagging fraud without casting too wide a net. Trilayered Neural Networks struck a balance between the two, offering reliable performance with broader pattern recognition.
But the breakthrough came when the team tried combining models.
“We realized no one model could do it all,” Oloniluyi said. “But if we ensemble the best performers—say, Naive Bayes for speed, SVMs for precision and Neural Nets for pattern recognition—we can create something stronger than the sum of its parts.”
The team shared their findings at the 2025 Ohio Supercomputer Center Research Symposium, where they presented a visual comparison of their models’ performance, highlighting improvements in accuracy, speed, and reliability. Their work has recently been published on multiple preprint servers like IEEE Techrxiv, Figshare, and Authorea.
Although they are still in the initial stages of presenting their research, Amamba, Oloniluyi, and Sikiru hope that their findings someday can be used by banks, credit card companies, and even third-party fraud detection services to play a significant role in reducing financial crimes.
“Some institutions are starting to bring fraud detection in-house,” Amamba said. “Others are still outsourcing it. We hope our work can help both groups refine and strengthen what they already have.”
In particular, Amamba emphasized the need to replace outdated models that may still "work" but do not deliver optimal results.
“We’ve talked to people in the industry who are hesitant to upgrade because their systems haven’t failed yet,” he said. “But we’re saying that you don’t have to wait until something breaks. You can make it better now.”
“At the end of the day, it’s about trust,” Sikiru added. “If people trust their banks, they’ll keep using them. But that starts with good systems. Systems that work quietly in the background, protecting you without getting in your way.”
While financial fraud continues to evolve, so must the systems that stand guard against it. With help from OSC, this Kent State team is helping to make fraud detection not just faster or smarter—but more human-aware, resilient, and ready for the future.
Written by Lexi Biasi
The Ohio Supercomputer Center (OSC) addresses the rising computational demands of academic and industrial research communities by providing a robust shared infrastructure and proven expertise in advanced modeling, simulation and analysis. OSC empowers scientists with the services essential to making extraordinary discoveries and innovations, partners with businesses and industry to leverage computational science as a competitive force in the global knowledge economy and leads efforts to equip the workforce with the key technology skills required for 21st century jobs.
