Temporal Graph Prototype-conditioned Conformal Prediction for Fraud Detection
This paper introduces ProtoCP, a conformal prediction framework for edge-level fraud detection on temporal graphs that leverages learned prototypes and temporal score diffusion to overcome calibration inefficiencies caused by benign-dominated neighborhoods and class imbalance, thereby achieving target coverage with significantly smaller prediction sets than state-of-the-art baselines.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the digital world, from online marketplaces to banking apps, fraud is a constant shadow. It hides in the complex web of transactions and interactions, where a single suspicious act can be buried among millions of legitimate ones. To catch these bad actors, computer systems use artificial intelligence to spot patterns that humans might miss. These systems are good at making a guess: is this transaction safe, or is it a scam? But a guess is not enough when money and trust are on the line. If a system is too sure of itself when it is wrong, it can cause real harm by blocking innocent people or letting thieves slip through. What is needed is a way for the computer to know when it is unsure, and to admit that uncertainty with a formal guarantee. This is the promise of a mathematical tool called conformal prediction. Instead of just giving a single answer, it offers a small list of possibilities that is mathematically guaranteed to contain the truth most of the time. It is like a safety net that ensures the system never claims to be certain when it is actually guessing.
The challenge, however, is that fraud is tricky. It often hides inside groups of normal behavior, making it hard to distinguish from the crowd. Furthermore, fraud is rare. In a sea of normal transactions, a few bad ones are like finding a needle in a haystack. When researchers tried to apply standard safety nets to this problem, they found the nets were too loose. The systems produced lists of possibilities that were so large they were useless, or they failed to catch the fraud at all because the rare bad examples were drowned out by the common good ones. The existing methods treated every neighbor in a transaction network as equally important, but in fraud, a suspicious transaction might be surrounded by innocent ones that are actually part of a disguise.
To solve this, a team of researchers at Emory University and the University of Illinois at Chicago developed a new approach called ProtoCP. Their goal was to build a safety net that is both tight enough to be useful and strong enough to be reliable, even when fraud is hiding in plain sight and is extremely rare. They realized that to make a good judgment, the computer needed to look at the right context. Instead of looking at a transaction and its neighbors as a messy, mixed-up pile, they taught the system to recognize specific "patterns" or prototypes of how fraud usually behaves. Imagine a fraudster who often uses a specific type of device or connects with a specific group of accounts; the system learns to spot these recurring patterns.
The researchers built a method that first identifies these patterns and then uses them to clean up the data. When the system examines a transaction to decide if it is safe, it looks at the surrounding network. But instead of listening to every single nearby transaction equally, it uses the learned patterns to filter out the noise. It focuses on the neighbors that actually look like the fraud patterns it has learned, and it ignores the neighbors that are just normal, harmless activity. This creates a clearer picture of the situation. Once the system has this cleaner view, it calculates a score to measure how strange the transaction looks. To make this score even more reliable, the system compares the transaction not just to a single number, but to how it differs from its immediate, filtered neighbors. If a transaction looks very different from its suspicious neighbors, it gets a higher warning score.
Finally, the researchers added a step to smooth out the results over time. Fraudsters change their tactics, and what looked like fraud yesterday might look different today. The new method spreads the information across time, so that a sudden, confusing spike in the data doesn't throw the whole system off balance. This allows the system to stay steady even as the fraudsters evolve. When the team tested this new method on four different real-world datasets, including financial transactions and online reviews, it worked better than any previous method. It managed to keep the safety net tight, producing small lists of possibilities that were easy for humans to review, while still catching the fraud at the rate they promised. In fact, for the rare fraud cases, the new method was much better at catching them without being overly cautious, whereas older methods either missed the fraud or created lists so large they were impractical.
The study shows that by teaching the system to recognize the specific shapes of fraud and to ignore the noise of normal behavior, we can make artificial intelligence much more trustworthy. The researchers found that their approach successfully handled the extreme rarity of fraud and the shifting nature of bad actors, proving that it is possible to have both high accuracy and a clear, honest measure of uncertainty. This means that in the future, online platforms could use these tools to flag suspicious activity with greater confidence, protecting users without causing unnecessary trouble for honest people. The work suggests that the key to catching fraud isn't just looking harder, but looking smarter at the right connections.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.