Empowering Credit Risk Detection in Weixin Pay with Billion-Scale Deep Graph Learning
This paper proposes a risk-aware overlapping subgraph learning framework that balances load distribution with the preservation of critical long-tail risk patterns through budget-constrained sampling and cross-subgraph consistency alignment, significantly improving billion-scale credit risk detection for Weixin Pay.
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
Imagine the internet as a giant, bustling city where everyone is connected. In this city, every person leaves behind a trail of digital footprints—what they buy, who they talk to, and where they go. Most of the time, these footprints tell a normal story. But sometimes, a few people try to sneak in and steal money or trick the system. This is called "credit fraud," and it's like a game of hide-and-seek where the bad guys are very good at hiding. To catch them, banks and payment apps used to look at a person's own footprints alone. But bad guys are smart; they often work in groups or hide in the shadows of other people's lives. So, scientists started using a special kind of map called a "graph." Think of a graph not as a chart, but as a massive web of strings connecting people to events. If you pull on one string, the whole web wiggles, revealing hidden patterns. This is the power of "Graph Neural Networks" (GNNs)—they are like super-smart detectives that can see how a person's connections might reveal if they are up to no good.
However, there's a huge problem with these maps. In the real world, like on Weixin Pay, the map is so big it has billions of people and connections. It's too heavy for a single computer to hold, like trying to carry the entire ocean in a teacup. So, engineers had to cut the map into smaller pieces to share the work among many computers. But here's the catch: when you cut a map into pieces, you often cut the very strings that connect the bad guys to the clues. It's like trying to solve a mystery by looking at only half the crime scene; you might miss the most important evidence. This paper introduces a new way to cut the map that keeps the important strings intact without making the pieces too heavy to carry.
The researchers, working with Tencent's Weixin Pay, developed a clever system called RAOS (Risk-Aware Overlapping Subgraph Learning). Imagine you are cutting a giant pizza to share with a group of friends. The old way was to slice it perfectly so no two friends got the same slice. But if a "bad topping" (a fraudster) was sitting right on the edge of a slice, the person holding that slice might miss it because the topping was technically on the neighbor's slice. The new method, RAOS, is like cutting the pizza with a little bit of overlap. You give each friend a slice, but you also let them hold a tiny piece of their neighbor's slice. This way, if a bad topping is on the edge, everyone who is near it gets a chance to see it.
But there's a risk: if you give everyone too much of the neighbor's slice, you end up with a lot of extra cheese and crust that doesn't help (this is called "noise" and "redundancy"). The team solved this by being very picky about which extra pieces to add. They focused on the "long-tail" connections—these are the quiet, less popular connections that often hold the secret clues to hidden fraud. They used a special filter (based on something called an "h-index") to find these important but overlooked pieces and added them to the slices, while ignoring the noisy, unimportant parts.
Furthermore, because the same person might appear on two different slices, the computer might get confused and think they are two different people with different stories. To fix this, the team added a "consistency alignment" step. It's like having a team meeting where all the friends compare their notes on the person who appears on both slices to make sure they all agree on who that person really is. This ensures the final picture is clear and accurate.
When they tested this new system on real data from Weixin Pay, which involves hundreds of millions of users, the results were impressive. The new method was much better at spotting fraudsters than the old ways of cutting the map. It didn't just catch the obvious bad guys; it was also much better at predicting who might become a fraudster in the future (a "forward risk" prediction). In fact, it improved the ability to catch risky users by over 10% compared to the standard method that just cuts the map without overlap. Even more surprisingly, their new method was actually better at finding fraud than trying to look at the entire map at once, because the old "whole map" approach got confused by too much noise.
In the real world, this means fewer people losing money to scams and a safer financial system for everyone. The team put this system to work in a live test with real users, and it successfully identified more risky users and prevented more financial loss than the previous systems. By carefully overlapping the pieces of the puzzle and making sure everyone agrees on the picture, they managed to solve a problem that was previously too big and too complex for computers to handle alone.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.