DUBLIN -- Many companies, especially retailers, are interested in machine learning, particular for its ability detect fraud. However, many of them are misunderstanding some of the fundamentals of the technology. Tuesday morning at MRC Dublin, an expert panel tried to clear up some of the confusion.
Panteha Pedram, director of risk for Ingenico ePayments, advised attendees to know what they want from machine learning. “There’s no recipe for everyone to follow,” she said. “Know how it fits your business.”
She noted that while machine learning can be used to root out fraud, that is not its primary function. The role of machine learning is to find a pattern that you are looking for. “When it comes to machine learning, people say, ‘It’s good for fraud.’ No… It is good for finding the pattern—including the fraud pattern,” she said.
Thierry Arrondo, managing director of Vendo Services advised merchants to have patience with machine learning, because you’re not going to get results immediately. “You have to invest; you have to find people that can help you develop the technology,” he said. “The results will come, but you have to be patient. A simple problem may take a week to solve. If you realize that, you’re going to be fine.”
Fred Sadaghiani, CTO of Sift Science, has observed that companies often have a lack of understanding of the mathematics and science behind machine learning. A common mistake is not having a principled evaluation. He advised attendees not to overreact if something goes wrong after they implement a machine learning system and their predictions go completely haywire. “That should be fine; expect that,” he said. “That’s an actual part of the process of machine learning. Wake up in the morning and expect your model to perform poorly, but have an evaluation by which you can measure how accurate your model is performing.”
He added that machine learning is very “needy” and it takes a lot of effort to ensure that the system is performing well. “That’s a side of machine learning that I don’t think people are aware of,” he said. “People see the headlines and think, ‘Wow! Look what this model can do!’ But there’s a lot of work behind it.”
Moderator Roger Lester, account director of Featurespace, added that merchants who don’t check their data and make sure it is clean are going to be “in a world of pain.”
Lester noted that machine learning takes away a lot of the manual processes away to allow your experts to focus on more strategic activities. However, he noted that the subject always brings up the question of whether there will still be a place for humans in the future.
Danny Grossman, head of data science for Forter, noted that particularly when it comes to using machine learning to detect fraud, a lot of it is about preparing yourself for what you haven’t seen. “It’s making your system ready to respond to anomalies,” he said. “Machine learning is the background of everything, and on top of that, you should have some person to handle the unknown. For that reason, humans will always have a very essential role.”
However, while Sadaghiani agreed that humans have to be the consumers of the results of the algorithm, he believes that machine learning is very much in the foreground rather than the background. “I think that we’re going to see more and more autonomous elements to fraud detection in this space and less need for humans to be involved,” he said. “There’s a lot of advancement happening in the research today, where the applicability and usability of machine learning in fraud is driving performance well beyond where we can go with existing algorithms. I think there’s going to be the same need for humans to say, ‘Yes, this makes sense,’ but the algorithms and the technology will continue to play a bigger and bigger part.”AFP 2018 has multiple sessions on machine learning in the Payments Track.