Machine Learning for Advanced Surveillance
How Alithya’s machine learning expertise gives its clients an edge
Adam Wisniewski, senior data scientist at Alithya, is a machine learning expert. In this post, he talks about what machine learning is, how it’s done, and how it can both improve a company’s performance and reduce its costs
What is machine learning?
Wisniewski defines machine learning as “a subtype of artificial intelligence where you use the data to automatically build models without having to explicitly program the logic of how the model works.”
Two main types of machine learning include: supervised learning and unsupervised learning.
In supervised learning, the machine learning techniques process data containing subjects that have been labeled with a numerical value or certain categories—say category A and category B- which is called a “training set”. The technique is then able to build a predicative model that predicts what category a new subject would fall into based on whether it matches more closely to the examples in category A or B that technique was trained on.
In unsupervised learning, the machine learning techniques look do not have labled subjects. These models identify patterns in the data set such as which subjects are the most similar to each other.
Giving a bank a competitive advantage
A great example of machine learning is a surveillance system that the Alithya team recently developed for a major Canadian bank. The system generates trading compliance alerts traditional, “if x, then y”, logic and categorizes them as suspicious or unsuspicious using machine learning. This significantly reduces the time compliance officers spend reviewing normal activity and focuses their effort on more suspicious/unusual behavior. “In the project, we used a supervised approach because we had explicit outcomes we were trying to predict, which was the level of suspicion of an alert,” explains Wisniewski.
Why this particular machine learning system is working so well
According to Wisniewski, this problem was well suited to machine learning because it involved tasks that are repetitive, with clear objectives. There are also millions of transactions and thousands of alerts giving a substantial training set that can be used to teach the models. By feeding in past examples of suspicious and non-suspicious alerts, the system is able to automatically build more sophisticated/accurate models from which to evaluate the suspiciousness of new alerts.
“Systems based solely on traditional logic were generating an overwhelming amount of false positives,” Wisniewski explains. “With our system, compliance officers no longer need to waste time reviewing activity that is obviously not suspicious, and they can focus their actions on the types of transactions and activity that pose most risk to the bank.” Wisniewski suggests that this type of system could also be used to identify terrorist financing, tax evasion and money laundering.
On the importance of working with relevant experts
For companies wanting to implement machine learning into their own processes, Wisniewski recommends first identifying the problem machine learning could solve, then deciding on the metrics to measure the solution’s success. He points out that machine learning programs take much less time to build when their accuracy rate doesn’t have to reach close to 100%. A program that is accurate in simpler cases but not in highly complex ones could still be used to tackle the simple problems and flag the complex ones for a human. “In some cases, you will see complete automation but we will also see tons of cases where humans and automation work together in collaboration to get the optimal result,” Wisniewski explains.
Wisniewski says companies should not simply work with machine learning experts, but also with experts in the data they’re analyzing. “On our team, we had people with experience working with the systems that hold that trade and order data, and we had people with the business experience to help us analyze the data,” he explains. “You can’t use the same system for finance as you can for geological exploration or retail analytics. Each of them have different measures that are important and only people in those industries understand which data are important and how to use them.”
Where machine learning is headed
Wisniewski expects the AI breakthroughs that have been promised, such as self-driving cars and programs that can diagnose health disorders, will be an everyday reality. This machine learning program for the bank was the first example of a financial institution using machine learning for capital markets compliance but Wisniewski also sees companies in numerous industries—from banking to energy to engineering—incorporating machine learning into their data analysis tasks. “Right now, very few companies have implemented this technology successfully, so we’re at the tip of that. But once they do, you’re going to see savings of millions of man hours that can be better spent on doing higher-value tasks.”