Combining rule-based and Machine Learning approaches in risk management is not old-fashioned
The wide array of anti-fraud companies build tools helping others to comply with AML, KYC, and GDPR, and to detect upcoming fraud patterns. Even so, fraudsters’ technologies forge ahead.
According to Javelin + LexisNexis’ latest research, in 2018 account takeover fraud rose 177%, existing-card fraud rose 8%, existing-non-card fraud rose 115%, new account fraud declined 80%, existing account fraud rose 16%, CNP fraud is 81% more prevalent than POS fraud, 61% of major banks’ fraud losses stem from identity fraud, and 20% of identity fraud incurred by larger banks is synthetic identity fraud.
It seems that risk professionals, with their rule-based approach, trail far behind fraudsters. Machine Learning (ML) becomes a solution here. This advanced AI technology prevents CNP fraud, account takeovers, detects synthetic identities and bots, etc.
However, at the moment it is inconceivable to completely abandon “if-then” intervention for ML technology. Each industry has unique features and problems, and machine learning models used by each fraud prevention system don’t adapt perfectly, allowing false positives and false negatives from time to time.
So, is it possible to achieve a balance between rule-based decision making and an ML approach?
Piece of cake. The rule-based approach, provided by risk analysts who simply create models comprehensible for humans, complements machines which can analyze large volumes of data.
Understanding the strengths and weaknesses of both approaches supports their connection in risk management and fraud prevention.
Strengths and weaknesses of the rule-based approach
The implementation of if-then rules is easy. However, being simple and strict, a rule-based approach becomes unwieldy when it comes to complex data points. Also, it isn’t “innovative;” rules cannot be changed as fast as the fraud world changes. They drift.
Strengths and weaknesses of the ML approach
Machine Learning is based on models, not on human expertise. While the rule-based approach shows clear determination, in machine learning statistics is in wide acceptance. An ML approach presumes that the output data can be described as a combination of input data with other facts. The input and output data are easily decoded, but the decision-making process usually seems like the “black box.” It can be both good and bad for businesses.
There are two types of learning in ML that combine rules and machines: supervised and unsupervised learning.
The supervised learning approach is more common. Within supervised learning, the risk analyst creates an ML model using historical data. Then, the algorithm combines old and new data to “signal” cases that are fraud and ones that are not fraud. The system collects all fraud signals and types and allows the algorithm to detect new patterns. The model constantly adjusts. In supervised learning, the risk analyst is always behind the process.
In contrast, unsupervised learning models derive patterns from a data complex without labeled outcomes. Unsupervised learning empowers risk analysts to resolve problems without knowing the end results ahead of time. They can receive structure from the dataset where they don’t know the exact impact of the variables. With unsupervised learning, the feedback is not based on the prediction.However, it can differentiate data on the basis of anomalous patterns. Risk analysts can then apply supervised learning approaches to these facts.
Unsupervised machine learning is more suitable for fraud prevention and risk management. It is easily combined with rule-based risk logic, and helps to solve issues when risk managers are one step behind the fraudsters.
Machine learning will not replace rules and can’t be imagined without humans, but it complements rule-based systems and expands the capabilities of risk management and fraud prevention platforms.
Pavel Gnatenko, Product Owner at Covery, Head of Risk at Maxpay
Biography of the contributor Pavel Gnatenko:
Pavel has a master’s degree in intellectual systems for decision-making. He is a risk management expert with more than seven years of experience in the FinTech. Currently, Pavel is focused on developing Covery – the next generation of risk management platforms.
Covery is a global risk management platform helping online companies solve fraud and minimize risk. We focus on the universality of our product and its adaptation to any type of business, based on the individual characteristics and customer needs using both rule-based and machine learning approaches.