November 14, 2017
In an increasingly connected world, real-time detection of potential threats is a huge challenge for financial institutions and other digital businesses. Consumers behave in very diverse ways that are hard to predict and fraudsters use new techniques to mimic legitimate user behavior. Consumer behavior, as well as fraud and risk patterns, are evolving fast. Businesses are increasingly relying on machine learning to produce an effective predictive model based on past transactional and behavior data.
Limitations of Traditional Machine Learning Approaches
Machine learning models offer capabilities that deliver a score but don’t provide insights into the key elements that drive the score. Algorithms like neural networks and large decision trees take a “black-box” approach by using models that are highly nonlinear and hard to decipher.
Apart from being hard to explain, these models generally degrade as fraud or consumer behavior evolves. This leads to a model that predicts fraud with low accuracy, leading to high friction. In this fast evolving space, not only is it critical to accurately score a transaction, but also to understand the underlying rationale to derive learnings about the changing environment.
The ThreatMetrix SmartLearning Approach
ThreatMetrix SmartLearning offers a clear approach that enables businesses to not only use the output of the model, but also understand the insights that drive the model. A proprietary algorithm reduces complexity without sacrificing the performance, meaning the model is translated into rules that can be read or modified by the users.
ThreatMetrix SmartLearning combines real-time actionable data from the Digital Identity Network with an organization’s truth data (feedback data provided by analysts or third parties), to generate rules and attributes that are optimal to solve the problem at hand. This is done while balancing model performance and complexity. Businesses can choose to use the risk score and set operational limits such as maximum percentage of challenged transactions, or to build their own complementary business rules and behavioral pattern detection logic using the insights generated by the model.
With this clear-box approach, ThreatMetrix provides insight into why the machine thinks the information is good or bad. This enables businesses to better understand their customers and use this information to influence other business decisions, while also increasing the efficiency of their fraud and risk teams.
ThreatMetrix SmartLearning Advantage
The ThreatMetrix machine learning system has several advantages over traditional supervised learning techniques for fraud detection:
- A Better Approach:
ThreatMetrix SmartLearning does the heavy-lifting in variable and feature generation and selection. It delivers high performing models with high accuracy and catch rates while reducing the size and complexity of the underlying model. This “clear box” approach to modeling enables businesses to build their own rules rather than just rely on a score from the model.
- Custom models based on global data:
SmartLearning delivers tailored scoring models based on business and use-case specific data and feedback while still leveraging shared intelligence from millions of daily consumer interactions. ThreatMetrix stitches together a customer’s true digital identity by analyzing the myriad connections between devices, locations and anonymized personal information.
- Operational context:
It’s not only about stopping fraud. SmartLearning scores and models are aligned to business operational objectives, such as increasing sales and managing the review rate or challenge rate that businesses want.
- Integrated Decision Analytics Platform:
ThreatMetrix SmartLearning can be complemented by customized business and behavior-based statistical rules that enable businesses to adapt quickly and achieve superior results.