Understanding Machine Learning Fraud Prevention
Machine learning is a type of artificial intelligence that enables computers to learn without being explicitly programmed. Using statistical techniques, machine learning algorithms analyze millions of combinations of inputs to uncover unique and subtle patterns in data. In the case of device intelligence, those inputs are data pertaining to the online transaction, including the device used in the transaction, as shown in Figure 6-1.
Every online transaction has hundreds of attributes associated with it, including multiple location-related attributes, device characteristics—including anomalies—and behavioral risks, like transaction velocities. While a fraud analyst may be able to look for patterns in one or two attributes (like an IP address and device type), machine learning can analyze millions of combinations of these attributes to determine the specific patterns that can predict risk.
For example, the pattern may be a mobile device with a specific version of Android OS, with a specific version of the browser, with a specific set of fonts and apps installed, with a transaction velocity of 30 per hour, originating from Los Angeles.
Machine learning technology constantly learns. Therefore, the patterns that are useful in predicting risk today may be different from the patterns used next week. In this way, your fraud prevention ability is constantly improving itself without any human intervention.
Two key components are needed to build a machine learning model. The first is a huge dataset, which is required because of the statistical nature of machine learning. We’re talking billions of data points. The second requirement is a way to train the model. Models need to understand what’s considered good and what’s considered bad. With device intelligence, we have a perfect means for this training: device reputation. By analyzing the attributes of devices and transactions that have been flagged as fraudulent (reputation), machine learning can be trained and then predict when new transactions will be fraudulent.
Not all machine learning models are the same. The best model for machine learning fraud prevention has the following characteristics:
- It uses a systematic means for training, such as device reputation.
- It looks at global datasets, rather than that of an individual business.
- It’s ready for use on Day 1.
Be cautious of machine learning models that require months of training before they are able to produce useful results.
Why You Need Machine Learning Fraud Prevention
We’re not exaggerating when we say that machine learning can transform your fraud prevention program. It can optimize your processes and improve your effectiveness, all while reducing your costs.
Consider your review queues. Machine learning can reduce the cost of manual reviews by predicting the trustworthiness or riskiness of a transaction. By knowing which transactions pose the greatest risk, you can focus your efforts on those first. In the meantime, you can fast-track good transactions and reserve more expensive verification methods for situations when they’re absolutely necessary.
Machine learning can also help reduce fraud losses. By analyzing every transaction, machine learning can predict which transactions will go bad, even if your own manual fraud rules do not. Machine learning can also stop fraud attempts that have been seen elsewhere by leveraging global fraud and risk insights from other businesses.
Finally, by identifying good customers, machine learning enables you to improve the customer experience. You can grow revenue faster by offering special incentives to new customers and high-value incentives to existing customers. You can also increase customer satisfaction by eliminating lengthy fraud reviews and delays.
Ready for the next step?
Spot user behaviors and device information that’s suspicious, and stop those fraudsters in real time. We track billions of devices and our fraud analysts add evidence to make this intelligence even more effective.