Unsupervised Machine Learning

Machine learning is a branch of artificial intelligence that enables algorithms to learn from existing data and then apply that knowledge to new data. Unsupervised machine learning (UML) is a major category of machine learning techniques that works without requiring labeled input data. Instead, it infers a function to describe the hidden structures of “unlabeled” input data points. UML is often used to discover patterns within large amounts of unlabeled data, and is especially effective for discovering new and unknown patterns.

Common UML approaches today broadly include anomaly detection techniques that attempt to identify outliers, and clustering/graph analysis techniques that focus on studying the relationships and connectivity among input data. The DataVisor UML Engine is developed based on the latter approach, combining clustering techniques and graph analysis algorithms together to discover correlated fraudulent or suspicious patterns from unlabeled data. By analyzing the distance and connectivity between data points that represent accounts and their activities across a large time period, the DataVisor UML Engine is able to  automatically discover new and emerging fraud patterns.

This approach has a much higher precision rate than single-user based detection. A good user will typically evidence diverse behavior patterns, but it is highly unusual for an entire large group of users to all share those same behaviors. So while an individual instance of a certain characteristic or behavior may not be enough to confidently act upon, when we see the behavior repeated across an entire group, we can have a much higher degree of confidence about the accuracy of our results. 

UML for Fraud Detection and Prevention

The signature advantage of unsupervised machine learning is its ability to operate without the need for labels; it can analyze data in real-time, with no prior legacy knowledge required. UML models are fundamentally self-tuning, and UML-powered solutions are free of the delays associated with Supervised Machine Learning and rules-based approaches. As such, UML is an ideal approach with which to challenge modern, fast-evolving fraud. 

The use of UML-based fraud management tactics can have specific and lasting financial impact, especially when a given organization is wrestling with many different fraud types—and especially when the attacks are new or previously unknown. 

There are many fraud use cases for which UML can be applied, including:

Application Fraud

Using UML, banks and financial institutions can analyze whole networks of applications to detect hidden connections that may appear legitimate when viewed in isolation. 

Transaction Fraud 

UML algorithms can be used to detect fraudulent accounts before those accounts can be used to conduct transactions that result in financial loss. 

Bot Attacks 

Using a UML-driven holistic data analysis approach, it is possible to analyze user histories, behavior changes, and suspicious patterns across millions of accounts. This enables the capture of significantly more bot-powered attacks. 

Promotion Abuse

UML solutions enable the captured of all members of a given fraud ring by identifying hidden linkages between fake account registrations and discovering unknown attacks without labels or training data.

Money Laundering 

UML algorithms can look at complex networks of transactions instead of individual ones, and can detect and eliminate launderers who deposit small denominations of funds to avoid CTR reporting. 

Unsupervised Machine Learning and DataVisor

At DataVisor, our approach combines cutting-edge AI and machine learning technologies to correlate fraudulent and suspicious patterns across billions of accounts in real time. Patented and proprietary unsupervised machine learning (UML) algorithms work without labeled input data to automatically detect new and previously unidentified fraud and abuse patterns.

The DataVisor Unsupervised Machine Learning Engine processes all events and account activities simultaneously to analyze patterns across hundreds of millions of accounts. This enables detection of suspicious connections between malicious accounts, even when those accounts are incubating, mimicking legitimate user activities, or changing attack techniques. This also allows the UML Engine to detect all the members of an attack ring at once, ensuring the attack is fully stopped.

The ability to accurately identify new and unknown attack types in real time with no need for historical data or lengthy training and retuning cycles gives businesses a defining advantage against adaptive and agile modern fraudsters. 

Learn more about how Unsupervised Machine Learning can reduce fraud and improve customer experience.