Fraud detection leverages machine learning so that businesses, banks and financial institutions can detect suspicious transaction activity quickly and at scale. Machine learning models are useful for flagging potential fraud for review and for spotting patterns of behavior before fraud actually occurs.

This means there is a lot of responsibility on the machine learning model itself. Fraud investigators and executives need to understand how it works and how well it is performing — both of which are easier said than done.

After all, machine learning (ML) is highly complex, performing calculations at far greater speeds and at much greater scale than humans can.

In this article, Chris Oakley Featurespace Subject Matter Expert, outlines how machine learning models are used to detect fraud today and what results we are seeing from the newest generation of ML models.

Why use machine learning models in fraud detection?

Fraud has existed for as long as money has existed. For most of the history of banking, fraud detection was a manual process, and often a tedious one.

In the late 20th century, technology gave organizations tools and processes for automating fraud detection. These were simpler, rules-based methods of spotting fraud e.g., “If transaction amounts exceed $10,000, flag for manual review”. By applying machine learning for better analysis, we can identify those transactions above $10,000 that can be of a greater risk.

The problem with rules-based fraud detection is their rigidity. Rules are typically retrospective and must be updated as circumstances and people’s behaviors change. That’s also a tedious and slow process, and it’s never quick enough to keep up with unfolding circumstances or customer behaviors.

Newer machine learning models solve that problem by bringing dynamism to fraud detection. For example, Featurespace’s Adaptive Behavioral Analytics introduced automated, self-learning algorithms that study transaction behaviors. They teach themselves what types of customer transactions are normal and what types are anomalous.

The more transactions that are examined, the better the model’s accuracy. Rather than having to periodically revise a set of static rules, Adaptive Behavioral Analytics constantly get smarter and more precise.

As a result, the model is able to achieve in real-time, a probability of whether the activity is fraudulent or genuine.

This generation of machine learning has made fraud detection:

  • Faster.
  • More scalable.
  • More accurate.
  • Less intrusive for banking customers.

How does machine learning work in fraud detection?

Machine learning can be hard to describe. It works at such speed and such scale that humans struggle to explain it in words. This is why model explainability — i.e. being able to describe conversationally what a machine learning model is actually doing — is so important in fraud detection.

Still, we can talk about fraud detection with machine learning as happening in three basic steps:

1. Data input

A machine learning model requires lots of data. Fortunately, finance and banking by their nature produce large amounts of data every day. In fraud detection, one such source of data would be transaction data, including who sent the money, who received the money, when the money was sent, etc.

2. Feature identification and extraction

A feature is some kind of intelligence gleaned from raw data. For example, a fraud detection algorithm might divide each transaction a customer makes by the average value of that person’s transactions. That’s a feature, and that is a kernel of intelligence the model can use to predict whether a transaction is suspicious.

The machine learning model then crunches historical data to determine whether a feature represents normal or anomalous behavior.

Take the feature above, transaction value divided by average transaction value. Let’s call that value X.

3. Feature weighting within the model

Now, imagine the model has found that, historically, transactions greater than or equal to X have been fraudulent 20 percent of the time. That information then factors into the model’s ultimate prediction of whether a given transaction is suspicious. Likewise, features may indicate that a given transaction is more likely to be genuine.

Depending upon the strength of the feature indicating whether a transaction is suspicious or genuine, a weighting is allocated to each feature used within the model, which collectively identify the probability of the transaction being fraudulent.

Classification is a process of trial and error. As the model crunches more and more data, it gets more precise in its predictions. This is how a machine learning model trains itself and ultimately achieves a level of performance that’s ready for deployment.

This optimization is achieved through repeating the above steps on a regular basis to allow iterative improvements in the model performance based upon new data and intelligence.

Where does deep learning fit into fraud detection?

For the last decade, data scientists have been pushing past the old boundaries of machine learning by exploring the field of deep learning.

Deep learning relies on neural networks, which are designed to recognize hidden patterns in the input data.

In traditional machine learning, the actual learning took place in the classification and model optimization step. With neural networks, the actual learning begins at the feature extraction step. This produces models that are both smarter than their predecessors and freer of the human biases that informed feature-extraction design.

Featurespace’s Automated Deep Behavioral Networks are an example of how deep learning can be applied to fraud detection. Early results from our deployments of Automated Deep Behavioral Networks show a significant uplift in fraud detection and fraud prevention.

How have Featurespace’s machine learning models helped businesses?

Numerous banks and financial institutions have deployed Featurespace’s ARIC™ Risk Hub to monitor transactions and detect anomalies via its Adaptive Behavioral Analytics.

Here are just a few examples:

  • Global payments company TSYS reduced fraud losses by 39 percent and reduced false positives by 34 percent by deploying ARIC Risk Hub’s real-time decision capabilities.
  • Leading UK bank NatWest deployed ARIC Risk Hub across the enterprise to get a better handle on fraud detection and prevention. Within the first 24 hours of operation, NatWest saw a rise in the number and value of scams detected, and a drop in the number of false positives.
  • Danske Bank, the largest bank in Denmark, uses ARIC Risk Hub’s machine learning capabilities to support card fraud detection and to reduce false positives, thereby ensuring the smoothest possible customer experience for its millions of cardholders.

To learn more about Featurespace’s model performance in enterprise deployments, have a look at the presentation below from Founder Dave Excell:

Get in touch

Featurespace’s machine learning models are built to help organizations detect and prevent fraud. Our ARIC Risk Hub monitors real-time customer data for anomalous behaviors and helps fraud investigation teams prioritize alerts.

To learn how ARIC Risk Hub’s machine learning models could support your own organization’s fraud detection and fraud prevention goals, book a demo today.