We’re living in the future – well at least the future we imagined not very long ago. Cars are beginning to drive themselves, voice assistants are able to complete tasks for us, and we can unlock our phones using our face. All of these improvements are advancements in technology powered by machine learning.
This is the essence of machine learning: We expose our statistical models to massive amounts of data without giving them explicit instructions, and the machine starts to make inferences about the relationships between the different data points. Spotify suggests new music based on your streaming history and the history of people who listen to similar music. Cars drive themselves by analyzing sensors, 3D maps and videos from previous trips to determine how to behave, like stopping for crossing pedestrians or red lights. With fraud detection, machine learning works similarly.
Leveraging a rich network of past orders, fraud detection models discern in real time whether or not an order being placed is legitimate or fraudulent. Just like other forms of machine learning, fraud detection systems rely on different data inputs to perform their task. These inputs are known as features.
Advanced statistical models use hundreds or even thousands of features to help perform their task. A feature can be something as simple as order value, or it can be more complex, like understanding the relationship between billing and shipping addresses. We can offer inputs like order value and the relationship between billing and shipping addresses for the model’s consideration, but it’s up to the model to decide how to leverage that feature in its decision.
After being exposed to millions of orders, the model may determine that an order value above $250 is riskier than a smaller order. Similarly, it may learn that billing addresses that match shipping addresses are more likely to be legitimate. Does this mean the model always approves orders with a billing and shipping match and always declines orders above $250? Certainly not. Over time, and with the input of industry experts overseeing the models, the machine learns how much weight to give to each feature of the model. That is why more data with more features makes for more accurate decisions.
Ecommerce fraud affects merchants across all industries. There are some features that will be specific to a particular industry, while others can inform any fraud detection case. Let’s look at examples of both general and industry-specific features that have improved fraud detection models.
General Features of Fraud
Certain criteria required for online transactions is consistent across industries. Whether you’re booking a flight or buying a pair of sneakers, there’s a standard set of information shoppers must provide. Here are some examples of features that are relevant for fraud detection no matter the industry:
Email Age: The age of an email account is typically available to a fraud detection system. In general, the longer established an email is, the safer it is to approve the order. An email account that has been active for 7 years is usually more legit than one created yesterday.
Express Shipping: Be wary of express shipping requests (unless you’re Amazon). Consumers don’t usually want to pay more for expedited shipping, but a fraudster doesn’t mind spending someone else’s money on the shipping fee and wants their goods faster.
Discount Orders: Items purchased on sale are usually safer than regular priced items. Discounted products aren’t as valuable to a fraudster, which is why fraud tends to be 50% lower on days like Black Friday, according to Riskified data.
While these features are standard across eCommerce, they need to be treated differently based on their context. Fraudsters are still capable of hacking into 7 year old email accounts to trick the system, and many newly created email accounts are perfectly safe. Some customers will be willing to pay for express shipping. This is why features are weighted differently in different models and why we add in industry-specific features to help further increase the model’s accuracy in detecting fraud.
We need industry-specific features because not all features will be relevant in every industry. Trips booked months in advance are typically deemed safer within travel, but time before departure can’t be defined for fashion. While reshipping would be an indicator of fraud for an electronics purchase, it wouldn’t necessarily be a red flag for someone purchasing high-end clothes with an international credit card. Here are some examples of industry-specific features that have been helpful in detecting fraud:
Relationship Between Age & Gift Card Type: Elderly individuals wouldn’t usually buy Xbox gift cards unless it’s a gift for their grandchild during the holidays, which means it’s more likely to be fraud. With enriched data from third-party sources, models can determine the age of someone who placed an order and learn to make inferences on whether the gift card category makes sense given the time of year.
Seller History on Sneaker Marketplaces: Many individuals are shopping for sneakers through online marketplaces that let anyone sell their product. Within the sneaker industry, knowing that the seller has sold to many customers and has a lower chargeback rate are stronger indicators to the model that a purchase made from this same seller is less likely to be fraudulent.
Distance Between IP Address and Billing Address: This is a feature that comes in handy for multiple industries but has different outcomes depending on the distance. A far distance between IP and billing addresses is not as alarming for a travel or ticket purchase online as it would be for a fashion or electronics purchase. For retail purchases, IP addresses that are more than 10 miles away from a billing address would be more suspicious. With travel and ticketing, it’s more likely that someone might be in another country or state when deciding to book another flight or attend an event.
Features are the building blocks of machine learning. Just as a tower with more interlocking bricks will be steadier, a statistical model with more complex features will improve accuracy. With more features, both general and industry-specific, the model learns how much weight to assign to each input, which builds more context into its decision-making. Take some of the examples from above and think about your own business and which features will detect fraud with the most precision. And then consider how much simpler it would be if you’d let a machine figure out the answers for you.