Machine learning is simply a form of artificial intelligence that enables computers to learn without being explicitly programmed. It’s especially good at recognising patterns in data and therefore equally good at spotting anomalies in those patterns. This makes it a great approach for preventing fraud.
Since machine learning is a rapidly expanding discipline there is a lot of innovation and experimentation with techniques, some old and some new. The cost and availability of computing power has made this experimentation affordable and the benefits are coming to the market now.
The basic idea is that fraudulent transactions have different characteristics to legitimate ones. Algorithms can be created based on those differences to predict whether a new transaction is likely be fraudulent before it is completed.
At Ravelin, we prioritise techniques that make it possible for us to explain how a fraud decision was arrived at (inspectability). Our customers tell us that they want to know why and how fraud is happening, so simply delivering a decision from a ‘black-box’ is not a valid approach for us. Therefore, we choose from the following approaches and blend them together to provide a probabilistic fraud score.
Logistic regression: This is a statistical technique where a merchant’s good transactions are compared with its chargebacks to create an algorithm to predict whether a new transaction is likely to be a chargeback or not. For very large merchants these models are specific to their customer base, but more usually general models will apply.
Decision tree: This is a mature machine learning algorithm family used to automate the creation of rules for classification tasks. They are essentially a set of rules which we have trained using examples of fraud that Ravelin's clients are facing. The creation of a tree ignores irrelevant features and does not require extensive normalisation of our data. A tree can be inspected and we can understand why a decision was made by following the list of rules triggers by a certain customer.
Random Forest: This is a technique of using an ensemble of multiple decision trees to improve the performance of the classification. It allows us to smooth the error which might exist in a single tree and increase our overall performance and accuracy while maintaining our ability to interpret the results and provide explainable scores to our users.
Neural networks: Neural networks attempt to mimic how the human brain learns and in particular how it sees patterns. They are exceptionally good at being trained on legitimate patterns and thereby flagging fraudulent ones. It is an excellent complement to other techniques and improves with exposure to data.
Machine learning is not a silver bullet for fraud prevention. It is a very useful technology that allows us to interrogate many more signals and inputs than we could manage using any other approach. It has some limitations however.
Cold start: it takes a significant amount of data for machine learning models to become accurate. For some merchants, this data volume is not an issue but for other is often better to apply a basic set of rules initially and allow the machine learning models to ‘warm up’ with more data. We often apply this approach with smaller datasets.
Inspectability: at Ravelin we want to be able to explain the reasons for a customer being flagged as a fraudster and prevented from using the system. We also need to do this so that our merchants can confirm fraud and therefore train the system. So while we have eliminated this problem through the techniques we have chosen, lack of inspectability can be a drawback of certain other machine learning-based approaches.
Connections: machine learning models score on actions, behaviour and activity. They are blind to connections in data. So a seemingly obvious connection (say a shared card between two accounts) would not be detected by a model. To counter this we enhance our models with Graph networks.
While there are limitations to machine learning, machines are much better at dealing with and processing large datasets than humans are. They are able to detect and recognise thousands of features on a user’s purchasing journey instead of the few that can be captured by creating rules. This ability to see deep into the data and make concrete predictions for large volumes of transactions is the reason why Ravelin uses machine learning as its primary method of preventing fraud for our merchants. For us the key benefits are:
Speed: Our merchants want results fast. In microseconds. Only machine learning techniques enable us to achieve that with the sort of confidence level needed to approve or decline a transaction.
Scale: rules-based programming and machine learning approaches have an inverse relationship with the size of datasets. Rules become less effective while machine learning approaches get better with larger datasets. So in order to avoid multiplying the cost of maintaining a fraud detection system as a merchant’s customer base gets larger, machine learning is the logical choice.
Efficiency: Machines love repetitive tasks, human hate them. The goal of an efficient fraud detection system is to only escalate decisions to people when their input adds insight. Our algorithms should do the heavy lifting.