Improving Fraud Detection Capabilities
Machine learning to improve fraud detection
01 October 2018
3 Minutes read time
Lots of hype continues to surround artificial intelligence (AI). And that means the trick is to get past that and identify real, practical applications that deliver tangible returns.
The good news is that lots of the processes financial institutions needing automation do not necessarily need the latest, cutting-edge machine learning. That would be like using a sledgehammer to crack a nut.
What is typically needed is a combination of technologies. For example, less complicated machine learning capabilities and a rules-based approach may well be sufficient for the transformation desired, rather than pure AI. This approach also means operational users, rather than IT, can define, train and test the services, thereby lowering the threshold to adoption.
Here are a couple of current examples.
One of our banking clients already uses Xceptor to scan payments to identify fraudulent transactions. Writing any rules to capture fraud is complex due to: the size of the data set; it being nigh on impossible to catch every fraudulent transaction and the prevalence of false positives. As fraudsters are relentless in evolving and finessing their approaches, financial institutions need to be as well.
By deploying a combination of both a rules-based approach and native machine learning, we can extend the level of automation in the fraud detection process. For example, we can apply hard rules for certain areas e.g. blacklisting known fraudsters as this isn't a huge data set and the rule can easily be written by a human user. For fuzzier areas, where we are looking to finesse more complex rules and there are large data sets, then we can overlay with machine learning models. The solution then learns and improves. With this approach the rules can be constantly fine tuned, the confidence score can be improved, as can the detection rate.
Another example is a financial institution that has 1000s of incoming client emails containing either netting or standard settlement instructions. Using natural language processing (NLP)-based machine learning, we can train the model to assess intent.
We do this by classifying the emails as either netting or SSI, assessing the right priority and sending to be reviewed by the right team.
Once again, a lot of the tasks in the overall classification process e.g. sending an email to a user to login, do not need machine learning. Deploying simple rules in conjunction with native machine learning enables the broader process to be automated.
Intelligent automation deploys a variety of techniques throughout a process. Machine learning is not a magic answer to automation. Nor is ripping everything out and replacing with AI. Deploying the best technique for each task throughout the process, enables a higher level of automation to be achieved and human intervention to happen as and when it is needed.