Ocado is developing cloud and machine learning technologies for use in detecting fraud.
The retailer, a Top50 retailer in IRUK Top500 research, is using technologies developed through its in-house Ocado Technology business and says initial tests with machine learning has made its fraud detection rate 15 times more precise. It believes this is the first example of a retailer using machine-learning to combat fraud.
Currently, its fraud agents, equipped with a rules-based detection system, might spot trends in transactions that later prove to be fraudulent. There might be a correlation between baskets containing large alcohol orders and confirmed instances of fraud, and they might look for this trend in future. But those carrying out the transactions could quickly shift to another area, making this a cat and mouse game that the agents must work hard to catch up with, says Ocado.
Once deployed, say Ocado, the new technology will make the job of those agents easier: its machine-learning model, which has been developed and running in the background over the last six months, will predict results in real-time and present the likelihood of a transaction being fraudulent. The agent must then decide simply whether, based on that probability, the order should be cancelled.
Roland Plaszowski, head of retail systems at Ocado Technology, said: “Removing fraud altogether is the final goal but it will be a long road. It’s worth remembering fraud is one part of this but it’s quite rare – it’s one in every thousand orders, so one part is to eliminate fraud which brings a lot of costs to the company but the other part is to make the user experience better.” The other part, he says, is in moving away from systems that flag up legitimate transactions, disappointing customers if their order is cancelled. “There’s a balance between removing fraud and doing so positively,” he said.
The graphic below shows how the machine-learning system works.
Customer order information is stored and analysed using Google Cloud BigQuery, before being processed using Dataflow. There the data is normalised into a format required for machine-learning algorithms such as the Deep Neural networks that use TensorFlow. Dataflow is used to transfer data to Google Cloud storage and Datastore. Cloud machine learning then uses data from cloud storage to produce models as APIs. The Ocado fraud detection model, powered by TensorFlow, then reads the data from Datastore and uses cloud machine learning APIs to make real-time predictions.
This latest example of machine-learning in retail follows Ocado’s previous use of the technology to improve its handling of customer queries, and in its fulfillment and logistics operations. It also uses the technology for tasks such as product recommendations, enabling it to avoid showing meat to vegetarians or gluten-containing products to coeliacs.
