A bodybuilder, a cyclist and a student.
They didn’t walk into a bar. But they did raise some hair-raising fraudulent insurance claims.
In 2017, a cyclist claimed £135,000 compensation after he falsely stated that he fell off his bike following a collision with a pothole. A bodybuilder claimed £150,000 for a back injury that wasn’t hindering him from the press-up challenge he went on to film. And a student thought his luck was in when he tried to claim £14,000 for the “loss” of some of his more expensive personal items while on a jolly holiday in Venice.
Insurance fraud cases cost the U.K. billions of pounds every year. On average, it boils down to over £10,000 per fraudulent claim — and results in consumers having to spend an extra £50 per policy.
Tackling the Big Issues
Insurance companies currently face two major challenges.
The first is the large number of calls they receive for fraudulent claims. The second is adapting to the recent GDPR law, which prohibits so-called black box policies. Instead, insurance companies have to be able to explain to their customers, as well as regulators, how decisions have been made.
In response, London-based Intelligent Voice has set out to develop a set of machine learning algorithms that can identify fraudulent behavior in real time. The goal is to make processes more efficient and effective, as well as reduce the fatigue experienced by call agents.
Intelligent Voice combines its machine learning and speech recognition skills with behavioral analytics knowledge from Strenuus, also based in London. The University of East London is working on adding an explainability layer to the technology that will determine how and when decisions were made in a particular case.
The team has shown that they can match human-level efforts in identifying potential fraud.
Their efforts are part of the U.K. government’s Next Generation Services Industrial Strategy Challenge Fund. The project will run for about two-and-a-half years.
Detecting Fraud Before the Payout
During calls to insurers, the system picks up signals of potential deception. These can take the form of specific words or phrases as well as tone of voice. A long short-term memory (LSTM) network has been trained to recognize the signals in real time, so call agents can respond to alerts immediately and change their responses accordingly.
Employee productivity gets a boost because calls flagged by the technology can be provided as a list noting potentially fraudulent markers. Call agents can jump directly to flagged sections for review.
Intelligent Voice’s machine learning algorithms are trained using hundreds of thousands of insurance calls, which have already been manually screened. To power this training, they use an assortment of NVIDIA GPUs and, in production, their software runs on NVIDIA Tensor Core V100 GPUs.
“From a technology perspective, we’ve not found anything which gives us the flexibility and performance that NVIDIA GPUs do,” said Nigel Cannings, CTO of Intelligent Voice. “The flexibility that CUDA offers, in particular, both on the programming side as well as supporting deep learning simultaneously, means that NVIDIA is the obvious choice for us.”