The financial services industry is often promoted as one of the earliest sectors to benefit from quantum machine learning, and an important use case is fraud detection. But how could it work to speed up the calculations classical computers struggle with and how can it help fraud detection operate in real time?
Machine learning on classical computers has significantly improved financial fraud detection capability. Speaking on the “Qubits and neurons to turbocharge the world of finance” panel at an IBM quantum and AI event last week, IBM quantum applications researcher Christa Zoufal explained that quantum computers can select the features that best train the machine learning model.
“In the financial industry, we have quite a lot of equipment and problems that have a very complex structure underneath, and with this complex structure it can get hard for classical algorithms to solve them,” she said.
“If you want to do fraud detection, for example, you can have a lot of data with features that represent the transactions. You don't just want to put thousands of features into your models because that would make the training very difficult. You want to collect a few very good ones to make the problem simpler to train faster, and to make it more understandable.”
Feature selection is a difficult problem and the number of solutions in the problem scales exponentially with the number of features that need to be considered. Most classical machine learning algorithms can, at most, look at correlations between two features, and even if there are 1,000 features, classical algorithms will look at correlations between one pair of these features at a time.
“What you can do with quantum machine learning in this particular context is run the algorithm using the quantum circuit and look at all of the correlations at the same time. In some instances, we have already found that with this quantum algorithm we can get a better picture with very simple models,” said Zoufal.
