Yota increased efficiency of fight against swindlers by 30% thanks to machine learning

Yota announced at the beginning of 2017 that the model created by her engineers for identification of swindlers on the basis of their activity in six months of use learned to define swindlers 30% more effective.

Engineers of Yota, using algorithms of machine learning and BigData technology, created the model capable to identify swindlers on the basis of their activity. For a short time the effectiveness of an algorithm grew from 15% to 30%.

"Each client, doing actions, usual for itself, creates a set of events and leaves the information mark. Our IT-systems carefully filter, accumulate and analyzes this statistics regarding abnormal activity. The used approach is characterized in the large volume of the processed information and with it we are helped by BigData technologies which we actively implement for increase in process performance of the company", - the CIO of Yota Bogdanov Andrey noted.

Yota selected one of a set of algorithms of machine learning - "decision tree". For training of model data on the actual swindlers which were detected earlier are used. On an input the aggregated depersonalized statistics on these users moves: the quantity of the entering and outgoing calls, average duration of calls, SMS and MMS activity what rates they use, what is the time are clients of the mobile operator and many other things. At each property the weight at final assessment of the user. At the correct processing and the analysis, they give probabilistic characteristic of the client as swindler.

Thus the model builds a profile of the potential swindler and looks for clients in base with a similar pattern of behavior. As well as in any analysis, there is a set of boundary and implicit values. Each such case gives new opportunities for training. The arriving information allows to update model and to increase its accuracy. After training a system learns and reacts to the new, earlier not revealed or not revealed by the simple analysis fraud methods from users.

As a result of use of an algorithm of loss from activity of swindlers decreased by 54%, and their quantity decreased due to fast detection. It breaks their business model and does senseless acceptances of a fraud.