Researchers from the Artificial Intelligence Laboratory of Sberbank and Skoltech have developed a method for training neural networks that enables algorithms to simultaneously consider local and global levels of data on banking transactions. The results of the study, supported by a grant from the Russian Science Foundation, are published in the International Journal of Information Management Data Insights.

The authors proposed a methodology for evaluating neural network models for event sequence processing tasks. The tasks were divided into three types: global, local, and dynamic. The global ones require an assessment of some general characteristics of the sequence, which almost does not change over the period under consideration: the client’s age, purchasing power, and satisfaction with the bank’s services. Local and dynamic ones rely on some characteristic that is constantly changing over time, for example, predicting the next event, and require the neural network to be able to quickly respond to sudden changes in customer behavior, for example, detecting a change of country of residence.

A wide range of best practices have been tested on all of the above tasks. Based on the results, a completely new methodology for analyzing sequential data has been developed. The analysis should include external contextual information, which is data about other clients, particularly those who are similar in a number of ways to the analyzed one. This helps to consider various global trends. This approach improves the metrics of the models for all proposed tasks, in some cases by 20%.

“One of the unique properties of neural networks is their versatility, the ability to adapt to different tasks without any additional cost. In our work, we have managed to describe a wide range of tasks and offer solutions that work well with all of them, even when user behavior changes over time. Also, I am proud that the model was able to take into account the behavior of similar users, which led to a further increase in the quality of the model. The work will continue beyond the publication, and we intend to apply the method to new types of data to increase the resistance of neural networks to anomalies,” commented Associate Professor Alexey Zaytsev, who heads the Skoltech-Sberbank Applied Research Laboratory at the Skoltech AI Center and manages a project supported by a grant from the Russian Science Foundation.

“We focused on finding algorithms that could handle local tasks, even though most of the tasks we worked on before starting this study were global. Surprisingly, now most of the tasks that we face now are more likely to be local. As it turned out, the practical need has just emerged, and we already have a suitable solution. In my opinion, this is one of the main advantages of the work, which sets it apart from most journal articles on artificial intelligence that are already somewhat outdated when published,” said Andrey Savchenko, Doctor of Technical Sciences, the scientific director of the Laboratory of Artificial Intelligence of Sberbank.