Development of Recommender System for the Bank's Clients Based on Transactional Activity of the Client's Environment

There are many ways to create recommender systems for data that have a rating scale, such as movies and music. The problem with such algorithms is that they cannot easily be standardized for data with non-scaled target values, such as data on customer purchases or their frequency. For example, ratings for songs and movies usually represent a scale on a five- or ten-point system. However, the purchase data is continuous and does not have an upper limit. This research describes an algorithm for developing a recommender system based on the transactional activity of the client's environment. The work contains a comparative analysis of various modifications and combinations of various approaches to find the best algorithm. The applicability of this model is demonstrated on real data, in particular, data on transactions of 1,300,000 Sberbank clients with verification in the next 1/3/6 months. Moreover, the paper also shows that recommendations made using the client graph (by his nearest neighbours) turned out to be more relevant than recommendations made on a random sample of 100/1000 clients. The presented results reflect the superiority of the obtained recommender system for the metric in question.