Abstract:

Lack of data and validation of their relevance are essential issues in the transition to online financial services. Increasing data volumes and fast processing enable financial intermediaries to expand retail, social media services. The options proposed in the article for introducing new datasets into recommender systems to improve the efficiency of recommendations. On the one hand, long-tail datasets represent missing data and improve the accuracy of recommendations; on the other hand, they allow you to validate recommendations prepared in a standard way.

In this article, we have proposed three schemes for using long-tail data in generative adversarial networks (GANs):

1) Input of data related to the normal distribution for the preparation of recommendations by generative networks (GN) and their validation using a discriminatory network (DN) into which the data of the "long tail" is input.

2 Expanding the datasets that the GAN works with by introducing tail data in both the GN and DN, thus increasing the relevance of the recommendations.

3) Fragmentation of the distribution of data by unique characteristics ("head" - "tail," and each group has different classes and levels of data by frequency of interest), the inclusion of distinct fragments (sets) of data in the GAN, and preparation of various versions of the recommendations.