Sberbank: RePlay (framework)

RePlay is a framework for creating and comparing recommendation systems and services on big data to generate recommendations offline and online.

2023: Opening access to RePlay

Sber November 30, 2023 announced that he had opened his own recommendation technologies to the entire developer community and. date scientists countries Bank posted the next version of its RePlay library in the public domain.

Now the library includes transformers - State-of-the-Art (SotA) algorithms for RecSys, which allow users to create their recommended end-to-end systems based on the best industrial solutions.

The updated version of RePlay supports the development of models not only on PySpark, but also on CPU (Pure Python), GPU and Multi-GPU. For developers, this is an opportunity to generate recommendations in any mode: from classic offline scenarios to the most advanced online scenarios, stressed in Sberbank.

"Sberbank is actively developing recommendation systems so that each client receives personalized experience when interacting with our products and services. Today we go even further and open up access to advanced SotA algorithms in the updated version of RePlay for all developers in our country. Models built on the basis of transforming algorithms have shown their effectiveness in creating ready-made recommendation systems in a wide range of scenarios and are the most advanced architecture of deep neural networks. The framework contains more than 15 algorithms for various recommendation tasks and more than 10 metrics for their evaluation. I am sure that the new version of the RePlay library will reduce the time of developers due to convenient tools for preprocessing data and comparing models, "said Ivan Kuzmin, Director of the Data and Recommendation Systems Department of B2C Sberbank.