Russian researchers cut neural network memory use on graphs of connections by 100x and tested the method on VKontakte
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Huge networks of connections can be processed on much less hardware, so such models will become more accessible.
VK's research team proposed a new way of storing data for graph neural networks, which work with connections between objects. According to the researchers, memory use falls 100-fold and data preparation speeds up 2.5-fold. On a test graph of 194 million users and 28 billion connections, the method was tried on friend recommendations on VKontakte: friend additions from suggestions rose 16% (11.5% in the detailed test). The work has been accepted at the CIKM 2026 conference.