Researchers at AI VK Research have developed a new method for scaling graph neural networks to handle massive datasets. The approach is designed for industrial applications involving hundreds of millions of objects and billions of connections, significantly reducing memory requirements and accelerating data preparation.
The team tested the technology on an anonymized social graph containing 194 million users and 28 billion friendship connections. The new data representation method reduced memory usage by more than 100 times. Additionally, the solution accounts for temporal changes in the graph and speeds up training data preparation by 2.5 times.
Practical implementation within the VKontakte social network yielded measurable results. The number of users adding friends from recommendations increased by 16%, while the proportion of users who added at least one recommended friend grew by 11.5%.
