Graph neural networks is a powerful generalization of a whole class of neural (and not only neural) architectures on the data, organized in a graph. The graph representation is on itself a powerful tool, allowing to explicitly show dependencies between features or attributes and abstract this prior from the architecture level to the level of the data.

Particular special cases of graph networks are already successfully used in particle physics and, therefore, implementing such tools for ROOT in general as a leading particle physics data analytics framework, and TMVA in particular is of utmost importance.

Implementing a generalized version of graph networks as we propose in this context, as opposed to implementing every possible architecture like Deep Sets or Message Passing NNs, will ensure the wide applicability and the ease of modification and adaptation for a particular task.



Stanislav Lukyanenko


  • Lorenzo Moneta
  • Sitong An
  • Sergei Gleyzer