We show how Interaction Networks could be used for jet tagging at the Large Hadron Collider.
We take as an example the problem of identifying high-pT H->bb decays exploiting both jet substructure and secondary vertices from b quarks. We consider all tracks produced in the hadronization of the two b’s and represent the jet both as a track-to-track and a track-to-vertex interaction. The representations of the two interactions are learned training two dense neural networks. The derived information is used to train a classifier of H->bb jets. Interaction networks achieve state-of-the-art discrimination performances, even when the training is prevented from learning to exploit the jet mass value as discriminating information.