Facing graph classification problems by a multi-agent information granulation approach

In this paper, we investigate an agent-based prototypical framework for graph classification by offering a comparison with a standard graph neural network on several heterogeneous datasets. The design of the algorithm is based on a swarm of agents orchestrated via evolutionary optimization in charge of finding meaningful substructures from unstructured training data. The algorithm is tailored to face pattern recognition problems by a suitable embedding from the graph domain to Euclidean space, also exploiting local metric learning. Results on five open-access datasets of fully labeled graphs show interesting performances in terms of accuracy, counterbalanced by relatively high computational complexity.