Network embedding for brain connectivity

In various contexts, network embedding techniques have been developed forperforming analysis on a single graph. This approach has been proven to workfor different applications, such as node classification or link prediction. A goodnetwork embedding algorithm is capable to capture only the relevant features ofthe graph and to reproduce them in a low-dimensional Euclidean space.In the context of neuroscience, networks are currently used for representing thesystem of connections in the brain with the purpose of determining the charac-teristics of a pathological brain. However, discriminating a healthy human brain connectivity network from a clinical one using common network descriptors couldbe misleading. In fact, a difference in the currently used graph measures couldnot be detected or could be insufficient for the discrimination. For this reason,we investigate network embedding and extend its fields of application to humanconnectivity network. Moreover, we use the embedding for performing graphs’comparison. Finally, we propose a definition of the representative network of aset of graphs. This representative captures the properties which are in common inthe group. In this way, we confirm the existence of a healthy signature, namely abrain structure which is shared by all healthy individuals.