A NewModelforearning in raphDomains
Dipartimento diIngegneria dell ' Informazione MarcoGori, Gabriele Monfardini, Franco Scarselli · 2005
Inseveral applications theinformation isnaturally represented bygraphs. Traditional approaches cope with graphi- caldata structures using apreprocessing phase which transforms thegraphs into a setofflat vectors. However, inthis way, important topological information maybelost andtheachieved results mayheavily depend onthepreprocessing stage. This paper presents anewneural model, called graph neural network (GNN), capable ofdirectly processing graphs. GNNsextends recursive neural networks andcanbeapplied onmostofthepractically useful kinds ofgraphs, including directed, undirected, labelled andcyclic graphs. A learning algorithm forGNNsisproposed andsomeexperiments arediscussed which assess theproperties ofthemodel.