A new neural network model for contextual processing of graphs

Abstract

We propose a novel simple approach to deal with fairly general graph structures by neural networks. Using a constructive approach, the model Neural Network for Graphs (NN4G) exploits the contextual information stored in the hidden units progressively added to the network, without introducing cycles in the definition of the state variables. In contrast to previous neural networks for structures, NN4G is not recursive but uses standard neurons (with no feedbacks) that traverse each graph without hierarchical assumptions on its topology, allowing the extension of structured domain to cyclic directed/undirected graphs. Initial experimental results, obtained on the prediction of the boiling point of alkanes and on the classification of artificial cyclic structures, show the effectiveness of this new approach. © Springer-Verlag Berlin Heidelberg 2006.