A New Neural Network Model for Contextual Processing of Graphs
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.
KeywordsSingular Value Decomposition Neural Network Model Hide Unit Output Unit Maximum Absolute Error
Unable to display preview. Download preview PDF.
- 1.Bianchini, M., Gori, M., Scarselli, F.: Recursive processing of cyclic graphs. In: Proc. of WCCI-IJCNN 2002, vol. 1, pp. 154–159 (2002)Google Scholar
- 3.Fahlman, S.E., Lebiere, C.: The cascade-correlation learning architecture. Technical Report CMU-CS-90-100, Carnegie Mellon (August 1990)Google Scholar
- 4.Frasconi, P., Gori, M., Küchler, A., Sperduti, A.: From sequences to data structures: Theory and applications. In: Kolen, J.F., Kremer, S.C. (eds.) A Field Guide to Dynamical Recurrent Networks, ch. 19, IEEE Press, Inc., Los Alamitos (2001)Google Scholar
- 5.Hammer, B., Steil, J.J.: Tutorial: Perspectives on learning with rnns. In: Proc. of ESANN 2002, pp. 357–368. D-side (2002)Google Scholar