Multi-layer perceptron learning in the domain of attributed graphs

We propose a multi-layer perceptron for learning on data represented in terms of attributed graphs. The approach is based on the idea to associate each simple perceptron with an attributed weight graph and to provide a concept similar to the inner product of vectors in the domain of graphs. This is achieved by the Schur-Hadamard inner product of graphs. To provide a supervised learning mechanism we customize the feed-forward pass, the error-back-propagation algorithm, and the error correcting rule. In first experiments, the proposed algorithm is successfully applied to function the regression and classification tasks. The results show better performance than support vector and nearest neighbor classifiers.