A Relational Cascade Correlation for Structured Outputs

We propose a Relational Neural Network defined as a special instance of the Recurrent Cascade Correlation. The proposed model is designed to deal with classification tasks where classes are organized into generic graphs (e.g., taxonomies, ontologies, etc.). The open challenge is to exploit the knowledge encoded in the relationships among the classes. This is particularly useful when there are many classes poorly represented by labeled examples. Exploiting the relationships we increase the bias, making the generalization more robust. The novelty of the proposed model can be seen from two different perspectives. On one hand, the temporal encoding of the standard recurrent networks is revised with a notion of non-stationary structural unfolding. On the other hand, it can be seen as a novel constructive algorithm that generates the neural network architecture exploiting the class structure. We present the results of an empirical evaluation on a hierarchical document classification task.