A Simple and Effective Neural Model for the Classification of Structured Patterns

Learning from structured data (i.e. graphs) is a topic that has recently received the attention of the machine learning community, which proposed connectionist models such as recursive neural nets (RNN) and graph neural nets (GNN). In spite of their sound theoretical properties, RNNs and GNNs suffer some drawbacks that may limit their application. This paper outlines an alternative connectionist framework for learning discriminant functions over structured data. The approach, albeit preliminary, is simple and suitable to maximum-a-posteriori classification of broad families of graphs, and overcomes some limitations of RNNs and GNNs. The idea is to describe a graph as an algebraic relation, i.e. as a subset of the Cartesian product. The class-posterior probabilities given the relation are reduced to products of probabilistic quantities estimated using a multilayer perceptron. Experimental comparisons on tasks that were previously solved via RNNs and GNNs validate the approach.