Inferring the Structure of Graph Grammars from Data

Graphs can be used to represent such diverse entities as chemical compounds, transportation networks, and the world wide web. Stochastic graph grammars are compact representations of probability distributions over graphs. We present an algorithm for inferring stochastic graph grammars from data. That is, given a set of graphs that, for example, correspond to a set of chemical compounds, all of which have some desirable property, the algorithm uncovers the structure shared by the graphs and represents it in the form of a stochastic graph grammar. The inferred grammar assigns high probability to the graphs from which it was learned and low probability to other graphs. We report results of preliminary experiments in which inferred graph grammars are compared to target grammars used to generated training data.