Abstract
We present Geo2DR, a Python library for unsupervised learning ongraph-structured data using discrete substructure patterns and neural languagemodels. It contains efficient implementations of popular graph decompositionalgorithms and neural language models in PyTorch which are combined to learnrepresentations using the distributive hypothesis. Furthermore, Geo2DR comeswith general data processing and loading methods which can bring substantialspeed-up in the training of the neural language models. Through this we providea unified set of tools and design methodology to quickly construct systemscapable of learning distributed representations of graphs. This is useful forreplication of existing methods, modification, or even creation of novelsystems. This work serves to present the Geo2DR library and perform acomprehensive comparative analysis of existing methods re-implemented usingGeo2DR across several widely used graph classification benchmarks. We show ahigh reproducibility of results in published methods and interoperability withother libraries useful for distributive language modelling.