Implementation of Learning Convolutional Neural Networks for Graphs by Mathias Niepert, Mohamed Ahmed, Konstantin Kutzkov, ICML 2014-2023. https://scholar.google.com.sg/scholar?oi=bibs&cluster=9917957179670149192&btnI=1&hl=en

Required Modules:

• networkx
• pynauty (See https://web.cs.dal.ca/~peter/software/pynauty/html/install.html for installation)
• tensorflow, keras
• numpy

Python 3 Implementation that uses the networkx library to construct input graphs, and transform them into receptive fields, which are tensors that form the inputs into a Convolutional Neural Network(CNN) for learning of subgraph properties and approximate isomoprhisms, based on the choice of node labelling procedures(e.g Between-ness Centrality). In particular, the algorithm does the following:

• Select node sequence according to the labelling procedure
• Assemble local neighbourhoods from subgraphs centered on the nodes selected the node sequence
• Perform graph normalization on neighbourhoods to form receptive fields
• Learn subgraph properties using receptive fields as inputs

The Conv1D, Dense, Dropout and softmax layers from keras are used to construct the deep CNN that performs classification of graphs according to their respective subgraph receptive fields. This is represented in the pscn.py file that contains the module using the Patchy-San algorithm with a CNN.

Various well known graph datasets for measuring performance on graph kernels were used in evaluating the PSCN algorithm. https://ls11-www.cs.tu-dortmund.de/staff/morris/graphkerneldatasets