Graph Convolutional Networks for Geometric Deep Learning

Graph convolutions are very different from graph embedding methods that were covered in the previous installment. Instead of transforming a graph to a lower dimension, convolutional methods are performed on the input graph itself, with structure and features left unchanged. Since the graph remains closest to its original form in a higher dimension, the relational inductive bias is therefore much stronger. There is a type of inductive bias in every machine learning algorithm. In vanilla CNNs for example, the minimum features inductive bias states that unless there is good evidence that a feature is useful, it should be deleted.