Imagine you could feed your data directly into a machine learning model and have it learn, without the need for any manual labelling. Imagine predictions in data-sparse environments could be improved without needing to label more data or even adjust the model structure. Thanks to Deep Graph Infomax — a graph machine learning algorithm that uses the graph structure to understand patterns in the data associated with each node — these are welcome realities.
Deep Graph Infomax is an unsupervised training procedure. A typical supervised task matches input data against input labels, to learn patterns in the data that are associated with the labels. Deep Graph Infomax skips the labels, and instead guides a model to learn from how input data points are connected as a graph, by understanding what should or shouldn’t be linked. It can thus be used for both unsupervised representations for these points and also pre-training a model to improve performance of semi-supervised tasks without much labelled data.
A graph is a collection of nodes and the edges between them, where the edges represent some connection or relationship between the nodes. Graph machine learning takes conventional machine learning models like multilayer perceptron and long short-term memory that apply to single data points (representing individual nodes) and augments them to use the information provided by each data point’s connections.