The usage of graphs to model and solve ML problems is becoming very popular. In this talk we'll review example implementations of Graph ML and how they assist by generating models that take advantage of the individual data point features combined with the graph structure. As an elaborate example, I will present how we use Graph ML to classify document sensitivity without looking at the content.
Many of the problems we face as scientists can be natively modelled on graph structures that captures both the attributes of objects (nodes) as well as the relationships between them (edges). Such structures include social networks, computer networks, code execution flows, molecular structures, and many more. We will skim briefly through examples (from the last decade) of how Graph ML has been applied successfully to problems including node classification, edge (link) prediction, and whole graph predictions (e.g. in classifying molecules).
We will then demonstrate how we at Authomize use Graph Neural Networks to classify different types of sensitive documents by looking at the document’s metadata. The metadata includes the title as well as relationships of the document (node) to other entities (such as employees, or its place in the hierarchy). GNNs allow us to classify these documents without inspecting the content. If time allows, we will dive into how models are delivered in production.