# Ljupčo Todorovski: Machine learning on graphs

Note: we meet in the non-standard room 3.04.

**Abstract:** As a follow-up to the previous seminar, which presented knowledge graphs obtained from libraries in the Agda programming language, we will look at machine learning approaches that learn from graphs.

A typical approach to learning from graphs embeds the nodes of a graph into a Euclidean space of predefined dimension. An embedding is deemed good when it maps "similar" nodes to similar vectors. Simple approaches rely on embeddings that only consider the properties of individual nodes. Advanced embeddings consider the position and connections of the embedded node in the graph.

A graph embedding transforms a set of nodes into a data table, which allows us to use classical machine learning algorithms and models to predict missing links. A limitation of predictive models learned from embeddings is that they cannot immediately explain their predictions. An alternative approach is direct machine learning from the nodes and connections in the graph, the results of which are interpretable models that can explicate their forecasts.