scroll
Research Area
Graph Neural Networks

Classical optimization frameworks and standard statistical approaches ignore the friction and dynamism of the real-world. Rarely is there an accurate map of how the supply-chain or health system functions, even as a static snapshot. We aim to do better by learning the map – directly from available data.

Classical optimization frameworks and standard statistical approaches ignore the friction and dynamism of the real-world. Rarely is there an accurate map of how the supply-chain or health system functions, even as a static snapshot. We aim to do better by learning the structure – directly from available data.

Defining a Graph

Supply chains are networks. The art world is a network. Ecosystems are networks. Networks are spatial, temporal and dynamic. MACRO-EYES advisory board member Stefanie Jegelka, PhD is an expert on graph neural networks. A graph is a structure that consists of entities [think: people] and the connections between entities [think: friendships]. A graph can consist of drugs, or even molecules; connections can be observed through the interactions that occur (and the adverse effects that may result) when a patient is prescribed more than one medication. It is difficult to describe mathematically the multiple connections between entities and how the links change in time and across space.

What if there’s no map? In many of the environments in which MACRO-EYES works, we know we’re dealing with a network – which if optimized, could save lives and resources – but no one knows exactly how it works end-to-end. Change is the constant in many of the systems that matter to MACRO-EYES. Each participant has a partial view and available data reflects this. It may be a network of networks. How then do the different layers of the network interact? There may be models of how the network should work, but ignoring the friction and ambiguity of how events unfold on the ground can be dangerous.

How to machine learn the relationships between entities from available data [satellite imagery to routinely collected data to data that must be scraped from social media]? How to encode the structure of those relationships? Take the example of a health system: which health facilities share resources? Are some facilities informal supply hubs? Might they have clients in common? Could the clients port information between sites? Do health workers spend time at more than one location? How do rumours spread between facilities? Which facilities are sources of information? Which are hubs for disinformation?

The more effectively a network can be described, the more likely we are to be able to dramatically improve its performance.