Learning optimal policies through causal discovery in social networks

This research aims to learn optimal policies by inferring the causal effects of intervening in a network of heterogeneous individuals. Network Analysis is a set of integrated techniques to depict relations among agents and analyse the social structures that emerge from the recurrence of these relations, but it does not provide a framework to describe what can happen under what-if scenarios. Inferring the causal effects of a network intervention can inform policy-makers of the best course of action. For example, when isolating individuals from society to stop the spread of COVID-19, politicians need to understand the healthcare and economic impact of these interventions.

The main challenge learning optimal such policies is that since the underlying causal graph is often unknown, it is not immediate how to infer the consequences of executing a given policy. This research presents a causal reinforcement learning framework that can "imagine" counterfactual scenarios to learn personalised policies from limited network data. This research will investigate efficient trajectory sampling techniques to model counterfactual outcomes. One potential approach is sampling trajectories from candidate policies given by a Generative Flow Network. We will also study how to insert human knowledge into the learning process in the form of structural shape constraints. These constraints are semantically meaningful regularises that guide the behaviour of the model in out-of-distribution regimes by introducing input-output relations that hold everywhere. For instance, monotonicity is a popular shape constraint that can express simple and interpretable constraints such as: "if age increases, then the risk of a heart attack increases", "if a treatment is more expensive, it will be less accessible to an individual with less purchasing power".