Abstract

Network data is central to understanding modern complex systems, from social interactions and epidemic spreading to financial transactions and logistics. However, conducting experiments on networks—such as A/B testing or clinical trials—presents unique challenges due to network interference, where a treatment applied to one node impacts neighbouring nodes. This work explores the crucial role of optimization in designing efficient experiments within these networked environments.

We use techniques from Design of Experiments to investigate how constrained optimization techniques, integer programming, and combinatorial optimization can be deployed to solve the "treatment assignment" problems. The primary objective is to maximize the statistical power of the experiment while minimizing the variance of the treatment effect estimators. By leveraging optimization algorithms, we can consider network structure to optimally assign treatments to mitigate the biasing effects of interference.

This work bridges the gap between graph theory, and statistical inference, providing a robust toolkit for modern data scientists and researchers working with connected data. We demonstrate with applications in agricultural experiments, and industrial marketing experiments.