Graphical models, in particular, that have the ability to validate hypothesized associations from multivariate data and discover novel, undocumented, causal associations have seen increasing applications in healthcare that has witnessed a surge in digitization and adoption of data-driven and evidence-based approaches to assist in decision making. Structure learning algorithms that form the core of graphical modeling and aim to decipher associations between the variables of interest from the given data fall under three broad categories, namely constraint-based, search and score, and hybrid. These computationally intensive algorithms need parallelization across HPC environments for enhanced performance. In this presentation, we will discuss the performance of established structure learning algorithms across POWER9 systems with and without GPUs across large data sets. Our approach will be open-sourced for enhanced transparency and reproducibility and adoption in education and research.
Continue the conversation in Slack