Advanced causal discovery with deep learning
Xun Zheng and colleagues’ DAGs with NO TEARS paper (Zheng et al., 2018), which we introduced in the previous chapter, ignited excitement in the causal discovery community and inspired a whole new line of research on gradient-based methods.
The fact that graph search could be carried out using continuous optimization opened up a path for integrating causal discovery with techniques coming from other deep learning areas.
One example of a framework integrating such techniques into the realm of causal discovery is DECI – a deep learning end-to-end causal discovery and inference framework (Geffner et al., 2022).
DECI is a flexible model that builds on top of the core ideas of the NO TEARS paper. It works for non-linear data with additive noise under minimality and no hidden confounding assumptions.
In this section, we’ll discuss its architecture and major components and apply it to a synthetic dataset, helping...
