Diviyan Kalainathan, Université Paris-Saclay
We created new types of causal models at the crossroad of Structural Equation Models and Deep Learning. Such models inherit the effective learning and generalization capabilities of modern machine learning tools and the explanatory power of causal models. Based on observational data only (as opposed to experimental/interventional data), we can create data generative models for (eventually large) sets of variables, whose structure is revealing of plausible mechanisms. Our framework, called Causal Generative Neural Network (CGNN) takes as input a un-oriented graph draft (skeleton) and tests various graph structures by minimizing the reproduction error of the joint distribution.
Once trained, such models may have various utilisations: (1) For scientific purposes, one may confront the structure with the opinion of experts and prioritize confirmatory experiments. (2) For research purposes, one may share the simulator or simulated data with a broader research community, while protecting the confidential nature of the original data. One may generate a lot more data than the original data to facilitate benchmarking algorithms. (3) For educational purposes, one may give access to the simulator to student so they can emulate experiments at low cost and without endangering human subject, environments, or disturbing any real system.