The data alone cannot differentiate between the fork, the pipe & the collider. You need a scientific model of how X/Y/Z work. Self-supervised generative deep learning/VAE/etc don't contain a scientific model & learn purely from data => the same limitation. Figure: @rlmcelreath