A new arXiv paper, 'Learning Non-Gaussian SEMs with Conditional Risk, and Why the Direction of Recursion Matters,' examines how structural equation models (SEMs) can be learned from data. The abstract notes that variance-based methods recover a Gaussian linear SEM when error variances are equal or suitably constrained, typically by recursively identifying source nodes with the smallest conditional variance.

The title points to two extensions: moving beyond Gaussian assumptions to non-Gaussian SEMs, and using conditional risk rather than variance alone. The paper also highlights that the direction of recursion—whether the algorithm proceeds top-down or bottom-up—affects the outcome.

Because the abstract is truncated, the full details of the proposed method are not available. But the central claim is that the direction of recursion matters for learning such models, a factor that variance-based methods may overlook.