Concomitant DAG Learning: On the Roles of Noise Adaptivity, Sparsity, and Non-negativity
CoLiDE introduces a novel convex score function for sparsity-aware Directed Acyclic Graph (DAG) inference that jointly estimates the DAG structure and unknown exogenous noise levels. This method is robust against heteroscedastic noise profiles by making the score function noise-adaptive, removing coupling between sparsity regularization and noise variance.