Causal Discovery with Unequal Edge Error Tolerance

We introduce an algorithm for causal discovery in the setting of linear additive Gaussian noise models that controls the false positive rate of individual edges. The ability to control individual edge errors has only recently started to gain attention. However, recent methods that make guarantees on the existence of edges can only do so in the limit of infinite samples or are computationally infeasible even for modestly sized graphs. By looking at the residual sum of squares between two vertices we can detect the presence of an edge while controlling the false positive rate for a given number of samples. We call our algorithm E-CUT and compare it to other state-of-the-art algorithms. In some instances, E-CUT achieves a 12 percent reduction in the false negative rate for a given false positive rate over the different algorlthms.