Causal order identification to address confounding: binary variables

This paper considers an extension of the linear non-Gaussian acyclic model (LiNGAM) that determines causal order among variables from a dataset when are expressed by set equations, including noise. In particular, we assume binary. The existing LiNGAM assumes no confounding is present, which restrictive in practice. Based on concept independent component analysis (ICA), this proposes extended framework mutual information noises minimized. Another significant contribution to reduce realization shortest path problem, distance between each pair nodes expresses associated value, and with minimum sum (KL divergence) sought. Although p! values should be compared, dramatically reduces computation present. proposed algorithm finds globally optimal solution, while approaches locally greedily seek based hypothesis testing. We use best estimator sense Bayes/MDL correctly detects independence for estimation. Experiments using artificial actual data show version achieves significantly better performance, particularly