Abstract
We address the problem of distinguishing cause from effect in bivariatesetting. Based on recent developments in nonlinear independent componentanalysis (ICA), we train nonparametrically general nonlinear causal models thatallow non-additive noise. Further, we build an ensemble framework, namelyCausal Mosaic, which models a causal pair by a mixture of nonlinear models. Wecompare this method with other recent methods on artificial and real worldbenchmark datasets, and our method shows state-of-the-art performance.
Quick Read (beta)
loading the full paper ...