Normalizing flows for conditional independence testing

Abstract Detecting conditional independencies plays a key role in several statistical and machine learning tasks, especially causal discovery algorithms, yet it remains highly challenging problem due to dimensionality complex relationships presented data. In this study, we introduce LCIT (Latent representation-based Conditional Independence Test)—a novel method for independence testing based on representation learning. Our main contribution involves hypothesis framework which test the between X Y given Z , first learn infer latent representations of target variables that contain no information about conditioning variable . The are then investigated any significant remaining dependencies, can be performed using conventional correlation test. Moreover, also handle discrete mixed-type data general by converting into continuous domain via variational dequantization. empirical evaluations show outperforms state-of-the-art baselines consistently under different evaluation metrics, is able adapt really well both nonlinear, high-dimensional, mixed settings diverse collection synthetic real sets.