Detecting Causality using Deep Gaussian Processes

摘要: Convergent cross mapping (CCM) is a state space reconstruction (SSR)-based method designed for causal discovery in coupled time series, where Granger causality may not be applicable due to separability assumption. However, CCM requires large number of observations and robust observation noise which limits its applicability. Moreover, variants, the SSR step mostly implemented with delay embedding parameters usually need selected using grid search-based methods. In this paper, we propose Bayesian version deep Gaussian processes (DGPs), are naturally connected neural networks. particular, adopt framework SSR-based carry out key steps DGPs within non-parametric probabilistic principled manner. The proposed approach first validated on simulated data then tested used obstetrics monitoring well-being fetuses, i.e., fetal heart rate (FHR) uterine activity (UA) signals last two hours before delivery. Our results indicate that UA affects FHR, agrees recent clinical studies.