Master Assignment
temporal causal discovery and structure learning with attention-based convolutional neural networks
Type: Master CS
Period: Feb, 2018 - Aug, 2018
Student: Nauta, M. (Meike, Student M-CS)
Date final project: August 29, 2018
Supervisors:
Abstract:
We present the Temporal Causal Discovery Framework (TCDF), a deep learning framework that learns a causal graph structure by discovering causal relationships in observational time series data. TCDF uses attention-based convolutional neural networks to detect correlations between time series and subsequently performs a novel validation step to distinguish causality from correlation. By interpreting the internal parameters of the convolutional networks, TCDF can also discover the time delay between a cause and the occurrence of its effect. Our framework can learn both cyclic and acyclic causal graphs, which can include confounders and instantaneous effects. The graph reduction step in TCDF removes indirect causal relationships to improve readability of the constructed graph. Using the representational power of deep learning, TCDF removes idealized assumptions upon the data that existing, usually statistical, causal discovery methods make. Experiments on actual and simulated time series data show state-of-the-art performance of TCDF on discovering causal relationships in continuous, noisy and (non-)stationary time series data. Furthermore, we show that TCDF can circumstantially discover the presence of hidden onfounders. Our broadly applicable framework can be used to gain novel insights into the causal dependencies in a complex system, which is important for interpretation, explanation, prediction, control and policy making.