Data Imputation with Adversarial Neural Networks for Causal Discovery from Subsampled Time Series

A relevant and challenging problem is causal discovery from time series data. This helps to understand dynamics events present in real world scenarios. However, causal interactions may occur at a timescale faster than the measurement frequency, resulting in a subsampled time series. This can lead to significant errors during causal discovery. We propose an approach based on imputing the missing data using adversarial neural networks to try to recover the true causal structure. The trained model is fed with the subsampled time series in order to generate data that behaves similarly to the original time series, so that the original causal structure can be recovered. The completed data series is then fed to a causal discovery algorithm. Experimental results on several synthetic dynamic models show that the imputed data time series is close to the original one, and that the causal structure derived from this data resembles the correct causal structure.