Fast and Accurate Causal Inference from Time Series Data
Yuxiao Huang, Samantha Kleinberg · The Florida AI Research Society · 2015
Causal inference from time series data is a key problem in many fields, and new massive datasets have made it more critical than ever. Accuracy and speed are primary factors in choosing a causal inference method, as they determine which hypotheses can be tested, how much of the search space can be explored, and what decisions can be made based on the results. In this work we present a new causal inference framework that 1) improves the accuracy of inferences in time series data, and 2) enables faster computation of causal significance. Instead of evaluating relationships individually, using only features of the data, this approach exploits the connections between each causal relationship's relative levels of significance. We provide theoretical guarantees of correctness and speed (with an order of magnitude improvement) and empirically demonstrate improved FDR, FNR, and computation speed relative to leading approaches.