Causal discovery for time series with constraint-based model and PMIME measure

Antonin Arsac (1) , Aurore Lomet (1) , Jean-Philippe Poli (1)

1 LI3A (CEA, LIST) - Intelligence Artificielle et Apprentissage Automatique (CEA, LIST)

Causality defines the relationship between cause and effect. In multivariate time series field, this notion allows to characterize the links between several time series considering temporal lags. These phenomena are particularly important in medicine to analyze the effect of a drug for example, in manufacturing to detect the causes of an anomaly in a complex system or in social sciences... Most of the time, these complex systems do not rely on the assumption of linearity required in many machine learning methods. To circumvent this problem, we present in this paper a new approach for discovering causality in time series data that combines a causal discovery algorithm with an information theoretic-based measure. Hence the proposed method allows inferring both linear and non-linear relationships and building the underlying causal graph. We evaluate the performance of our approach on several simulated datasets, showing promising results.

Keywords: artificial intelligence; fuzzy logic; online learning; machine learning; complex systems; causality; causal graph; statistical analysis

Domains: Information Theory [math.IT]; Statistics [math.ST]; Machine Learning [stat.ML]

Antonin Arsac, Aurore Lomet, Jean-Philippe Poli. Causal discovery for time series with constraint-based model and PMIME measure. When Causal Inference meets Statistical Analysis, Apr 2023, Paris, France. ⟨cea-04138529⟩