Temporal Causal Inference with Time Lag.

Area of Science:

• Statistics
• Machine Learning
• Time Series Analysis

Background:

• Causal inference in time series is crucial for understanding temporal variable interactions.
• Time lag is a common challenge, where past events influence future outcomes with a delay.
• Existing methods often require predefining a fixed time window, which is not always feasible.

Purpose of the Study:

• To develop a method for simultaneously learning causal relationships and time lags from time series data.
• To address the limitations of fixed time window approaches in modeling time-varying lags.
• To improve the accuracy of causal inference in the presence of temporal delays.

Main Methods:

• Proposed a probabilistic decomposed slab-and-spike (DSS) model.
• Utilized a pair of decomposed spike-and-slab variables for model coefficients.
• Employed an efficient expectation propagation (EP) algorithm for parameter inference.

Main Results:

• The DSS model successfully learns both causal relations and time lags simultaneously.
• Experimental results on synthetic and real-world data demonstrate the method's effectiveness.
• Identified time lags were validated by domain knowledge in real-world applications.

Conclusions:

• The proposed DSS model offers an effective approach for causal inference with time lags in time series.
• Simultaneous learning of causality and lag improves understanding of temporal dynamics.
• The method provides a flexible alternative to fixed time window assumptions.