Collective causal inference with lag estimation

Abstract

Abstract Temporal causal inference is a fundamental task in time series analysis and has attracted increasing attention in recent years. In many applications, we are presented with multiple target variables rather than a single one, and the relationships exist among these target variables. Unfortunately, most existing techniques focus on the causalities of the single response variable and assume the data to be i.i.d. over time, ignoring value information. In terms of temporal modeling, an unavoidable issue is the existence of time lag among different time series. That is, a past evidence would take some time to cause a future effect instead of an immediate response. Several methods have been developed to explicitly construct temporal causal models. However most of the previous researches aiming to learn the causal relationships only treat the time lag as a predefined constant, which may miss important information or include noisy information if the time lag is set too small or too large. We address these issues by introducing a method of collective causal inference with lag estimation. Extensive empirical evaluations are conducted on both synthetic and two real-world datasets, and the results demonstrate that our proposed model is superior to the traditional methods in learning the temporal causal relationships.