GC-Mixer: A Novel Architecture for Time-varying Granger Causality Inference

Abstract

The neural network has emerged as a practical approach to evaluate the Granger causality in multivariate time series. However, most existing studies on Granger causality inference are based on time-invariance. In this paper, we propose a novel MLP architecture, Granger Causality Mixer (GC-Mixer), which extracts parameters from the weight matrix and imposes the hierarchical group lasso penalty on these parameters to infer time-invariant Granger causality and automatically select time lags. Furthermore, we extend GC-Mixer by introducing a multi-level fine-tuning algorithm to split time series automatically and infer time-varying Granger causality. We conduct experiments on the VAR and Lorenz-96 datasets, and the results show that GC-Mixer achieves outstanding performances in Granger causality inference.