Interpretable Multi-Scale Neural Network for Granger Causality Discovery

We propose a novel multi-scale neural network for Granger causality discovery (MSNGC) in multivariate time series. Compared with existing counterparts, our model avoids the explicit data segmentation between series and between time lags for the first time. By extracting diverse causal information from the data with different delay ranges and then integrating them effectively via the learned attention weights, it can capture the complete causal relationships between all the series and provide accurate weighted adjacency matrix estimation. Thus, further interpretable inference can be better supported. Specifically, we propose a consistency-based thresholding algorithm for binary causal structure inference, an effect sign detection method to distinguish the positive causal effects from the negative ones, and a self-adaptive lag discovery algorithm to identify the lagged time points. Experiments on multiple benchmarks demonstrate that our model significantly outperforms the state-of-the-art methods.