Abstract
With the advancement of neural networks, diverse methods for neural Grangercausality have emerged, which demonstrate proficiency in handling complex data,and nonlinear relationships. However, the existing framework of neural Grangercausality has several limitations. It requires the construction of separatepredictive models for each target variable, and the relationship depends on thesparsity on the weights of the first layer, resulting in challenges ineffectively modeling complex relationships between variables as well asunsatisfied estimation accuracy of Granger causality. Moreover, most of themcannot grasp full-time Granger causality. To address these drawbacks, wepropose a Jacobian Regularizer-based Neural Granger Causality (JRNGC) approach,a straightforward yet highly effective method for learning multivariate summaryGranger causality and full-time Granger causality by constructing a singlemodel for all target variables. Specifically, our method eliminates thesparsity constraints of weights by leveraging an input-output Jacobian matrixregularizer, which can be subsequently represented as the weighted causalmatrix in the post-hoc analysis. Extensive experiments show that our proposedapproach achieves competitive performance with the state-of-the-art methods forlearning summary Granger causality and full-time Granger causality whilemaintaining lower model complexity and high scalability.