A Gradient-based Causal Discovery Framework with Applications to Complex Industrial Processes

The paper introduces Gradient Regularization-based Neural Granger Causality (GRNGC), a novel framework that utilizes $L_{1}$ regularization on the gradient between input and output to infer Granger causality without the need for separate models for each time series. GRNGC demonstrates enhanced flexibility by being compatible with various architectures, including KAN, MLP, and LSTM, and shows superior performance over existing models in numerical simulations and real-world applications, particularly in reconstructing gene regulatory networks. This approach significantly reduces computational overhead, making it valuable for practitioners dealing with complex industrial processes and large datasets.