A Gradient-based Causal Discovery Framework with Applications to Complex Industrial Processes
Researchers have developed a new gradient-based causal discovery framework called GRNGC, designed to overcome limitations in existing neural network-based Granger causality models. GRNGC reduces computational costs by using a single time series prediction model instead of component-wise models and enhances the capture of complex interactions by applying L1 regularization to the gradient between the model's input and output. This flexible framework can be implemented with various architectures like KAN, MLP, and LSTM, and has demonstrated superior performance and reduced overhead on multiple benchmark datasets, including DREAM and CausalTime, as well as real-world gene regulatory network datasets. AI
IMPACT Introduces a more computationally efficient and flexible method for causal discovery, potentially improving applications in complex systems like gene regulatory networks.