Introduction
Continuous-time dynamics models, particularly Neural Ordinary Differential Equations (NODEs), have emerged as powerful tools for modeling complex time-series data. By treating system evolution as a continuous flow, they offer a physically intuitive alternative to discrete-step models. However, these models often function as "black boxes," providing high-accuracy forecasts without revealing the underlying causal mechanisms. In critical domains like tsunami forecasting or climate science, understanding why a prediction was made is as important as the prediction itself. This work introduces a framework to extract Granger causality directly from these continuous-time models.
Methodology: Neural Granger Causality
The researchers propose a mechanism to "mine" causality by enforcing sparsity in the input layers of the neural network. The core concept relies on the definition of Granger causality: a variable $x_j$ is causal for $x_i$ if its past values help predict the future of $x_i$.
To implement this, the framework utilizes:
- Component-wise Architectures: Instead of a single monolithic network, separate multi-layer perceptrons (MLPs) represent the derivative of each state variable.
- Sparsity-Promoting Loss: A group Lasso-style $L_2$ norm penalty is applied to the columns of the first weight layer ($W^{i,1}$). This forces the model to ignore non-causal inputs by driving their weights to zero.
- Neural Delay Differential Equations (NDDEs): To account for physical transport delays (e.g., wave travel time), the model is extended to handle lagged inputs.
Experimental Validation
The approach was validated across three benchmarks:
- Lorenz-96 System: A standard climate modeling benchmark used to test spatial dependency recovery.
- Mackey-Glass Equation: A chaotic system with inherent time delays.
- Tsunami Dataset: Real-world synthetic simulations in the Strait of Juan de Fuca, involving surface elevation readings from multiple gauges.
Results and Discussion
The proposed method successfully recovered known causal structures while maintaining high forecasting accuracy. In the Lorenz-96 system, the model perfectly identified the "neighbor-only" dependencies inherent in the governing equations.
In the tsunami application, the NDDE models outperformed standard NODEs. The "mined" causality aligned with physical wave propagation: Gauge 702 (at the strait's entrance) was found to influence interior gauges (901 and 911) with specific time delays (approx. 24 and 60 minutes). These delays correspond to the physical "sloshing" or seiche effects within the local topography.
Conclusion
By integrating sparsity-promoting penalties into continuous-time neural architectures, this research bridges the gap between predictive performance and interpretability. The ability to automatically detect both causal variables and their associated time lags makes this framework particularly valuable for high-stakes physical systems where understanding the "why" behind a forecast is essential for decision-making.
Summary updated