Linear scaling causal discovery from high-dimensional time series by dynamical community detection
Proposes a linear-scaling causal discovery framework based on dynamical community detection for high-dimensional time series.
Key Findings
Methodology
The approach employs the differentiable Information Imbalance (DII) metric to automatically identify dynamical communities by optimizing variable weights via gradient descent. It involves three steps: inferring autonomous sets, detecting communities, and constructing a community causal graph. Multiple time delays τ are incorporated to improve detection in complex systems. The method scales linearly with variable count, demonstrated on systems with up to 80 variables, maintaining high accuracy and efficiency.
Key Results
- In high-dimensional systems with 80 variables, the algorithm accurately reconstructed ground-truth causal structures, with adjusted mutual information (AMI) exceeding 0.95. Tests on 15 coupled Logistic maps, 5 Lorenz oscillators, and two Lorenz 96 systems showed consistent community detection and causal links with minimal errors, and computational time scaled linearly.
- Varying the threshold ε, the community structures remained stable, with high validation scores. Compared to state-of-the-art methods like PCMCI, the proposed approach demonstrated superior scalability and accuracy, especially beyond 50 variables, effectively capturing multi-variable interactions.
- Incorporating multiple delays τ enhanced robustness against nonlinearity and non-stationarity, confirming broad applicability across diverse dynamical systems.
Significance
This work addresses the critical challenge of scalable causal inference in high-dimensional data, providing a computationally efficient framework that captures macro-level causal hierarchies. It advances understanding in fields like neuroscience, climate modeling, and finance by enabling large-scale causal network reconstruction. The method’s ability to reveal system-wide causal structures with linear complexity marks a significant step forward, facilitating analysis of complex systems previously limited by computational constraints.
Technical Contribution
The key technical innovation is the development of a differentiable information imbalance (DII) metric optimized via gradient descent, allowing automatic detection of autonomous variable sets and dynamical communities. The hierarchical community structure clarifies causal pathways at multiple levels, improving interpretability. The approach combines theoretical guarantees with practical efficiency, outperforming traditional conditional independence tests and extending applicability to nonlinear, non-stationary systems.
Novelty
This is the first framework to integrate dynamical community detection with linear scalability for causal discovery in high-dimensional time series. Unlike existing methods relying on exhaustive conditioning sets, it employs an optimization-based approach to identify causal groups efficiently. The hierarchical community structure provides a new macro-level perspective, revealing layered causal relationships that were previously inaccessible in large systems.
Limitations
- Assumes causal sufficiency, neglecting unobserved confounders, which may bias inferred causal links.
- Sensitive to noise, especially at extreme threshold values, risking missed weak connections.
- Performance may degrade in highly nonlinear or non-stationary systems without further robustness enhancements.
Future Work
Future directions include integrating Bayesian confidence measures to quantify link reliability, extending multi-scale and multi-delay detection, and addressing unobserved confounders. Enhancing robustness against noise and nonlinearity, as well as applying the framework to real-world large datasets, will broaden its impact across scientific domains.
AI Executive Summary
In the era of big data, understanding the causal architecture of complex systems remains a formidable challenge. Traditional methods like Granger causality and Transfer Entropy struggle with high-dimensional data due to exponential growth in computational cost. To overcome this, the present study introduces a novel framework based on dynamical community detection, which achieves linear scaling with the number of variables. Central to this approach is the differentiable information imbalance (DII), a metric optimized via gradient descent to automatically identify autonomous variable sets and group them into dynamical communities. These communities represent macro-level units of influence, simplifying the causal network structure.
The methodology involves three key steps: first, inferring autonomous sets for each variable; second, detecting minimal dynamical communities through hierarchical clustering; and third, constructing a community causal graph that encodes the causal hierarchy among groups. Incorporating multiple time delays τ enhances detection in nonlinear and non-stationary systems. Extensive experiments on simulated data—including 15 coupled Logistic maps, 5 Lorenz oscillators, and two Lorenz 96 systems—demonstrate the method’s high accuracy, with adjusted mutual information exceeding 0.95 and errors comparable to ground-truth structures. The results show that the approach maintains efficiency and robustness even as the system size grows.
This work significantly advances causal inference by enabling large-scale, interpretable causal network reconstruction. Its linear complexity and ability to uncover hierarchical causal structures make it highly suitable for applications in neuroscience, climate science, and finance. Despite some limitations—such as assumptions of causal sufficiency and sensitivity to noise—the framework opens avenues for further improvements, including confidence estimation and multi-scale analysis. Overall, this method provides a powerful new tool for scientists seeking to decode the causal fabric of complex, high-dimensional systems.
Deep Dive
Abstract
Understanding which parts of a dynamical system cause each other is extremely relevant in fundamental and applied sciences. However, inferring causal links from observational data, namely without direct manipulations of the system, is still computationally challenging, especially if the data are high-dimensional. In this study we introduce a framework for constructing causal graphs from high-dimensional time series, whose computational cost scales linearly with the number of variables. The approach is based on the automatic identification of dynamical communities, groups of variables which mutually influence each other and can therefore be described as a single node in a causal graph. These communities are efficiently identified by optimizing the Information Imbalance, a statistical quantity that assigns a weight to each putative causal variable based on its information content relative to a target variable. The communities are then ordered starting from the fully autonomous ones, whose evolution is independent from all the others, to those that are progressively dependent on other communities, building in this manner a community causal graph. We demonstrate the computational efficiency and the accuracy of our approach on time-discrete and time-continuous dynamical systems including up to 80 variables.