Network granger causality with inherent grouping structure

Sumanta Basu, Ali Shojaie, George MichailidisJMLR 2015 · OpenReview

TL;DR

This paper introduces a novel framework for learning complex network structures from time-series data, particularly when the underlying entities (like genes or financial assets) can be naturally grouped. By extending Granger causality with a "group lasso" approach, they achieve more robust and accurate network inference, especially in high-dimensional settings where traditional methods struggle.

What Problem Does This Paper Address?

Estimating the intricate relationships within complex systems, such as gene regulatory networks or financial markets, from time-series data is a significant challenge. These systems often involve a large number of interacting entities (high-dimensional), and traditional methods for inferring these interactions, like Granger causality, can become unreliable. Furthermore, these entities often possess inherent structures, like genes belonging to specific biological pathways or financial assets within industry sectors. Ignoring this "grouping structure" can lead to suboptimal performance. This paper aims to address these limitations by developing a method that leverages both the temporal dynamics and the inherent group structure of the data to learn these networks more effectively.

Methods

The core of this work lies in adapting the Granger causality framework, which determines if one time series can predict another, to a network setting with multiple time series. They propose using a group lasso regression regularization framework. This approach is particularly suited for high-dimensional problems where the number of potential connections is vast. The "group" aspect means that the penalty is applied to entire groups of connections, encouraging sparsity at the group level.

To handle situations where the predefined groups might not be perfectly accurate, they introduce a thresholded variant of group lasso. This method allows for more fine-grained selection within groups, even if the initial grouping is imperfect. A key theoretical contribution is the introduction of "direction consistency," a novel concept that measures how well the estimated direction of influence within a group aligns with the true direction, even when the group itself is misspecified. This concept is crucial for developing the thresholded method.

Datasets

The paper evaluates its proposed methods using:

  • Extensive simulation studies: These studies are designed to mimic various scenarios, including balanced and unbalanced group sizes, and misspecified group structures.
  • Functional genomics data: Specifically, data related to T-cell activation, where genes are naturally grouped by biological function.
  • Financial econometrics data: Data on banking balance sheets, where variables are categorized into assets (loans and securities) and liabilities (deposits).

Key Insights and Ideas

  • Leveraging Group Structure is Crucial: Explicitly incorporating known grouping structures (like biological pathways or industry sectors) significantly improves the accuracy and robustness of network inference compared to methods that ignore this information.
  • Group Lasso for Network Inference: The group lasso penalty is a powerful tool for learning sparse network structures in high-dimensional settings, especially when applied to Granger causality. It effectively handles the simultaneous estimation of connections and the order of the time series model.
  • Direction Consistency for Robustness: The novel concept of "direction consistency" provides a theoretical foundation for developing methods that can still infer meaningful relationships even when the predefined group structures are not perfectly accurate. This is a significant step towards more practical applications.
  • Thresholding Enhances Performance: The proposed thresholded group lasso method offers improved performance, particularly in scenarios with misspecified groups, by allowing for within-group variable selection.

Key Results

  • Improved Accuracy: The proposed group lasso and thresholded group lasso methods consistently outperform standard lasso in terms of precision, recall, and Matthew's Correlation Coefficient (MCC) across various simulation settings, especially when group structure is known and correctly specified. For instance, in one simulation, group NGC estimators recovered about 71% of true edges with a sample size where lasso recovered only 31%.
  • Reduced Error in VAR Order Estimation: Group variants of NGC estimators show substantially lower "ERR LAG" (error associated with incorrect estimation of VAR order) compared to lasso, indicating better handling of temporal dynamics.
  • Richer Network Structures: In real-world applications (T-cell activation and banking balance sheets), the group lasso estimates reveal richer and more nuanced network structures, identifying known regulatory mechanisms and financial relationships that were missed by the standard lasso approach.
  • Theoretical Guarantees: The paper establishes theoretical properties of the proposed estimators, including norm consistency and variable selection consistency, under mild assumptions.

Limitations and Future Work

  • Sensitivity to Grouping: While the thresholded method helps with misspecified groups, the performance of group lasso still relies on the quality of the initial grouping. If the grouping is severely misspecified, performance can degrade.
  • Tuning Parameter Selection: The performance of the methods depends on the choice of tuning parameters. While cross-validation is mentioned, efficient methods for selecting these parameters, especially for the thresholded variants, could be explored further.
  • Pure Time Series Data: The paper acknowledges that in many real-world scenarios, replicates at each time point might not be feasible (e.g., stock returns). Future work could address the technical challenges of applying these methods to pure time-series data where only dependent observations are available.
  • Computational Complexity: While the proposed thresholded method is described as "easy to compute," the paper notes that some alternative penalized regression procedures for within-group sparsity can involve nonconvex optimization problems and be computationally expensive. Further investigation into the computational efficiency of these methods would be valuable.

Summary updated

Abstract

The problem of estimating high-dimensional network models arises naturally in the analysis of many biological and socio-economic systems. In this work, we aim to learn a network structure from temporal panel data, employing the framework of Granger causal models under the assumptions of sparsity of its edges and inherent grouping structure among its nodes. To that end, we introduce a group lasso regression regularization framework, and also examine a thresholded variant to address the issue of group misspecification. Further, the norm consistency and variable selection consistency of the estimates are established, the latter under the novel concept of direction consistency. The performance of the proposed methodology is assessed through an extensive set of simulation studies and comparisons with existing techniques. The study is illustrated on two motivating examples coming from functional genomics and financial econometrics.
BibTeX
@article{basu2015network,
  author = {Sumanta Basu and Ali Shojaie and George Michailidis},
  title = {{Network granger causality with inherent grouping structure}},
  journal = {J. Mach. Learn. Res.},
  year = {2015},
  url = {https://openreview.net/forum?id=7N11OXJUffS},
}
Figure