Uncertainty Quantification for Inferring Hawkes Networks

Haoyun Wang, Liyan Xie, Alex Cuozzo, Simon Mak, Yao Xie

Abstract

Hawkes processes are essential for modeling self-exciting event sequences, yet quantifying the uncertainty of inferred network connections remains challenging. Traditional asymptotic methods often fail with limited data, leading to over-conservative or inaccurate results. This paper presents a non-asymptotic framework for uncertainty quantification (UQ) using martingale concentration inequalities, providing tighter and more reliable confidence intervals for multivariate Hawkes processes.

Introduction

Hawkes processes are a powerful mathematical tool used to model "self-exciting" event sequences—where one event increases the probability of future events. In a network setting, these processes are used to infer Granger causality, determining if an event at one node causes an event at another. While point estimation methods for these connections are well-developed, the field has lacked robust methods for answering a fundamental statistical question: how certain are these estimates?

Existing methods typically rely on asymptotic confidence intervals, which assume an infinite data regime. In practical applications like neuroscience or medicine, data is often sparse. Asymptotic intervals in these contexts can be dangerously over-confident or uselessly wide, making it difficult to distinguish true causal links from random noise. This research addresses these limitations by developing a non-asymptotic approach to UQ.

Methodology

The proposed framework constructs confidence sets for the Maximum Likelihood Estimate (MLE) of a Hawkes network based on three mathematical pillars:

  1. Martingale Construction: The authors leverage the fact that the score function (the gradient of the log-likelihood) of a Hawkes process behaves like a continuous-time martingale. This allows for tracking the evolution of estimation error over time.
  2. Concentration Inequalities: Instead of assuming error follows a standard bell curve, the method uses concentration bounds. These bounds provide a "safety envelope" that remains valid even for finite, short-duration datasets.
  3. Polyhedral Confidence Sets: The method constructs a polyhedron in high-dimensional space rather than a simple box. This shape is more flexible and can be solved efficiently using linear programming.

A key innovation is the use of "adapted z," which allows the model to use data from earlier time steps to refine the shape of the confidence interval for future events, resulting in tighter bounds as more data is processed.

Experimental Setup

The method was evaluated using synthetic neuronal spike train data generated via the PyNN Python package and the NEURON simulator. The researchers created a biologically plausible network of 32 neurons, consisting of both excitatory and inhibitory types. To test the UQ method's performance under data constraints, they simulated a "long" sequence (2000s) to establish ground truth parameters and a "short" sequence (400s) for testing the inference and uncertainty bounds.

Results

The proposed non-asymptotic method was compared against the traditional asymptotic approach. The results highlighted a significant trade-off between coverage and precision.

  • Coverage Accuracy: For a nominal 95% confidence level, the proposed method achieved a 96.8% coverage rate. While the asymptotic method achieved a 100% coverage rate, this was due to its intervals being excessively wide and lacking practical utility.
  • Precision and Sparsity: The proposed method produced significantly narrower confidence intervals for non-zero connections. This precision allowed for the identification of subtle causal relationships that asymptotic methods would dismiss. Furthermore, the method correctly identified many estimated edges as zero, helping recover the true, sparse structure of the neuronal network.
Comparison of true and estimated networks alongside confidence interval (CI) performance metrics.

As shown in the figure above, the proposed method (96.8% coverage) provides a much more "honest" and precise representation of the network compared to the over-conservative asymptotic baseline.

Discussion and Conclusion

This research demonstrates that non-asymptotic UQ is both safer and more precise for Hawkes process inference. By utilizing the "intrinsic variance" of the process, the authors created dynamic bounds that adapt to data volatility.

However, limitations remain. The current framework assumes the functional form (kernel) of the influence function is known. Future work will need to address scenarios where the kernel shape is unknown and extend these bounds to non-linear Hawkes processes. Additionally, while polyhedral sets are efficient for medium-sized networks, scaling to massive social or biological networks will require further algorithmic optimization.

Summary updated

BibTeX
@inproceedings{wang2020uncertainty,
  author = {Haoyun Wang and Liyan Xie and Alex Cuozzo and Simon Mak and Yao Xie},
  title = {{Uncertainty Quantification for Inferring Hawkes Networks}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2020},
}
Figure