TL;DR
This paper introduces the first exact, frequentist maximum likelihood estimation method for multivariate Hawkes processes that can model both exciting and inhibiting interactions between events. This new approach allows for more accurate inference of complex event dynamics, particularly in fields like neuroscience where inhibition plays a crucial role.
What Problem Does This Paper Address?
Hawkes processes are powerful tools for modeling event sequences where past events influence future ones. Traditionally, these models focused on "excitation" – one event making another more likely. However, many real-world phenomena involve "inhibition," where an event decreases the likelihood of subsequent events. Existing methods for multivariate Hawkes processes struggle to accurately capture these inhibiting effects, often leading to biased parameter estimates or an inability to identify the true interaction structure. This paper tackles the challenge of developing a robust and exact inference framework for multivariate Hawkes processes that can handle both excitation and inhibition simultaneously.
Methods
The core contribution is a novel maximum likelihood estimation (MLE) approach for multivariate exponential Hawkes processes. Key innovations include:
- Exact Log-Likelihood Computation: The authors derive a closed-form expression for the log-likelihood, which is crucial for efficient and accurate estimation. This involves defining and computing "restart times" to precisely determine when the conditional intensity becomes non-zero after inhibition.
- Identifiability Proof: They establish conditions under which the model parameters can be uniquely identified from observed data, a critical theoretical foundation.
- Variable Selection: A post-hoc thresholding step is introduced to identify which interaction parameters are statistically significant, effectively recovering the "connectivity graph" of interactions between different event types.
- Goodness-of-Fit Tests: The framework allows for rigorous assessment of the model's fit to the data, enabling model selection and validation.
The proposed methods are compared against existing approaches, including those designed for exciting processes but adapted to handle negative parameters.
Datasets
The paper evaluates its methods on:
- Synthetic Datasets: Multiple simulated datasets are generated to test the performance of the proposed estimators under various scenarios, including different dimensions (2D and 10D) and varying degrees of inhibition and null interactions.
- Neuronal Activity Dataset: A real-world dataset of spike trains from 223 neurons is used to demonstrate the practical application of the method in inferring neuronal connectivity.
Key Insights and Ideas
- Inhibition is Crucial: The study strongly suggests that accurately modeling inhibiting interactions is essential for understanding complex systems, as demonstrated by the improved performance of their method on synthetic data and the insights gained from the neuronal data.
- Exactness Matters: Deriving an exact log-likelihood, rather than relying on approximations, leads to more reliable parameter estimates, especially in scenarios with strong inhibition where approximations can fail.
- Connectivity Matters: The ability to identify which interactions are significant (the "connectivity graph") is vital for interpreting the underlying system dynamics, whether it's social media trends, seismic activity, or neuronal networks.
- Identifiability is Achievable: Proving identifiability under reasonable assumptions provides theoretical backing for the proposed estimation procedure.
Key Results
- The proposed MLE method with variable selection outperforms existing approaches on synthetic data, particularly in scenarios with strong inhibiting effects.
- The method successfully identifies both exciting and inhibiting interactions in a neuronal activity dataset, revealing a complex connectivity structure.
- Goodness-of-fit tests are shown to be effective in detecting when parameters are incorrectly estimated, especially for methods that don't properly handle inhibition.
- The computational cost, while higher than simpler methods, is deemed "quite reasonable" for the proposed approach, even with selection steps.
Limitations and Future Work
- Exponential Kernel Assumption: The current method is limited to exponential kernel functions. Future work could explore extensions to other kernel shapes (e.g., power-law kernels, as briefly investigated in robustness tests) or more general parametric forms.
- Delay Factor Assumption: The assumption that delay factors ($\beta_i$) depend only on the receiving process is a simplification. Relaxing this would likely require numerical integration, increasing computational complexity.
- Computational Cost: While reasonable, the computational cost can still be a factor for very high-dimensional or long time series. Further research into computational efficiency is suggested.
- Theoretical Guarantees: Extending the theoretical study of the estimator's asymptotic behavior to the multivariate case is an ongoing area of research.
- Interpretability in Partially Observed Networks: The paper acknowledges the challenge of interpreting interactions when only a subset of a network is observed and points to recent work on addressing this.
Summary updated