Abstract
Hawkes processes are essential for modeling self-exciting event sequences, such as seismic activity or social media interactions. However, real-world datasets often suffer from missing or unobserved events, which can bias standard inference methods. This paper introduces a robust Markov Chain Monte Carlo (MCMC) sampling framework to handle such latent data. By incorporating auxiliary variables and reversible-jump techniques, the method effectively "sees the unseen," uncovering hidden structures and improving predictive performance.
Introduction
Hawkes processes model "self-excitatory" phenomena where the occurrence of one event increases the probability of future events. While widely used, these models traditionally assume a complete record of all events. In practice, data is frequently incomplete due to sensor limitations, unrecorded interactions, or the presence of hidden actors. Ignoring these missing events leads to inaccurate parameter estimation and poor predictive power. This research addresses this gap by developing an efficient way to infer the existence and timing of unobserved events.
Methodology: Inferring the Latent Structure
The core contribution is a sophisticated MCMC sampler designed to estimate the posterior distribution of unobserved events. The approach leverages three primary techniques:
- Parent Auxiliary Variables: The Hawkes process is viewed as a branching tree where each event is either a "root" (caused by background rates) or a "child" (triggered by a previous event). Tracking these relationships simplifies the underlying mathematics.
- Virtual Event Auxiliary Variables: To navigate the continuous-time search space for hidden events, the authors introduce "virtual events." These act as placeholders that the sampler can promote to actual events or demote as needed.
- Reversible-Jump MCMC: Since the number of hidden events is unknown, the sampler employs reversible-jump moves. This allows the model to dynamically add or remove events during the sampling process while maintaining mathematical consistency.
Experimental Evaluation
The proposed method was tested on both synthetic and real-world datasets to evaluate its efficiency and practical utility.
Synthetic Benchmarks
The researchers created "easy" and "hard" simulated datasets using exponential and power-law kernels. The MCMC sampler was compared against Likelihood Weighting (LW), a standard baseline. In "hard" scenarios where LW failed to converge, the MCMC sampler demonstrated robust performance, achieving a steady $O(1/n)$ variance reduction in error.
Real-World Application: Chicago Gang Homicides
The model was applied to a dataset of 2,195 gang-related homicides in Chicago (1965–1995). By introducing hidden event labels, the researchers aimed to capture unobserved "turf war" dynamics. The results were significant:
- Prediction Accuracy: Adding five hidden labels improved the accuracy of predicting the neighborhood of the next homicide by 22% (from 5.3% to 6.5%).
- Structure Discovery: The inferred hidden labels mapped closely to documented gang territories and conflicts that were not explicitly present in the original police records, providing a richer sociological understanding of the data.
Discussion and Conclusion
This work demonstrates that treating missing data as latent variables can significantly enhance the modeling of self-exciting processes. While the MCMC sampler requires a notable "burn-in" period for complex problems, its ability to uncover hidden influences makes it a powerful tool for analyzing messy, real-world event data. Future research may focus on automated model selection for the number of hidden event types and optimizing performance across a broader range of trigger kernels.
Summary updated