In many real-world systems, events do not occur in isolation. An earthquake is often followed by a cluster of aftershocks; a single viral tweet triggers a cascade of retweets; and a localized COVID-19 outbreak increases the risk of infection in neighboring blocks. Traditional statistical models, like the standard Poisson process, fail here because they assume events are independent.
To capture the "bursty" nature of these phenomena, researchers use self-exciting point processes, most notably the Hawkes process. This framework treats every event as a potential trigger for future ones, allowing us to model the hidden momentum of social, biological, and physical systems.
The Intuition of Self-Excitation
The core of a Hawkes process is its intensity function, $\lambda(t)$, which represents the instantaneous probability of an event occurring at time $t$. Unlike a static model, this intensity is dynamic and history-dependent. It is typically decomposed into two parts:
- The Background Rate (Exogenous): A baseline probability of an event occurring due to external factors (e.g., the general likelihood of a crime in a neighborhood).
- The Triggering Kernel (Endogenous): A function that "spikes" the intensity immediately after an event occurs and then decays over time.
As Aram Galstyan, Yoon-Sik Cho, and colleagues demonstrate in Latent Point Process Models for Spatial-Temporal Networks, this mechanism allows the model to account for the temporal clustering of interactions even when the specific participants are unknown. By treating past events as "parents" of future "children," the model creates a branching structure that naturally mirrors how information or disease spreads.
Distinguishing Feedback from Shocks
A critical challenge in modeling complex systems is determining whether an event happened because of a permanent environmental factor or a temporary chain reaction. Minkyoung Kim and Dean Paini refer to this as reflexivity.
In Modeling Reflexivity of Social Systems in Disease Spread, they use the Latent Influence Point Process (LIPP) to separate external "shocks" (like international travelers arriving with a virus) from internal "feedback loops" (local transmission). This distinction is vital for policy: if an outbreak is driven by external shocks, border controls are effective; if it is driven by internal reflexivity, local social distancing is required.
Similarly, in the context of urban safety, Yunyi Li and Tong Wang developed the Next Hit Predictor. Their model combines stationary background trends with a dynamic "triggered" component to predict the next location in a serial crime spree, recognizing that a perpetrator's recent success often dictates their next target.
Scaling to Space and Networks
While the basic Hawkes process handles time, real-world events also have locations and connections. Modern research extends these processes into spatiotemporal and graph-based domains.
- Spatial Dispersion: In Cox-Hawkes: doubly stochastic spatiotemporal Poisson processes, Xenia Miscouridou, George Mohler, and their team use Gaussian spatial kernels to model how the influence of an event (like a gunshot) diffuses geographically over time.
- Network Reciprocity: In social networks, interactions are often reciprocal. Jiasen Yang and Vinayak Rao introduced the Dual Latent Space (DLS) model in Decoupling Homophily and Reciprocity with Latent Space Network Models. By embedding nodes in a latent space, they use Hawkes processes to model how a single message increases the probability of a reply, effectively capturing the "bursty" nature of conversation.
- Mobility Flows: Sirisha Rambhatla and Yan Liu showed in Towards Accurate Spatiotemporal COVID-19 Risk Scores that integrating origin-destination flow data into a Hawkes framework allows for much more accurate localized risk scores than traditional density-based models.
Integration with Deep Learning
The most recent frontier involves merging the statistical rigor of point processes with the representational power of deep learning.
For example, Shixiang Zhu and Yao Xie demonstrate in Crime Linkage Detection by Spatio-Temporal-Textual Point Processes how to incorporate unstructured data, like police narratives, into the triggering kernel. This allows the model to calculate the probability that two crimes are linked not just by their timing, but by the "Modus Operandi" described in text.
Furthermore, Yunpu Ma and Volker Tresp have explored Learning with Temporal Knowledge Graphs, where self-exciting processes are integrated into Graph Neural Networks. This enables the prediction of when a new edge will form in a graph by identifying the causal chains and temporal clusters inherent in historical activity.
By shifting from "if" an event will happen to "when and where" it will be triggered, these models provide a mathematical lens for the interconnectedness of modern data.