In many real-world systems, events do not occur in isolation or at a constant, predictable rate. Instead, they arrive in clusters or "bursts." A single tweet triggers a flurry of retweets; an earthquake is followed by a cascade of aftershocks; a sudden stock price drop sparks a wave of frantic trading.
Traditional statistical models, such as standard Poisson processes, fail to capture this behavior because they assume that the occurrence of an event is entirely independent of its history. To capture the real world's conversational, financial, and physical momentum, researchers rely on self-exciting point processes, most notably the Hawkes process.
The Core Intuition: Conditional Intensity
To understand why Hawkes processes work, we must look at how they define the probability of an event happening at any given millisecond. This is governed by the conditional intensity function, denoted as $\lambda(t)$. You can think of $\lambda(t)$ as the instantaneous rate or "speed" at which events are expected to appear.
In a standard, memoryless Poisson process, this rate is either constant or changes based only on an external clock. In a self-exciting point process, however, the rate is explicitly driven by history:
$$\lambda(t) = \mu(t) + \sum_{t_i < t} \gamma(t - t_i)$$
This equation reveals the two core forces driving the system:
- The Background Rate ($\mu$): This represents the baseline risk or exogenous propensity for an event to happen due to outside factors. For instance, a user deciding to post an original update on social media.
- The Triggering Kernel ($\gamma$): This is the self-exciting engine. Every time an event occurs at time $t_i$, it injects a sudden "bump" or surge into the intensity function. This surge means that future events are immediately much more likely to happen.
Crucially, this excitation is not permanent. The triggering kernel $\gamma(t - t_i)$ typically decays over time—often exponentially or via a power law. As time passes without another event, the intensity cools down, gradually returning toward the baseline background rate. As shown by Takeshi Kurashima et al. in their work on human action sequences, Modeling Interdependent and Periodic Real-World Action Sequences, utilizing exponential decay kernels allows models to elegantly capture how recent actions dictate immediate future probabilities while preserving long-term patterns.
Intensity λ(t)
^
| | (Event 1) | (Event 2)
| v v
| /| /|
| / | / |
| / | / | /\
| / | / | / \
| / | / | / \
|_/_____|_________/_____|/______\____> Baseline μ
+------------------------------------> Time (t)
The Branching Structure Viewpoint
Why does this mathematical formulation map so well to real-world phenomena? A powerful way to build intuition for the Hawkes process is to view it as a latent branching structure.
Under the hood, every event in a Hawkes process can be categorized into one of two types:
- Immigrants: Events generated spontaneously by the background rate $\mu(t)$. These are truly exogenous shocks.
- Offspring: Events triggered directly by a previous event via the kernel $\pi$. These are endogenous reactions.
When an immigrant event occurs, it can spawn multiple offspring events. Each of those offspring can, in turn, spawn its own offspring, creating a genealogical tree of activity. This branching perspective simplifies the mathematics of tracking cascades. As demonstrated by Christian Walder et al. in Efficient Non-parametric Bayesian Hawkes Processes, treating the process as a latent branching tree allows researchers to break down complex event dependencies into independent Poisson processes, making it possible to learn flexible, non-parametric triggering kernels efficiently.
Furthermore, this structural view helps solve real-world data imperfections. In Hawkes Process Inference With Missing Data, the author Zhen Qin leverages this parent-child relationship to reconstruct event cascades even when some intermediate events are completely unobserved or hidden.
Scaling to Networks: Multivariate and Mutual Excitation
Events rarely happen in a single, isolated stream. In a social network, a financial market, or a biological system, there are thousands of interacting streams. To model this, the framework scales up to multivariate Hawkes processes.
In a multivariate setting, instead of a single intensity function, every node or entity $i$ has its own intensity $\lambda_i(t)$. Crucially, an event happening at node $j$ can increase the intensity function of node $i$. This is known as mutual excitation.
$$\lambda_i(t) = \mu_i + \sum_{j} \sum_{t_k^j < t} \alpha_{ji} \cdot \gamma(t - t_k^j)$$
Here, the matrix $\alpha_{ji}$ represents the directed influence or "infectivity" that node $j$ exerts over node $i$. By observing only the timestamps of events across different nodes, we can mathematically back out the hidden network structure.
This mechanism has been applied across diverse fields:
- Social Influence: Hongyuan Zha et al. show in Learning Social Infectivity in Sparse Low-rank Networks Using Multi-dimensional Hawkes Processes how the timing of recurrent events can uncover the true, invisible threads of social influence, moving past mere static friend lists to find who actually drives behavior.
- Network Topology Recovery: Emmanuel Bacry et al. demonstrate in Sparse and low-rank multivariate Hawkes processes that combining low-rank and sparse assumptions with multivariate Hawkes processes allows for the data-driven recovery of massive network structures using nothing but event timestamps.
- Conversational Dynamics: In communication networks, interactions are driven by both homophily (connecting with similar others) and reciprocity (replying to a message). In Decoupling Homophily and Reciprocity with Latent Space Network Models, Jiasen Yang et al. combine point processes with latent space embeddings to isolate these two distinct drivers, capturing how a single message triggers an immediate, decaying spike in the probability of a reciprocal response.
Deep Learning and Modern Extensions
While classical Hawkes processes rely on fixed parametric assumptions (like exponential decay), modern applications often require capturing more complex, non-linear relationships. The latest frontier in this research direction involves embedding self-exciting point processes directly into deep learning architectures.
Graph Neural Networks and Transformers
Instead of treating the background rate or the network influence as static parameters, researchers use neural networks to dynamically compute them. For example, Zhihao Wen et al. developed TREND: TempoRal Event and Node Dynamics for Graph Representation Learning, which integrates point processes into graph neural networks to capture how historical edge formations dynamically boost the probability of future connections.
Similarly, in text and sequence modeling, Anthony Hills et al. introduce a Hawkes-inspired layer into a Transformer architecture in Exciting Mood Changes: A Time-aware Hierarchical Transformer for Change Detection Modelling. This allows the model to weight past social media posts not just by their textual content, but by the frequency and temporal burstiness of the user's posting behavior.
Incorporating High-Dimensional Marks
Events often come with rich metadata, known as "marks"—such as the text of a news article, the geographic location of a tweet, or a user's profile. Modern frameworks treat these marks as integral components that modulate excitation.
In Hierarchical Dirichlet Gaussian Marked Hawkes Process for Narrative Reconstruction in Continuous Time Domain, Yeon Seonwoo et al. treat high-dimensional text embeddings as marks. This allows their model to determine the likelihood of a new event based on both the time elapsed and the semantic similarity of the text, successfully reconstructing hidden narrative threads across online discussions.
By modeling the continuous-time, self-exciting nature of cascades, Hawkes processes provide the mathematical scaffolding needed to move from static snapshots to truly dynamic, history-dependent models of the world.