Traditional statistical models often treat sequential events like flips of a coin or ticks of a metronome. In a standard Poisson process, the probability of an event occurring in the next minute depends only on a baseline rate, completely oblivious to whether an event happened one second ago. As noted by Yiming Xu et al. in Nonparametric Embeddings of Sparse High-Order Interaction Events, this independence assumption fails to capture the reality of real-world data.
In practice, events cluster. A single email triggers a flurry of replies; an initial social media post sparks a cascade of retweets; a clinical symptom precedes a cluster of medical interventions. Charles Blundell et al. emphasize in Modelling Reciprocating Relationships with Hawkes Processes that the timing of the most recent interaction is often the single most critical predictor of the next. When we assume independence, we smooth out these vital "bursts," rendering our models blind to the causal chains that drive human behavior, ecological spread, and network evolution.
The Anatomy of Self-Excitation
To capture this burstiness, researchers leverage self-exciting point processes, most notably the Hawkes process. The core intuition relies on a dynamic conditional intensity function, $\lambda(t)$, which represents the instantaneous arrival rate of events at time $t$.
Instead of remaining static or varying purely based on time, $\lambda(t)$ is history-dependent. It can be conceptualized as:
$$\lambda(t) = \mu(t) + \sum_{t_i < t} \alpha \cdot \kappa(t - t_i)$$
This equation reveals why the model works by breaking the event rate into two distinct components:
- The Baseline ($\mu(t)$): The background rate at which events naturally occur without external provocation.
- The Triggering Kernel ($\kappa(t - t_i)$): Every time an event occurs at a past time $t_i$, it injects a sudden spike ($\alpha$) into the intensity function. This spike represents the "excitation." Crucially, this effect decays over time according to a kernel function $\kappa$ (often an exponential decay).
An event makes subsequent events momentarily much more likely. If a second event occurs during this heightened window, it adds its own spike to the intensity, creating a cascading feedback loop—a burst. As the excitement decays back toward the baseline, the burst subsides.
Amrita Gupta et al. demonstrate the power of this formulation in Discrete Interventions in Hawkes Processes with Applications in Invasive Species Management, showing how tracking these localized triggering effects allows for highly accurate predictions of how phenomena propagate over time. Similarly, Aram Galstyan et al. apply this to social interactions in Latent Point Process Models for Spatial-Temporal Networks, capturing how a single interaction triggers a burst of subsequent related events even when the exact participants are latent or unknown.
Integrating Self-Excitation with Representation Learning
While classical Hawkes processes work well for simple, isolated event streams, real-world data is multi-dimensional, structured, and highly complex. The frontier of this research lies in embedding self-exciting dynamics directly into advanced representation learning frameworks.
Graphs and Networks
In social networks or knowledge graphs, interactions are rarely isolated random events; they are driven by contact history and relational structures, as argued by Jiangjiang Cheng et al. in Multidimensional opinion dynamics with heterogeneous bounded confidences and random interactions. To capture this, Zhihao Wen et al. developed the TREND framework in TREND: TempoRal Event and Node Dynamics for Graph Representation Learning, which explicitly integrates triggering effects into graph neural networks to better predict future node activity. On a macro scale, Yunpu Ma et al. show in Learning with Temporal Knowledge Graphs that modeling these self-exciting causal chains is essential for forecasting complex international relations and global trade events.
Tensors and High-Order Interactions
When interactions involve multiple entities simultaneously (such as a multi-party meeting or a multi-entity transaction), data is often represented as a tensor. Shandian Zhe et al. in Stochastic Nonparametric Event-Tensor Decomposition combined Gaussian Processes with Hawkes Processes to capture both complex nonlinear relationships and mutual triggering dynamics across multiway data. This allows models to uncover hidden clusters in sparse datasets by looking at who triggers whom over time.
Deep Generative Models
To handle the immense randomness and variety of human actions, researchers have fused self-exciting processes with deep generative architectures like Variational Autoencoders (VAEs). For instance, Nazanin Mehrasa et al. introduced the APP-VAE in A Variational Auto-Encoder Model for Stochastic Point Processes to move away from simple exponential distributions and better predict both what action happens next and when.
This hybrid approach is highly practical for generating realistic synthetic data. Chufan Gao et al. utilized a similar VAE-Hawkes combination in TrialSynth: Generation of Synthetic Sequential Clinical Trial Data to accurately mimic the intricate, history-dependent timelines of patient health events.
Expanding the Intensity Function
A key challenge in these models is ensuring the intensity function remains flexible yet mathematically sound. While traditional approaches rely on rigid parametric forms, Gabriel Loaiza-Ganem et al. introduced Deep Random Splines in Deep Random Splines for Point Process Intensity Estimation to estimate smooth, non-negative intensity functions using deep neural networks.
By pairing highly flexible functional forms with self-exciting temporal dependencies, modern frameworks can optimize real-world objectives. For example, Mohammad Reza Karimi et al. leverage temporal point processes in Smart broadcasting: Do you want to be seen? to maximize a broadcaster's visibility on social media, demonstrating that moving beyond independent event assumptions yields concrete algorithmic advantages in dynamic environments.
Key Foundational Work
Modelling Reciprocating Relationships with Hawkes Processes (2012) — contributed 2 concept elements to this direction.
Latent Point Process Models for Spatial-Temporal Networks (2013) — contributed 1 concept elements to this direction.