When a piece of news breaks or a meme goes viral, it doesn’t spread at a constant rate. Instead, information moves in "bursts"—a single post triggers a flurry of retweets, which in turn spark their own sub-cascades. Traditional models often treat popularity as an aggregate growth curve, but this misses the fundamental mechanism of virality: every interaction is a "spark" that increases the probability of future events.
To capture this, researchers have turned to Hawkes Processes, a class of self-exciting point processes. Unlike standard Poisson processes where events occur independently, a Hawkes process assumes that the arrival of one event temporarily increases the "intensity" or likelihood of subsequent events.
The Intuition of Self-Excitation
The core idea, as explored by Huawei Shen and colleagues in their work on microblog popularity, is that information diffusion is recursive. In their Self-Excited Hawkes Process (SEHP) model, they treat each individual share not just as a tally mark in a total count, but as a discrete event with its own "triggering strength."
Imagine a campfire: a Poisson process is like rain falling at a steady rate; a Hawkes process is like the embers. One ember pops and lands on dry leaves, starting a small new fire. This captures the "bursty" nature of social media, where activity is characterized by long periods of silence interrupted by rapid-fire cascades. Koon‐Kiu Yan notes that this stochastic, multiplicative approach explains why scientific paper views follow a lognormal distribution—early traction significantly increases the probability of future dissemination (The Spread of Scientific Information).
Mapping Influence Across the Web Centipede
Information rarely stays on one platform. A conspiracy theory might start on a fringe board like 4chan’s /pol/, migrate to a specific subreddit, and eventually explode on Twitter. To track this, researchers use Multivariate Hawkes Processes (MHP).
In an MHP, events on one "dimension" (e.g., Reddit) can increase the intensity of events on another (e.g., Twitter). Savvas Zannettou, Emiliano De Cristofaro, and their collaborators used this framework in The web centipede to quantify the influence of different communities. By measuring the "background rate" of a topic versus the "triggered rate" caused by other platforms, they identified that fringe communities act as powerful incubators that disproportionately shape the mainstream news ecosystem. This same methodology has been applied to track state-sponsored troll farms and the migration of conspiracy theories like QAnon.
Strategic Mitigation of Misinformation
If we can model how a "fire" of misinformation spreads, we can theoretically determine where to place "firebreaks." Ahmed Abouzeid, Ole-Christoffer Granmo, and Morten Goodwin have developed frameworks that pair Hawkes processes with Learning Automata to fight fake news (Learning Automata-based Misinformation Mitigation).
The challenge is that a network’s "budget" for debunking is limited. By modeling the self-exciting nature of both the misinformation and the potential "real news" correction, these researchers treat mitigation as a stochastic optimization problem. Instead of shouting into the void, the system identifies the specific users or moments where injecting a factual correction will have the highest "triggering" potential to dampen the misinformation cascade. Mahak Goindani further refined this by incorporating political bias into the model, recognizing that a "spark" of truth only catches if the "fuel" (the user) is receptive to it (Social Reinforcement Learning to Combat Fake News Spread).
Adding Depth with Marks and Nesting
Recent advances have moved beyond simple event timing to include the content of the events, known as Marked Hawkes Processes. Jianxin Li and Peiyuan Sun use these "marks" to infer hidden, multi-layered social structures, where different topics (marks) trigger different diffusion patterns across the same set of users (Inferring Multiplex Diffusion Network).
Furthermore, online discussions often have a hierarchical structure—threads and replies. Guangmo Tong introduced the Nested Temporal Point Process (NesTPP) to model this (NesTPP: Modeling Thread Dynamics). In this view, the intensity of a top-level thread is not just a function of other threads, but is "fed" by the volume of nested replies within it. This formalizes the intuition that a highly active discussion acts as a catalyst for new, related conversations to emerge.
By treating digital interactions as a series of self-exciting triggers rather than static data points, these models provide a mathematically rigorous way to understand, predict, and ultimately influence the complex flow of information in the modern world.