Hawkes Point Processes

Description

Models behavioural events as a point process where each event temporarily raises (or suppresses) the chance of further events, capturing bursts and cascades in interaction data.

A Hawkes process has a conditional intensity equal to a baseline rate plus decaying contributions from past events, so events cluster in self-exciting bursts, as in earthquake aftershocks or trading activity. The Behavioral Spectral Framework proposes representing behavioural events as point processes such as Hawkes models to capture self-excitation and suppression, as part of handling irregular sampling intelligently, alongside Lomb-Scargle periodograms and survival models of inactivity. The appeal for UX is that clicks, retries and messages are event-driven: a failed attempt may trigger a rapid cascade of retries (excitation), while a confusing screen may suppress further action. The playbook mentions the technique as a direction rather than a worked protocol, so the steps below follow standard practice.

Attributes

Loss type
Unmigrated
Stage
Post-launch
Effort
Days

Steps

  1. Collect event timestamps per user, with event types if relevant.
  2. Specify a baseline rate and an excitation kernel (e.g. exponential decay).
  3. Fit the model by maximum likelihood.
  4. Interpret the branching ratio (how many follow-on events each event triggers) and decay time.
  5. Compare parameters across cohorts or design variants.
  6. Validate with goodness-of-fit checks and qualitative review of bursts.

Sources

  • Vivek Gopinath (2025), From Stars to Signals: Translating Matteo Paz's Fourier-Wavelet Model into Behavioral and UX Research Methodology

Also known as

  • Self-exciting point processes
  • Hawkes models

Categories

  • Analytics & Behaviour
  • Statistics & Modelling