Most classical models of time assume that events happen at a steady, predictable rate. If you are modeling radioactive decay or the arrival of calls at a massive switchboard, you use a Poisson process. In this world, the time between events follows an exponential distribution: most gaps are average-sized, and extremely long or extremely short silences are rare.
But humans do not behave like atoms or switchboards. As R. Dean Malmgren, Duncan J. Watts, and colleagues argue in Characterizing individual communication patterns, human activity is defined by burstiness. We experience intense volleys of activity—sending ten emails in ten minutes—followed by long, irregular periods of silence.
This research direction seeks to replace the "steady-rate" assumption with models that capture the heavy-tailed, self-exciting, and rhythmic nature of how we actually live.
The Geometry of Waiting
The core observation across this field is that the "inter-event time" (the gap between two actions) follows a power-law or heavy-tailed distribution rather than an exponential one. In Universal and Distinct Properties of Communication Dynamics: How to Generate Realistic Inter-event Times, Christos Faloutsos and his co-authors identify that this pattern is universal across SMS, phone calls, and web comments.
Why does this happen? Intuition suggests two primary drivers:
- Task Prioritization: When an emergency or a deadline hits, we stop routine tasks to focus on a single cluster of activity. Yi Luo and others show in Human emergency behaviour and psychological stress characteristic mining based on large-scale emergencies that during crises, communication shifts from random distributions to extreme power-law patterns, signaling a collective shift in focus.
- Circadian Rhythms: We sleep, work, and eat. Jalal Mahmud notes in When Will You Answer This? Estimating Response Time in Twitter that "one-size-fits-all" models fail because they ignore these non-homogeneous rhythms.
Mechanisms of Self-Excitation
If a Poisson process is "memoryless," the models proposed here are "memory-heavy." The most prominent tool used is the Hawkes process, a self-exciting point process where the occurrence of one event increases the probability of another event happening soon after.
In Mutually exciting point process graphs for modelling dynamic networks, Nicholas A. Heard and Francesco Sanna Passino use this to model network edges. The intuition is simple: if I email you now, the "intensity" of our relationship spikes, making a reply or a follow-up much more likely in the next hour than in the next week.
This "recency effect" is a cornerstone of the Dynamic Nonparametric Network Distribution (DNND) proposed by Elahe Ghalebi and colleagues in A Nonparametric Bayesian Model for Sparse Temporal Multigraphs. By accounting for the fact that we tend to interact with those we have contacted recently, these models can predict network growth and "bursty" cascades far more accurately than static graphs.
Beyond the Individual: Social Influence
Burstiness isn't just an internal psychological state; it is socially contagious. Greg Ver Steeg and Aram Galstyan demonstrate in Modeling Temporal Activity Patterns in Dynamic Social Networks that users are not "isolated islands." A burst of activity from a friend often triggers a burst in the user.
This social triggering is what creates "activity cascades." In Understanding Collective Social Behavior in OSS Communities, Lisi Qarkaxhija and Stefan Menzel show that in open-source development, one developer's edit acts as a catalyst, sparking a chain reaction of collaborative edits.
Why Temporal Precision Matters
Understanding when things happen changes our understanding of how they spread. Traditional models of disease or information spreading often use "wall-clock" time, assuming a steady rate of contact. However, Ciro Cattuto argues in Activity clocks: spreading dynamics on temporal networks of human contact that we should use "activity clocks."
Because human contact is bursty, information doesn't spread linearly. It jumps during bursts and stalls during lulls. Failing to account for these "long silences" leads to massive errors in predicting how quickly a virus or a viral tweet will move through a population. As Bruno Ribeiro and colleagues show in Quantifying the effect of temporal resolution on time-varying networks, even the "influence" of a node changes depending on the time window you use to observe these bursts.
By shifting from Poisson-based "averages" to burst-aware "intensities," researchers like Junchi Yan have even improved industrial applications, such as predicting B2B sales wins in On Machine Learning towards Predictive Sales Pipeline Analytics. The intuition remains the same: the timing of the activity is often more informative than the activity itself.