Characterizing Individual Communication Patterns

The increasing availability of electronic communication data, such as that arising from e-mail exchange, presents social and information scientists with new possibilities for characterizing individual behavior and, by extension, identifying latent structure in human populations. Here, we propose a model of individual e-mail communication that is sufficiently rich to capture meaningful variability across individuals, while remaining simple enough to be interpretable. We show that the model, a cascading non-homogeneous Poisson process, can be formulated as a double-chain hidden Markov model, allowing us to use an efficient inference algorithm to estimate the model parameters from observed data. We then apply this model to two e-mail data sets consisting of 404 and 6,164 users, respectively, that were collected from two universities in different countries and years. We find that the resulting best-estimate parameter distributions for both data sets are surprisingly similar, indicating that at least some features of communication dynamics generalize beyond specific contexts. We also find that variability of individual behavior over time is significantly less than variability across the population, suggesting that individuals can be classified into persistent “types”. We conclude that communication patterns may prove useful as an additional class of attribute data, complementing demographic and network data, for user classification and outlier detection–a point that we illustrate with an interpretable clustering of users based on their inferred model parameters.

💡 Research Summary

The paper addresses the growing availability of electronic communication logs—specifically email exchange records—and asks how one can quantitatively characterize the behavior of individual users while still retaining interpretability. Existing work largely focuses on aggregate statistics (e.g., overall volume, network topology) or on simple homogeneous Poisson models that ignore two well‑documented features of human email activity: (1) strong daily and weekly cycles, and (2) bursty periods in which many messages are sent in rapid succession. To capture both phenomena simultaneously, the authors propose a “cascading non‑homogeneous Poisson process.” The baseline intensity λ(t) is a time‑varying function, modeled as a sum of sinusoidal components that encode daily and weekly rhythms. Whenever an email is sent, the process may transition into a “burst” state with probability p; in the burst state inter‑event times follow a geometric distribution, producing a rapid cascade of additional messages before returning to the baseline state.

Mathematically, this two‑state mechanism can be expressed as a double‑chain hidden Markov model (HMM). The first chain governs the time‑dependent baseline intensity, while the second chain captures the burst/no‑burst hidden state. This formulation enables the use of an efficient expectation‑maximization (EM) algorithm within a variational Bayes framework: the E‑step computes posterior probabilities over the hidden states given the observed timestamps, and the M‑step updates the sinusoidal coefficients of λ(t), the burst transition probability p, and the geometric parameter governing burst length. Because the model is constrained to two chains, the computational cost scales linearly with the number of observed events (O(N·K), where K=2), making it feasible for large email logs.

The methodology is evaluated on two independent university email datasets. The first comprises 404 users from a U.S. university collected in 2004; the second contains 6,164 users from a Swiss university collected in 2005. Each user generated on the order of a few hundred messages over a period of several months, providing sufficient data to estimate individual parameters. After fitting the model to every user, the authors compare the distributions of the inferred parameters across the two populations. Remarkably, the amplitude and phase of the daily/weekly sinusoidal component of λ(t) are nearly identical in both datasets, and the burst transition probability p clusters tightly around 0.15–0.22. This similarity suggests that, despite cultural and temporal differences, fundamental aspects of email communication dynamics are universal.

To assess temporal stability, the authors re‑fit the model to the same users across multiple months. Within‑individual parameter variation is substantially smaller than the variation observed across the whole population, indicating that each person’s communication style is relatively persistent over time. Leveraging this stability, the authors perform K‑means clustering on the three‑dimensional parameter vectors (daily amplitude, weekly amplitude, burst probability). The resulting clusters correspond to intuitive “types”: high‑frequency short‑burst users, low‑frequency long‑burst users, and users with pronounced daily rhythms but modest burstiness. These types are orthogonal to traditional demographic attributes (e.g., department, seniority) and therefore provide an additional, behavior‑based axis for user profiling.

The discussion highlights several practical implications. First, the model’s parameters are directly interpretable, allowing analysts to reason about a user’s typical workload, responsiveness, or propensity for collaborative bursts. Second, outlier detection becomes straightforward: users whose parameters lie far from any cluster may represent spammers, compromised accounts, or atypical work patterns, enabling early security interventions. Third, the framework could be extended to other communication channels (instant messaging, social media) where similar rhythmic and bursty patterns exist.

In conclusion, the authors demonstrate that a cascading non‑homogeneous Poisson process, cast as a double‑chain hidden Markov model, offers a parsimonious yet expressive representation of individual email communication. The model successfully captures both circadian/weekly cycles and burst dynamics, yields stable individual signatures, and reveals cross‑cultural regularities. By providing interpretable parameters that can be clustered into meaningful user types, the approach opens new avenues for augmenting demographic and network data with behavioral attributes for classification, recommendation, and anomaly detection tasks.