Characterizing individual communication patterns

R. Dean Malmgren, Jake M. Hofman, Luís A. Nunes Amaral, Duncan J. WattsKDD 2009 · DOI · arXiv

Introduction

Traditional social science often classifies individuals using demographic attributes like age, gender, and income, or through their positions within a social network. However, demographics are frequently poor predictors of actual behavior, and standard network analysis often overlooks the temporal "rhythm" of human activity. Human communication is inherently "bursty"—we do not send emails at a steady rate. Instead, activity occurs in concentrated "sessions" separated by long pauses, all modulated by circadian and weekly cycles. This paper introduces a generative model designed to characterize individuals based on the timing of their communications, identifying persistent behavioral "types" that remain stable over several years.

Methodology: The Double-Chain Hidden Markov Model

To capture these complex dynamics, the authors propose a cascading non-homogeneous Poisson process formulated as a double-chain hidden Markov model (DCHMM). In this model, an individual transitions between two hidden states:

  • Active State ($z_n = 1$): The user is in a communication session, sending emails at a high, constant rate ($\rho_a$).
  • Passive State ($z_n = 0$): The user is between sessions. The rate of initiating a new session, $\rho(t)$, is non-homogeneous, varying according to the time of day and day of the week to capture rhythmic patterns.
The DCHMM framework: (A) state transitions, (B) non-homogeneous session initiation rate $\rho(t)$, and (C) the graphical model structure.

Unlike standard HMMs where only hidden states are coupled, the DCHMM also couples observed timings. This structure allows for the use of a highly efficient Expectation-Maximization (EM) algorithm. This approach provides a significant computational speedup—$O(N)$ time complexity compared to $O(N^2 \ln N)$ for previous optimization methods—enabling the analysis of thousands of users in minutes.

Data and Analysis

The study utilizes two large-scale email datasets from different eras and locations:

  1. EMS Dataset: 3,188 accounts at a European university (83 days), collected before widespread home internet.
  2. KW Dataset: 122,133 accounts at a US university (two years).

After filtering for active users (at least one email every two days), the final analysis focused on 404 users from EMS and 6,164 from KW.

Key Insights and Results

The model reveals that an individual's communication rhythm is a remarkably persistent trait. Within-individual variability over time was found to be significantly lower than the variability across the general population for over 80% of individuals.

Probability density distributions for inferred model parameters across the EMS and KW datasets, showing universal patterns in communication behavior.

Universal Patterns and Behavioral Types

Despite differences in era and geography, the distribution of communication parameters was surprisingly similar across both datasets. By clustering users based on their "start of day" ($\tau_{0d}$) and "end of day" ($\tau_{1d}$) parameters, the authors identified two distinct behavioral groups:

  • "Day Laborers": Users who strictly communicate during standard 9-to-5 working hours.
  • "E-mailaholics": Users whose activity spans nearly their entire waking day (e.g., 9:00 AM to 1:00 AM).

Stability was a key finding: approximately 75% of individuals remained in the same behavioral cluster over the two-year study period.

Conclusion and Future Work

The DCHMM provides a rigorous yet interpretable way to differentiate individuals based on how they communicate rather than just who they talk to. While currently focused on email, the model's efficiency and ability to detect outliers suggest broad applications, including identifying non-human behavior (bots/spammers) and predicting organizational roles based on temporal rhythms. Future work may extend this analysis to other bursty mediums like instant messaging or web browsing.

Summary updated

BibTeX
@inproceedings{malmgren2009characterizing,
  author = {R. Dean Malmgren and Jake M. Hofman and Luís A. Nunes Amaral and Duncan J. Watts},
  title = {{Characterizing individual communication patterns}},
  booktitle = {ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year = {2009},
  doi = {10.1145/1557019.1557088},
}
Figure