Rhythm and Randomness in Human Contact

There is substantial interest in the effect of human mobility patterns on opportunistic communications. Inspired by recent work revisiting some of the early evidence for a L'evy flight foraging strategy in animals, we analyse datasets on human contact from real world traces. By analysing the distribution of inter-contact times on different time scales and using different graphical forms, we find not only the highly skewed distributions of waiting times highlighted in previous studies but also clear circadian rhythm. The relative visibility of these two components depends strongly on which graphical form is adopted and the range of time scales. We use a simple model to reconstruct the observed behaviour and discuss the implications of this for forwarding efficiency.

💡 Research Summary

The paper investigates the statistical properties of human inter‑contact times (ICTs) using three real‑world Bluetooth trace datasets: MIT Reality Mining (100 smartphones over nine months), Cambridge Haggle‑2 (36 iMotes over eleven days) and INFOCOM 2006 (77 iMotes over three days). The authors first review prior work that models ICTs with truncated Lévy walks (TLWs), reporting stability exponents α in the range 0 – 0.9, and they highlight methodological criticisms of those studies (over‑reliance on log‑log histograms, limited scaling ranges, biased exponent estimation, and lack of alternative hypotheses).

For each dataset the authors compute all pairwise ICTs and present the distributions using four graphical forms: (a) rank‑order (double‑log) plots, (b) probability density functions (log‑log axes with exponentially spaced bins), (c) linear‑bin histograms (semi‑log), and (d) log‑log histograms (linear bins) limited to 12 h. Across all visualisations a heavy‑tailed region is evident at short time scales (minutes to a few hours), consistent with a power‑law tail p(t) ∝ t^{-(1+α)} where α varies between 0.3 and 0.9 depending on the experiment. However, the log‑log histograms also reveal pronounced peaks at integer multiples of 24 h (and a weekly pattern in the MIT data), indicating a strong circadian rhythm that cannot be captured by a pure power‑law model.

To explain these observations the authors construct two simulation models. The first generates ICTs from a pure TLW: independent Pareto samples (parameter α) are summed to produce contact times within the experiment duration L, and ICTs are obtained as successive differences. This model reproduces the short‑scale power‑law tail but fails to generate any of the daily peaks seen in the empirical data. The second model augments the TLW with a periodic “activity window”: only contacts that fall within a predefined daily working period (e.g., 09:00–18:00) are retained, while the rest of the day is treated as a blackout period. This hybrid model reproduces both the heavy tail and the 24‑hour peaks, matching the empirical histograms far more closely.

Statistical fitting is performed using maximum‑likelihood estimation (MLE) for the three candidate distributions (pure power‑law, exponentially truncated power‑law, and the hybrid model). Goodness‑of‑fit is assessed with Kolmogorov‑Smirnov tests and Akaike Information Criterion (AIC). The hybrid model consistently achieves the lowest AIC, confirming that a combination of Lévy‑flight randomness and deterministic circadian constraints best describes human contact patterns in the studied settings.

The paper then discusses implications for opportunistic networking. Many DTN routing protocols (e.g., Spray‑and‑Wait, Prophet) assume a stationary heavy‑tailed ICT distribution; under that assumption they predict relatively short expected delays. Incorporating the observed daily rhythm shows that a large fraction of contacts occur during a limited window each day, so messages waiting for a contact outside that window experience much longer delays. Simulations indicate that routing strategies that prioritize forwarding during active periods and defer or buffer during inactive periods can significantly reduce average delivery latency and increase delivery ratios.

In summary, the authors demonstrate that human inter‑contact times are not purely Lévy‑flight driven but are better modeled as Lévy walks constrained by a periodic (circadian) domain. The choice of graphical representation dramatically influences the perceived dominance of the power‑law versus the periodic component. The findings call for future mobility models, epidemic spread simulations, and DTN protocol designs to explicitly incorporate time‑varying contact rates rather than assuming stationary heavy‑tailed processes.