Rhythm and Randomness in Human Contact
Abstract
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 Lévy flight foraging strategy in animals, we analyse datasets on human contact from real world traces. By analysing the distribution of intercontact 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.
User population  Intel  Cambridge 1  INFOCOM 2005  Toronto  UCSD  Dartmouth  Europe 
Source  Chaintreau et al. (2006) [3]  Gonzalez et al. (2009) [10]  
Device  iMote  iMote  iMote  PDA  PDA  Laptop/PDA  Mobile phone 
Network type  Bluetooth  Bluetooth  Bluetooth  Bluetooth  WiFi  WiFi  Mobile phone 
Granularity  120 seconds  120 seconds  120 seconds  120 seconds  120 seconds  300 seconds  N/A 
Duration  3 days  5 days  3 days  16 days  77 days  114 days  6 months 
Devices participating  8  12  41  23  273  6648  100,000 
Number of internal contacts  1,091  4,229  22,459  2,802  195,364  4,058,284  16,364,308 
Approximate extent of power law region  4min  14min  10min  30min  10min  10h  2min  6min  20min  1day  10min  1h  100s  8h 
Quoted power law exponent  0.9  0.9  0.4  0.9  0.3  0.3  0.9 +/ 0.1 
Type of distribution plotted  Tail df (ccdf)  Tail df (ccdf)  Tail df (ccdf)  Tail df (ccdf)  Tail df (ccdf)  Tail df (ccdf)  Logbinned pdf 
Inferred stability exponent  0.9  0.9  0.4  0.9  0.3  0.3  0.1 +/ 0.1 
I Introduction
Digital traffic flows not only over the wired backbone of the Internet or network of mobile phone masts, but also in small leaps through physical space as people pass one another on the street [14]. Thus opportunities for a new communication paradigm via wirelessenabled devices are emerging, which communicate directly with other devices within their range and without a costly and inflexible planned infrastructure (e.g., [9]). To improve communication efficiency and prevent the spread of wireless viruses in this new generation of communication requires new insights and quantitative models of human interaction. Of fundamental importance in this case is the time sequence of human contacts, as well as other properties of complex networks, such as smallworldness, etc. (e.g., the special issue of Science on Complex Systems and Networks, July 24, 2009).
Recently, the emergence of human interaction traces from online and pervasive environments is allowing us to understand details of human activities. For example, the MIT Reality Mining project [6] collected proximity, location and activity information, with nearby nodes being discovered through periodic Bluetooth scans and location information from cell tower IDs. Several other groups have performed similar studies. Some have used Bluetooth to measure device connectivity [6], [9], [18], while others rely on WiFi [11], GPS [22, 23, 15], or the position of cell towers [10]. The duration of experiments has varied from 2 days to over one year, and the numbers of participants has also varied from to .
It has been suggested that the probability density function (pdf) of times between human contact is well approximated by a truncated power law i.e. over some range. This is so whether the contact is by physical proximity (i.e., detectability of wireless access points or Bluetooth devices, or closeness of GPS locations [3], [13], [22]) or by telecommunication (i.e., mobile phone call [10] or email [16]), and whether one or both contacting devices are in motion (e.g., both Bluetooth, one Bluetooth and fixed wireless access points, mobile phone and fixed masts).
A summary is given in Table I of studies in which the stability exponent has been inferred from an intercontact time (ICT) distribution, together with the approximate range of applicability. From the quoted values, is inferred to be in the interval which is within the allowable range () for the tails of a Lévy (stable) distribution [20, 17] (except possibly for the marginal case of the Europe study of mobile phone contact which could actually be a gamma distribution.) Consequently it has been argued that human mobility patterns resemble truncated Lévy walks (TLW). The TLW paradigm represents a development of the Lévy flight, which was a random walk comprising steps drawn from a Lévy distribution, rather than a Gaussian as occurs in the more familiar Brownian random walks [24]. The first modification, to a finite constant velocity, was dubbed a Lévy walk. Subsequently the limitation to a finite domain was described as truncation [17]. More recently some researchers have also considered the velocity to be a variable (e.g. [23]).
Similar movement patterns have also been inferred for animals [25], and it has been proposed that Lévy foraging is an optimal strategy under at least some circumstances [26]. Debate continues as to the extent to which a Lévy strategy could be universal and insensitive to the details of the environment and of the physiology and motivation of the individual (e.g. [8, 21], and references therein). However, the statistical analysis methods which have most frequently been used to infer empirical support for the truncated Lévy walk hypothesis have recently been criticised, both in the ecology literature and more generally [8, 7, 4, 27]. Key problems identified have included: (1) The widespread inference of power law pdfs by the graphical method of straight line fitting to histograms with double logarithmic axes; (2) the difficulty of inferring power laws over very limited ranges; (3) the use of intrinsically biased methods (such as (1)) for estimating the power law exponent; and, perhaps most importantly, (4) inadequate, or even a complete lack of, alternative hypotheses.
In the light of this, it is worthwhile to consider how these problems might apply to the human mobility studies cited above and summarised in Table I. For example, some simply compare their distributions with a straight line on a loglog plot with unavoidable bias and spread for the inferred power law exponent [3, 13]. In addition, referring to Table I, the inference of a possible power law region is very weak for the Intel, Cambridge 1 and Toronto experiments because the region is so limited ( decade), presumably related to the small samples ( contacts). The evidence is more convincing for the larger samples (INFOCOM 2005, UCSD, Dartmouth and Europe) with wider apparent power law regions. Alternative hypotheses to the pure power law null model have been considered, such as the exponentiallytruncated power law [13, 10], but only one study [15] has actually fitted and quantitatively compared several alternative models to ICT distributions (albeit simulated), using the less biased maximum likelihood estimate to infer the model parameters such as the power law exponent and Akaike weights [1] to compare the goodness of fits. Thus, at present, the inference of a truncated power law ICT distribution directly from experiment is limited.
Indeed, it would be surprising if a truncated power law was a complete description of human ICT considering our prior knowledge about the social habits and structures of humans, such as the working day and family and community responsibilities [16]. In fact it has been recognised that the ICT distribution is not stationary and changes with the time of day [13]. Spatial movement distributions also exhibit daily patterns [10] and Fourier analyses of proximity edges have daily and weekly periodicities [6]. Similarly a fundamental semidiurnal periodicity was identified in an early study claiming a Lévy strategy for animal foraging [25]. This suggests that alternative models combining nontrivial randomness and periodic rhythms should be investigated. At present such more complicated models are challenging to test rigorously (e.g., by MLE) but progress can nevertheless be made by closer examination of the experimental ICT distribution using different graphical methods and modelling.
In this paper we consider three similar human contact experiments of varying durations (section 2).
We analyse and model them to identify regularities that modify the underlying Lévy walk
behaviour (section 3). Then we briefly compare these analyses with others in the literature and
discuss how this hybrid behaviour may be modelled and will affect the efficiency of adhoc
communication (section 4).
Ii Datasets
We analyse trace data from the Haggle project [9] and Crawdad database [5],
collected using Bluetooth communication in a conference environment and two university study
environments. The configuration of data collection is summarised in
Table II.
MIT: in the MIT Reality Mining project [6], 100 smart phones were deployed
to students and staff at MIT over a period of 9 months. These phones were running software that
logged contacts.
Cambridge 2: in the Cambridge Haggle project [9], 36 iMotes (Intel Mote
ISN100BA) were deployed to 1st year and 2nd year undergraduate students for 11 days to detect
proximity using Bluetooth. The iMote runs TinyOS and is equipped with an ARM7TDMI processor
operating at 12MHz, with 64kB of SRAM, 512kB of flash storage, and a multicoloured LED, and a
Bluetooth 1.1 radio, which has a radio range around 30 meters.
INFOCOM 2006: also in the Cambridge Haggle project,
77 iMotes were deployed at the INFOCOM 2006 conference for 3 days.
Experimental data set  MIT  Cambridge 2  INFOCOM 2006 

Device  Phone  iMote  iMote 
Network type  Bluetooth  Bluetooth  Bluetooth 
Duration (days)  246  11  3 
Granularity (seconds)  600  100  100 
Number of Devices  97  36  77 
Number of Contacts  54,667  10,873  191,336 
Average # Contacts / Day  0.024  0.345  6.7 
The logged data from the above experimental studies are used to build timedependent network information to study the distribution of contact times, intercontact times, community structure and their statistical properties, where we constructed discrete event traces of pair interactions of 10 to 600 seconds intervals. We have aggregated raw data within 100 or 600 second time windows to avoid uncertainty of device detection from a complex Bluetooth communication protocol.
A complex operation is required to collect accurate connectivity traces
using Bluetooth communication, as the device discovery protocol may limit
detection of the devices in radio proximity. Bluetooth uses a special
physical channel for devices to discover each other.
A device becomes discoverable by entering the inquiry substate where it can
respond to inquiries from other devices. The inquiry scan substate is used
to discover surrounding devices. The discovering device iterates (hops)
through all possible inquiry scan channel frequencies in a pseudorandom
fashion.
For each frequency, it broadcasts an inquiry and listens for responses.
Therefore, a Bluetooth
device cannot scan for other devices when the device cannot be in
discoverable. Bluetooth inquiry can only happen in second intervals.
It is reported that an interval of seconds gives a
more than 90% chance of finding a device. However, there is no available
data for situations where many devices are present, and no precise study has
been reported. The Bluetooth standard recommends being in the inquiry scan
substate for 10.24 seconds in order to collect all responses in an
errorfree environment. A 10.24 seconds alternation may cause missing links,
and we therefore deploy 5.12 seconds for inquiry.
The power consumption of Bluetooth is also a critical limitation for the
scanning interval. The iMote connectivity traces in Haggle [9] use
a scanning interval of approximately 2 minutes, while the Reality Mining
project in MIT [6], with cell phones, uses 5 minutes. The
ratio of devices with Bluetooth enabled to the total number of devices is
around only an average 15%  20% of population. The range of Bluetooth
varies between 10m and 80m, which depends on the device class such as cell
phones or laptops. In cell phones, the Bluetooth range is usually 5  10m.
We have observed that the devices can be detected in a 20m range if there
are no obstacles, while with obstacles such as a thick wall the range drops
to 5m (see more detail in [18][19]).
Iii Rhythm and randomness
In each of the experiments we calculated all possible intercontact times between any two
nodes, where ICT is defined as the time between the end of contact between two nodes and the start
of next contact between the same two nodes. Figures 13 summarise the ICT distribution for the
three experiments. In each case, the distribution is plotted as (a) a rank order plot with double
logarithmic axes, (b) a probability density function with logarithmic coordinate (probability
density) axis and logarithmic ordinate (intercontact time) axis using exponentially spaced bins
(i.e., equal bin width in logarithm space = 0.1 decade), (c) a histogram with logarithmic
coordinate (frequency) axis and linear ordinate (intercontact time) axis using 100 equallyspaced
bins (equivalent to a pdf with linearly spaced bins to within a constant), and (d) a histogram for
intercontact times up to 12 hours with logarithmic coordinate (frequency) axis and logarithmic
ordinate (intercontact time) axis using equal bin widths at the granularity s or
600 s. (The inset in Figure 3c shows a double logarithmic histogram using equal bin widths of
1800 s.)
Iiia Truncated power law distribution
Considering the rank order plots in Figure 13, we might suggest as others have done that the ICT tail distribution of all three experiments roughly resembles a restricted range power law with exponent (cf Figure 1 and 2 of [3]). To illustrate this, we performed the following simulation:

A set of contact times is calculated for Lévy walks in a domain bounded by the duration of the experiment (see Table II). Specifically we calculate the cumulative sum, , where is a set of iid samples chosen from the Pareto distribution with pdf in the range to . The samples are generated by picking iid samples from the uniform distribution in the range (0,1] and then inverting the analytical equation for the Pareto cumulative probability distribution to find the value that yields the value .

Divide the contact times into individual trials (i.e., trial number ).

Calculate the set of intercontact times from the time differences between neighbouring contact times, , omitting intercontact times that straddle trials.
Figure 4a shows a simulated ICT probability
distribution (solid line) choosing and other parameters corresponding to the
configuration of the INFOCOM 2006 experiment – s, days, and .
It is clear that the simulated distribution is only a crude approximation to the actual INFOCOM 2006 distribution
(dashed line) and other structure is evident.
IiiB Circadian rhythm
This is also obvious in the other experiments (e.g., the histograms in figures 1c, 2c and 3c.) where there are significant deviations about any candidate monotonic function. Closer inspection reveals much of this deviation to be associated with a circadian rhythm, as evidenced by the alignment of peaks in the histogram/PDF at integer multiples of 24 hours. (Note also a weekly rhythm in Figure 3.)
Nevertheless, the INFOCOM 2006 and MIT distributions are well approximated by a power law on time scales much
less than a day (e.g., hours, see Figure 1d and 3d). (This is less obvious in the Cambridge 2 experiment
(Figure 2d) due to a min periodicity which is likely an experimental artefact). This suggests that a better
null model of ICT in these experiments is a
Lévy walk in a periodic domain. To investigate this we performed the following simulation:

A set of contact times is calculated for Lévy walks in a domain bounded by the duration of the experiment (see Table II). Specifically we calculate the cumulative sum, , where is a set of iid samples chosen from the Pareto distribution with pdf in the range to . The samples are generated by picking iid samples from the uniform distribution in the range (0,1] and then inverting the analytical equation for the Pareto cumulative probability distribution to find the value that yields the value .

Divide the contact times into days and retain only contact times that fall within a working day, defined to start at h and end at h (i.e., ).

Divide the contact times into individual trials (i.e., trial number ).

Calculate the set of intercontact times from the time differences between neighbouring contact times, , omitting intercontact times that straddle trials.
Figure 4b shows a simulated ICT probability distribution (solid line) choosing and other parameters corresponding to the configuration of the INFOCOM 2006 experiment: s, days, and . The simulated distribution compares favourably with the actual INFOCOM 2006 distribution (dashed line), supporting the null model over a Lévy walk confined within the domain but not within the working day.
Iv Conclusions and Implications
The distribution of human intercontact times from three experiments of differing durations has
been analysed using different graphical presentations. This has revealed three essential properties
of human contact:
Random, scalefree. On sufficiently short time scales, the ICT distribution is
approximated by a power law consistent with the return times of a Lévy flight. The value of the
stability exponent () implies no characteristic ICT in the absence of other
constraints.
Truncated. At some time scale the power law component is truncated by a constraint on
intercontact time. One artificial constraint is the experiment itself which prohibits recording
ICTs longer than the experiment duration. This is demonstrated in the simulated ICT distribution in
Figure 4a and should be considered in comparing results from experiments of differing durations.
More significantly, another constraint is the removal of agents from the contact domain. An example
of this is movement from work to home which suppresses ICTs between agents in the same work group
on times scales beyond the working day. This is demonstrated in the simulated ICT distribution in
Figure 4b by the truncation of the power law component at s.
Periodic. Environmental, biological, and social constraints may have rhythms that
encourage repeated encounters such as the daily toing and froing between work and home. This is
demonstrated in the simulated ICT distribution in Figure 4b by the peak at s and s (i.e., separated by 24 hours).
These three properties have been previously surmised by various different means but evidence of
their coexistence in the ICT distribution has been overlooked. In particular, closer examination
of previously published ICT distributions (e.g., [10]) reveals deviations about a truncated
power law consistent with a circadian rhythm. Recognition of this rhythm in the empirical
distribution is important otherwise models of human movement and behaviour may be unrealistically
modified to generate only the scalefree property (e.g., [2]). It also has
significant implications for building efficient routing algorithms and functionality on top of
opportunistic networks. As a very simple example, clearly a rhythm of period that removes
agents from each other for a time reduces the average number of contacts by 50% over
multiple cycles. But its determinism might also be exploited to increase communication efficiency.
For example, the time of the next encounter could be estimated at the node and thus selection of
the next hop could be determined based on the expected shortest time to the next encounter. The
periodic behaviour of nodes could indicate moving from one network partition to another and this
could be used for temporal clustering of nodes, where temporalbased communities could be used as a
backbone of logical network structure for forwarding [12]. By these means,
mobilityassisted forwarding can take advantage of patterns arising in the distribution of nodes in
time and space. One alternative movement model is suggested that combines the Lévy walk model
with models such as the Home Cell Mobility Model (e.g., [2]) that incorporate the
influence of social structure. However the development of more complicated models will also present
challenges in testing them and distinguishing between competing models.
Acknowledgement
The research is part funded by the EU grants for the Haggle project, IST4027918, the SOCIALNETS project, 217141, and the EPSRC DDEPI Project, EP/H003959. MPF and NWW are supported by the Natural Environment Research Council.
References
 [1] D. R. Anderson and K. P. Burnham. Understanding information criteria for selection among capturerecapture or ring recovery models, S1421. Bird Study, 46 (suppl.), 1999.
 [2] C. Boldrini, M. Conti, and A. Passarella. The sociable traveller: human travelling patterns in socialbased mobility. In Proc. MOBIWAC09, 2009.
 [3] A. Chaintreau, P. Hui, J. Crowcroft, C. Diot, R. Gass, and J. Scott. Impact of human mobility on the design of opportunistic forwarding algorithms. In Proc. INFOCOM, April 2006.
 [4] A. Clauset, C. R. Shalizi, and M. E. J. Newman. Powerlaw distributions in empirical data. SIAM Review, 51, 2009.
 [5] Dartmouth College. A community resource for archiving wireless data at dartmouth, http://crawdad.cs.dartmouth.edu/index.php, 2007.
 [6] N. Eagle and A. Pentland. Reality mining: sensing complex social systems. Personal and Ubiquitous Computing, V10(4):255–268, May 2006.
 [7] A. M. Edwards. Using likelihood to test for Lévy flight search patterns and for general powerlaw distributions in nature. Journal of Animal Ecology, 77, 2008.
 [8] A. M. Edwards, R. Phillips, N. W. Watkins, M. P. Freeman, and Others. Revisiting Lévy flight search patterns of wandering albatrosses, bumblebees and deer. Nature, 449, 2007.
 [9] EU FP6 Haggle Project. http://www.haggleproject.org, 2010.
 [10] M. Gonzalez, C. Hidalgo, and A. Barabasi. Understanding individual human mobility patterns. Nature, 453(7196), 2008.
 [11] T. Henderson, D. Kotz, and I. Abyzov. The changing usage of a mature campuswide wireless network. In Proc. Mobicom, 2004.
 [12] P. Hui, J. Crowcroft, and E. Yoneki. BUBBLE Rap: Social Based Forwarding in Delay Tolerant Networks. In MobiHoc, 2008.
 [13] T. Karagiannis, J.Y. L. Boudec, and M. Cojnovic. Power law and exponential decay of inter contact times between mobile devices. In Proc. MOBICOM, 2007.
 [14] J. Kleinberg. The wireless epidemic. Nature, 449(20), 2007.
 [15] K. Lee, S. Hong, and Others. Slaw: A mobility model for human walks. In Proc. INFOCOM, 2009.
 [16] R. D. Malmgren, D. B. Stouffer, A. E. Motter, and L. A. N. Amaral. A Poissonian explanation for heavy tails in email communication. PNAS, 105, 2008.
 [17] R. N. Mantegna and H. E. Stanley. An introduction to econophysics: Correlations and complexity in finance. Cambridge University Press, 2000.
 [18] T. Nicolai, E. Yoneki, N. Behrens, and H. Kenn. Exploring social context with the wireless rope. In Proc. Workshop MONET: LNCS 4277, 2006.
 [19] E. Nordstrom, C. Diot, R. Gass, and P. P. Gunningberg. Experiences from measuring human mobility using bluetooth inquiring devices. In Proc. MobiEval, 2007.
 [20] W. Paul and J. Baschnagel. Stochastic processes from physics to finance. Springer, 1999.
 [21] A. M. Reynolds and C. J. Rhodes. The Lévy flight paradigm: random search patterns and mechanisms. Ecology, 90, 2009.
 [22] I. Rhee, M. Shin, and Others. Human mobility patterns and their impact on routing in humandriven mobile networks. In Proc. HotnetsVI, 2007.
 [23] I. Rhee, M. Shin, and Others. On the Lévywalk Nature of Human Mobility. In Proc. INFOCOM, 2008.
 [24] M. F. Shlesinger, G. M. Zaslavsky, and U. Frisch. Lévy flights and related topics in physics: Lecture Notes in Physics 450. Springer, 1995.
 [25] G. M. Viswanathan and Others. Lévy flight search patterns of wandering albatrosses. Nature, 381, 1996.
 [26] G. M. Viswanathan and Others. Optimizing the success of random searches. Nature, 401, 1999.
 [27] E. P. White, B. J. Enquist, and J. L. Green. On estimating the exponent of powerlaw frequency distributions. Ecology, 89, 2008.