ArticlePDF Available

On data processing required to derive mobility patterns from passively-generated mobile phone data

Authors:

Abstract and Figures

Passively-generated mobile phone data is emerging as a potential data source for transportation research and applications. Despite the large amount of studies based on the mobile phone data, only a few have reported the properties of such data, and documented how they have processed the data. In this paper, we describe two types of common mobile phone data: Call Details Record (CDR) data and sightings data, and propose a data processing framework and the associated algorithms to address two key issues associated with the sightings data: locational uncertainty and oscillation. We show the effectiveness of our proposed methods in addressing these two issues compared to the state of art algorithms in the field. We also demonstrate that without proper processing applied to the data, the statistical regularity of human mobility patterns-a key, significant trait identified for human mobility-is over-estimated. We hope this study will stimulate more studies in examining the properties of such data and developing methods to address them. Though not as glamorous as those directly deriving insights on mobility patterns (such as statistical regularity), understanding properties of such data and developing methods to address them is a fundamental research topic on which important insights are derived on mobility patterns.
Content may be subject to copyright.
On data processing required to derive mobility patterns from
passively-generated mobile phone data
Feilong Wang and Cynthia Chen*
Department of Civil and Environmental Engineering, University of Washington, Seattle, United
States
Abstract
Passively-generated mobile phone data is emerging as a potential data source for transportation
research and applications. Despite the large amount of studies based on the mobile phone data,
only a few have reported the properties of such data, and documented how they have processed the
data. In this paper, we describe two types of common mobile phone data: Call Details Record
(CDR) data and sightings data, and propose a data processing framework and the associated
algorithms to address two key issues associated with the sightings data: locational uncertainty and
oscillation. We show the effectiveness of our proposed methods in addressing these two issues
compared to the state of art algorithms in the field. We also demonstrate that without proper
processing applied to the data, the statistical regularity of human mobility patterns—a key,
significant trait identified for human mobility—is over-estimated. We hope this study will
stimulate more studies in examining the properties of such data and developing methods to address
them. Though not as glamorous as those directly deriving insights on mobility patterns (such as
statistical regularity), understanding properties of such data and developing methods to address
them is a fundamental research topic on which important insights are derived on mobility patterns.
Keywords
Human mobility trajectory; Locational uncertainty; Oscillation problem; Representativeness issue;
Incremental clustering method; Time-window-based method; Statistical regularity
1. Introduction
In the past decade, passively-generated mobile phone data (hereafter called “mobile phone
data”) has risen as a viable data source for human mobility analysis and transportation
applications. It is the by-product of the primary, often non-transportation related purposes
such as billing and operations (Alexander et al., 2015a; Chen et al., 2014, 2016; Iqbal et al.,
2014; Toole et al., 2015). Often location- and time-stamped, mobile phone data offer a
tremendous opportunity to revolutionize the transportation field that goes beyond behavior
analysis, traffic operations and safety analysis (Vlahogianni et al., 2015). Indeed, a search in
the literature using keyword combinations of mobile phone data, mobility, and travel
behavior resulted in more than 1000 articles published in journals across different disciplines
*Corresponding author. qzchen@u.washington.edu (C. Chen).
HHS Public Access
Author manuscript
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Published in final edited form as:
Transp Res Part C Emerg Technol
. 2018 February ; 87: 58–74. doi:10.1016/j.trc.2017.12.003.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
(Ahas et al., 2010a; Becker et al., 2013; Calabrese et al., 2013; Candia et al., 2008; Chen et
al., 2014, 2016; Gao et al., 2013; Song et al., 2010b; Wang et al., 2014). These articles cover
a wide range of topics including, for example, estimating mobility patterns (Csáji et al.,
2013; González et al., 2008; Song et al., 2010a), inferring OD matrix (Calabrese et al.,
2011b; Iqbal et al., 2014), finding anchor locations (Dong et al., 2015; Isaacman et al.,
2011), inferring activity types (Jiang et al., 2017; Widhalm et al., 2015) and travel modes
(Qu et al., 2015; Wang et al., 2010).
Unlike household travel survey data that is generated by asking subjects record an entire
day’s activities and travels, thus resulting in complete information on one’s spatio-temporal
movement, mobile phone data generates one or more records when a user connects to a
network (Chen et al., 2016; Wang et al., 2014). Each record typically includes the device ID,
the timestamp and the location information. Two kinds of mobile phone data have been
frequently used today: CDR (Call Detail Record) data and sightings data (Calabrese et al.,
2013; Chen et al., 2014), where the former (CDR data) comprises a set of phone call records
along with the time information and the location information of the cell tower that channels
the call, and the sightings data contains sightings that are results of the triangulation of cell
towers. Hereby we call either a record or a sighting as a trace. Neither data covers the full
spectrum of a user’s movement in a day. Since both rely on the phone usage, both are
temporally sparse: for sightings data, the average time between two consecutive network-
connection activities could be longer than one hour and for CDR data the interval could be
even longer (Calabrese et al., 2013). This issue is in addition to their much coarser spatial
resolution in location estimation compared with data from household travel surveys and GPS
(Global Positioning System)-based surveys. The precision of location estimation is reported
to be approximately 300 m on average in urban area for sightings data (Jiang et al., 2013)
and between several hundred meters to several kilometers for CDR data, depending on the
density of cellular towers (Chen et al., 2016; Järv et al., 2014). Another issue is purely
signaling related: some traces are resulted from signaling activities, as opposed to users’
movements. That is, when a user stays at one location, his/her phone may be handed over to
different cell towers due to load-balancing or other operational purposes. This results in
traces suggesting that the user bounces between multiple stations. This is known as
oscillation (Lee and Hou, 2006; Qi et al., 2016).
Additionally, because a trace in mobile phone data represents either a phone call (for CDR
data) or a sighting when a phone is seen on the network (for sightings data), a trace in the
mobile phone data cannot be viewed directly as an activity location where a person spends
time for an activity. Because of this and the previously noted issues such as uncertainty and
oscillation, the mobile phone data must be processed first before it can be used for mobility
analysis. The word “processed” here specificially refers to the process that must be applied
to mobile phone data to address issues such as uncertainty and oscillation in order to infer
activity and trip related information. This process is uniquely different from the cleaning
process typically applied to household travel survey data whose records directly represent
activities and trips conducted by the subjects. Clearly we do not need to emphasize the
importance of this processing task: the results of processing directly affect the results of
mobility analysis using such data (Horn et al., 2014). We demonstrate this in Section 5.3 of
the paper.
Wang and Chen Page 2
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Surprisingly, among the many studies that have used mobile phone data, only a few have
reported the properties and issues associated with such data, and documented how they have
processed the data and how the results compare to external data sources that may provide
some validations (Palchykov et al., 2014; Serok and Blumenfeld-Lieberthal, 2015; Yin et al.,
2017). This paper focuses on mobile phone sightings data, in particular, seeking to illustrate
and address two prominent issues of the dataset: (1) locational uncertainty and (2)
oscillation. The sightings data has now generated many papers in the field (Calabrese et al.,
2011a, 2013; Jiang et al., 2013), is already being used by several planning agencies and
departments of transportation in the US (Milone, 2015; Stabler and Sikder, 2014) and the
data provider is now widely advertising its products (e.g., OD matrices) derived from such
data (Milone, 2015). Thus, having a clear understanding of such data and the techniques that
can be applied to process such data is not only important for researchers but also for
practitioners who may be interested in using such data or their derived products. The first
issue (locational uncertainty) is unique to the sightings data, as every location estimate
contained in the data is unique, as the results of triangulation. Methods are designed to deal
with this issue so that activity locations (where people spend time to conduct activities) can
be identified from the sightings data. A common solution is to define a radius and a
minimum duration corresponding to the positioning errors and the time spent on activities
respectively (Jiang et al., 2013). Then an activity location is identified if a sequence of traces
meets both the spatial and temporal constraints. The second issue—oscillation—exists in
both sightings and CDR data. The phenomenon cannot be ignored as traces resulted from
oscillation take up a substantial fraction of the total number of records: approximately 30%
(Lee and Hou, 2006). Solutions, however, especially for sightings data, are lacking and are
mostly discussed in computer science communities (Bayir et al., 2010; Lee and Hou, 2006;
Shad et al., 2012). Two approaches have been commonly used to detect oscillation
sequences in the traces: pattern-based and hybrid methods. The former examines trace
sequences and the one that exhibits a specific switching pattern will be identified as the
oscillation case (Lee and Hou, 2006). The latter extends pattern-based methods by utilizing
temporal and/or spatial information to consider velocity or other measurements (Iovan et al.,
2013). Our paper reviews these existing methods and proposes new solutions. We show that
our proposed solutions exceed the performance of the existing methods in addressing these
issues.
The rest of the paper is organized as follows. After reviewing related works in Section 2,
Section 3 introduces the sightings data used in our study. In Section 4, we construct a data-
processing framework that provides solutions for the two problems. The framework consists
of two stages: the first is to address the issue of locational uncertainty by proposing a revised
incremental clustering algorithm; and the second is to handle the oscillation problem with a
time-window-based method in addition to a modified pattern-based method. Through
analyzing the processed data, we show advantages and effectiveness of our framework.
Section 5 covers this part of work. Lastly in Section 6, we provide a discussion on related
questions that are not addressed within the scope of this paper. The representativeness issue,
which is related to the issue of temporal sparsity, is discussed in this section. We also discuss
how our developed framework and solutions can be applied to the CDR data. We hope our
work will stimulate many more studies addressing this critical topic that will determine not
Wang and Chen Page 3
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
only the accuracy of the results from using such data but also its wide applications in the
future.
2. Related works
2.1. Understanding passively generated mobile phone data
As noted earlier, two types of mobile phone data have been commonly studied: Call Details
Record (CDR) data (Ahas et al., 2010a; Järv et al., 2014; Xu et al., 2015) and the sightings
data(Calabrese et al., 2011b; Chen et al., 2016; Wang, 2014). In the former, each record
represents a device-driven activity (e.g., phone calls, messaging services and Internet
service) and contains information such as the initial time of the activity, the duration, and the
ID and location information of the cell tower that channels the activity.
The sightings data differs with the CDR data in two aspects: (1) more records: in addition to
just recording device-driven activities, data is also generated from other network-driven
interactions, such as handovers when a device switches between two neighboring cell areas
(Demissie et al., 2013; Hard et al., 2016; Iovan et al., 2013); and (2) higher spatial accuracy:
instead of cell-level location information as in the CDR data, the location information
contained in the sightings data is an estimation of the device’s location through triangulation
of multiple cells (Calabrese et al., 2011b; Chen et al., 2016).
Passively generated, both types are characterized with spatio-temporal uncertainties. From
the temporal perspective, depending on device usage pattern, the temporal distribution of
records can be irregular and heterogeneous (Iovan et al., 2013). Consequently, the time
interval between two consecutive records is uncertain and could be too long to capture users'
movements that took place during the long time gap. The interval for a sample CDR data is
reported with a mean of as long as eight hours (González et al., 2008). And a median of
slightly longer than one hour is reported for a sample sightings data (Calabrese et al.,
2011b). One way to address the temporal issues is to apply a data-filtering process, such as
only selecting highly active users (Calabrese et al., 2013; González et al., 2008; Song et al.,
2010b; Zhao et al., 2016). However, results from those studies may be affected by the
selected filtering process due to potential correlations between individual mobility and
device usage, e.g., higher communication frequency is likely to lead to more trips derived
(Iovan et al., 2013; Yuan et al., 2012). Though not directly addressed in the current study, we
discuss it in Section 6.
From the spatial perspective, several uncertainty issues exist in the mobile phone data, which
affect individual mobility derived from data. The first issue existing in both types of data is
the relatively low “spatial accuracy”. Depending on the density of cellular towers, the
accuracy of location estimates varies. The accuracy is reported to be about 300 m on average
in urban area for sightings data (Calabrese et al., 2011b; Hard et al., 2016; Jiang et al., 2013)
and between several hundred meters to several kilometers for CDR data corresponding to the
coverage of a tower/antenna (Järv et al., 2014). There are efforts to improve spatial accuracy,
for example, via interpolation methods (Järv et al., 2017; Louail et al., 2014). The second
uncertainty related issue that exists in both types of data is the oscillation phenomenon: due
to dynamic factors including varying transmission conditions that influence received signal
Wang and Chen Page 4
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
strength and load balancing policies, instead of connecting to the nearest cell tower, a mobile
phone switches its connection between different towers (typically within a short period of
time) even though the device itself is not moving (Calabrese et al., 2011b; Wu et al., 2014).
The oscillation phenomenon generates a considerable number of records that do not reflect
devices' actual movements (Lee and Hou, 2006). The third issue is specifically related to
sightings data: location estimates for a stay at the same location change over time. This is
related to the method of triangulation, as measurements on factors used for triangulation
(e.g., number of surrounding cell towers, received signal strength) are subject to fluctuations,
leading to distinct estimations for the same location (Calabrese et al., 2011a). Aggregation
on these varying location estimates is required so that we can identify activity locations and
calculate the time spent at those locations (Hard et al., 2016).
2.2. Addressing locational uncertainty by extracting activity locations
For CDR data with location information at the cell level, frequently visited locations (e.g.
home, work places) by an individual are usually indicated by locations of network cells that
were most frequently visited (Ahas et al., 2010b; Isaacman et al., 2011; Järv et al., 2014;
Wang et al., 2012; Xu et al., 2016). By clustering cell towers that are spatially close, these
locations could also be represented by clusters containing towers that frequently channel call
activities (Jiang et al., 2017; Xu et al., 2016).
However, for sightings data, as the location estimation is the result of triangulation among
multiple towers (Chen et al., 2016; Widhalm et al., 2015), each of them is unique. In other
words, even when a user stays at the same location, locations recorded in the data differ from
time to time (though they may scatter in close proximity) (Fig. 1). This means the methods
for identifying activity locations from the CDR data, such as those relying on counting call
activity frequency with cell towers, are not applicable for the sightings data. The common
way to address this locational uncertainty and to reveal activity locations from the sightings
data is to aggregate these distinct location estimates by applying a clustering algorithm
(Calabrese et al., 2011b; Chen et al., 2016; Jiang et al., 2013). Typically, the centroid of the
outputting clusters is used as representing the activity location.
Conventional clustering methods are usually not applicable mainly because parameters in
these methods need to be pre-determined, which is infeasible in mobile phone data. For
example, popular methods such as k-means and DBSCAN (Ester et al., 1996; Kanungo et
al., 2002) require the number of clusters or density-related parameters be known, which
cannot be specifically identified because of the variety of mobile phone usage and travel
behaviors that exist across individuals (Chen et al., 2014).
Some new methods aggregate traces by segmenting one trajectory into several sequences of
traces. Here, one trajectory of a user refers to the user’s available traces of one day. In the
trajectory-segmentation methods, an activity location is defined as a sequence of consecutive
traces bounded by both temporal and spatial constraints (Calabrese et al., 2011a, 2013; Jiang
et al., 2013). Given the trajectory {d0, d1, d2, …, dk}, one cluster of traces {dm, dm+1, …,
dn} (0 ≤ m, n ≤ k) will be defined as an activity location, if two conditions are met: (1) the
maximal geographical distance between any two location estimates Ldi and Ldj (m ≤ i, j ≤ n)
should not exceed Rc (here Ldi, Ldj are location records of traces di and dj respectively and
Wang and Chen Page 5
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Rc is the spatial constraint (e.g., 300 m); and (2) the time difference between the last and the
first trace of this sequence (i.e. tdn–tdm) should be at least Tc (here tdn and tdm are
timestamps of traces dn and dm respectively and Tc is the temporal constraint (e.g., 10 min)).
Thus, in the trajectory-segmentation method, activity locations can be identified by applying
these two conditions and segmenting the trajectories into sequences. This method is intuitive
and easy to apply. But, the first condition, requiring any two traces in the sequence to satisfy
the distance constraint, can pose challenges if outliers exist. For example, given a sequence
of traces {dm, dm+1, …, dn} that are spatially close, the insertion of an outlier doutlier may
result into three clusters: {dm, dm+1, …, dm+k}, {doutlier} and {dm+k+1, …, dn},
consequently leading to biased estimation on the number and the types of activity places one
has visited.
Chen et al. (2014) suggested a model-based clustering method. Rather than relying on a pre-
determined threshold value for all devices, this method determines the optimal number of
clusters for each device through a statistical model, more specifically, by searching through a
finite set of possible numbers of clusters and finding the one that generates the highest
Bayesian Information Criterion (BIC) (Fraley, 1998; Fraley and Raftery, 2002). The centroid
of an identified activity cluster (defined as a sequence of traces that are considered to be
activity places) is used as activity locations for subsequent modeling processes. The
advantage of applying the model-based clustering is obvious: there is no requirement for
pre-determined clustering parameters. However, there are limitations. This method is
sensitive to the spatial density of the traces, which are often of low density. Additionally, the
size of outputting clusters cannot be controlled, meaning that several far-away outliers may
be clustered together, causing the resulting cluster to stray away (Fig. 2): in other words, a
cluster without spatial constraint makes the clustering process less meaningful.
Another method to aggregate traces is the incremental clustering method (Alexander et al.,
2015a; Hariharan and Toyama, 2004; Wang et al., 2015; Widhalm et al., 2015). Given traces
{d0, d1, d2, …, dk} in one trajectory, the clustering is performed as follows: (1) starting from
trace d0, one new cluster C0 is created and d0 is the center; (2) each trace that is not clustered
will be checked and the trace within a distance Rc to the center of C0 is aggregated to the
cluster; (3) every time the cluster grows, its center is updated; (4) if no trace could be
aggregated in the current cluster, one new cluster will be created containing the non-
clustered trace. This procedure repeats itself until all traces are clustered. After the clustering
process, clusters with a time duration of no less than Tc will be considered as activity
locations. The cluster duration is calculated as the largest time difference between any two
traces in the cluster. This clustering method is relatively robust to outliers present among
traces, as the mechanism of updating centers (step 3) improves its tolerance to outliers.
Most of above mentioned clustering methods are applied to individual days of traces. In data
of multiple days, however, the same activity location may appear distinct in terms of their
coordinates. This creates challenges for the identification of common activity locations, such
as home and work locations (Xu et al., 2015), and observations on mobility patterns such as
the regularity property (González et al., 2008). Therefore, it would be necessary to develop
methods to identify common activity locations where users conduct activities on multiple
days. By dividing the study area into small grids, activity locations on multiple days are
Wang and Chen Page 6
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
regarded as a common location if they fall into the same grid (Andrienko et al., 2011; Zheng
et al., 2010). Jiang et al. identified the common locations by aggregating activity locations of
multiple days via the agglomerative clustering algorithm (Jiang et al., 2013).
2.3. Removing oscillation sequences
By applying the clustering algorithms, the issue of locational uncertainty is dealt with so that
activity locations could be identified. However, the results can still be biased if traces
generated from oscillations are not removed (Qi et al., 2016; Wu et al., 2014). When
oscillation occurs, some devices may be observed switching between two (or more) far-away
locations with high frequencies. In most cases, an oscillation pattern forms, such as L0-L1-
L0-L1-L0, where L0 and L1 are distinct location estimates. Table 1 provides an example
oscillation case. The Euclidean distance between locations L0 and L1 is 2.7 km, but the
user’s location switches between them three times within four minutes. If the switching
speeds are calculated, they are incredibly high, which should not reflect the actual movement
of the user. Since oscillation is a common phenomenon, a data processing step must be
applied.
A few heuristic rules were proposed for CDR data to detect oscillations among cellular
towers. In general, they are based on the observation that oscillation generates trace
sequences showing unique patterns (e.g., L0–L1–L0–L1) and usually lead to an illusion that
users travel with an incredibly high speed (e.g. 1000 km/h) (Bayir et al., 2010; Lee and Hou,
2006; Shad et al., 2012; Yin et al., 2017).
Lee and Hou (2006) defined a sequence of consecutive location estimates with pattern L0
L1–L0–L1 or L0–L1–L2–L0 as an oscillation sequence. A similar method was introduced by
Bayir et al. (2010). According to their definition, an oscillating sequence is detected if at
least triple switches are found in one trip (One switch refers to a one-time switch between
two locations). For example, the sequence of location estimates {L0–L1–L0–L2–L3–L2–L1}
is regarded as an oscillation sequence because of the triple switches between L0 and L1. This
method was designed for a unique dataset collected by using an application installed on
subjects’ mobile phones and requesting subjects to put semantic tags to locations they
visited. Shad et al. (2012) investigated the solution to oscillation using the same data. The
similar pattern-based method they proposed relies on cellular towers’ Location Area Code
and radius information. For datasets where semantic tags and radius information are
missing, it is hard to implement these methods. Meanwhile, relying on the unique patterns
detected in the sequence of traces risks mistaking actual travels for occurrences of
oscillation. In other words, an individual may indeed travel between two places multiple
times in a typical day and traces generated from this actual travel behavior could be
mistaken as oscillation traces.
A few methods have been designed to improve the detection precision by utilizing the
temporal information available in the data (Iovan et al., 2013; Qi et al., 2016; Widhalm et al.,
2015; Wu et al., 2014). In an oscillation sequence, the frequent switches between distinct
locations give the illusion that people travel at an abnormally high speed. This problem is
considered in the speed-based methods, where the switching speed is calculated. An
oscillation sequence is identified if the switching speed exceeds a given threshold Vc. Iovan
Wang and Chen Page 7
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
et al. (2013) set Vc as 200 km/h. In their work, they extended the definition with one more
condition: for an oscillation case, the difference between the heading directions of two
consecutive displacements should be 180 degrees, which corresponds to two switches (i.e.
L0–L1–L0) in the pattern-based methods. Thus, their work can be regarded as a hybrid
method – a combination of pattern-based and speed-based method. But the criterion of two
switches is too simple to detect complex patterns that involve more than two locations.
Meanwhile, given the low accuracy of location estimation, the method involving the speed
measurement could be misleading, especially if the time intervals of switches are small
(Demissie et al., 2013).
Wu et al. elaborated several heuristic rules to detect oscillation traces (Wu et al., 2014).
Working on CDR data, they begin with finding suspicious sequences that contain at least
three traces from multiple cellular towers within a short time period (e.g., one minute). Then
a suspicious sequence is confirmed as an oscillation sequence if it contains a circular event
(i.e. a tour L0–L1–…–L0). In this rule, the limit of a short time window is to account for the
observation that oscillation is a short time event and the circular-event criterion is to rule out
suspicious sequences resulting from fast movements.
Following the detection of oscillation sequences, a meaningful location needs to be
identified to represent the device’s location when oscillation occurs. This meaningful
location could be the one, among all locations in the sequence, which is visited most
frequently (Iovan et al., 2013). Wu et al. (2014) introduced a score for each location in the
oscillation sequence by considering both the visiting frequency and the average distance to
other locations. The one that receives the highest score is selected as the meaningful
location.
Oscillation solutions for sightings data are limited (Alexander et al., 2015a; Dong et al.,
2015; Palchykov et al., 2014; Yin et al., 2017). Calabrese et al. (2011b) and Widhalm et al.
(2015) addressed two problems (both locational uncertainty and oscillation) simultaneously
using clustering methods. Their clustering methods could eliminate some of oscillation
traces, since oscillation traces falling into a same cluster are aggregated and replaced by the
centroid of the cluster. The problem is: oscillation traces could be too faraway (e.g., several
kilometers away) to be compressed into a constraint-specific cluster (see our results in
Section 5.2). Additionally, Yin et al. (2017) mentioned the use of the pattern-based method
to detect oscillation traces in their applied sightings data, but details (e.g., what parameters
to use) are not present.
3. The sightings data
The mobile phone sightings data used in this study consist of sightings from about one
million users in the Buffalo metropolitan area during the month of April 2014. Sightings are
derived by an intermediate company from records generated by phone operators for
operation and billing purposes so that phone users maintain connections to the network
when needed (for calling, texting and activities that require accessing the internet) and are
correctly billed. Each sighting reveals the trace of a user, containing an encrypted mobile
identification number, a time stamp and a location estimate expressed in latitude and
Wang and Chen Page 8
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
longitude (Table 2). The data include devices that were observed in the Buffalo area and
served by the cellular network carriers contracted with the data provider in April 2014.
According to the data provider, users in the data could be resident workers, home workers,
inbound commuters, outbound commuters, short-term visitors, or long-term visitors, etc.
The location information of traces is estimated through the data provider’s proprietary
technology and its accuracy is reported to be about 300 m on average in the urban area. For
each sighting, a variable called certainty radius is calculated, which is a measurement in
meters representing the radius of a circular area where the actual location of the device can
be found with a 90% probability. The center of the area gives the location estimate of each
trace in the data in the format of latitude and longitude. Fig. 3 shows the spatial distribution
of certainty radius. Area with smaller certainty radius is shown in lighter shade, indicating
relatively higher spatial accuracy. We note that generally downtown area shows higher
accuracy than the rest of the study area.
Fig. 4a shows the distribution of the number of days observed in the data. Half of the users
are observed for fewer than seven days, and only 6797 users (less than 1% of all users) have
at least one trace everyday in the entire month. As another test on the temporal sparsity of
the data, we divide a day into 24 hourly slots and, for each trajectory of a device, we count,
among 24 slots, the number of slots in which the device is sighted at least once (Widhalm et
al., 2015) (Fig. 4b). The median is six slots, indicating that half of trajectories have no more
than six hourly slots with locations revealed by the mobile phone network.
4. Methodology
4.1. An overview
In this section, we provide a framework to process sightings data. As noted earlier, the
framework consists of two stages, corresponding to two issues of the data. In the first stage,
the issue of locational uncertainty is addressed by aggregating locations into clusters using a
revised incremental clustering algorithm. We show that the revised incremental clustering
algorithm can maintain the spatial precision of the original data and is adaptive to the spatial
density of traces in rural and urban area (see details in Section 5). In the second stage,
oscillation sequences are detected by a time-window-based detection method and a modified
pattern-based method, which are able to identify complex oscillation patterns and to reduce
the risk of mistaking real trips. Here, the order of these two stages matters, which will be
discussed at the end of this section. The framework is summarized in Fig. 5 and described in
subsequent sections.
4.2. Addressing locational uncertainty toward extracting activity locations
Our method to deal with the locational uncertainty in the sightings data contains three steps:
(1) applying the revised incremental clustering algorithm to aggregate traces; (2) finding a
suitable parameter for the incremental clustering algorithm; and (3) applying the k-means
clustering algorithm to deal with an order problem that is not addressed by the incremental
clustering algorithm (see details below).
Wang and Chen Page 9
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
4.2.1. Incremental clustering algorithm—We first identify activity locations. As noted
earlier, in sightings data, location estimates for a single activity location could be unique
because of the locational uncertainty. Thus, duration at a location cannot be directly inferred.
In principle, these location estimates shall be closely distributed in space. Therefore, we
could aggregate these closely distributed locations and represent them with one cluster so
that the duration associated with each potential activity location could be calculated.
The other issue caused by the locational uncertainty is that the common activity location that
a user visits on different days may appear distinct in their expressed format of longitude and
latitude. If we put together all the available location estimates of a potential activity location,
we will find that they scatter closely in a cluster as well. Thus, common activity locations
could be revealed by putting traces of multiple days together and clustering them without
regarding their time ordering.
In this study, we develop a revised incremental clustering algorithm to aggregate traces of
multiple days for each user without considering their time ordering. This method enables us
to identify common activity locations by aggregating traces that are close in space but may
be far away in time (e.g., several days). Details of this algorithm are provided in Fig. 6.
4.2.2. Determining a suitable Rc via trial and error—The incremental clustering
method requires a spatial constraint Rc as an input. Without information to choose a suitable
Rc to cluster our data, we find it via trial and error. More specifically, we try different values
of Rc and find the suitable one by counting the total number of activity locations identified
in each trial and analyzing its relationship with Rc. To do this, we first need to identify
activity locations, which is described bellow.
(1) Duration and activity locations: Given a setting of Rc (e.g., 0.5 km), traces are
aggregated into clusters without considering the temporal information. To determine the
duration, we use individual trajectories. Through scanning every trajectory, we find
consecutive traces that belong to the same cluster and aggregate them into a newly created
trace, which contains several fields (Table 3): the original user ID; the initial time given by
the timestamp of the first trace of the cluster; the duration that is calculated as the difference
of the first and last timestamp in the cluster; and the location information that is given by the
centroid of the cluster. Then, an activity location is identified if its duration exceeds Tc. We
set Tc as five minutes. This five-minute threshold also follows the rule used in many
household travel surveys to define what counts an activity (Transportation Research Board,
2005). Additionally, it is an appropriate threshold for an activity location in the activity
based modeling context (Yin et al., 2017).
(2) Determining a suitable Rc: In Fig. 7, we show the average number of distinct activity
locations visited per day per person
nAL
as a function of Rc. If Rc is small, one actual
activity location may be split into several clusters. With increasing Rc, small clusters start to
merge into meaningful representations of actual activity locations. Fig. 7 suggests the
suitable Rc appears at 1 km. Recall that the identification of activity locations is subject to
both spatial and temporal constraints. Above 1 km, it is possible that some clusters may
absorb far-away outliers (passing-by points), therefore lengthen their duration and meet the
Wang and Chen Page 10
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
temporal constraint (i.e. 5 min), consequently leading to the growth of the number of activity
locations. This is witnessed by the slight increase of
nAL
and by the steep rise of cluster
duration (per cluster) when Rc is above 1 km (Fig. 7).
4.2.3. K-means clustering algorithm—We notice an order problem that could not be
handled by the incremental clustering algorithm alone—clustering results are subject to the
order that how traces are clustered, which can result in unreasonable clusters. Fig. 8a gives
one example, where fourteen traces are to be clustered. Following the incremental clustering
algorithm, a new cluster is created at L0, and consequently we obtain two strange clusters:
one consists of locations {L0, L1, L2, L3, L4} and the other one contains other locations. For
this case, however, a more reasonable output might be: one cluster contains L0 alone and the
other one compasses the remaining.
We apply the k-means algorithm to address this problem. This algorithm enables us to
aggregate traces into k clusters in which each trace belongs to the cluster with the nearest
distance to the centroid. We address the above order problem by initializing the k-means
clustering using the results yielded from the incremental clustering algorithm. More
specifically, the number of resulting clusters and their centroids are required for the
initialization (Kanungo et al., 2002). In other words, the above order problem is resolved by
applying the k-means clustering algorithm to revise the clustering results from the
incremental clustering algorithm. We notice that the clustering error in the example is
corrected (Fig. 8b). Following this revision, the duration and centroid of clusters are
updated.
4.3. Removing oscillation sequences
In the second stage of data processing, we remove oscillation sequences. Our solution
consists of four steps: (1) detecting oscillation sequences using a time-window-based
method; (2) determining the suitable time window (required input parameter of previous
step); (3) applying a modified pattern-based method to detect oscillation sequences with low
switching frequency; and (4) removing oscillation sequences detected.
4.3.1. Time-window-based method—Inspired by the method proposed by Wu et al.
(2014), we develop a time-window-based method to detect oscillation sequences. It scans
trajectories with a short time window Tw (defined in the following section), which always
starts at the ending time of a trace d0. Here, the ending time of a trace is the time given by
the sum of the initial time and duration of the trace (Table 3). A sequence of traces returns
after the time window is applied (note that this sequence may contain a single trace). Among
the sequences returned, we find the suspicious sequence which includes more than one
distinct location estimates. And then oscillation sequences are con-firmed as those
containing at least one circular event. A circular event refers to a tour that one device is
initially found at location L0, later goes somewhere else (e.g., L1, L2), and at last returns to
location L0. Since the time window is defined quite short (e.g., several minutes), it is less
likely for anyone to make a tour within such a short time period. Therefore, we consider that
the oscillation sequences detected contain traces resulted from oscillation phenomenon. Fig.
9 illustrates how this time-window-based method works. The oscillation case shown in Table
Wang and Chen Page 11
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
1 is detected, because a circular event L0-L1- L0 is found in the sequence {d0, d1, d2, d3},
which is returned after the time-window is applied.
4.3.2. Determining a suitable Tw via trial and error—To determine a suitable Tw, we
test a set of time-windows of different lengths. Then we calculate the average oscillation
ratio as a function of Tw (Fig. 10), where the oscillation ratio for a user is defined as the ratio
of the number of detected oscillation traces over the total number of traces associated with
that user. Fig. 10 shows that when Tw is small, the average oscillation ratio increases
sharply; but beyond a certain value, the tangent of the curve remains stable, indicating five
minutes as a reasonable choice to separate oscillation cases from real trips. More
specifically, five-minute is chosen as a longer Tw does not yield much better performance on
detecting oscillation traces. Instead, a longer Tw may be at the risk of mistaking more actual
trips for oscillations. When Tw is five minutes, the estimated average oscillation ratio is
0.17, suggesting that on average approximately 17% of traces are generated due to pure
signaling activities.
4.3.3. Modified pattern-based method—The time-window-based method is effective
in identifying most oscillation sequences. But, we found that due to temporal sparsity, the
time window may be too short to capture the circular event even though the oscillation
occurs. Thus, we further process the result using a modified pattern-based method. We first
detect sequences with a twice switches pattern, i.e. L0–L1–L0–L1. As the pattern-based
method alone may mistake actual trips for oscillation, for each sequence detected, we look
for additional evidence: only if a time interval, among all time intervals between any two
consecutive traces in the sequence, is found shorter than Tw, we identify it as an oscillation
sequence. In Section 5, we show that the simple two-switch criterion is effective, though
more complex ones could be designed.
As the oscillation detection methods rely on identifying circular events, it is necessary to
address the issue of locational uncertainty first. Without addressing the problem of locational
uncertainty, location estimates in sightings data may be unique from time to time, preventing
us from identifying circular events and thus the detection of oscillation.
4.3.4. Removing oscillation sequences detected—For each oscillation sequence
detected, we remove oscillation traces by finding and replacing them with a meaningful
location. An oscillation sequence may reoccur at different times of a day and/or on different
days. If the meaningful location is determined from each oscillation sequence independently
(Bayir et al., 2010; Wu et al., 2014), we probably yield different results for the same
oscillation sequence that appears at different time. For example, we may obtain two different
meaningful locations L0 and L1 for one oscillation sequence {L0–L1–L0…} that occurs
twice in one day (e.g., in the morning and evening). To address this issue, we determine the
meaningful location by examining the traces of the entire month. Specifically, the
meaningful location is identified as the one where the user spends most of time during the
entire month. The time spent at one location is computed by summing up durations that is
associated with the location (Table 3).
Wang and Chen Page 12
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
5. Results
5.1. Locational uncertainty
We first evaluate the effectiveness of our framework in terms of addressing the issue of
locational uncertainty by comparing resulting clusters with those using commonly used
agglomerative clustering method (Hariharan and Toyama, 2004; Jiang et al., 2013).
Specifically, the radius of outputting clusters is computed and compared. Since location
estimates belonging to one cluster are replaced by the cluster’s centroid, the radius of the
cluster could be a measure of the precision of new location estimates: the new location
records generated from large clusters have lower precision. Fig. 11 shows the distribution of
the radius of clusters produced by different methods using the same spatial constraint (Rc =
1 km). We notice that comparing with the uniform distribution of the radius of clusters
obtained by agglomerative clustering, the radius of clusters resulting from our method is
more heterogeneously distributed. In addition, approximately 80% of these clusters have a
radius no longer than 500 m (compared with 50% in the output of agglomerative clustering
method), meaning that our method performs better in terms of preserving location resolution
while conducting the clustering to extract activity locations.
Interestingly, compared with the agglomerative clustering that outputs clusters with a clear
cutoff on their size, our method produces some clusters whose size exceeds Rc (Fig. 11).
This suggests that the clustering method proposed here is also adaptive to address cases
where traces belonging to the same cluster can scatter far from each other, exceeding the
given spatial constraint Rc. Note that this adaptive feature could be useful if traces of one
activity location scatter broadly (i.e. the associated locational uncertainty is large) because of
the low density of cellular towers that are involved in estimating locations.
5.2. Oscillation
We then evaluate our solution to addressing the oscillation phenomenon. Since the
oscillation sequences give us the illusion that a user travels back and forth within a short
time, we may overestimate the trip rates of users whose records include oscillation traces.
Here, one trip is defined as a sequence of traces starting and ending at two consecutive
activity locations. Without removing oscillation sequences, we notice a large number of
trajectories with very high trip rates. The Buffalo travel survey suggests that about 95% of
population conduct no more than eight trips a day (Greater Buffalo-Niagara Transportation
Survey, 2002). We thus assume if a trajectory includes more than eight trips, it is identified
as a potentially abnormal trajectory because of possible presence of oscillation traces. We
found that before removing oscillation sequences, the number of abnormal trajectories is
high – about 12% of the total number of trajectories. After oscillation traces are removed,
this value drops to about 0.7%. This resulting value is much smaller than 5% from the travel
survey data, which could be attributed to the missing short-distance trips that are shorter than
Rc and the related representativeness issues that are not dealt with in this paper but discussed
in Section 6.
Wang and Chen Page 13
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
In Fig. 12, we plot the percentage of abnormal trajectories as a function of Tw. Before and
after the implementing the modified pattern-based method, the substantial drop in the
number of abnormal trajectories testifies the effectiveness of the method.
The time-window-based method is capable of identifying oscillation sequences with
complex patterns that involve more than two locations. For example, it can detect oscillation
sequences with pattern L0–L1–L2–L3–L0–L1–L2–L1–L0, meaning a device is observed
switching among four locations. Fig. 13a gives the distribution of different oscillation
patterns this method detected in the data. Here,
n
locations involved in oscillation means
devices are observed oscillating among
n
locations. We see that though in most of cases
oscillation only involves two locations, mobile phone data contains complex oscillation
patterns and they could be detected with our framework. Specifically, we notice that
approximately 3% of oscillation sequences involve more than four locations. Meanwhile,
Fig. 13b shows the distribution of distance between oscillation traces (detected using our
method), indicating that some oscillation traces are farther apart than 1 km. This suggests
that, as we noted earlier in Section 2.3, a clustering method with a specific spatial constraint
may fail to detect faraway oscillation traces.
5.3. Trajectory
As a part of our evaluation, in Fig. 14 we visualize two trajectories of one user before and
after the data is processed. The trajectories on the left are drawn from the original data and
the right ones are from the processed data. We note that common places visited in multiple
days are revealed and the trajectories become smoother after we remove oscillation
sequences.
Previous studies show that individuals’ trajectories are characterized by statistical regularity:
preferable returns to a few locations and occasional explorations to other faraway locations
(González et al., 2008; Pappalardo et al., 2015; Song et al., 2010a; Susilo and Axhausen,
2014). As the measurement of mobility regularity, the visiting frequency
f
of the k-th most
visited locations is observed following a scaling law:
fk
k
ξ
. Here, as the indicator of
mobility regularity, a larger exponential factor
ξ
suggests a higher level of regularity: less
exploration into new places and more returns to familiar locations. The mobility analysis in
the previous study, which is based on sightings data as well, shows
ξ
= 1.04 (Jiang et al.,
2013). In Fig. 15, we test this scaling property with the datasets before and after removing
oscillation sequences. We notice that though the regularity of mobility pattern could be
revealed from both datasets, the exponential factors are different: the data processed by the
first stage of our framework only (i.e. clustering algorithms) yields
ξ
= 1.02, which is
consistent with Jiang et al.’s study (2013); while after removing the oscillation sequences,
ξ
decreases significantly to 0.88. Therefore, the oscillation phenomenon leads to the non-
negligible overestimation of the regularity of individuals’ mobility.
6. Discussions
As noted earlier, despite the rapid surge of studies using mobile phone data, our
understanding on the nature of the data is very limited—few describes its issues and
documents the procedures applied to such data to derive mobility trajectories. This paper is
Wang and Chen Page 14
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
to fill this gap. Our study proposes a data-processing framework to specifically address two
problems in the mobile phone sightings data: locational uncertainty and oscillation. For each
problem, we extend existing works and propose our algorithms. These algorithms are simple
to implement. We illustrate the effectiveness of the proposed framework by comparing to the
results using the state of the art algorithms. As shown in Fig. 15, without applying those data
processing steps, a fundamental characteristic of human mobility patterns—regularity
(González et al., 2008; Jiang et al., 2013; Song et al., 2010a)—is overestimated. Chen et al.
(2016) reviewed two lines of research for human mobility research—traditional travel
behavior analysis using household travel surveys and human mobility analysis using big data
such as mobile phone data. They noted that though regularity and variety are found in both
lines of works, variety is much more pronounced in those studies using household travel
surveys. Fig. 15 suggests one potential underlying cause of this discrepancy between the two
lines of research may be due to our lack of understanding of the nature of mobile phone data
and insufficient processing on them.
One known issue of mobile phone data that is not addressed in the paper relates to
representativeness. Put it more plainly, since users in a mobile phone data are not the result
of a probabilistic sample (as for household travel surveys) and all users voluntarily select
into being the subscribers to certain service networks, there is an issue of selection bias
(Arai et al., 2015; Wesolowski et al., 2013). We do not address it here as it itself is a
complicated issue and addressing it at requires a complete separate study. We briefly discuss
a number of complications here. First, the underlying population of mobile phone data is
often unknown, as the data provider is often unwilling to release such information. Second,
there is lack of ground truth information, an issue related to the unknown underlying
population. Because of these two reasons, the usual means of validating against external,
ground truth data is infeasible (Chen et al., 2014) and new methods and metrics must be
developed for validation purposes. Third, there is a need to uncouple the strong correlation
between trip rates and communication frequencies: more connections to communication
networks, more sightings resulted and thus more trips derived (Iovan et al., 2013; Yuan et al.,
2012). This correlation also suggests that for those studies relying only on active users, their
results are likely affected by this issue (Alexander et al., 2015b; Wang et al., 2010). Lastly,
because of the temporal sparsity and spatial uncertainty naturally associated with mobile
phone data, there is a need for new thinking and new interpretations on the scale and
meanings of representativeness (Lu et al., 2017; Zhao et al., 2016). In other words, a day of
trajectory from mobile phone data may not be comparable to a trajectory derived from a
household travel survey; rather, to be comparable to the latter, it may take trajectories of
multiple days. Taking it further, since a key advantage of mobile phone data is its
longitudinal nature, new metrics may be developed to capture how representative mobility
patterns evolve over time (Chikaraishi et al., 2009, 2011).
Though the reported data-processing framework is developed for the sightings data, some of
the steps within the framework can also be applied to the CDR data. As noted earlier and
described in more detail in Chen et al. (2016), the location estimates in the CDR data are
those of cell towers, rather than distinct coordinates for the devices in sightings data. This
suggests that the clustering step as described in our paper (Section 4.3.2) is not needed.
However, retaining locations at the cell tower level may miss trips within a single cell area.
Wang and Chen Page 15
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
In this study, after clustering the sightings data, the duration of clusters is then computed.
This duration represents the time a user spends at the location. For CDR data, one would
find the estimation of the activity duration is similar to the one in sightings data. The second
stage of our framework (detecting and removing oscillation traces) can be applied to CDR
data directly, as CDR data also has the oscillation issue. Given that the CDR data is
temporally sparser than the sightings data (Chen et al., 2016), there may be more difficulty
in differentiating oscillations from actual movements, as the detection of oscillations relies
on the temporal distribution of traces. Therefore, a careful selection of time-window Tw is
required.
It is our hope that the current study will spark many more studies discussing the nature of
the data and its issues, documenting and reporting the various data processing procedures
that are used, and addressing the representativeness issue. It is undoubted that such studies
are of great importance—they are the foundations of those resulting trajectories on which
new discoveries are made. This is particularly true if one is looking to use such data to
inform our transportation policies.
Acknowledgments
The study is supported by the funding provided by the U.S. National Institute of Health (NIH)
(1R01GM108731-01A1). The authors are grateful to Dr. Jae Hyun Lee’s help in generating Fig. 3 as well as his
participation in related discussions. Comments from three anonymous reviewers also greatly strengthen our paper.
Views expressed in the paper do not represent those of NIH and the authors are responsible for all errors that may
remain.
References
Ahas R, Aasa A, Silm S, Tiru M. Daily rhythms of suburban commuters’ movements in the Tallinn
metropolitan area: Case study with mobile positioning data. Transp Res Part C Emerg Technol.
2010a; 18:45–54. http://dx.doi.org/10.1016/j.trc.2009.04.011.
Ahas R, Silm S, Järv O, Saluveer E, Tiru M. Using mobile positioning data to model locations
meaningful to users of mobile phones. J Urban Technol. 2010b; 17:3–27. http://dx.doi.org/
10.1080/10630731003597306.
Alexander L, Jiang S, Murga M, González MC. Origin–destination trips by purpose and time of day
inferred from mobile phone data. Transp Res Part C Emerg Technol. 2015a; 58:240–250. http://
dx.doi.org/10.1016/j.trc.2015.02.018.
Alexander L, Jiang S, Murga M, González MC. Origin–destination trips by purpose and time of day
inferred from mobile phone data. Transp Res Part C Emerg Technol, Big Data Transport Traffic
Eng. 2015b; 58 (Part B):240–250. http://dx.doi.org/10.1016/j.trc.2015.02.018.
Andrienko, G., Andrienko, N., Hurter, C., Rinzivillo, S., Wrobel, S. Visual Analytics Science and
Technology (VAST), 2011 IEEE Conference on. IEEE; 2011. From movement tracks through events
to places: extracting and characterizing significant places from mobility data; p. 161-170.
Arai, A., Witayangkurn, A., Horanont, T., Shao, X., Shibasaki, R. Understanding the unobservable
population in call detail records through analysis of mobile phone user calling behavior: a case
study of greater Dhaka in Bangladesh; 2015 IEEE International Conference on Pervasive
Computing and Communications (PerCom). 2015. p. 207-214.http://dx.doi.org/10.1109/PERCOM.
2015.7146530
Bayir MA, Demirbas M, Eagle N. Mobility profiler: a framework for discovering mobility profiles of
cell phone users. Pervasive Mob Comput. 2010; 6:435–454. http://dx.doi.org/10.1016/j.pmcj.
2010.01.003.
Wang and Chen Page 16
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Becker R, Cáceres R, Hanson K, Isaacman S, Loh JM, Martonosi M, Rowland J, Urbanek S,
Varshavsky A, Volinsky C. Human mobility char-acterization from cellular network data. Commun
ACM. 2013; 56:74–82. http://dx.doi.org/10.1145/2398356.2398375.
Calabrese F, Colonna M, Lovisolo P, Parata D, Ratti C. Real-time urban monitoring using cell phones:
a case study in Rome. IEEE Trans Intell Transp Syst. 2011a; 12:141–151. http://dx.doi.org/10.1109/
TITS.2010.2074196.
Calabrese F, Diao M, Di Lorenzo G, Ferreira J, Ratti C. Understanding individual mobility patterns
from urban sensing data: a mobile phone trace example. Transp Res Part C Emerg Technol. 2013;
26:301–313. http://dx.doi.org/10.1016/j.trc.2012.09.009.
Calabrese F, Lorenzo GD, Liu L, Ratti C. Estimating origin-destination flows using mobile phone
location data. IEEE Pervasive Comput. 2011b; 10:36–44. http://dx.doi.org/10.1109/MPRV.
2011.41.
Candia J, González MC, Wang P, Schoenharl T, Madey G, Barabási AL. Uncovering individual and
collective human dynamics from mobile phone records. J Phys Math Theor. 2008; 41:224015.
http://dx.doi.org/10.1088/1751-8113/41/22/224015.
Chen C, Bian L, Ma J. From traces to trajectories: how well can we guess activity locations from
mobile phone traces? Transp Res Part C Emerg Technol. 2014; 46:326–337.
Chen C, Ma J, Susilo Y, Liu Y, Wang M. The promises of big data and small data for travel behavior
(aka human mobility) analysis. Transp Res Part C Emerg Technol. 2016; 68:285–299. http://
dx.doi.org/10.1016/j.trc.2016.04.005. [PubMed: 27182125]
Chikaraishi M, Fujiwara A, Zhang J, Axhausen K. Exploring variation properties of departure time
choice behavior by using multilevel analysis approach. Transp Res Rec J Transp Res Board. 2009;
2134:10–20. http://dx.doi.org/10.3141/2134-02.
Chikaraishi M, Fujiwara A, Zhang J, Axhausen K, Zumkeller D. Changes in variations of travel time
expenditure. Transp Res Rec J Transp Res Board. 2011; 2230:121–131. http://dx.doi.org/
10.3141/2230-14.
Csáji BC, Browet A, Traag VA, Delvenne JC, Huens E, Van Dooren P, Smoreda Z, Blondel VD.
Exploring the mobility of mobile phone users. Phys Stat Mech Appl. 2013; 392:1459–1473. http://
dx.doi.org/10.1016/j.physa.2012.11.040.
Demissie MG, de Almeida Correia GH, Bento C. Intelligent road traffic status detection system
through cellular networks handover information: an exploratory study. Transp Res Part C Emerg
Technol. 2013; 32:76–88. http://dx.doi.org/10.1016/j.trc.2013.03.010.
Dong H, Wu M, Ding X, Chu L, Jia L, Qin Y, Zhou X. Traffic zone division based on big data from
mobile phone base stations. Transp Res Part C Emerg Technol Big Data Transport Traffic Eng.
2015; 58(Part B):278–291. http://dx.doi.org/10.1016/j.trc.2015.06.007.
Ester M, Kriegel HP, Sander J, Xu X. A density-based algorithm for discovering clusters in large
spatial databases with noise. Proceedings of the Second International Conference on Knowledge
Discovery and Data Mining. 1996:226–231.
Fraley C. Algorithms for model-based Gaussian hierarchical clustering. SIAM J Sci Comput. 1998;
20:270–281.
Fraley C, Raftery AE. Model-based clustering, discriminant analysis, and density estimation. J Am
Stat Assoc. 2002; 97:611–631. http://dx.doi.org/10.1198/016214502760047131.
Gao S, Wang Y, Gao Y, Liu Y. Understanding urban traffic-flow characteristics: a rethinking of
betweenness centrality. Environ Plan B Plan Des. 2013; 40:135–153. http://dx.doi.org/10.1068/
b38141.
González MC, Hidalgo CA, Barabási AL. Understanding individual human mobility patterns. Nature.
2008; 453:779–782. http://dx.doi.org/10.1038/nature06958. [PubMed: 18528393]
Greater Buffalo-Niagara Transportation Survey. 2002, 2003. <https://gbnrtc-mpo-
osl2.squarespace.com/s/GBNRTC_Onboard_Transit_Study.pdf>
Hard, E., Byron, C., Praprut, S., Steve, F., Darrell, B., Lisa, G. Synopsis of New Methods and
Technologies to Collect Origin-Destination (O-D) Data (No FHWA-HEP-16-083). 2016. Federal
Highway Administration report FHWA-OR-15-01<https://www.fhwa.dot.gov/planning/tmip/
publications/other_reports/>
Wang and Chen Page 17
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Hariharan, R., Toyama, K. Project Lachesis: Parsing and Modeling Location Histories. In: Egenhofer,
MJ.Freksa, C., Miller, HJ., editors. Geographic Information Science, Lecture Notes in Computer
Science; Presented at the International Conference on Geographic Information Science; Berlin,
Heidelberg: Springer; 2004. p. 106-124.http://dx.doi.org/10.1007/978-3-540-30231-5_8
Horn C, Klampfl S, Cik M, Reiter T. Detecting outliers in cell phone data. Transp Res Rec J Transp
Res Board. 2014; 2405:49–56. http://dx.doi.org/10.3141/2405-07.
Iovan, C., Olteanu-Raimond, AM., Couronné, T., Smoreda, Z. Moving and calling: mobile phone data
quality measurements and spatiotemporal uncertainty in human mobility studies. In:
Vandenbroucke, D.Bucher, B., Crompvoets, J., editors. Geographic Information Science at the
Heart of Europe. Springer; International Publishing, Cham: 2013. p. 247-265.
Iqbal MS, Choudhury CF, Wang P, González MC. Development of origin–destination matrices using
mobile phone call data. Transp Res Part C Emerg Technol. 2014; 40:63–74. http://dx.doi.org/
10.1016/j.trc.2014.01.002.
Isaacman, S., Becker, R., Cáceres, R., Kobourov, S., Martonosi, M., Rowland, J., Varshavsky, A.
International Conference on Pervasive Computing. Springer; 2011. Identifying important places in
people’s lives from cellular network data; p. 133-151.
Järv O, Ahas R, Witlox F. Understanding monthly variability in human activity spaces: a twelve-month
study using mobile phone call detail records. Transp Res Part C Emerg Technol. 2014; 38:122–
135. http://dx.doi.org/10.1016/j.trc.2013.11.003.
Järv O, Tenkanen H, Toivonen T. Enhancing spatial accuracy of mobile phone data using multi-
temporal dasymetric interpolation. Int J Geogr Inf Sci. 2017; 31:1630–1651. http://dx.doi.org/
10.1080/13658816.2017.1287369.
Jiang, S., Ferreira, J., Gonzalez, MC. Activity-based human mobility patterns inferred from mobile
phone data: a case study of Singapore; IEEE Trans Big Data. 2017. p. 1-1.http://dx.doi.org/
10.1109/TBDATA.2016.2631141
Jiang, S., Fiore, GA., Yang, Y., Ferreira, J., Jr, Frazzoli, E., González, MC. Proceedings of the 2nd
ACM SIGKDD International Workshop on Urban Computing, UrbComp ’13. ACM; New York,
NY, USA: 2013. A review of urban computing for mobile phone traces: current methods,
challenges and opportunities; p. 2:1-2:9.http://dx.doi.org/10.1145/2505821.2505828
Kanungo T, Mount DM, Netanyahu NS, Piatko CD, Silverman R, Wu AY. An efficient k-means
clustering algorithm: analysis and implementation. IEEE Trans Pattern Anal Mach Intell. 2002;
24:881–892. http://dx.doi.org/10.1109/TPAMI.2002.1017616.
Lee, JK., Hou, JC. Proceedings of the 7th ACM International Symposium on Mobile Ad Hoc
Networking and Computing, MobiHoc ’06. ACM; New York, NY, USA: 2006. Modeling steady-
state and transient behaviors of user mobility: formulation, analysis, and application; p.
85-96.http://dx.doi.org/10.1145/1132905.1132915
Louail T, Lenormand M, Ros OGC, Picornell M, Herranz R, Frias-Martinez E, Ramasco JJ,
Barthelemy M. From mobile phone data to the spatial structure of cities. Sci Rep. 2014; 4:5276.
http://dx.doi.org/10.1038/srep05276. [PubMed: 24923248]
Lu S, Fang Z, Zhang X, Shaw SL, Yin L, Zhao Z, Yang X. Understanding the representativeness of
mobile phone location data in characterizing human mobility indicators. ISPRS Int J Geo-Inf.
2017; 6:7. http://dx.doi.org/10.3390/ijgi6010007.
Milone, R. Presented at 15th TRB National Transportation Planning Applications Conferrence.
Metropolitan Washington Council of Governments; 2015. Preliminary Evaluation of Cellular
Origin-Destination Data as a Basis for Forecasting Non-Resident Travel. <https://
www.trbappcon.org/2015conf/Program.aspx>
Palchykov V, Mitrović M, Jo HH, Saramäki J, Pan RK. Inferring human mobility using communication
patterns. Sci Rep. 2014; 4:6174. http://dx.doi.org/10.1038/srep06174. [PubMed: 25146347]
Pappalardo L, Simini F, Rinzivillo S, Pedreschi D, Giannotti F, Barabási AL. Returners and explorers
dichotomy in human mobility. Nat Commun. 2015; 6:8166. http://dx.doi.org/10.1038/
ncomms9166. [PubMed: 26349016]
Qi, L., Qiao, Y., Abdesslem, FB., Ma, Z., Yang, J. Proceedings of the First Workshop on Mobile Data,
MobiData ’16. ACM; New York, NY, USA: 2016. Oscillation resolution for massive cell phone
traffic data; p. 25-30.http://dx.doi.org/10.1145/2935755.2935759
Wang and Chen Page 18
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Qu, Y., Gong, H., Wang, P. Transportation mode split with mobile phone data; 2015 IEEE 18th
International Conference on Intelligent Transportation Systems. 2015. p. 285-289.http://dx.doi.org/
10.1109/ITSC.2015.56
Serok N, Blumenfeld-Lieberthal E. A Simulation model for intra-urban movements. Plos One. 2015;
10:e0132576. http://dx.doi.org/10.1371/journal.pone.0132576. [PubMed: 26161640]
Shad SA, Chen E, Bao T. Cell oscillation resolution in mobility profile building. 2012 ArXiv Prepr.
ArXiv12065795.
Song C, Koren T, Wang P, Barabási AL. Modelling the scaling properties of human mobility. Nat Phys.
2010a; 6:818–823. http://dx.doi.org/10.1038/nphys1760.
Song C, Qu Z, Blumm N, Barabási AL. Limits of predictability in human mobility. Science. 2010b;
327:1018–1021. http://dx.doi.org/10.1126/science.1177170. [PubMed: 20167789]
Stabler, B., Sikder, S. Development of the Idaho Statewide Travel Demand Model Trip Matrices Using
Cell Phone OD Data and Origin Destination Matrix Estimation. 2014. TREC Friday Seminar
Series<https://pdxscholar.library.pdx.edu/trec_seminar/30/>
Susilo YO, Axhausen KW. Repetitions in individual daily activity–travel–location patterns: a study
using the Herfindahl-Hirschman Index. Transportation. 2014; 41:995–1011. http://dx.doi.org/
10.1007/s11116-014-9519-4.
Toole JL, Colak S, Sturt B, Alexander LP, Evsukoff A, González MC. The path most traveled: Travel
demand estimation using big data resources. Transp Res Part C Emerg Technol Big Data Transport
Traffic Eng. 2015; 58(Part B):162–177. http://dx.doi.org/10.1016/j.trc.2015.04.022.
Transportation Research Board. Does the Built Environment Influence Physical Activity?: Examining
the Evidence – Special Report 282. Institute of Medicine of the National Academies; 2005.
Vlahogianni EI, Park BB, van Lint JWC. Big data in transportation and traffic engineering. Transp Res
Part C Emerg Technol, Big Data Transport Traffic Eng. 2015; 58 (Part B):161. http://dx.doi.org/
10.1016/j.trc.2015.08.006.
Wang, H., Calabrese, F., Lorenzo, GD., Ratti, C. Transportation mode inference from anonymized and
aggregated mobile phone call detail records; 2010 13th International IEEE Conference on
Intelligent Transportation Systems (ITSC). 2010. p. 318-323.http://dx.doi.org/10.1109/ITSC.
2010.5625188
Wang, M. Understanding Activity Location Choice with Mobile Phone Data (Thesis). 2014. <https://
digital.lib.washington.edu/researchworks/handle/1773/26523>
Wang M, Chen C, Ma J. Time-of-day dependence of location variability: application of passively-
generated mobile phone dataset. Presented at the Transportation Research Board 94th Annual
Meeting. 2015
Wang P, Hunter T, Bayen AM, Schechtner K, González MC. Understanding road usage patterns in
urban areas. Sci Rep. 2012; 2 http://dx.doi.org/10.1038/srep01001.
Wang T, Chen C, Ma J. Mobile phone data as an alternative data source for travel behavior studies.
Presented at the Transportation Research Board 93rd Annual MeetingTransportation Research
Board. 2014
Wesolowski A, Eagle N, Noor AM, Snow RW, Buckee CO. The impact of biases in mobile phone
ownership on estimates of human mobility. J R Soc Interface. 2013; 10:20120986. http://
dx.doi.org/10.1098/rsif.2012.0986. [PubMed: 23389897]
Widhalm P, Yang Y, Ulm M, Athavale S, González MC. Discovering urban activity patterns in cell
phone data. Transportation. 2015; 42:597–623. http://dx.doi.org/10.1007/s11116-015-9598-x.
Wu, W., Wang, Y., Gomes, JB., Anh, DT., Antonatos, S., Xue, M., Yang, P., Yap, GE., Li, X.,
Krishnaswamy, S., Decraene, J., Nash, AS. Oscillation resolution for mobile phone cellular tower
data to enable mobility modelling; 2014 IEEE 15th International Conference on Mobile Data
Management. 2014. p. 321-328.http://dx.doi.org/10.1109/MDM.2014.46
Xu Y, Shaw SL, Zhao Z, Yin L, Fang Z, Li Q. Understanding aggregate human mobility patterns using
passive mobile phone location data: a home-based approach. Transportation. 2015; 42:625–646.
http://dx.doi.org/10.1007/s11116-015-9597-y.
Xu Y, Shaw SL, Zhao Z, Yin L, Lu F, Chen J, Fang Z, Li Q. Another tale of two cities: understanding
human activity space using actively tracked cellphone location data. Ann Am Assoc Geogr. 2016;
106:489–502. http://dx.doi.org/10.1080/00045608.2015.1120147.
Wang and Chen Page 19
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Yin, M., Qiao, SM., Feygin, S., Paiement, J-F., Pozdnoukhov, A. IEEE Transactions in ITS,
MobiData ’16. ACM; New York, NY, USA: 2017. A generative model of urban activities from
cellular data; p. 25-30.http://dx.doi.org/10.1145/2935755.2935759
Yuan Y, Raubal M, Liu Y. Correlating mobile phone usage and travel behavior – a case study of
Harbin, China. Comput Environ Urban Syst, Special Issue: Geoinform. 2012; 36:118–130. http://
dx.doi.org/10.1016/j.compenvurbsys.2011.07.003.
Zhao Z, Shaw SL, Xu Y, Lu F, Chen J, Yin L. Understanding the bias of call detail records in human
mobility research. Int J Geogr Inf Sci. 2016; 30:1738–1762. http://dx.doi.org/
10.1080/13658816.2015.1137298.
Zheng, VW., Zheng, Y., Xie, X., Yang, Q. Proceedings of the 19th International Conference on World
Wide Web, WWW ’10. ACM; New York, NY, USA: 2010. Collaborative location and activity
recommendations with GPS history data; p. 1029-1038.http://dx.doi.org/
10.1145/1772690.1772795
Wang and Chen Page 20
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Fig. 1.
Locational uncertainty. (The inset figure is the zoom-in of closely distributed location
estimates of one activity location).
Wang and Chen Page 21
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Fig. 2.
A problematic classification example of model-based clustering. (Different clusters are
shown in different symbols of different colors. The size of the ellipse overlaid on a cluster
reflects the size of that cluster. Here locations as far as nine kilometers are aggregated into
one cluster, making the clustering less meaningful.)
Wang and Chen Page 22
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Fig. 3.
Spatial distribution of certainty radius.
Wang and Chen Page 23
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Fig. 4.
(a) Distribution of the number of days observed. (b) Distribution of the number of hourly
slots with at least one trace.
Wang and Chen Page 24
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Fig. 5.
Framework of processing sightings data.
Wang and Chen Page 25
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Fig. 6.
Revised incremental clustering algorithm.
Wang and Chen Page 26
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Fig. 7.
Number of distinct activity locations per day and associated duration as a function of Rc.
Wang and Chen Page 27
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Fig. 8.
Illustration of the order problem and the solution. (a) Classification without k-means
algorithm. (b) Resulting classification after implementing k-means algorithm. (Following the
incremental clustering algorithm, we obtain two strange clusters: one in triangle and the
other one in circle. After applying k-means clustering, we have a more reasonable output.)
Wang and Chen Page 28
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Fig. 9.
Illustration of time-window-based detection method. (This figure corresponds to the
oscillation case in Table 1. The implementation of the time-window returns a sequence of
traces {d0, d1, d2, d3}, containing the circular event (L0–L1–L0)).
Wang and Chen Page 29
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Fig. 10.
Average oscillation ratio as a function of Tw.
Wang and Chen Page 30
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Fig. 11.
Distribution of the radius of clusters.
Wang and Chen Page 31
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Fig. 12.
Percentage of abnormal trajectories as a function of Tw.
Wang and Chen Page 32
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Fig. 13.
(a) Distribution of different oscillation patterns detected. (b) The distribution of distances
between oscillation traces detected (Tw = 5 min).
Wang and Chen Page 33
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Fig. 14.
Trajectories of one sample user. (a) before the data processing and (b) after the data
processing. (We show two trajectories in different colors. For a clear observation, only traces
with a duration (i.e.
duration
> 0) are shown here.)
Wang and Chen Page 34
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Fig. 15.
The location visiting frequency (Tw = 5 min).
Wang and Chen Page 35
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Wang and Chen Page 36
Table 1
An oscillation case.
Trace Location Time Distance (km) Switching speed (km/h)
d0L012:21:48 \ \
d1L112:22:01 2.7 748
d2L012:25:20 2.7 49
d3L112:25:39 2.7 512
(The switching speed of two consecutive traces is calculated as the division of the distance between two location estimates by the time interval of
these two traces. A space-time visualization of this oscillation case is given in Fig. 8.)
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Wang and Chen Page 37
Table 2
A sample of mobile phone sightings data.
ID Time*Location estimate**
36cc5
**
77ca 1,396,381,131 42.951554|−78.665607
36cc5
**
77ca 1,396,385,839 42.929892|−78.624486
36cc5
**
77ca 1,396,386,054 42.928625|−78.607994
36cc5
**
77ca 1,396,386,319 42.929892|−78.624436
36cc5
**
77ca 1,396,386,567 42.951554|−78.665607
36cc5
**
77ca 1,396,386,653 42.929892|−78.624436
36cc5
**
77ca 1,396,386,876 42.955625|−78.606594
*
In Unix time – defined as the number of seconds that have elapsed since 00:00:00 Coordinated Universal Time, 1/1/1970.
**
Location estimates in the table are modified to protect privacy. The same is applied as for Table 3.
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Wang and Chen Page 38
Table 3
A sample of data after clustering.
ID Initial time Duration (second) Location estimate
36cc5**77ca 1,396,377,202 10,246 42.949429|−78.670337
36cc5**77ca 1,396,388,619 26 42.972285|−78.776146
36cc5**77ca 1,396,388,890 134 42.998380|−78.791654
36cc5**77ca 1,396,389,805 73 43.039335|−78.860423
36cc5**77ca 1,396,392,466 1310 42.949429|−78.670337
Transp Res Part C Emerg Technol
. Author manuscript; available in PMC 2018 February 01.
... Tus, the data can register the phone user's trajectories in the base station networks. On the basis of the potential mapping relationship between the real trip trajectories and the connection sequences of base stations, inferring the user's trip details using cellular signaling data can be expected [4,5]. Compared with travel survey and GPS data, cellular signaling data have the advantages of 24-hour uninterrupted collection, wide spatial coverage, and high sampling rate. ...
Article
Full-text available
Cellular signaling data have become increasingly indispensable in analyzing residents’ travel characteristic. Especially with the enhancement of positioning quality in 4G-LTE and 5G wireless communication systems, it is expected that the identification accuracy of fine-grained travel modes will achieve an optimal level. However, due to data privacy issues, the empirical evaluation of the performance of different identification methods is not yet sufficient. This paper builds a travel mode identification model that utilizes the gated recurrent unit (GRU) neural network. With 24 features as input, this method can identify four traffic modes, including walking, bicycle, car, and bus. Moreover, in cooperation with the operator, we organized an experiment collecting cellular signaling data, as well as the corresponding GPS data. Using the collected dataset as ground-truth data, the performance of the method presented in this paper and other popular methods is verified and compared. The results indicate that the GRU-based method has a better performance, with a precision, recall, and F score of 90.5%. Taking F score as an example, the outcome of the GRU-based method is about 6% to 7% higher than methods based on other machine learning algorithms. Considering the identification accuracy and model training time comprehensively, the method suggested in this paper outperforms the other three deep learning-based methods, namely, recurrent neural network (RNN), long short-term memory network (LSTM), and bidirectional long short-term memory network (Bi-LSTM). This study may provide some insights for the application and development of cellular signaling-based travel information collection technology for residents in the future.
... Each contains data from a range of partner apps on an informed consent basis, with data limited to users aged 16+. Data is collected when an app records the time and location of a device based on the most accurate location sensor available at the time, including GPS, Bluetooth, cellular tower, Wi-Fi or a combination of sources (Wang & Chen, 2018). Due to a lack of transparency from the commercial providers, the specific applications included in the datasets are unknown to researchers. ...
Article
Full-text available
Emerging forms of mobile phone data generated from the use of mobile phone applications have the potential to advance scientific research across a range of disciplines. However, there are risks regarding uncertainties in the socio-demographic representativeness of these data, which may introduce bias and mislead policy recommendations. This paper addresses the issue directly by developing a novel approach to assessing socio-demographic representativeness, demonstrating this with two large independent mobile phone application datasets, Huq and Tamoco, each with three years data for a large and diverse city-region (Glasgow, Scotland) home to over 1.8 million people. We advance methods for detecting home location by including high-resolution land use data in the process and test representativeness across multiple dimensions. Our findings offer greater confidence in using mobile phone app data for research and planning. Both datasets show good representativeness compared to the known population distribution. Indeed, they achieve better population coverage than the 'gold standard' random sample survey which is the alternative source of data on population mobility in this region. More importantly, our approach provides an improved benchmark for assessing the quality of similar data sources in the future.
... which is available to the Urban Big Data Centre (UBDC, http://www.ubdc.ac.uk/) for non-commercial academic research. The data is an example of Location-Based Service data which is generated when a MP application updates the location of a mobile device using the most accurate available location sensor including single or a combination of GPS, Bluetooth, cellular tower signals, or Wi-Fi (Wang and Chen, 2018). This type of MP application data provides the point locations of the device with a certain degree of error (typically 10s of metres). ...
Conference Paper
Full-text available
This research advances the home location detection of new forms of mobile phone data by using high-resolution land-use data and compares resulting home location estimates with public socio-demographic data. The research allows us to identify potential biases in mobile phone data which may arise through uneven population coverage. Results show that the number of mobile phone users estimated across the Glasgow City Region in 2020 are proportional to the working population across different socio-demographic groups based on the Scottish Index of Multiple Deprivation.
Article
Transportation research has shown that socio‐demographic factors impact people's mobility patterns. During the COVID‐19 pandemic, some of these effects have changed in accordance with changing mobility needs adapting to the pandemic, including restrictions on in‐person gatherings, closure of in‐person businesses, and working from home. We investigate two gaps in current knowledge in this area of transportation research: to what extent the associations between socio‐demographic factors and mobility metrics have changed, and how these associations vary across geographic space. We used aggregate deidentified cell tower location data to measure two mobility metrics—movement time and radius of gyration—and socio‐demographic data from the 2016 Canadian Census to model these associations across Ontario, Canada in 2020 using a linear model and a geographically weighted regression model. We find that certain associations between socio‐demographics and mobility have changed from what we previously observed before the pandemic, and we can see the variation of these associations across space. These findings will improve our understanding of how socio‐demographic factors affect mobility patterns in different communities and demonstrate the importance of measuring these associations at a more fine‐grained level using models that consider spatial variation to best reflect the nature of these associations.
Article
Along with the development of information and positioning technologies, there emerges passively collected location data that contain location observations with time information from various types of mobile devices. Passive location data are known for their large sample size and continuous behavior observations. However, they also require careful and comprehensive data processing and modeling algorithms for privacy protection and practical applications. In the meantime, the travel demand estimation of origin–destination (OD) tables is fundamental in transportation planning and analysis. There is a lack of national OD estimation that provides time-dependent travel behaviors for all travel modes. Passively collected location data appeal to researchers for their potential of serving as the data source for estimation and monitoring of large-scale multimodal travel demand. This research proposes a comprehensive set of methods for passive location data processing including data cleaning, activity location and purpose identification, trip-level information identification, social demographic imputation, sample weighting and expansion, and demand validation. For each task, the paper evaluates the state-of-the-practice and state-of-the-art algorithms and develops an applicable method jointly considering different features of various passive location data sources, imputation accuracy, and computation efficiency. The paper further examines the viability of the method kit in a national-level case study and successfully derives the multimodal national-level OD estimates with additional data products, such as trip rate and vehicle miles traveled, at different geographic levels and temporal resolutions.
Article
Vehicle volume serves as a critical metric and the fundamental basis for traffic signal control, transportation project prioritization, road maintenance planning, and more. Traditional methods of quantifying vehicle volume rely on manual counting, video cameras, and loop detectors at a limited number of locations. These efforts require significant labor and cost for expansions. Researchers and private sector companies have also explored alternative solutions, such as probe vehicle data, although this still suffers from a low penetration rate. In recent years, along with the technological advancement in mobile sensors and mobile networks, the quantity of mobile device location data (MDLD) has been growing dramatically in spatiotemporal coverage of the population and its mobility. This paper presents a big-data driven framework that can ingest terabytes of MDLD and estimate vehicle volume over a larger geographical area with a larger sample size. The proposed framework first employs a series of cloud-based computational algorithms to extract multimodal trajectories and trip rosters. A scalable map matching and routing algorithm is then applied to snap and route vehicle trajectories to the roadway network. The observed vehicle counts on each roadway segment are weighted and calibrated against ground truth control totals, that is, annual vehicle-miles traveled and annual average daily traffic. The proposed framework is implemented on the all-street network in the State of Maryland using MDLD for the entire year of 2019. The results demonstrate that our proposed framework produces reliable vehicle volume and also its transferability and generalization ability.
Article
Full-text available
Novel digital data sources allow us to attain enhanced knowledge about locations and mobilities of people in space and time. Already a fast-growing body of literature demonstrates the applicability and feasibility of mobile phone-based data in social sciences for considering mobile devices as proxies for people. However, the implementation of such data imposes many theoretical and methodological challenges. One major issue is the uneven spatial resolution of mobile phone data due to the spatial configuration of mobile network base stations and its spatial interpolation. To date, different interpolation techniques are applied to transform mobile phone data into other spatial divisions. However, these do not consider the temporality and societal context that shapes the human presence and mobility in space and time. The paper aims, first, to contribute to mobile phone-based research by addressing the need to give more attention to the spatial interpolation of given data, and further by proposing a dasymetric interpolation approach to enhance the spatial accuracy of mobile phone data. Second, it contributes to population modelling research by combining spatial, temporal and volumetric dasymetric mapping and integrating it with mobile phone data. In doing so, the paper presents a generic conceptual framework of a multi-temporal function-based dasymetric (MFD) interpolation method for mobile phone data. Empirical results demonstrate how the proposed interpolation method can improve the spatial accuracy of both night-time and daytime population distributions derived from different mobile phone data sets by taking advantage of ancillary data sources. The proposed interpolation method can be applied for both location- and person-based research, and is a fruitful starting point for improving the spatial interpolation methods for mobile phone data. We share the implementation of our method in GitHub as open access Python code.
Article
Full-text available
The advent of big data has aided understanding of the driving forces of human mobility, which is beneficial for many fields, such as mobility prediction, urban planning, and traffic management. However, the data sources used in many studies, such as mobile phone location and geo-tagged social media data, are sparsely sampled in the temporal scale. An individual's records can be distributed over a few hours a day, or a week, or over just a few hours a month. Thus, the representativeness of sparse mobile phone location data in characterizing human mobility requires analysis before using data to derive human mobility patterns. This paper investigates this important issue through an approach that uses subscriber mobile phone location data collected by a major carrier in Shenzhen, China. A dataset of over 5 million mobile phone subscribers that covers 24 h a day is used as a benchmark to test the representativeness of mobile phone location data on human mobility indicators, such as total travel distance, movement entropy, and radius of gyration. This study divides this dataset by hour, using 2- to 23-h segments to evaluate the representativeness due to the availability of mobile phone location data. The results show that different numbers of hourly segments affect estimations of human mobility indicators and can cause overestimations or underestimations from the individual perspective. On average, the total travel distance and movement entropy tend to be underestimated. The underestimation coefficient results for estimation of total travel distance are approximately linear, declining as the number of time segments increases, and the underestimation coefficient results for estimating movement entropy decline logarithmically as the time segments increase, whereas the radius of gyration tends to be more ambiguous due to the loss of isolated locations. This paper suggests that researchers should carefully interpret results derived from this type of sparse data in the era of big data.
Article
Full-text available
In this study, with Singapore as an example, we demonstrate how we can use mobile phone call detail record (CDR) data, which contains millions of anonymous users, to extract individual mobility networks comparable to the activity-based approach. Such an approach is widely used in the transportation planning practice to develop urban micro simulations of individual daily activities and travel; yet it depends highly on detailed travel survey data to capture individual activity-based behavior. We provide an innovative data mining framework that synthesizes the state-of-the-art techniques in extracting mobility patterns from raw mobile phone CDR data, and design a pipeline that can translate the massive and passive mobile phone records to meaningful spatial human mobility patterns readily interpretable for urban and transportation planning purposes. With growing ubiquitous mobile sensing, and shrinking labor and fiscal resources in the public sector globally, the method presented in this research can be used as a low-cost alternative for transportation and planning agencies to understand the human activity patterns in cities, and provide targeted plans for future sustainable development.
Technical Report
Full-text available
This report provides an overview and detail on the use of cellular, GPS, and Bluetooth technologies for origin-destination (O-D) data. It discusses what each technology represents and its capabilities and limitations in relation to accuracy, sample saturation, and frequency. It includes takeaways and lessons learned from numerous studies in recent years that have used cell, GPS, and/or Bluetooth to collect O-D data. The report provides a comparison between the technologies in relation to the ability to provide O-D data by external trip types, by non-commercial and commercial vehicle categories, and by supplemental attributes such as residency status and routing. It discusses the suitability of each technology for planning versus operational O-D studies and for different geographic scales such as urban, regional, or statewide. The report provides potential users of O-D data sourced from cell, GPS, or Bluetooth general guidance on which technology or combinations of technologies is best suited for different O-D study types, sizes, and objectives.
Article
Full-text available
The last decade has witnessed very active development in two broad, but separate fields, both involving understanding and modeling of how individuals move in time and space (hereafter called “travel behavior analysis” or “human mobility analysis”). One field comprises transportation researchers who have been working in the field for decades and the other involves new comers from a wide range of disciplines, but primarily computer scientists and physicists. Researchers in these two fields work with different datasets, apply different methodologies, and answer different but overlapping questions. It is our view that there is much, hidden synergy between the two fields that needs to be brought out. It is thus the purpose of this paper to introduce datasets, concepts, knowledge and methods used in these two fields, and most importantly raise cross-discipline ideas for conversations and collaborations between the two. It is our hope that this paper will stimulate many future cross-cutting studies that involve researchers from both fields.
Article
Full-text available
One major problem of using location data collected from mobile cellular networks for mobility modelling is the oscillation phenomenon. An oscillation occurs when a mobile phone intermittently switches between cell towers instead of connecting to the nearest cell tower. For the purpose of mobility modeling, the location data needs to be cleansed to approximate the mobile device's actual location. However, this constitutes a challenge because the mobile device's true location is not known. In this paper, we study the oscillation resolution problem. We propose an algorithm framework called DECRE (Detect, Expand, Check, Remove) to detect and remove oscillation logs. To make informed decisions DECRE includes four steps: Detect, to identify log sequences that may contain oscillation using a few heuristics based on the concepts of stable period and moving at impossible speed, Expand, to look before and after suspicious records to gain more information, Check, to check whether a cell tower is observed repeatedly (which is a strong indication of oscillation), and Remove, resolving oscillation by selecting a cell tower to approximate the mobile device's actual location. Our experimental results on travel diaries show that our oscillation resolution approach is able to remove records that are far from mobile device's ground-truth locations, improve the quality of the location data, and performs better than an existing method. Our performance study on large scale cell tower data shows that the MapReduce implementation of our approach is able to process 1 Terabyte of cell tower data in five hours using a small cluster.
Article
Full-text available
Activity space is an important concept in geography. Recent advancements of location-aware technologies have generated many useful spatiotemporal data sets for studying human activity space for large populations. In this article, we use two actively tracked cellphone location data sets that cover a weekday to characterize people's use of space in Shanghai and Shenzhen, China. We introduce three mobility indicators (daily activity range, number of activity anchor points, and frequency of movements) to represent the major determinants of individual activity space. By applying association rules in data mining, we analyze how these indicators of an individual's activity space can be combined with each other to gain insights of mobility patterns in these two cities. We further examine spatiotemporal variations of aggregate mobility patterns in these two cities. Our results reveal some distinctive characteristics of human activity space in these two cities: (1) A high percentage of people in Shenzhen have a relatively short daily activity range, whereas people in Shanghai exhibit a variety of daily activity ranges; (2) people with more than one activity anchor point tend to travel further but less frequently in Shanghai than in Shenzhen; (3) Shenzhen shows a significant north–south contrast of activity space that reflects its urban structure; and (4) travel distance in both cities is shorter around noon than in regular work hours, and a large percentage of movements around noon are associated with individual home locations. This study indicates the benefits of analyzing actively tracked cellphone location data for gaining insights of human activity space in different cities.
Conference Paper
Mode split is an important step in the estimation of travel demand. Beside traditional costly surveys, many mode split methods, which employ the new emerging large-scale social signal data, are recently proposed. In this paper, we develop a mode split model based on the widely available mobile phone data and transportation networks' geographical data. The shares of three transportation modes (car, public transportation, walking) in each census tract are estimated for Boston central and suburb area. Finally, the proposed model is validated with real mode share data obtained from the U.S. census.
Conference Paper
Cellular towers capture logs of mobile subscribers whenever their devices connect to the network. When the logs show data traffic at a cell tower generated by a device, it reveals that this device is close to the tower. The logs can then be used to trace the locations of mobile subscribers for different applications, such as studying customer behaviour, improving location-based services, or helping urban planning. However, the logs often suffer from an oscillation phenomenon. Oscillations may happen when a device, even when not moving, does not only connect to the nearest cell tower, but is instead unpredictably switching between multiple cell towers because of random noise, load balancing, or simply dynamic changes in signal strength. Detecting and removing oscillations are a challenge when analyzing location data collected from the cellular network. In this paper, we propose an algorithm called SOL (Stable, Oscillation, Leap periods) aimed at discovering and reducing oscillations in the collected logs. We apply our algorithm on real datasets which contain about 18.9~TB of traffic logs generated by more than 3~million mobile subscribers covering about 21000 cell towers and collected during 27~days from both GSM and UMTS networks in northern China. Experimental results demonstrate the ability and effectiveness of SOL to reduce oscillations in cellular network logs.
Article
In recent years, call detail records (CDRs) have been widely used in human mobility research. Although CDRs are originally collected for billing purposes, the vast amount of digital footprints generated by calling and texting activities provide useful insights into population movement. However, can we fully trust CDRs given the uneven distribution of people’s phone communication activities in space and time? In this article, we investigate this issue using a mobile phone location dataset collected from over one million subscribers in Shanghai, China. It includes CDRs (~27%) plus other cellphone-related logs (e.g., tower pings, cellular handovers) generated in a workday. We extract all CDRs into a separate dataset in order to compare human mobility patterns derived from CDRs vs. from the complete dataset. From an individual perspective, the effectiveness of CDRs in estimating three frequently used mobility indicators is evaluated. We find that CDRs tend to underestimate the total travel distance and the movement entropy, while they can provide a good estimate to the radius of gyration. In addition, we observe that the level of deviation is related to the ratio of CDRs in an individual’s trajectory. From a collective perspective, we compare the outcomes of these two datasets in terms of the distance decay effect and urban community detection. The major differences are closely related to the habit of mobile phone usage in space and time. We believe that the event-triggered nature of CDRs does introduce a certain degree of bias in human mobility research and we suggest that researchers use caution to interpret results derived from CDR data.