Spatial Economic Analysis

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/rsea20

A pre-processing and network analysis of GPS

tracking data

Antonino Abbruzzo , Mauro Ferrante & Stefano De Cantis

To cite this article: Antonino Abbruzzo , Mauro Ferrante & Stefano De Cantis (2020): A
pre-processing and network analysis of GPS tracking data, Spatial Economic Analysis, DOI:
10.1080/17421772.2020.1769170

To link to this article: https://doi.org/10.1080/17421772.2020.1769170

Published online: 03 Jul 2020.

Submit your article to this journal

Article views: 4

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=rsea20
SPATIAL ECONOMIC ANALYSIS
https://doi.org/10.1080/17421772.2020.1769170

A pre-processing and network analysis of GPS tracking

data
a b c
Antonino Abbruzzo , Mauro Ferrante and Stefano De Cantis

ABSTRACT
Global Positioning System (GPS) devices afford the opportunity to collect accurate data on unit movements
from temporal and spatial perspectives. With a special focus on GPS technology in travel surveys, this paper
proposes: (1) two algorithms for the pre-processing of GPS data in order to deal with outlier identification and
missing data imputation; (2) a clustering approach to recover the main points of interest from GPS
trajectories; and (3) a weighted-directed network, which incorporates the most relevant characteristics of
the GPS trajectories at an aggregate level. A simulation study shows the goodness-of-fit of the imputation
data algorithm and the robustness of the clustering algorithm. The proposed algorithms are then applied
to three cases studies relating to the mobility of cruise passengers in urban contexts.
KEYWORDS
cluster-based method, network analysis, spatio-temporal data, global positioning systems

HISTORY Received 12 July 2019; in revised form 28 April 2020

INTRODUCTION

Global Positioning System (GPS) devices can record unit travel times and the spatial coordinates
of locations to a high degree of temporal precision. Nowadays, GPS devices are compact, equipped
with significant autonomy and, of the utmost importance, they can memorize the geographical
coordinates of a statistical unit at a given moment in time. The widespread availability of GPS
data permits the analysis of social behaviour across time and geographical scales (Butz & Torrey,
2006). Given the wide interest in understanding human mobility, applications of GPS tracking
technologies in social science research may be found in various disciplines, including urban econ-
omics (McCabe et al., 2013; Theo, 2011), urban geography (Ahas et al., 2010; Raanan & Shoval,
2014), population geography (Silm & Ahas, 2010) and environmental health (Elgethun et al.,
2003; Zenk et al., 2011), to mention but a few. The availability of high-quality data, in terms
of temporal (seconds) and spatial (metres) resolution, demands statistical methods that facilitate
the extraction of knowledge from a copious amount of information relating to the various unit
trajectories.

CONTACT
a [email protected]
Department of Economics, Business and Statistics, University of Palermo, Palermo, Italy
b (Corresponding author) [email protected]
Department of Culture and Society, University of Palermo, Palermo, Italy.
c [email protected]
Department of Economics, Business and Statistics, University of Palermo, Palermo, Italy

© 2020 Regional Studies Association

2 Antonino Abbruzzo et al.

Despite the accuracy of GPS tracking data, they require a pre-processing phase to correct out-
liers and missing data points. Therefore, after summarizing the main steps required in the analysis
of GPS tracking data (and with a special focus on the context of travel surveys), the contribution of
this paper to the literature is threefold: first, two algorithms for the pre-processing of GPS data
will be proposed in order to deal with outlier identification and missing data imputation; second,
a clustering approach based on an existing algorithm (namely the DBSCAN algorithm) (Ester
et al., 1996; Palma et al., 2008) will be adapted herein to identify points of interests from GPS
trajectories at the individual level; and third, individual-level points of interest will be collated
into a weighted directed network to summarize the most relevant characteristics at an aggregate
level of the GPS trajectories. The goodness-of-fit and sensitivity of the proposed algorithms
will be evaluated through simulation experiments.
From an empirical perspective, the approach proposed in this paper will be applied to three
cases studies relating to cruise passengers’ trajectories in the cities of Palermo (Sicily, Italy),
Dubrovnik (Croatia) and Copenhagen (Denmark). The results obtained may assist in identifying
the most visited places, in addition to identifying the network of relationships among the various
points of interest, with pertinent implications from a destination management perspective.
From a more general perspective, this paper fits into the ongoing debate as to how GPS track-
ing data can be processed and analysed, and which type of information and methodology can be
extracted from this type of data. In the analysis of this type of data, different approaches regarding
various problems and research aims will be presented and methodological proposals presented,
according to specific analytical and research aims. These regard the pre-processing of GPS track-
ing data and the identification and analysis of points of interests in trajectories, in terms of their
relationships, interpreted as a network.
The paper is structured as follows. The next section provides a background of the use of GPS
technology in social research. The third section describes the statistical methods used to pre-pro-
cess, locate the points of interest and construct the network from GPS data. The fourth section
evaluates the goodness-of-fit and sensitivity of the proposed algorithms through a simulation
study. Finally, the fifth section expounds three empirical applications using the proposed methods.

GPS TRACKING DATA ANALYSIS IN TRAVEL SURVEY: A REVIEW

The widespread availability of GPS tracking devices have facilitated the collection of spatio-tem-
poral information on human movements, providing new challenges and opportunities in social
and economic research. GPS technology was initially in travel surveys used in the late 1990s
and this use has developed rapidly over the past decade. Shoval and Ahas (2016) present a sum-
mary and evaluation of the development and progress in this emerging field of scholarship within
the field of tourism research. Before the introduction of the GPS technology, conventional travel
surveys were generally conducted by paper-and-pencil interviews (PAPI), computer-assisted tele-
phone interviews (CATI) or computer-assisted self-interviews (CASI). However, these tra-
ditional survey methods may have been burdensome to respondents since it was necessary to
recall detailed information (including trip timing, travel modes and trip purposes) for each trip
during the survey period (Wolf et al., 2003). GPS was used to record the location of the partici-
pants and their arrival and departure times so that the obtaining of the boarding information, the
routes of the trips and transit trip times would be more precise. Owing to the fact that GPS devices
are very accurate at recording time and positional characteristics of a given journey, GPS surveys
can improve the accuracy and extent of travel survey data and correct any trip-misreporting issues
caused by respondents (Murakami & Wagner, 1999; Wolf et al., 2001). As a result, these tech-
nologies have become a critical data collection tool in human mobility studies (Shoval, 2008).
It is widely acknowledged that GPS surveys can report accurate data, which is unbiased by
users’ perceptions (Blanchard et al., 2010; Wolf et al., 2003). Given that GPS devices have the

SPATIAL ECONOMIC ANALYSIS

 A pre-processing and network analysis of GPS tracking data 3

potential to provide very large quantities of data, it is necessary to develop specific algorithms to
deal with the main issues that may arise in the analysis of these data. However, the literature on the
analysis of GPS tracking data are rather fragmented among various disciplinary fields. The latter
ranges from computational geometry and data-mining to specific applications, such as animal
tracking, road network, vehicle movements monitoring and hurricane formation, to mention
but a few (Bermingham & Lee, 2017). Moreover, research typically deals with a particular pro-
blem and related solutions arising from the analysis of this type of data. On the other hand, com-
parative approaches regarding existing and proposed algorithms (for the same purpose) and a
general framework for the analysis of space–time data (derived from GPS tracking devices) are
lacking (for a survey of problems and methods in spatio-temporal data, see Atluri et al., 2018).
In the context of travel surveys, steps in the processing of GPS tracking data and the main
topics of interest in their analysis can be summarized as follows:

. Pre-processing of the information:

. Outlier detection.
. Missing data imputation.
. Smoothing of the series.
. Extraction and synthesis of relevant information, such as:
. Activity identification and change detection.
. Cluster identification:
– at the individual level;
– at the aggregate level.
. Network relationships:
– at the individual level;
– at the aggregate level.
. Analysis of trajectories.

Moreover, spatio-temporal information – when available – may be related to other types of

information derived from questionnaire-based surveys or other source of information, such as
map layers containing destination-related attributes, socioeconomic or environmental data (Van
der Spek et al., 2009). This will facilitate the analysis of the relationships between space–time
behaviour and other individual and contextual-level information.
Before conducting any meaningful data analysis, a set of pre-processing operations, aimed at
reducing the magnitude of data problems, must be implemented. Stopher et al. (2008) describe
various reformatting steps required to derive relevant information from the basic set of data,
which have been collected from the GPS tracking device. These include (Stopher et al., 2008),
for example, a calculation of local time and date, latitude–longitude correction, distance compu-
tation from the previous observation, time interval from the previous observation, and cumulative
time intervals from the beginning of recording.
A crucial stage in pre-processing GPS data involves the identification of anomalies in data
(also referred to as an outlier), and missing data imputation due to signal loss and outlier obser-
vations. Outlier observations may occur for several reasons, such as a cold or a warm start.
These usually occur at the beginning of tracking operations (i.e., a cold start) or when the GPS
device switches from sleep mode to working mode after a person stops for one or two hours (i.e.,
a warm start) (Shen & Stopher, 2014). Signal loss is generally determined by closed areas,
urban canyons, etc. The latter are formed by roads cutting through blocks of tall buildings,
which leads to a loss in signal (Stopher et al., 2008). They can have an impact on GPS signal
reception and cause GPS data loss. Signal problems can result in missing or partial trips, and
those generating spurious trips (a sequence of points generated by a stationary GPS device, incor-
rectly identified as a trip).

SPATIAL ECONOMIC ANALYSIS

4 Antonino Abbruzzo et al.

In the context of GPS tracking data, the joint presence of spatial and temporal components
needs to be considered in the identification of outlier observations (Atluri et al., 2018), and several
approaches have been proposed in the literature (e.g., Erenoglu & Hekimoglu, 2010; Třasák &
Štroner, 2014). Atluri et al. (2018) propose a review of the main contributions dealing with
anomaly detection in the context of spatio-temporal data-mining, according to the different
type of data considered (point, trajectory, raster).
Once outlier observations have been identified and eliminated, it is necessary to implement
imputation techniques relating to missing observations. Although several studies have discussed
issues relating to signal loss and outlier observations (Gong et al., 2012; Tsui, 2005), only a
few authors (Chen et al., 2010) have suggested how to deal with errors in GPS tracking data.
The latter is still, therefore, a considerable challenge in GPS studies (Shen & Stopher, 2014).
The final step in pre-processing of GPS data may involve reducing the noise level by means of
smoothing techniques (Jun et al., 2006). Smoothing splines (Castro et al., 2006), Kernel smooth-
ers (Schuler et al., 2014) and Kalman filters (Jun et al., 2006) are often used in the pre-processing
stage of GPS data, and other approaches have also been proposed (Chazal et al., 2011). Jun et al.
(2006) compare various smoothing techniques for vehicle GPS data and evaluate the performance
of the considered algorithms in terms of their capability at minimizing the impact of random
errors on the estimation of speed, acceleration and distance.
After pre-processing the information, the extraction and synthesis of relevant information
from the large quantity of data provided by GPS is required. Invaluable information regarding
individual movements is related to activity detection, which is in turn mainly related to the identi-
fication of the different travel modes by which movements can occur (Siła-Nowicka et al., 2016).
The most commonly variables to consider regarding travel mode detection are average speed, aver-
age absolute acceleration and travel distance. This step is generally based on the segmentation of
trajectories, which requires the identification of segments characterized by different behaviour
types in terms of mobility variables. Several approaches have been proposed for travel mode identi-
fication, such as, for example, machine-learning algorithms (Torrens et al., 2011), neural networks
(Hu et al., 2004), Bayesian networks (Xiao et al., 2015). However, the availability of contextual
information regarding the location of movements may also be effectively used for travel mode
identification (Sester et al., 2012; Siła-Nowicka et al., 2016).
Having described trajectory features, the identification of stop locations is a relevant step in the
extraction and synthesis of relevant information from tracking data, since they may indicate rel-
evant locations or points of interest for the subject. Gong et al. (2015) review the main research
results on stop location identification and propose a classification of methods according to five
groups, namely centroid-based methods, speed-based methods, duration-based methods, den-
sity-based methods and hybrid methods. For a review of this topic, see Gong et al. (2018). None-
theless, in evaluating any method for identifying stop locations, the characteristics of the data set
used, as well as the research aims, study context and related assumptions, must be taken into
consideration.
By aggregating individual-level stops, attractive locations visited by many users in the region
being analysed may be identified. In the context of tourism, this step may reveal popular places,
such as tourist attractions, restaurant locations or shopping centres. This approach has been
undertaken to classify and rank places (Tiwari & Kaushik, 2013), evaluate the shared character-
istics of users (Andrade & Gama, 2019), and even for the determination of optimal meeting points
for a set of users (Khetarpaul et al., 2012).
Having obtained the main points of interest in a given location, a further step in the extraction
of relevant information may consider each point of interest as the vertices (or nodes) of a graph,
with the aim of analysing the network relationships among these nodes (Gilad, 2001). At the indi-
vidual level, individuals’ mobility can be described as a graph, where each node indicates various
points of interests for each unit; the edges of the graph may be characterized by features, such as

SPATIAL ECONOMIC ANALYSIS

 A pre-processing and network analysis of GPS tracking data 5

the duration and frequencies of transitions (Lin & Hsu, 2014; Zheng et al., 2009). Similarly, at
the aggregate level, the nodes on the graph are represented by relevant locations for a given number
of units, and the edges of the graph may be characterized in terms of the number of units passing
from one node to another. Applications of this approach can be found in an analysis of the travel
route decision process (Knapen et al., 2020; Knapen et al., 2016). For example, Lin and Hsu
(2014) contain summaries of studies regarding graph-based trajectory-mining using GPS tracking
data.
Finally, a wide range of applications of GPS tracking data dealt with the analysis of GPS tra-
jectories. These trajectories are represented as a set of stops and moves (Spaccapietra et al., 2008).
A stop is characterized by consecutive GPS data points within a predefined time and distance
threshold. Several studies have focused on models for efficiently analysing moving objects (Brakat-
soulas et al., 2004; Du Mouza & Rigaux, 2005; Güting et al., 2006; Wolfson et al., 1998). The
main focus of enquiry in Wolfson et al. (1998) relies on the geometric properties of trajectories,
while Brakatsoulas et al. (2004) and Du Mouza and Rigaux (2005) consider semantics and back-
ground geographical information. Moving patterns have been extracted from data by Du Mouza
and Rigaux (2005), who define the patterns a priori. For instance, they suggest locating all the
trajectories that move from zone A to zone B, crossing zone Z. They state that moving patterns
are those trajectories that follow a given pattern.
Alvares et al. (2007) extract moving patterns from data that are unknown a priori but whose
minimal number of trajectories are frequent. An algorithm named SMoT (Stops and Moves of
Trajectories) in Alvares et al. has been provided to extract stops and moves from trajectory sample
points. This work also demonstrates how simple trajectory data analysis develops when this
semantic model is used. Moving patterns have been extracted from stops and moves by Campello
et al. (2013), and the resulting patterns modelled in the geographical conceptual schema in order
to visualize trajectory patterns in geographical space. The user specifies the places of interest in all
the previous approaches, given that stops and moves are defined from an application point of view.
The main drawback of this assumption is that important places, which may lead to the discovery of
interesting patterns, can be missed if they are unknown to the user.

STATISTICAL METHODS

Trajectories have been generally considered as a path, which is followed by an object moving in
space and time (Güting et al., 2006; Wolfson et al., 1998). Each point on this path represents
one position in space at a given instant of time. Typically, trajectory data are obtained from
GPS devices that capture the position of an object at specific time intervals. An analysis of
GPS trajectories usually involves several steps (Van der Spek et al., 2009). This section describes
the pre-processing of data, which relates to the set of procedures implemented for outlier detection
and missing data imputation from raw GPS data. In a second step, similar to Palma et al. (2008), a
statistical approach based on a density-based cluster algorithm, the DBSCAN (Ester et al., 1996),
is proposed in order to recover the main points of interest of a trajectory. Finally, a directed net-
work approach is defined to summarize the density-based clusters for every unit, and to identify
the most significant characteristics of their movements. At the aggregate level and in order to sum-
marize the behaviour of all the units, a weighted directed network is derived by means of a second
implementation of the DBSCAN algorithm with reference to the set of all the individual points of
interest.
Before describing the statistical methods used, let us formally define what is a trajectory and
the quantities that can be derived from the trajectory.
Definition 3.1:. Let Dj(i) = (x(i) (i) (i)
j , yj , tj ), where x is the longitude, y is the latitude, and t rep-
resents the time, be a sample point from a GPS tracker for unit i at position j. The set

SPATIAL ECONOMIC ANALYSIS

6 Antonino Abbruzzo et al.

D(i) = {Dj(i) , j = 1, . . . , ni }, where t1(i) , . . . , tn(i)i , is the temporal–spatial movement or trajec-

tory of the i-th unit. Let dj(i) denotes the Euclidean distance between (x(i) (i) (i) (i)
j , yj ), (x j+1 , y j+1 ),
Dtj = t j+1 − tj the time between two consecutive observed spatial points, Tj = t j+1 − t1(i)
(i) (i) (i) (i) (i)

the cumulative time, and v(i) (i) (i)

j = dj /Dtj the velocity to pass from the spatial point j to the
point j + 1.
Figure 1 shows an example of quantities, which can be derived from a trajectory D(i) . Specifi-
cally, the distance time series is shown on the left, and the velocity time series can be seen on the
right.

Pre-processing of GPS data

According to the framework presented in the second section, the first step in GPS data analysis
regards pre-processing, which involves outlier detection and the imputation of missing obser-
vations from the trajectories (Stopher et al., 2008). Figure 2 shows missing data points (on the
left) and an outlier for a statistical unit in a hypothetical trajectory. Algorithm 1 is proposed to
detect the outliers, together with Algorithm 2 to impute the missing spatial–temporal points.

Outlier detection
An outlier is a point whose coordinate exceeds a certain distance with respect to the true coordi-
nate. As described in the second section, several approaches can be adopted to detect outlier obser-
vations (Shen & Stopher, 2014). In this paper we will propose the application of changepoint

Figure 1. Example of quantities derived from a trajectory. The distance time series is on the left; and the
velocity time series is on the right.

Figure 2. Example of missing data and outlier from a raw hypothetical trajectory.

SPATIAL ECONOMIC ANALYSIS

 A pre-processing and network analysis of GPS tracking data 7

analysis – generally used in time series to discover multiple segments (Haynes et al., 2017a, 2017b)
– to the time series of the velocities. The lower and upper limits for each segment were constructed
in order to identify outlier observations. A changepoint will be detected if there is a change in the
mean or in the velocity variance at some point Tj(i) . Algorithm 1 describes the steps for identifying
and dealing with outliers.
Algorithm 1: Outlier detection
(1) Use changepoint analysis for a given time series to discover multiple segments s = {s1 , . . . , sk }
in the velocity time series.
(2) Compute the quantities for each segment sl of the series lower limit(sl ) = Q1 (sl ) − l∗IQR(sl )
and upper limit(sl ) = Q3 (sl ) + l∗IQR(sl ), where Q1 (sl ) is the first quartile, Q3 (sl ) is the third
quartile and IQR(sl ) = Q3 (sl ) − Q1 (sl ) is the interquartile range, l . 0 is a tuning parameter.
(3) Remove the coordinate Dj(i) (sl ) if v(i) (i)
j (sl ) . upper limit(sl ) or vj (sl ) , lower limit(sl ).

A tuning parameter l . 0 was used in step 2 of Algorithm 1 in order to detect outliers.

According to Tukey (1977), l = 1.5 indicates an ‘outlier’ and l = 3 indicates that the data are
‘far out’. Figure 3 shows an example of the application of Algorithm 1. The circled points indicate
detected ‘far out’ points, and the horizontal lines indicate the mean of each segment.
In the segment in which two ‘far out’ points have been discovered, it seems that the unit was
walking at a constant speed and, at some point, the velocity increased sharply twice at around 15
km/h. A similar argument can be used in the last segment where another ‘far out’ point was
detected.

Missing data imputation

The second step in pre-processing of GPS data regards missing data imputation. Missing data
derives from outlier removal and signal loss. The latter becomes particularly important in the
identification of the stops in a trajectory. A loss of signal can be easily detected since GPS devices
are generally set to record spatial locations at every c seconds. Algorithm 2 reconstructs a trajectory
D(i) where D(i) (i) (i)
j = t j+1 − tj is equal to a constant c. The rationale of this algorithm is to impute
missing points by assigning several points proportional to the time lost at the last point, which had
been observed before losing the signal. Some of these points will then be used to reconstruct a
linear path from the point of signal loss to the next available signal point. The new reconstructed
path will be a regularized trajectory in the sense that D(i) j will be constant for all j. This idea is
illustrated in Figure 4. The raw data can be observed on the left and the pre-processed data on
the right. A total of 25 signal points were lost in this example: 11 were used to reconstruct the

Figure 3. Example of outlier detection.

SPATIAL ECONOMIC ANALYSIS

8 Antonino Abbruzzo et al.

4219950
4219950

Point j Point j

4219900
4219900

4219850
4219850

4219800
4219800

4219750
4219750

Point j+1 Point j+1

4219700
4219700

355900 356000 356100 356200 356300 356400 356500 355900 356000 356100 356200 356300 356400 356500

Figure 4. Example of the trajectory of a pre-processed unit. The raw data are on the left; and the pre-
processed data are on the right. A total of 25 signal points were lost in this example: 11 were used to
reconstruct the linear path between j and j + 1; and 14 were imputed randomly close to the spatial
point j. Points j and j + 1 are indicated by triangles; and random and linear imputed points are indicated
by pluses.

linear path between j and j + 1; and 14 were imputed randomly close to the spatial point j where
the signal had been lost.

Algorithm 2: Missing data imputation for trajectory regularization.

Algorithm 2a: Find the number of points to be imputed

1. Given a unit i, denote with D∗(i) the regularized trajectory;

2. FOR j in 1:ni Dtj(i)
(imp points)
. Compute n
j = ⌊+; ⌉, where ⌊+; x⌉ is the nearest integer function of real
c
number x, that indicates the number of lost points between j and j + 1;
(imp points) (i)
o IF nj = 1 then D∗j = Dj(i)
(i) (i) .. (i) (i) (i)
o ELSE D∗j = [D1j .D2j ] where D1j and D2j are recovered in Algorithm 2b
END FOR
3. Repeat the steps 1 and 2 for i = 1, . . . N .

Algorithm 2b: Create set of linear and random points between two points of a spatial–tem-
poral trajectory

(imp points)
1. Given nj . 1 missing points, create two sets of points of length n(linear)
j and n(random)
j as
follows:
. Estimate a regression linear model where y = (yj , y j+1 ) and x = (xj , x j+1 )

j+4
vi=j
j+4
. Compute v j = i=j−4 8 where vi indicates the velocity, then d j = vj ∗ i=j−4 Dti=j is
the averaged distance of the unit in a neighbourhood of j.
= ⌊dj /d j ⌉ and n(random)
(linear) (imp points)
. Set n
j j = nj − n(linear)
j
(imp points)
.
Create a set of times of length nj from tj to t j+1 and denote with t1 the first n(random)
j
time points, and with t2 the remaining
2. Create an equidistant sequence ximp from xj to x j+1 of length n(linear)
j and obtain the predicted
(i)
values ŷimp from the linear regression. Then D2j = (ximp , ŷimp , t2 )

SPATIAL ECONOMIC ANALYSIS

 A pre-processing and network analysis of GPS tracking data 9

3. Create two sets uimpx and uimpy of length n(random)

j from U  Unif (−a, a) then
(i)
D1j = (x + uimpx , y + uimpy , t1 )

Density-based cluster algorithm for GPS data

Let D(i) be the pre-processed trajectory for unit i. A cluster can be defined as a minimum set of
temporal–spatial points, which are sufficiently close (regarding distance) to form a cluster. For
example, in a tourism application, the cluster may indicate some points of interest, such as: a
monument, a museum, a hotel or restaurant. Note that these places may have any geometrical
shape.
The DBSCAN is a density-based algorithm (Ester et al., 1996), which is designed to discover
arbitrary-shaped clusters, where the clusters are sets of spatial points which fall within a certain
distance. Concurrently, the algorithm can identify the noise points, which are spatial points not
belonging to any cluster. Each cluster in this study was interpreted as a point of interest, that
is, a virtual place where the unit spent a certain amount of time.
Let p = (x, y) be a point in the trajectory of a generic unit. The e-neighbourhood of a point p is
defined by nee (p) = {q [ D:|| p, q|| ≤ e [ R+ }, where || p, q|| is a distance function. If the car-
dinality of an e-neighbourhood of a point p, that is, |nee (p)|, is at least greater than a minimum
number (minpts [ R+ ) then p is a core point.
A point p is directly density-reachable from the object q with respect to e and minpts if
p [ nee (q) and |nee (q)| ≥ minpts.
A point p is density-reachable from the object q with respect to e and minpts if there is a chain
p1 , . . . , pl , p1 = q, pl = p such that pi+1 is directly density-reachable from pi .
An object p is density-connected to object q, with respect to e and minpts, if there is an object o
such that both p and q are density-reachable from o with respect to e and minpts.
Definition 3.2:. A cluster C is a non-empty subset of D satisfying the following requirements:

. ∀ p, q: if p [ C and q is density-reachable from p with respect to e and minpts, then q [ C;

. ∀ p, q [ C: p is density-connected to q with respect to e and minpts.
Let C1 , . . . , Ck be the clusters of D with respect to e and minpts, then p [ D is a noise point if
it does not belong to any cluster Ci . The algorithm starts with the first point p in the database D,
and it retrieves all the neighbours of a point p with respect to e and minpts. If p is a core point, this
procedure will yield a cluster concerning e and minpts. If p is not a core point, no points will be
density-reachable from p, and the DBSCAN algorithm will proceed to consider the next point of
the database. The DBSCAN has been deployed in the DBSCAN R package (Hahsler et al.,
2018).

Example: from a trajectory to clusters

Set minpts = 30 (5 min) and e = 40 (metres). The clusters for the trajectory D of a hypothetical
unit are illustrated in Figure 5 and Table 1.
The implementation of this algorithm may provide important information. In this example we
observe that the number of cluster for the considered unit is five, with a tour length of 219.66 min
(3.65 h), and an average time per cluster of 13.2 min.

Choice of the tuning parameter e and minpts

The DBSCAN algorithm requires two parameters, e and minpts, to be selected. The easier-to-set
parameter of DBSCAN is the minpts parameter, whose choice is also related to the minimum
amount of type required to characterize a cluster as meaningful. On the other hand, and as
suggested by Schubert et al. (2017), the parameter e is often harder to set, and the authors

SPATIAL ECONOMIC ANALYSIS

10 Antonino Abbruzzo et al.

Figure 5. Example of clusters relating to the trajectory of a unit: Geographical location, and direction of
the trajectory represented by the scale ofgrey from white (starting point) to black (end-point).

Table 1. Example of clusters relating to the trajectory

of the unit in Figure 5: time spentat each cluster.
Cluster Time (min) % Time

Noise points 153.66 69.98%

1 5.00 2.28%
2 25.33 11.54%
3 13.66 6.22%
4 16.33 7.44%
5 5.60 2.55%
Total 219.58 100.00%

have made various recommendations in this regard. Several internal validity cluster measures can
be used to select the tuning parameter e (Desgraupes, 2013). Such measures quantify the quality of
a clustering relying only on properties intrinsic to the data. Examples of internal measures include
the silhouette, Davies-Boldin and Calinski–Harabasz measures. However, none of these indices
has been proved to be appropriate for selecting the tuning parameter in the context of GPS
data analysis (Van Craenendonck & Blockeel, 2015). Consequently, the problem of selecting
the tuning parameter e is still an open-ended issue, worthy of a separate paper.
In an ideal scenario, there exists domain knowledge for selecting this parameter, which is based
on the application domain. In the scenario described in this paper, that is, in clustering GPS
locations, a domain expert may decide that objects within q metres (e) should be neighbours if
there are at least a minimum number of points p (minpts). In the application outlined in this
paper, the minpts was chosen such that the time spent at cluster j was at least x min. For example,
if points were collected at 10-s intervals, then minpts = 30 guaranteed that the time spent in each

SPATIAL ECONOMIC ANALYSIS

 A pre-processing and network analysis of GPS tracking data 11

cluster was at least x = 5 min. Given the parameter minpts, we suggest to let the parameter e
depend on the distance function {dj }, j = 1, . . . , n, which denotes the Euclidean distance
between the coordinates (xj , yj ). The simulation study described in the fourth section indicates
that a choice of e between 10 and 40 provides robust results.

Networks of DBSCAN objects

The application of the DBSCAN algorithm to a unit trajectory D(i) produces a set of clusters
C1(i) , . . . , Ck(i)i and a set of noise points. These noise points are coordinates that can be removed
from the analysis. The centroid of each cluster can be seen as its core location for a given unit.
In order to synthesize each trajectory, it is proposed to construct a network that includes the con-
cepts of clusters and moves. Specifically, a directed network can be defined as follows.
Definition 3.3:. Let C  (i) be the centroids of the clusters of a trajectory D(i) . A directed
 (i) , . . . , C
1 ki
network for a generic unit i is the set G = (V , E), where V is a set of nodes with cardinality
|V | = ki ; and E is a set of links (moves between clusters for the unit), where ekl = 1 if the unit
goes from the cluster k to the cluster l , and 0 otherwise.
Note that each cluster originally contained a set of spatial points but in Definition 3.3 a cen-
troid was associated with each cluster. The cluster centroids can be used as coordinates to represent
the network nodes graphically. Moreover, the cardinality of each cluster provides information
regarding the time spent by the unit at each cluster; it also permits the recovery of the direction
of the link between two clusters. The latter information is obtained by virtue of the temporal
ordering of the spatial movements.
Figure 6 shows an example of application of the DBSCAN to a hypothetical trajectory, dis-
played on the left, with the recovered directed network displayed on the right. In this example,
the implementation of DBSCAN algorithm led to an estimate of k = 5 clusters. Consequently,
the directed network has five vertices V = {1, . . . , 5} and the directed links show the path of
the unit. Moreover, it can be seen that unit one starts from node 1, selecting the following points
of interest 1− . 2− . 3− . 4− . 5− . 4− . 1. Note that the positions of each node in the
network correspond to the centroid of each cluster. According to this approach, the information

Figure 6. Raw GPS data and clusters obtained from the DBSCAN algorithm (left). Nodes and links
(right) represent points of interest (cluster) and moves between clusters, respectively, for the considered
unit. The nodes of the network are placed in the centroid of each cluster.

SPATIAL ECONOMIC ANALYSIS

12 Antonino Abbruzzo et al.

relating to the noise points (i.e., the streets which the units followed during their moves) has been
lost, not being of interest in this step of the analysis.

A weighted directed network of clustered trajectories

Thus far, an approach for deriving a set of clusters and directed networks from the temporal–
spatial GPS data D(i) for each unit i, i = 1, . . . , N has been described. This analysis has pro-
duced the set of clusters centroids C (i) = {C
 (i) , . . . , C
 (i) }. Let C
 = {C
 (1) , . . . , C
 (N ) } be the set
1 ki
of centroid clusters obtained by applying the DBSCAN. By reapplying the DBSCAN algor-
 a new set of clusters is obtained. The latter can be interpreted as the points of
ithm to the set C,
interest for the units sampled, and a weighted directed network can be reconstructed as
follows.
Definition 3.4:. A weighted directed network is a triplet G = (V , E, W ), where the set V rep-
resents the points of interest of the collected sample units, which have been obtained by applying
the DBSCAN algorithm to C.  The set of edges E represents the transitions from one point of
interest to another. An edge elk = 1 if a unit goes from cluster l to cluster k and 0 otherwise.
The weighted matrix W represents the number of units transferring from one node to another.
Specifically, wlk indicates the number of units that have been transferred from node l to node
k; and the Wll diagonal represents the number of units that spent a certain amount of time at
cluster l .
It is proposed to indicate the procedures described in the third section as the ClusNet
algorithm.

SIMULATION STUDY

A simulation study was performed in order to: (1) validate the goodness-of-fit of the pre-proces-
sing Algorithm 2; and (2) asses the sensitivity of the DBSCAN algorithm with respect to the
choice of the tuning parameter e. The simulations were conducted making use of the R statistical
software (R Development Core Team, 2008).

Goodness-of-fit of the pre-processing Algorithm 2

In order to evaluate the performance of Algorithm 2, we used 50 collected trajectories for which all
data points were observed. To evaluate the goodness-of-fit of Algorithm 2 in missing data impu-
tation, a set of points from the collected trajectories was removed. Considering that these points
may be lost both in a given segment of the trajectory, and between a segment of the trajectory and a
cluster in empirical applications, a two-factor scheme was developed. The first factor was related to
the number of clusters in the trajectory, which provides information regarding the positions in
which several points is removed. The second was given by the number of points to be removed
in each of these positions.
In order to determine the positions from which a constant number of points is removed, nrm ,
the number of clusters C = {C1 , . . . , CK } was estimated for each trajectory by applying the
DBSCAN algorithm (with minpts = 30 and e = 30). This step permitted the definition of
D = {D1 , C1 , D2 , C2 , . . . , DK , CK , Dk+1 }. The second factor nrm = 10, 20 indicates the number
of data points to be removed in each of the K positions of every trajectory. Specifically, if
|Dj | . nrm /2 + 1, where | · | denotes the cardinality of the set, the last nrm /2 data points are
removed from Dj and the first nrm /2 points from Cj . Note that |Cj | ≥ 30 since the authors set
minpts = 30 in the DBSCAN algorithm. Otherwise, only the first nrm points are removed only
from the cluster Cj .

SPATIAL ECONOMIC ANALYSIS

 A pre-processing and network analysis of GPS tracking data 13

The set observations having been removed according to the above described criteria, Algor-
ithm 2 described in the third section was implemented to impute the missing data. The imputed
(i)
trajectory is denoted by Dimp . To evaluate the goodness-of-fit of the missing value imputation,
pre-processing algorithm, for any considered i-trajectory, a measure of distance between the
true and the imputed trajectory, was considered via the dynamic time warping (DTW) algorithm
(Rabiner & Juang, 1999), as follows:


t=1
df (D, Dimp ) = d (fD (t), fDimp (t))mf (k)/Mf , (1)
T

where the warping curve f(t), t = 1, . . . , T , and fD (t) [ {1, . . . , N } and fDimp [ {1, . . . , M}
are wrapper functions remapping the time indices of D and Dimp , mf (t) is a per step weighting
coefficient; and Mf is the corresponding normalization constant, which ensures that the accumu-
lated distortions are comparable along different paths. At the core of equation (1) lies a wrapping
curve technique. See Giorgino (2009) for further details on the application of DTW for compar-
ing GPS trajectories.
The results of the simulation study regarding the missing data imputation algorithm are
reported in Figure 7 for a value of nrm = 10 on the left and of nrm = 20 on the right. The
value of the total number of points removed from each trajectory, which depends on the number
of clusters K identified in each trajectory, is reported on the x-axis; the value of the average dis-
tance (m) among the true and the imputed trajectory is reported on the y-axis, as determined from
equation (1).
With reference to the first scenario, where nrm = 10 and k [ [1, 22], the minimum is
df = 0.13 m, the median is 0.46 m and the maximum is 1.83 m. In the second scenario with
nrm = 10 and k [ [1, 22], the minimum is df = 0.32 m, the median is 1.41 m and the maximum
is 3.36 m. Although the distance between the true and the imputed trajectory generally increases as
the number of points removed increases, other factors (as expected) may affect the accuracy of the
imputation procedure; this is mainly related to the path followed by the considered unit. None-
theless, the very low average distances between the true and imputed trajectories for all 50 units
under consideration suggest that the algorithm proposed is a high performer in the estimation
of missing observations.

Figure 7. On the left side the first scenario, nrm = 10; and on the right side, nrm = 20. The straight line
shows the linear increasing of the distance df when the number of removed points increases.

SPATIAL ECONOMIC ANALYSIS

14 Antonino Abbruzzo et al.

Sensitivity of the DBSCAN algorithm with respect to the choice of the tuning
parameter
In order to evaluate the sensitivity of the DBSCAN algorithm with respect to the tuning par-
ameters, a set of trajectories with a certain (known) number of clusters was simulated for cluster
identification in GPS tracking data. A trajectory can be generated with two theoretical models:
correlated random walks in which the turning angle of each step is the direction of the previous
step + some error; and directed walks, or compass-based navigation, in which the angular errors
at each step are added to the ‘ideal’ or compass direction. In the simulation study we assumed cor-
related random walks with landscape resistance raster from the R SiMRiv package. The former
takes into account the characteristics of the city in terms of a street configuration. Furthermore,
the clusters were simulated from a mixture of Gaussian distributions and added to the correlated
random walk path. The pseudo-algorithm for generating trajectories with clusters is as follows:

. Generate a correlated random walks with landscape resistance raster from the package R
SiMRiv. Denote the generated trajectory with D.
. Take K positions from the trajectory D. These positions indicate the position of the K clusters.
. Generate a vector of m = (m1 , . . . , mK ) and an array of S = S[1], . . . , S[K ].
. Use the function gen.mix to generate K clusters from a mixture of bivariate Gaussian distri-
butions with mean m and variance–covariance S.
. Reorder the cluster points such that the first point of the cluster will be close to the last point of
the trajectory before the cluster and the last point of the cluster will be closer to the first point of
the trajectory after the cluster. The R function solve_TSP with method ‘nearest_insertion’ was
used to obtain the reordering.
. Add these K ordered clusters in the K position of the trajectory D.

Algorithm 3: Generate trajectories with clusters.

The simulations were conducted making use of R, by considering a 2 × 2 factorial design result-
ing from the combination of two factors: number of clusters K = 5, 20, and autocorrelation of
the random walk r = 0.70, 0.90. The numbers of points for each cluster were kept fixed at
n clust = 40. Each scenario was repeated 100 times. We used the DBSCAN algorithm
described in the third section in which the path of e [ [10, 70] and the minimum number of
points for each cluster was fixed at minpts = 30. The sensitivity of the tuning parameter e to
the set of values was evaluated by two indices: the distance between the centroid of the true clus-
ters and the centroids of the estimated clusters (averaged across the 100 replicates); and the
difference in the absolute value between the number of true clusters and those estimated (aver-
aged across the 100 replicates).
The best e is that which minimizes the distance between the true and estimated centroids, and
the difference between the true and estimated number of clusters. This indicates that both lines
should be as low as possible. The two plots in the upper part of Figure 8 indicate that the
DBSCAN is quite robust with a choice of e between [10, 20]. The two plots in the lower part
of Figure 8 indicate that the DBSCAN is quite robust with a choice of e between [10, 30].

EMPIRICAL APPLICATION

The data sets collected comprise 303, 51 and 73 GPS tracks relating to cruiser passengers disem-
barking from Palermo (2014), Dubrovnik (2015) and Copenhagen (2018), respectively. The
details of data collection procedures and other survey characteristics may be found in (Ferrante
et al., 2018). After the pre-processing step, the ClusNet algorithm is applied to the three samples
of cruise passengers who disembarked in Palermo, Copenhagen and Dubrovnik. Since this

SPATIAL ECONOMIC ANALYSIS

 A pre-processing and network analysis of GPS tracking data 15

Figure 8. Results from simulated trajectories with correlated random walks 0.7 and 0.9, a landscape
resistance raster, and 5 and 10 clusters.

application relates to tourism data sets, it can be assumed that the clusters are point of interests
visited by tourists, and the moves form a virtual trajectory used to travel from one attraction to
another. The application of the ClusNet algorithm, as described in this paper, can reveal the
paths of individual tourists through the directed network (defined in Definition 3.3) or a global
tourist path, through the weighted directed network, as defined in Definition 3.4. As discussed
in the third section, the tuning parameters were heuristically fixed by minpts and e. Specifically,
the minpts was fixed at 30 and e = 20 in order to form individual clusters. minpts = 10 and
e = 15 were fixed in the second step of the algorithm. Note that in this second step, the
minpts represents the minimum number of subjects required to form a cluster.

Palermo
This section will outline the application of the proposed ClusNet algorithm to GPS tracks relating
to a sample of cruise passengers disembarking in Palermo. Based on the ClusNet algorithm, 27

SPATIAL ECONOMIC ANALYSIS

16 Antonino Abbruzzo et al.

clusters were estimated. Figure 9 shows the centroids (circles) for 23 over 27 detected clusters on a
map. Note that a zoom feature has been used over the city centre of Palermo in order to improve
the clarity of the points of interest, even though four more places resulted from the algorithm:
Piazza di Monreale, Monreale Cathedral, Santuario di Santa Rosalia and the Catacombs. The
bigger the circle, the greater the number of tourists who visited the point of interest.
As indicated in Table 2, the top-10 tourist points of interest (by tourist visits) in Palermo
are: the Cathedral, Teatro Massimo, Piazza Pretoria, Piazza Bellini, Piazza San Domenico,
Santuario di Santa Rosalia, Teatro Politeama, the Catacombs, Cappella Palatina and San Mat-
teo. But the area of disembarkation also appears, that is, where tourists start and end their visit
to the city.
The ClusNet algorithm recovers the weighted directed network of the clusters (attraction
points), which have been visited by the units of the sample (Figure 10). Note that the complete
adjacency matrix (on the left of Figure 10) and a reduced network (where the directed links with
fewer than 10 moves from one cluster to another have been removed; on the right of Figure 10)
are both shown. The circle size of each node is proportional to the number of units detected at
each point of interest. The link size between clusters i and j is proportional to the number of
units that have been transferred from i to j. The links should be read in pairs and not as a
sequence of movements. It can, for example, be stated that there have been numerous moves
from the Cathedral to Teatro Massimo, but it is not possible to specify a path between the points
of interests.

Figure 9. Circled points indicate the clusters recovered from the DBSCAN algorithm for the city of
Palermo with a zoom shot of the town centre. The larger the circle, the higher the number of tourists
who visited the point.

SPATIAL ECONOMIC ANALYSIS

 A pre-processing and network analysis of GPS tracking data 17

Table 2. Tourist points of interest in Palermo recovered by the

DBSCAN algorithm.
Points of attraction names Frequency %Frequency

Cathedral 196 66.44%

Disembarkation 2 132 44.75%
Harbour 1 107 36.27%
Via Amari 1 95 32.20%
Teatro Massimo 87 29.49%
Piazza Pretoria 80 27.12%
Bus Stop 75 25.42%
Piazza Bellini 45 15.25%
Piazza San Domenico 43 14.58%
Santuario di Santa Rosalia 37 12.54%
Teatro Politeama 35 11.86%
Catacombs 34 11.53%
Disembarkation point 1 25 8.47%
Cappella Palatina 25 8.47%
Shopping 1 24 8.14%
Church San Matteo 23 7.80%
Piazza Marina 23 7.80%
Via Principe di Belmonte 23 7.80%
Via Amari 2 17 5.76%
Piazza Casa Professa 15 5.08%
Shopping 2 13 4.41%
Piazza Castelnuovo 12 4.07%
Cattedrale di Monreale 12 4.07%
Piazza Bologni 11 3.73%
Piazza di Monreale 10 3.39%
Via Amari 3 10 3.39%
Disembarkation point 3 10 3.39%
Note: Frequencies represent the number of subjects in each cluster. The names of
the points of attraction were obtained by adding the centroid coordinates of each
cluster in Google maps.

Dubrovnik and Copenhagen

This section will outline the application of the proposed ClusNet algorithm to GPS tracks, which
are related to two samples of cruise passengers disembarking in Dubrovnik and Copenhagen.
Based on the ClusNet algorithm, six and 10 clusters were estimated for Dubrovnik and Copenha-
gen, respectively.
The right part of Figure 11 shows the weighted directed networks of the main points of inter-
est, which were visited by the units of the sample. The size of each cluster in this case is also pro-
portional to the number of units visiting that particular point of interest. Table 3 shows the
frequency of the cluster units of the top-six attractions in Copenhagen (left-hand side) and
Dubrovnik (right-hand side).

SPATIAL ECONOMIC ANALYSIS

18 Antonino Abbruzzo et al.

Figure 10. Weighted directed network recovered from the ClusNet algorithm. The graph derived by the
complete adjacency matrix is shown on the left; and the reduced network (where the directed links with
fewer than five moves have been removed) is shown on the right.

Figure 11. Centroid cluster and weighted directed networks recovered by using the ClusNet algorithm.
On the upper part are the clusters and network of the city of Dubrovnik; and on the lower part are the
clusters and network of the city of Copenhagen.

SPATIAL ECONOMIC ANALYSIS

 A pre-processing and network analysis of GPS tracking data 19

Table 3. Top-six attractions identified through the ClusNet algorithm in Copenhagen and Dubrovnik.
Copenhagen Dubrovnik
Attraction Frequency % Frequency Attraction Frequency % Frequency

Little Mermaid 50 70.42% Tour agencies 47 25.76%

Canal tours 33 46.48% Onofrio’s Fountain 46 92.16%
Disembarking point 21 29.58% Tower museum 16 31.37%
Marmorkirken 21 29.58% Restaurant area 14 27.45%
Real Gefion’s Fountain 20 28.17% Old Town 10 19.61%
Restaurant area 1 17 23.94% Maritime Museum 10 19.61%

CONCLUSIONS

Any attempt to understand human mobility is a challenging topic notwithstanding its economic
and social relevance; the latter includes the provision of transportation services, the spatial distri-
bution of facilities and, more generally, regional and urban planning and destination management
(Siła-Nowicka et al., 2016). The widespread availability of GPS tracking data, which can be
derived from various location-aware sensors (such as smartphones, GPS tracking devices and
web-based information), offer great potential in social science research. Nonetheless, the variety
and complexity of this type of data poses new challenges in terms of research questions and related
methods, which may arise in various fields of application. It has been demonstrated in this paper
that the analysis and synthesis of GPS tracking data involve several steps that require the
implementation of appropriate analytical techniques and statistical methods. The complexity
and performance of the various techniques to be implemented depends on the quality and the
structure of data available, the context under consideration and the specific research aim.
It has been shown in the literature that GPS tracking data are generally affected by the chal-
lenges of missing data and outlier observations. Such irregularities could introduce bias into any
analysis of GPS tracking data, and this needs to be rigorously addressed. Moreover, there is a
need to provide appropriate statistical methods capable of extracting relevant information from
copious quantities of information relating to individual trajectories.
Departing from these considerations, we contend that the contribution of this paper to the lit-
erature is threefold. First, it proposed two algorithms regarding outlier identification and missing
data imputation, with which to tackle the pre-processing of GPS tracking data. The proposed
algorithms make use of various known techniques, which have been applied in other fields of
study. An example is changepoint detection for the analysis of time series, here adapted for the
identification of outlier observations in the context of GPS tracking data. On the other hand,
the missing data imputation algorithm proposes an innovative procedure for imputing missing
observations when these are lost in a segment of the trajectory and in proximity of a point of inter-
est. The simulation results described in this paper have provided satisfactory results regarding the
proposed algorithm, thereby enhancing its utility in solving common problems in the pre-proces-
sing stage of this type of data.
Second, an existing spatial clustering algorithm, namely the DBSCAN, was implemented for
the identification of clusters in GPS tracking data. The aim here is to identify relevant locations at
the individual and aggregate levels. Thereafter, the analysis was integrated in an original way by
identifying the relationships between the identified cluster in terms of the network at the individ-
ual and aggregate levels. The sensitivity of the algorithm to the tuning parameters was also ana-
lysed by means of a simulation study. The latter provided insights regarding the robustness of the
results according to the value of the tuning parameter. The simulation study also indicated ways of

SPATIAL ECONOMIC ANALYSIS

20 Antonino Abbruzzo et al.

implementing this procedure, which, however, will depend on the research aim and the empirical
context being studied. We believe the proposed methodology to be easily replicable in different
contexts.
Third and from an empirical perspective, the policy implications of the knowledge to be
obtained from an analysis of GPS tracking data in the tourism context are of relevance to destina-
tion management. The determination of the number of visits and information regarding the
length of stay at each attraction or other areas of the destination is fundamental for service man-
agement. This information may be used to improve the administration of existing attractions,
enhance the efficacy of their marketing, or plan to introduce new attractions to the tourist market
(Lew & McKercher, 2006). Of great promise, knowledge about network relationships between
the various places of interest (visitor flow) may orient transportation planning and promotion
strategies, to mention but one of the potential applications of analysing GPS tracking data.
Future research could include the modelling of network relationships among points of interest
regarding individual and contextual information, to include the characteristics of the various
locations. Furthermore, the effectiveness of policy actions may be measured in terms of their capa-
bility of changing the structure of relationships of a given network. The increasing availability of
spatio-temporal information describing human movements in general and tourist behaviour
specifically make imperative the identification of new research questions. The latter will demand
relevant methodological solutions, a challenging topic for future research indeed.

ACKNOWLEDGEMENTS

The authors gratefully acknowledge the invaluable comments of the editor and reviewers. They are
also grateful to Professor Szilvia Gyimóthy at the Copenhagen Business School, Department of
Marketing; the Department Internationalisation and Tourism at the Municipality of Copenha-
gen; and the Copenhagen Cruise Network at VisitCopenhagen, Copenhagen, for data collection
operations in Copenhagen.

DISCLOSURE STATEMENT

No potential conflict of interest was reported by the authors.

ORCID

Antonino Abbruzzo http://orcid.org/0000-0003-2196-3570

Mauro Ferrante http://orcid.org/0000-0003-1287-5851
Stefano De Cantis http://orcid.org/0000-0002-5068-6421

REFERENCES

Ahas, R., Silm, S., Järv, O., Saluveer, E., & Tiru, M. (2010). Using mobile positioning data to model locations
meaningful to users of mobile phones. Journal of Urban Technology, 17(1), 3–27. https://doi.org/10.1080/
10630731003597306
Alvares, L. O., Bogorny, V., Kuijpers, B., de Macedo, J. A. F., Moelans, B., & Vaisman, A. (2007, November 7–9).
A model for enriching trajectories with semantic geographical information. Proceedings of the 15th International
Symposium on Advances in Geographic information Systems ACM GIS 2007 A, Seattle, Washington,
USA (pp. 1–8).
Andrade, T., & Gama, J. (2019). Identifying Points of Interest and Similar Individuals from Raw GPS Data. In:
Cagáňová D., Horňáková N. (Eds) Mobility Internet of Things 2018. Mobility IoT 2018. EAI/Springer

SPATIAL ECONOMIC ANALYSIS

 A pre-processing and network analysis of GPS tracking data 21

Innovations in Communication and Computing. Springer, Cham. https://doi.org/10.1007/978-3-030-30911-

4_21.
Atluri, G., Karpatne, A., & Kumar, V. (2018). Spatio-temporal data mining: A survey of problems and methods.
ACM Computing Surveys (CSUR), 51(4), 1–41. https://doi.org/10.1145/3161602
Bermingham, L., & Lee, I. (2017). A framework of spatio-temporal trajectory simplification methods. International
Journal of Geographical Information Science, 31(6), 1128–1153. https://doi.org/10.1080/13658816.2017.
1290250
Blanchard, R. A., Myers, A. M., & Porter, M. M. (2010). Correspondence between self-reported and objective
measures of driving exposure and patterns in older drivers. Accident Analysis & Prevention, 42(2), 523–529.
https://doi.org/10.1016/j.aap.2009.09.018
Brakatsoulas, S., Pfoser, D., & Tryfona, N. (2004, July 9). Modeling, storing and mining moving object databases.
Proceedings: International database Engineering and applications Symposium, 2004. IDEAS’04, Coimbra,
Portugal.
Butz, W. P., & Torrey, B. B. (2006). Some frontiers in social science. Science, 312(5782), 1898–1900. https://doi.
org/10.1126/science.1130121
Campello, R. J., Moulavi, D., & Sander, J. (2013). Density-based clustering based on hierarchical density estimates.
Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in
Bioinformatics), 7819 LNAI(PART 2), 160–172. https://doi.org/10.1007/978-3-642-37456-2_14
Castro, M., Iglesias, L., Rodríguez-Solano, R., & Sánchez, J. A. (2006). Geometric modelling of highways using
global positioning system (GPS) data and spline approximation. Transportation Research Part C: Emerging
Technologies, 14(4), 233–243. https://doi.org/10.1016/j.trc.2006.06.004
Chazal, F., Chen, D., Guibas, L., Jiang, X., & Sommer, C. (2011, November 1–4). Data-driven trajectory smooth-
ing. Proceedings of the 19th ACM SIGSPATIAL International Conference on Advances in Geographic
Information Systems, Chicago, IL, USA.
Chen, C., Gong, H., Lawson, C., & Bialostozky, E. (2010). Evaluating the feasibility of a passive travel survey
collection in a complex urban environment: Lessons learned from the New York city case study.
Transportation Research Part A: Policy and Practice, 44(10), 830–840. https://doi.org/10.1016/j.tra.2010.08.004
Desgraupes, B. (2013). Clustering indices. University of Paris Ouest-Lab Modal?X, 1, 34. Retrieved June 2020, from
https://cran.biodisk.org/web/packages/clusterCrit/vignettes/clusterCrit.pdf
Du Mouza, C., & Rigaux, P. (2005). Mobility patterns. GeoInformatica, 9(4), 297–319. https://doi.org/10.1007/
s10707-005-4574-9
Elgethun, K., Fenske, R. A., Yost, M. G., & Palcisko, G. J. (2003). Time–location analysis for exposure assessment
studies of children using a novel global positioning system instrument. Environmental Health Perspectives, 111
(1), 115–122. https://doi.org/10.1289/ehp.5350
Erenoglu, R., & Hekimoglu, S. (2010). Efficiency of robust methods and tests for outliers for geodetic adjustment
models. Acta Geodaetica et Geophysica Hungarica, 45(4), 426–439. https://doi.org/10.1556/AGeod.45.2010.4.3
Ester, M., Kriegel, H.-P., Jorg, S., & Xu, X. (1996, August 2–4). A density-based clustering algorithms for dis-
covering clusters. In E. Simoudis, J. Han, and U. Fayyad (Eds.), KDD-96: Proceedings of the second international
conference on knowledge discovery and data mining, 96(34), Portland, Oregon, (pp. 226–231). Retrieved June
2020, from https://www.aaai.org/Papers/KDD/1996/KDD96-037.pdf
Ferrante, M., De Cantis, S., & Shoval, N. (2018). A general framework for collecting and analysing the tracking
data of cruise passengers at the destination. Current Issues in Tourism, 21(12), 1426–1451. https://doi.org/10.
1080/13683500.2016.1194813
Gilad, E.-T. (2001). Graph theory applications to gps networks. GPS Solutions, 5(1), 31–38. https://doi.org/10.
1007/PL00012874
Giorgino, T. (2009). Computing and visualizing dynamic time warping alignments in R: The dtw package. Journal
of Statistical Software, 31(7), 1–24. https://doi.org/10.18637/jss.v031.i07
Gong, H., Chen, C., Bialostozky, E., & Lawson, C. T. (2012). A GPS/GIS method for travel mode detection in
New York city. Computers, Environment and Urban Systems, 36(2), 131–139. https://doi.org/10.1016/j.
compenvurbsys.2011.05.003

SPATIAL ECONOMIC ANALYSIS

22 Antonino Abbruzzo et al.

Gong, L., Sato, H., Yamamoto, T., Miwa, T., & Morikawa, T. (2015). Identification of activity stop locations in
GPS trajectories by density-based clustering method combined with support vector machines. Journal of
Modern Transportation, 23(3), 202–213. https://doi.org/10.1007/s40534-015-0079-x
Gong, L., Yamamoto, T., & Morikawa, T. (2018). Identification of activity stop locations in GPS trajectories by
dbscan-te method combined with support vector machines. Transportation Research Procedia, 32, 146–154.
https://doi.org/10.1016/j.trpro.2018.10.028
Güting, R. H., de Almeida, T., & Ding, Z. (2006). Modeling and querying moving objects in networks. The VLDB
Journal—The International Journal on Very Large Data Bases, 15(2), 165–190. https://doi.org/10.1007/s00778-
005-0152-x
Hahsler, M., Piekenbrock, M., Arya, S., & Mount, D. (2018). Package ‘dbscan’: Density Based Clustering of
Applications with Noise (DBSCAN) and Related Algorithms. R package version. https://github.com/
mhahsler/dbscan.
Haynes, K., Eckley, I. A., & Fearnhead, P. (2017a). Computationally Efficient changepoint detection for a range of
Penalties. Journal of Computational and Graphical Statistics, 26(1), 134–143. https://doi.org/10.1080/
10618600.2015.1116445
Haynes, K., Fearnhead, P., & Eckley, I. A. (2017b). A computationally efficient nonparametric approach for chan-
gepoint detection. Statistics and Computing, 27(5), 1293–1305. https://doi.org/10.1007/s11222-016-9687-5
Hu, W., Xie, D., & Tan, T. (2004). A hierarchical self-organizing approach for learning the patterns of motion
trajectories. IEEE Transactions on Neural Networks, 15(1), 135–144. https://doi.org/10.1109/TNN.2003.
820668
Jun, J., Guensler, R., & Ogle, J. (2006). Smoothing methods designed to minimize the impact of GPS random
error on travel distance, speed, and acceleration profile estimates. Transportation Research Record, 1972, 141–
150. https://doi.org/10.1177/0361198106197200117
Khetarpaul, S., Gupta, S., Subramaniam, L. V., & Nambiar, U. (2012, August 8–10). Mining GPS traces to rec-
ommend common meeting points. Proceedings of the 16th International Database Engineering &
Applications Sysmposium, Prague (pp. 181–186). https://doi.org/10.1145/2351476.2351497
Knapen, L., Hartman, I. B.-A., & Bellemans, T. (2020). Using path decomposition enumeration to enhance route
choice models. Future Generation Computer Systems, 107, 1077–1088. https://doi.org/10.1016/j.future.2017.
12.053
Knapen, L., Hartman, I. B.-A., Schulz, D., Bellemans, T., Janssens, D., & Wets, G. (2016). Determining struc-
tural route components from gps traces. Transportation Research Part B: Methodological, 90, 156–171. https://
doi.org/10.1016/j.trb.2016.04.019
Lew, A., & McKercher, B. (2006). Modeling tourist movements: A local destination analysis. Annals of Tourism
Research, 33(2), 403–423. https://doi.org/10.1016/j.annals.2005.12.002
Lin, M., & Hsu, W.-J. (2014). Mining gps data for mobility patterns: A survey. Pervasive and Mobile Computing,
12, 1–16. https://doi.org/10.1016/j.pmcj.2013.06.005
McCabe, S., Kwan, H., & Roorda, M. J. (2013). Comparing GPS and non-GPS survey methods for collecting
urban goods and service movements. International Journal of Transport Economics/Rivista Internazionale di
Economia dei Trasporti, 40(2), 183–205. Retrieved June 2020, from https://www.jstor.org/stable/42748309
Murakami, E., & Wagner, D. P. (1999). Can using global positioning system (GPS) improve trip reporting?
Transportation Research Part C: Emerging Technologies, 7(2–3), 149–165. https://doi.org/10.1016/S0968-
090X(99)00017-0
Palma, A. T., Bogorny, V., Kuijpers, B., & Alvares, L. O. (2008, March 16–20). A clustering-based approach for
discovering interesting places in a single trajectory. SAC ’08: Proceedings of the 2008 ACM symposium on
Applied computing, Fortaleza, Ceará, Brazil (pp. 863–868), https://doi.org/10.1145/1363686.1363886
Raanan, M. G., & Shoval, N. (2014). Mental maps compared to actual spatial behavior using GPS data: A new
method for investigating segregation in cities. Cities, 36, 28–40. https://doi.org/10.1016/j.cities.2013.09.003
Rabiner, L. R., & Juang, B.-H. (1999). Fundamentals of speech recognition. Tsinghua University Press.
R Development Core Team. (2008). R: A language and environment for statistical computing [Computer software
manual]. Vienna, Austria. Retrieved from http://www.R-project.org

SPATIAL ECONOMIC ANALYSIS

 A pre-processing and network analysis of GPS tracking data 23

Schubert, E., Sander, J., Ester, M., Kriegel, H. P., & Xu, X. (2017). DBSCAN revisited, revisited: Why and how
you should (still) use DBSCAN. ACM Transactions on Database Systems, 42(3), 1–21. https://doi.org/10.1145/
3068335
Schuler, K. L., Schroeder, G. M., Jenks, J. A., & Kie, J. G. (2014). Ad hoc smoothing parameter performance in
kernel estimates of GPS-derived home ranges. Wildlife Biology, 20(5), 259–266. https://doi.org/10.2981/wlb.
12117
Sester, M., Feuerhake, U., Kuntzsch, C., & Zhang, L. (2012). Revealing underlying structure and behaviour from
movement data. KI-Künstliche Intelligenz, 26(3), 223–231. https://doi.org/10.1007/s13218-012-0180-9
Shen, L., & Stopher, P. R. (2014). Review of GPS travel survey and GPS data-processing methods. Transport
Reviews, 34(3), 316–334. https://doi.org/10.1080/01441647.2014.903530
Shoval, N. (2008). Tracking technologies and urban analysis. Cities, 25(1), 21–28. https://doi.org/10.1016/j.cities.
2007.07.005
Shoval, N., & Ahas, R. (2016). The use of tracking technologies in tourism research: The first decade. Tourism
Geographies, 18(5), 587–606. https://doi.org/10.1080/14616688.2016.1214977
Siła-Nowicka, K., Vandrol, J., Oshan, T., Long, J. A., Demšar, U., & Fotheringham, A. S. (2016). Analysis of
human mobility patterns from GPS trajectories and contextual information. International Journal of
Geographical Information Science, 30(5), 881–906. https://doi.org/10.1080/13658816.2015.1100731
Silm, S., & Ahas, R. (2010). The seasonal variability of population in Estonian municipalities. Environment and
Planning A: Economy and Space, 42(10), 2527–2546. https://doi.org/10.1068/a43139
Spaccapietra, S., Parent, C., Damiani, M. L., de Macedo, J. A., Porto, F., & Vangenot, C. (2008). A conceptual
view on trajectories. Data and Knowledge Engineering, 65(1), 126–146. https://doi.org/10.1016/j.datak.2007.
10.008
Stopher, P., FitzGerald, C., & Zhang, J. (2008). Search for a global positioning system device to measure person
travel. Transportation Research Part C: Emerging Technologies, 16(3), 350–369. https://doi.org/10.1016/j.trc.
2007.10.002
Theo, L. (2011). Simplifying central place theory using GIS and GPS. Journal of Geography, 110(1), 16–26. https://
doi.org/10.1080/00221341.2010.511244
Tiwari, S., & Kaushik, S. (2013, March 25–27). Mining popular places in a geo-spatial region based on GPS data
using semantic information. In A. Madaan, S. Kikuchi, & S. Bhalla (Eds.), Databases in networked information
systems, 8th international workshop DNIS 2013, Aizu-Wakamatsu, Japan (pp. 262–276). https://doi.org/10.
1007/978-3-642-37134-9_20
Torrens, P., Li, X., & Griffin, W. A. (2011). Building agent-based walking models by machine-learning on diverse
databases of space–time trajectory samples. Transactions in GIS, 15, 67–94. https://doi.org/10.1111/j.1467-
9671.2011.01261.x
Třasák, P., & Štroner, M. (2014). Outlier detection efficiency in the high precision geodetic network adjustment.
Acta Geodaetica et Geophysica, 49(2), 161–175. https://doi.org/10.1007/s40328-014-0045-9
Tsui, J. B.-Y. (2005). Fundamentals of global positioning system receivers: A software approach (Vol. 173). John Wiley
& Sons.
Tukey, J. (1977). Exploratory data analysis (Vol. 2). Addison-Wesley.
Van Craenendonck, T., & Blockeel, H. (2015). Using internal validity measures to compare clustering algorithms.
Benelearn 2015 Poster presentations (online), 1–8.
Van der Spek, S., Van Schaick, J., De Bois, P., & De Haan, R. (2009). Sensing human activity: GPS tracking.
Sensors, 9(4), 3033–3055. https://doi.org/10.3390/s90403033
Wolf, J., Guensler, R., & Bachman, W. (2001). Elimination of the travel diary: Experiment to derive trip purpose
from global positioning system travel data. Transportation Research Record: Journal of the Transportation Research
Board, 1768(1), 125–134. https://doi.org/10.3141/1768-15
Wolf, J., Oliveira, M., & Thompson, M. (2003). Impact of underreporting on mileage and travel time estimates:
Results from global positioning system-enhanced household travel survey. Transportation Research Record:
Journal of the Transportation Research Board, 1854(1), 189–198. https://doi.org/10.3141/1854-21

SPATIAL ECONOMIC ANALYSIS

24 Antonino Abbruzzo et al.

Wolfson, O., Xu, B., Chamberlain, S., & Jiang, L. (1998, July 3). Moving objects databases: Issues and solutions.
Proceedings of the Tenth International Conference on Scientific and Statistical Database Management (cat.
no. 98tb100243) (pp. 111–122). https://doi.org/10.1109/SSDM.1998.688116
Xiao, G., Juan, Z., & Zhang, C. (2015). Travel mode detection based on GPS track data and Bayesian networks.
Computers, Environment and Urban Systems, 54, 14–22. https://doi.org/10.1016/j.compenvurbsys.2015.05.005
Zenk, S. N., Schulz, A. J., Matthews, S. A., Odoms-Young, A., Wilbur, J., Wegrzyn, L., Gibbs, K., Braunschweig,
C., & Stokes, C. (2011). Activity space environment and dietary and physical activity behaviors: A pilot study.
Health & Place, 17(5), 1150–1161. https://doi.org/10.1016/j.healthplace.2011.05.001
Zheng, Y., Zhang, L., Xie, X., & Ma, W.-Y. (2009, April 20–24). Mining interesting locations and travel sequences
from GPS trajectories. WWW’09: Proceedings of the 18th international conference on World Wide Web,
Madrid, Spain (pp. 791–800). https://doi.org/10.1145/1526709.1526816

SPATIAL ECONOMIC ANALYSIS