=Paper=
{{Paper
|id=Vol-1743/paper16
|storemode=property
|title=DAZIO: Detecting Activity Zones based on Input/Output Call and SMS Activity
|pdfUrl=https://ceur-ws.org/Vol-1743/paper16.pdf
|volume=Vol-1743
|authors=Miguel Nuñez-del-Prado,Ana Luna,Romain Gauthier
|dblpUrl=https://dblp.org/rec/conf/simbig/Nunez-del-Prado16a
}}
==DAZIO: Detecting Activity Zones based on Input/Output Call and SMS Activity==
https://ceur-ws.org/Vol-1743/paper16.pdf
DAZIO: Detecting Activity Zones based on Input/Output call and SMS
activity
Miguel Nuñez-del-Prado-Cortez Ana Luna Romain Gauthier
Universidad del Pacı́fico Universidad del Pacı́fico Instersec Labs
Av. Salaverry 2020 Av. Salaverry 2020 Parı́s - France
Lima - Perú Lima - Perú romain.gauthier@intersec.com
m.nunezdelpradoc@up.edu.pe ae.lunaa@up.edu.pe
Abstract of installing new infrastructure, to provide a better
QoS or to plan their infrastructure.
Mobile telecoms operators possess an Therefore, in the present paper, we will identify
enormous quantity of data, which could be activity and transit zones to monitor and to predict
used to reduce the cost of installing new the activity levels in the telecoms operators net-
infrastructure, to provide a better QoS or work. These monitoring and prediction are based
to plan their infrastructure. Thus, they are on the SMS and calls input/output activity levels
concerned to model, understand and pre- issued from the Telecoms Italia Big Data Chal-
dict SMS and calls activity levels in their lenge1 . The results of the present study is directly
infrastructures. Besides, SMS and call applied for: (1) targeting advertisement to activ-
activities analysis can open new business ity zones; (2) proposing a suitable place to open a
opportunities for geomarketing as well as new store in a city or (3) planning where to add
trade area analysis. In the present effort, cell towers to improve QoS.
we detected activity zones with a differ- In the present effort, we describe a methodology to
ence of only 0.5 km from the reference detect activity and transit zones. More precisely,
activity areas extracted from Geo-tweets. the contribution of this work is twofold. On one
We also used Markov chains to represent hand, we present an Activity Markov chain model
and predict SMS and call activity lev- to represent activity levels. On the other hand, we
els, achieving a prediction success rate be- predict future activity levels using the aforemen-
tween 80% and 90%. tioned model. The rest of the paper is organized
as follows. First, Section 2 describes the related
1 INTRODUCTION works on activity zones detection. Then, Section
3 presents the datasets we use for experiments.
Telecoms data is a rich information source for Next, Section 4 introduces our technique to detect
many purposes, ranging from urban planning and model activity zones as well as the approach
(Toole et al., 2012), human mobility patterns to forecast activity levels. Section 5 shows corre-
(Ficek and Kencl, 2012; Gambs et al., 2011), lation measurements of activity levels versus pol-
points of interest detection (Vieira et al., 2010), lution and weather conditions. Finally, Section 6
epidemic spread modeling (Lima et al., 2013), concludes the paper and depicts some future direc-
community detection (Morales et al., 2013), disas- tions.
ter planning (Pulse, 2013) and social interactions
(Eagle et al., 2013). 2 RELATED WORK
One common effort for these applications is to de- Dense areas detection has been studied from a
termine dense areas where many users stay for a human mobility point of view, using fine grain
significant amount of time, namely activity zones. and coarse-grained location data. As an example
Another task is to identify contiguous zones relay- of fine-grained location, the work of (Gambs et
ing activity zones (i.e., transit zones). Thus, in our al., 2011) use mobility traces of 172 Yellow Cabs
context, the detection of activity and transit zones 1
Telecoms Italia Big Data Challenge web-
as well as the interaction between identified ac- site: www.telecomitalia.com/tit/en/
tivity zones are crucial tasks for reducing the cost bigdatachallenge.html
130
Taxis, issued from GPS, in San Francisco Bay (Pi- pared to the ground truth.
orkowski et al., 2009) to detect taxi’s point of in- Another, more refined technique to identify dense
terests (POI). These POIs are equivalent to activity areas respecting natural tessellation is presented
zones, which tend to be zones with high pedes- by (Vieira et al., 2010). Authors use CDR loca-
trian presence. The authors rely on the begin- tions from calls of one million users during four
end heuristic (Gambs et al., 2010) and cluster- months over an area of 80 000 km2 . They propose
ing algorithms, such as Density Joinable (Zhou et a method composed of three phases: the first step
al., 2004), Density Time (Hariharan and Toyama, is the graph construction, which relies on Delau-
2004) and Time Density clustering (Gambs et al., nay triangulation (Dobkin and Laszlo, 1987). The
2010) to detect POIs in San Francisco city. An- triangulation algorithm makes connexions (edges)
other fine-grain data used to detect activity areas between near antennas (vertex) maximizing the
are issued from Geo-social networks. size of the angles of the triangles. Once the graph
The work of (Qu and Zhang, 2013) uses is built, all edges are weighted by the total activ-
Foursquare’s check-ins from 446 users during ten ity and by the number of users of both connected
months for identifying trade areas. They rely on antennas. The second phase is the computation of
four different techniques like Center of Mass lo- dense areas based on a maximum spanning tree
cation, the most commonly checked-in location, build using the Kruskal algorithm (Kruskal, 1956).
the place with the highest check-in density and Taking as input the weighted graph G, the idea
the center of mass of the most frequently visited behind this algorithm is to find a subgraph of G,
location cluster. The algorithm use for cluster- which maximizes the density and does not con-
ing is DBSCAN (Ester et al., 1996). Once they tain any cycle. At last, the post-processing phase
have identified the activity centers, they mark the uses (Shiloach and Vishkin, 1982) algorithm to es-
boundary of the area using drive-time/distance tablish groups of antennas representing dense ar-
polygon (Kures and Pinkovitz, 2011). This tech- eas. Thus, the algorithm groups adjacent vertex
nique consists of computing the decay distance from previously computed sub-graph to find a set
from a given store to home or work (authors as- of close vertex (a set of antennas). The authors val-
sume that the two most checked-in places are idate their results empirically based on the subway
home and work). The drawback of this approach structure of the region under study. Inspired by
is that the selected users are conditioned to check- these works, we propose a novel ad-hoc method-
in in the store under study. Thus, the dataset is ology to find activity (dense) zones as well as to
biased. model and forecast their activity levels using the
data provided by the Telecom Italia Big Data Chal-
Other works use coarse grain location from Call lenge (TIM challenge). We describe this dataset in
Data Records (CDR). For instance, the work of the next section.
(Isaacman et al., 2011) uses Hartigan’s leader clus-
tering algorithm (Hartigan, 1975) to identify dense 3 DATASET
areas. First, authors sort antennas by the amount
of time that phones contact the antenna. Once data Datasets provided by the TIM challenge were
is sorted, the clustering algorithm takes the first collected in the cities of Milan and Trento over
antenna as the centroid of a cluster. Then, it veri- November and December 2013. Our study only
fies if the next antenna is within a distance d from takes into account the dataset gathered from Milan
the centroid. If it is not the case, the antenna be- city. Nevertheless, all the analysis and the method-
comes the centroid of a new cluster. In the case ology could be generalized for any city. In the next
the antenna is within the distance, the algorithm subsections, we will describe the datasets provided
computes the new centroid as the weighted aver- for Milan city. These datasets were used to detect
age. They repeat the process until all antennas be- and model the activity zones (primary datasets)
long to a cluster. Researchers use CDR locations and to find some correlation between activity and
of 97 and 71 thousand unique users in Los An- other measures like air quality and weather condi-
geles and New York cities collected over 2 and a tions (auxiliary datasets). It should be noted that
half months as well as 19 volunteers as the ground space was discretized in a grid, and all measures
truth to validate their results. They were able to are normalized to correspond in one square of the
estimate dense areas with an error of 3 miles com- grid.
131
3.1 Main datasets number of vehicles that belongs to Milan, the
The datasets of Milan city we use to detect and number of vehicles that is not from Milan,
model the activity zones are: the number vehicles with engine ignition
systems, in movement and stopped (c.f.
Milan Grid is a geographical segmentation over Table 3).
the city to aggregate the measurements of the
other datasets. The area of each square is Id Time Direction
55 225 m2 , and it has 10 000 squares in the 60 17/12/13 18:00 WEST
Avg speed Std speed Mi plates
form of a point (x, y) and the latitude and 24 95 21
longitude belonging to this x, y position. An Non-Mi plates Ignition Mov/Stopped
example of the described data as well as the 62 2 2/0
grid over Milan are introduced in Table 1 and
Table 3: Example of private mobility data in Milan
Figure 1, respectively.
Point Latitude Longitude 3.2 Auxiliary datasets
x1 , y1 9.011 45.568 We also used additional datasets to analyze activ-
ity zones, dynamics, and correlations, as we show
Table 1: Example of the Milan grid data
in the following paragraphs.
Telecommunications - MI to MI provides infor-
mation regarding the directional interaction
strength, between the city of Milan and dif-
ferent areas based on the calls exchanged be-
tween Telecom Italia Mobile users. More
precisely, this dataset contains the origin and
destination Id squares, the time and the direc-
tional interaction strength i.e., Activity (c.f.
Table 4)
Figure 1: Milan grid
Id 1 Id 2 Time Activity
Telecommunications (SMS, Call and Internet) 1 3 1383345474 0.24
provides information about the activity of a Table 4: Example of activity data between Milan zones
square concerning received and sent SMS,
incoming and outcoming calls as well as
Precipitation describes the intensity and the pre-
internet usage. This data is temporal ag-
cipitation type over the city of Milan. In
gregated in timeslots of ten minutes and
more detail, the dataset uses a coarse spa-
provides the measure of the activity of a
tial aggregation by dividing Milan city into
given event as well as the square id (c.f. Table
four quadrants (northeast, northwest, south-
2). This kind of information is organized in a
east and southwest). The intensity value of
way that SMS-in and SMS-out activity scale
the phenomenon is between 0 and 3, the per-
are given in arbitrary units, and their values
cent of coverage of a given quadrant and the
range from 0 to 1.
precipitation type between 0 and 2, where 0
Id Time Country SMS-in
means absence of precipitation, 1 is rain and
1 1383265200 39 0.24 3 is snow (c.f. Table 5).
SMS-out Call-in Call-out Data
0.16 0.108 0.026 6.83
Time Id Intensity Coverage Type
Table 2: Example of activity data in Milan 201311060220 1 1 45 1
201311060220 2 0 0 0
201311060220 3 0 0 0
Private Transportation (Cobra Telematics) 201311060220 4 2 78 1
gives information about the private mobility
in Milan city by measuring the speed, the Table 5: Example of activity data between Milan zones
132
Air Quality describes the air pollution monitor- 4.1 Detecting activity levels
ing system of Milan city obtained by us- The basic idea behind this method is to have a
ing various types of sensors located within good representation of the activity variation levels
the city limits. This environmental dataset over the time. Activity levels could be classified
measures a different kind of contamination in three different degrees, low, medium and high.
agents, such as Ammonia, Nitrogen Dioxide, Cumulative distribution of incoming/outcoming
Total Nitrogen, Particulate Matter 2.5 µm SMS and call activity levels illustrated by a Heat
(PM2.5), Particulate Matter 10 µm (PM10), map over Milan city, as shown in Figure 2, were
Benzene, Sulphur Dioxide, Black Carbon, used to analyze data. The objective is to, empir-
Carbon Monoxide and Ozone. An exam- ically, find a suitable threshold to distinguish a
ple of pollution measure is given in Table square with high activity level from a square with
6, where the characteristics of this particular medium or low activity levels represented as green
sensor are in Table 7. and red in Figure 3, respectively.
Sensor id Time Measure
5823 2013/12/30 04:00 1.9
Table 6: Example of air quality measure
Sensor id Lat/Lon Pollution
5823 45.24/9.27 Carbon Monoxide
Table 7: Description of the sensor 5823
Social Pulse contains data derived from an analy-
sis of geolocalized tweets originated in Mi-
lan. This dataset provides a user id, DB-
Pedia entity, tweets language, municipality,
time, timestamp and location (c.f. Table 8).
User Entities Language
5fa4b1cc71 Halloween En
Figure 2: Milan activity heat map
Municipality Timestamp Lat, Lon
Milan 1383260474 9.21, 45.49
Figure 3 depicts the cumulative distribution of
Table 8: Example of Social Pulse data the aggregated incoming and outcoming SMS and
call activity of the telecommunications dataset (c.f.
Based on the aforementioned datasets, we have Subsection 3.1). The Heat map is built for an ac-
implemented our experiments using the main and tivity threshold of 25 units. Based on this visual-
the auxiliary datasets. These experiments are de- ization technique, that amount of units seems to
tailed in sections 4 and 5, respectively. be a good trade-off between compact and well-
separated activity zones. Heat maps were used
4 EXPERIMENTS
to represent represent tourist activity as shown by
In the present section, we describe our methodol- Olteanu et al. (Olteanu et al., 2011).
ogy to discover, model and predict the behavior In order to detect groups of squares representing
of a zone. We distinguish two different areas, the an activity zone, we can use a high activity thresh-
activity zone, where people stay on a regular ba- old (c.f. Subsection 4.2). In addition, we study, in
sis for a significant amount of time and the transit detail, the activity over work hours to analyze the
zone which is the area used by individuals to go difference between busy and idle squares. Figure
from one activity zone to another. In the next sub- 4 shows the difference in activity levels between
sections, we describe how to recognize an activity activity and transit zones. From 8 AM to 8 PM the
zone from a transit zone (Subsection 4.1), how to activity is considerably high, that is why that area
model activity levels (Subsection 4.2) and finally is composed of a vast number of squares during
how to predict them (Subsection 4.3). the day. On the other hand, transit zones display a
133
Figure 3: Cumulative distribution of the sms/call activity with
heat maps
Figure 5: Vehicles speed heat map based on Cobra dataset
much lower activity level and a higher fluctuation
throughout the day.
Figure 4: Activity over time in workdays between 8AM-
8PM. Activity zone (left) and transit zone (right)
Taking into account the elements as mentioned
earlier, we use the Heat map and activity threshold
presented in Figure 3 for detecting high activity
zones. Nevertheless, we need to define the borders
of these activity zones. From Figure 4, we can in-
fer that this irregularity of transit zone represents Figure 6: Centroids of the activity zones in Milan city (blue),
movement. Thus, the Cobra dataset gathers the centroid of the clusters from Geo tweets (green) and com-
modities issued from Foursquares check-ins shops (red) and
information about the movement of vehicles and restaurants (purple).
we depicted this information in a Heat map over
Milan city in Figure 5. The speed combined with
the activity level allowed us to detect the activity geolocalized tweets dataset to verify the accuracy
zones, as well as their borders. As the result of of our methodology. Applying DBSCAN (Ester
the combination of this two variables, we obtained et al., 1996) clustering algorithm with at least 5
28 activity areas and the centroids of these activity points per cluster within a radius of 3 km over
zones which are shown in Figure 6 in blue color. 8 282 users (i.e.,109 762 geolocalized tweets). We
Since we do not have the ground truth, we used the obtained 24 clusters depicted as green points in
134
Figure 6. Thus, some groups are close to the iden- (transition probabilities). In our case, an Activ-
tified activity zones in the northeast and south. In ity Markov chain is a probabilistic automaton (PA)
Northwestern, activity zones are represented by model that represents, in a compact way, the occu-
only one cluster due to the proximity of geolocal- pation (activity) of a square or activity zone. The
ized tweets and the approach of DBSCAN to build nodes symbolize the state (low, medium or high)
clusters. The Downtown area has many clusters of the squares or zones ([activity level] [zone
due to commodities concentration. To verify this code], ex: L A) and edges, weighted with a prob-
fact, we have included two categories of check-ins ability, represent the transition from one state to
from Foursquare, like shops (red) and restaurants another over time windows. This model could be
(purple). One thing that surprised us was the de- expressed in the form of a graph or the form of a
tection of a large group of geolocated tweets in transition matrix (c.f. Figure 9).
the southern outskirts of Milans grid. Using these The process for building an Activity Markov
data we found that the distance between the cen- model is divided into two stages. The first one
troids of the clusters and activity zones is 0.5 km is basically to order the events in a chronological
closer. These results validated the accuracy of our way. Then we classify them as low, if they have
heuristic method to find activity zones. Finally, we less than 15 activity units, as medium if activity
present the centroid of the detected activity zones units are between 15 and 25 and as high if there
in Figure 7. These regions are used in the next are more than 25 activity units. Then the transi-
subsection for predicting purpose. tion matrix is built by counting the variations from
one level to another, taking care to avoid loops.
When events are not recorded anymore, the matrix
is normalized to obtain the transition probabilities.
As shown in Figure 8, we have divided the time
into 4 different windows depending on the range
of time studied, each one has 6 hours and starts
at 6:00 am; for both weekdays and weekends; giv-
ing a total number of eight windows. Furthermore,
in each time window interactions between differ-
ent levels of activity are also modeled. Moreover,
we matched, after a model processing, an activity
zone of a time window with another (blue arrows).
We finally conclude, from the stationary vector of
Markov chains, that activity areas are occupied in
only 11% and free in 71%. It is important to point
out that, for improving the accuracy of the anal-
ysis, it would be better to divide the time slots in
less intervals instead of 6 hours.
Figure 7: Centroids of the activity zones in Milan. Until this point we showed that we are able to
identify high activity squares, activity and transit
Up to this point, we are able to identify activity zones. In the next subsections, we detail how to
zones. Thus, the next tasks are to model the behav- predict, in an unprecedented way and with a very
ior activity levels in the detected regions as well as acceptable rate of success, not only activity levels
to predict the activity levels in the identified activ- but also their possible changes.
ity zones.
4.2 Modeling activity levels 4.3 Prediction of activity levels
Relying on the thresholds mentioned above, we Anticipating high activity levels within an “ac-
can build an Activity Markov chains model to rep- tivity area” allows Telecoms operators to plan or
resent the change of activity levels over time. An avoid unnecessary investment in infrastructure, as
Activity Markov chain model is a stochastic pro- well as ensure the QoS or to start a new en-
cess where the changes of states are related to a trepreneurship. In this paper, our prediction, in-
probability associated with various state changes spired from the work of Gambs et al. (Gambs et
135
Figure 8: Example of an activity Markov chain (activity over time windows)
al., 2012), is performed using the transition matrix inputs, the algorithm returns the maximal outgo-
and allows us to obtain changes of activities be- ing probability from the transition matrix taking
tween temporary windows and within them. Then, into account only columns corresponding to the
predictions of activity changes within the same same time windows of the index row (local tran-
window and estimations can be made. For exam- sition from line 1 to 4 of the Algorithm 1). For in-
ple, we wonder what is the probability of passing stance, in Figure 8, if the actual state is medium on
from a low activity state to a high one or what is the time windows from 18:00h to 0:00h on week-
the likelihood for a given activity to remain in the days, the prediction algorithm will give an output
same state when we consider the next window and in the high level in the same time window. Another
the same range of time. To answer these questions, kind of prediction is to take into account others
we rely on the algorithm presented in Algorithm 1. columns instead of those that belong to the same
time window (inter time windows transition from
Algorithm 1: Prediction algorithm line 6 to 8 of the Algorithm 1). Given the low in
Data: TransitionMatrix, indexRow, the time windows from 18h to 0h on weekdays in
inTimeWindows Figure 8, the algorithm will output the low state on
Result: indexColumn time windows from 0h to 6h on weekends. In the
1 if inTimeWindows then case of the output we have the same probability,
2 //predict next state in the time windows ties are break randomly.
3 i=maxOutgoingProbOnWin(indexRow)
4 indexColumn = i
5 else
6 //predict the next state when the time
window change
7 i=maxOutgoingProbOnNextWin(indexRow)
8 indexColumn = i
9 return indexColumn Figure 9: Transition matrix example. Where H=high,
M=medium, L=low, W=Weekday, We=Weekend and be-
gin end hours. As a way of example we show two temporary
More precisely, Algorithm 1 takes as input a tran- windows framed with purple numbers inside.
sition matrix (transitionMatrix). Where the index
of the row in the transition matrix corresponds to To validate the accuracy of the predictions, we
the actual state of the system (indexRow) and a used data from the whole month of November as
boolean value is used to indicate whether the pre- training set and the first 16 days of December as
diction is local (inTimeWindows). Based on these testing set (we did not take into account New Year
136
we are going to use the auxiliary datasets to ana-
lyze activity zones interaction and to study possi-
ble correlation with the levels or evaluate how the
weather impact the utilization of the Telecoms ser-
vice.
5 PLAYING WITH OTHER DATASETS
In the present section, we will study the interac-
tion between detected activity zones (Subsection
5.1); the correlation between activity levels versus
pollution measures (Subsection 5.2) and the influ-
ence of the weather on the activity levels in the
Telecoms operator infrastructure (Subsection 5.3).
Figure 10: Success rate of prediction. Where local refers to
prediction inside a time windows and transition (trans) refers 5.1 Interaction between activity zones
to prediction over time windows.(#) means prediction value.
Using the directional interaction activity dataset
square zone between zones in the area of Milan, we plotted a
Local 0.86 0.7 graph to visualize the communication exchange,
Trans. 0.86 0.89 as well as various activity levels as we can appre-
ciate in Figure 11, where the width of the edges
Table 9: Table summarizing results of Figure 10 accounts for the logarithm of the aggregated activ-
ity for the whole month of November. To extend
celebration to avoid special dates that separate our the semantic of this graph, we modulated the size
study from normal behavior). The results are de- of the nodes according to the amount of tweets
picted in Figure 10, where the success rate (Equa- emitted from the corresponding zone taking into
tion 1) is the ratio between the correct prediction account global pulses dataset (c.f. Subsection 3.2).
and the total number of predictions and the num-
ber of overall forecast. Note that the number of
predictions is indicated in parentheses at the bot-
tom part of each bar.
#goodprediction
success rate = (1)
#predictions
We observe, from Table 9, the success rate for
both kinds of predictions, namely (1) within a
time window (local) and (2) in different time win-
dows (trans.). It is important to note that there
are two distinct scenarios; the first one considers
both local and trans. predictions based on Activity
Markov chains models which were built from ac-
tivity levels of squares; while the second scenario
takes into account activity levels from detected ac-
tivity zones to forecast future values. We did not
performed k-fold cross validation since the train-
ing set of a month is representative of the mobil- Figure 11: Interaction graph of activity zones.
ity pattern. Thus, adding more mobility traces to
the training test does not contribute to increase the We observe that Milan city has a star topology,
success rate. where there is a central node that communicates
So far, we are able to model, identify and predict with the other peripheral nodes. Another interest-
activity levels in activity zones. In the next section, ing fact is that small nodes tend to communicate
137
to the central node. Nevertheless, there are a few
exchanges between small contiguous nodes.
5.2 Forecasting pollution through activity
In this subsection, we study the correlation be-
tween the activity level presented on the telecom-
munication dataset and air quality measurements
(both described in Subsection 3.2) to forecast the
pollution level of the Milan city based on the ac-
tivity of the telecom operator antennas. Figure 12
shows the results of the correlation of the activity
with respect to different polluting gasses as well
as the number of vehicles in movement, ignition
or stopped.
Figure 13: Correlation of the activity of a square with radia-
tion sensors.
Figure 12: Correlation of the activity with pollution measures
and private mobility (Cobra).
Figure 14: Activity level of outgoing SMS and call.
We found out that the activity has a positive cor-
relation with PM10 (particulates matter with a di- red and light blue bars). Nevertheless, people tend
ameter of 10 microns or less) and PM2.5 pollution to call or send less SMS when it is snowing (dark
measures (fine particles of 2.5 micrometers of di- and light green bars).
ameter or less). From Figure 13 we can visualize
that the activity levels has a positive correlation 6 CONCLUSION AND FUTURE
with the radiation measures and a negative corre- DIRECTIONS
lation with the relative humidity. Our purpose in this research is to understand the
activity of the telecommunication network by an-
5.3 Influence of weather on the activity alyzing several aspects of Milan’s phone traffic
We compare the outgoing SMS and call activity flows. We were interested in the definition and
in the presence of different weather phenomena’s morphology of the activity and transit zones; the
scenarios, like rain, snow or the absence of both. prediction of activity levels over different regions
For this purpose, we used the Precipitation dataset with a success rate between 80% and 90%; the in-
(c.f. Subsection 3.2). teractions between the different activity zones and
From Figure 14, we can observe that people send the influence of the weather and the pollution on
more SMS and give more calls in presence of rainy that activity. Thus, our results offer a new way of
weather, even if rain is slight (blue, red, yellow looking at the telecommunication traffic data by
and green bars) than in normal conditions (dark, examining the various connections between appar-
138
ently uncorrelated datasets, providing insights to Alexander Varshavsky. 2011. Identifying important
manage and to optimize the whole network. For places in peoples lives from cellular network data. In Per-
vasive Computing, June.
business opportunities, this means (1) new geo-
marketing opportunities through a better under- J. B. Kruskal. 1956. On the Shortest Spanning Subtree of a
Graph and the Traveling Salesman Problem. In Proceed-
standing of users communication patterns, (2) new ings of the American Mathematical Society, 7.
trade area analysis, (3) cheaper network load bal-
ancing as well as (4) improved QoS. In the future, M Kures and Ryan Pinkovitz. 2011. Downtown and business
district market analysis. http://fyi.uwex.edu/
we would like to study and analyze in detail the downtown-market-analysis/.
opinions (sentiment analysis) discussed by users
Antonio Lima, Manlio De Domenico, Veljko Pejovic, and
and generators of tweets and identifying the geo- Mirco Musolesi. 2013. Exploiting cellular data for dis-
location of these activity areas. Another line of ease containment and information campaigns strategies in
investigation would include coarse mobility Call countrywide epidemics. Computing Research Repository,
abs/1306(4534):1–7, Jun.
Data Record (CDR) to take into account as an ad-
ditional element for detecting activity and traffic J. Morales, W. Creixell, J. Borondo , J. C. Losada , and
areas. R. M. Benito. 2013. Understanding ethnical interactions
on ivory coast. In Data for Development (D4D) challenge,
pages 115–120, May.
References Ana-Maria Olteanu, Roberto Trasarti, T Couronn, Fosca Gi-
D. P. Dobkin and M. J. Laszlo. 1987. Primitives for the ma- annotti, Mirco Nanni, Zbigniew Smoreda, and Cezary
nipulation of three-dimensional subdivisions. In Proceed- Ziemlicki. 2011. Gsm data analysis for tourism applica-
ings of the Third Annual Symposium on Computational tion. In Proceedings of the 7th International Symposium
Geometry, pages 86–99, New York, NY, USA. on Spatial Data Quality (ISSDQ).
Nathan Eagle, Alex S. Pentland, and David Lazer. 2013. Michal Piorkowski, Natasa Sarafijanovic-Djukic, and
Inferring social network structure using mobile phone Matthias Grossglauser. 2009. CRAWDAD data set
data. Proceedings of the National Academy of Sciences, epfl/mobility (v. 2009-02-24). Downloaded from
106(36):15274–15278., September. http://crawdad.org/epfl/mobility/, February.
Martin Ester, Hans-Peter Kriegel, Joerg Sander, and Xiaowei United Nations Global Pulse. 2013. Mobile phone
Xu. 1996. A density-based algorithm for discovering network data for development. http://www.
clusters in large spatial databases with noise. In Second unglobalpulse.org/research, October.
International Conference on Knowledge Discovery and
Yan Qu and Jun Zhang. 2013. Trade area analysis using user
Data Mining, pages 226–231.
generated mobile location data. In International Confer-
M. Ficek and L. Kencl. 2012. Inter-call mobility model: ence on World Wide Web, pages 1053–1064.
A spatio-temporal refinement of call data records using
Yossi Shiloach and Uzi Vishkin. 1982. An o(log n) parallel
a gaussian mixture model. In Conference on Computer
connectivity algorithm. Journal of Algorithms, 3(1):57–
Communications, pages 469–477, March.
67.
S. Gambs, M.-O. Nuñez Killijian, and M.. del Prado Cortez. Jameson L. Toole, Michael Ulm, Marta C. González, and Di-
2010. Gepeto: A geoprivacy-enhancing toolkit. In Ad- etmar Bauer. 2012. Inferring land use from mobile phone
vanced Information Networking and Applications Work- activity. In Proceedings of the ACM SIGKDD Interna-
shops (WAINA), 2010 IEEE 24th International Conference tional Workshop on Urban Computing, pages 1–8, New
on, pages 1071–1076, April. York, NY, USA.
Sébastien Gambs, Marc-Olivier Killijian, and Miguel M.R. Vieira, V. Frias-Martinez, N. Oliver, and E. Frias-
Núòez del Prado Cortez. 2011. Show me how you move Martinez. 2010. Characterizing dense urban areas from
and i will tell you who you are. Transition on Data Pri- mobile phone-call data: Discovery and social dynamics.
vacy, 4(2):103–126, August. In Second International Conference on Social Computing,
pages 241–248, August.
Sébastien Gambs, Marc-Olivier Killijian, and Miguel Núñez
del Prado Cortez. 2012. Next place prediction using mo- Changqing Zhou, Dan Frankowski, Pamela Ludford, Shashi
bility markov chains. In Proceedings of the First Work- Shekhar, and Loren Terveen. 2004. Discovering personal
shop on Measurement, Privacy, and Mobility, page 3. gazetteers: An interactive clustering approach. In Inter-
ACM. national Workshop on Geographic Information Systems,
pages 266–273, November.
Ramaswamy Hariharan and Kentaro Toyama. 2004. Project
lachesis: parsing and modeling location histories. In In
Geographic Information Science, pages 106–124.
John A. Hartigan. 1975. Clustering Algorithms. John Wiley
& Sons, Inc., New York, NY, USA, 99th edition.
Sibren Isaacman, Richard Becke, Ramn Cceres, Stephen
Kobourov, Margaret Martonosi, James Rowland, and
139