Environment and Planning B: Planning and Design2014, volume 41, pages 11131131

doi:10.1068/b130047p

An analysis of visitors behavior in The Louvre Museum:

astudy using Bluetooth data

Yuji Yoshimura, Stanislav Sobolevsky, Carlo Ratti

SENSEable City Laboratory, Massachusetts Institute of Technology,
77MassachusettsAvenue, Cambridge, MA 02139, USA;
email:yyoshi@mit.edu, stanly@mit.edu, ratti@mit.edu
Fabien Girardin
Near Future Laboratory, CP242, 3960 Sierre, Switzerland;
email:fabien@nearfuturelaboratory.com
Juan Pablo Carrascal, Josep Blat
Information and Communication Technologies Department, Universitat Pompeu Fabra,
RocBoronat, 138, Tanger Building 08018 Barcelona, Spain;
email:jp.carrascal@upf.edu, josep.blat@upf.edu
Roberta Sinatra
Center for Complex Network Research and Department of Physics, Northeastern
University, 110 Forsyth Street, Boston, MA 02115, USA; email:r.sinatra@neu.edu
Received 7 April 2013; in revised form 13 August 2013; published online 31 July 2014

Abstract. Museums often suffer from socalled hypercongestion, wherein the number of
visitors exceeds the capacity of the physical space of the museum. This can potentially be
detrimental to the quality of visitors experiences, through disturbance by the behavior
and presence of other visitors. Although this situation can be mitigated by managing
visitors flow between spaces, a detailed analysis of visitor movement is required to
realize fully and apply a proper solution to the problem. In this paper we analyze visitors
sequential movements, the spatial layout, and the relationship between them in a large-
scale art museumThe Louvre Museumusing anonymized data collected through
noninvasive Bluetooth sensors. This enables us to unveil some features of visitor behavior
and spatial impact that shed some light on the mechanisms of museum overcrowding.
The analysis reveals that the visiting styles of short-stay and long-stay visitors are not as
significantly different as one might expect. Both types of visitors tend to visit a similar
number of key locations in the museum while the longer-stay visitors just tend to do so
more time extensively. In addition, we reveal that some ways of exploring the museum
appear frequently for both types of visitors, although long-stay visitors might be expected
to diversify much more, given the greater time spent in the museum. We suggest that these
similarities and dissimilarities make for an uneven distribution of the number of visitors
in the museum space. The findings increase the understanding of the unknown behaviors
of visitors, which is key to improving the museums environment and visitor experience.

Keywords: Bluetooth tracking, visitor behavior, museum studies, human mobility,

building morphology

1 Mesoscopic research of visitors sequential movement in an art museum

Falk and Dierking argue that a major problem at many museums is crowding, and crowds
are not always easy to control (1992, page145). Museums and their exhibits, along with
their own spectacular architecture, become some of the most popular destinations for the
tourists, thus triggering hypercongestion (Krebs etal, 2007), as the number of visitors often
exceeds the capacity of spaces, which results in the museum becoming overcrowded.
1114 Y Yoshimura, S Sobolevsky, C Ratti and coworkers

Congestion in museums shows, on the one hand, high attractiveness and vitality, resulting
in positive economic impact. On the other hand, the increased number of visitors implies
potential negative effects which are detrimental to the quality of visiting conditions and
the visitors experience can be disturbed by the behavior and presence of other visitors
(Maddison and Foster, 2003, pages173174). In an age when museums play an important
role in mass cultural consumption and with urban regeneration and the promotion of the
image of cities (Hamnett and Shoval, 2003), museums are expected to achieve seemingly
contradictory objectives at the same time; that is, toincrease the number of visitors and also
enhance the quality of their experience by achieving comfortable visiting conditions through
management of the flow of visitors.
Visitors movement and circulation patterns in museums are recognized as an important
topic for research (Bitgood, 2006, page463). However, most of these studies conducted in art
museums have been done for only two extreme cases: (a)visitor patterns at the macroscale
to investigate the basic demographic composition of the museums visitors (Schuster, 1995),
along with psychographic factors which influence visit motives and barriers (Hood, 1983);
and (b)at the microscale to research visitor circulation in the individual exhibition rooms,
limited galleries, or other areas. This often results in revealing that: (1)the visitors attributive
features from a sociocultural point of view (ie,highly educated people and wealthy upper- or
middle-class people tend to visit more frequently than people from the lower social classes)
(Hein, 1998, page115116); and (2)there is a local interaction between the layout of the
exhibits displayed in the galleries and the visitors behavior in those spaces (Klein, 1993;
Melton, 1935; Parsons and Loomis, 1973; Weiss and Boutourline, 1963). This polarized
research resulted in a shortage of mesoscopic empirical analysis of visitors in large-scale art
museums, which have different research targets compared with a single exhibition, small or
medium-sized museums (Serrel, 1998; Trndle etal, 2012), or other types of museum (Kanda
etal, 2007; Laetsch etal, 1980; Sparacino, 2002).
Space syntax (Hillier, 1996; Hillier and Hanson, 1984) applies a different approach to
analyzing the influences of the spatial layout and design of buildings using visitors movement
and behavior by describing the overall configuration of the museum setting (for a review, see
Hillier and Tzortzi, 2006). This type of knowledge is key to producing patterns of exploration
and interaction of visitors, and the copresence and coawareness that exists between visitors in
the museum environment as a whole (Choi, 1999).
Yet all these studies rely on a spatially and temporally limited dataset, which often results
in providing just a snapshot of a limited area in the built environment. Even a simulation-
based analysis uses a simplification of human behavior to estimate visitors behavior rather
than revealing actual patterns of movement with real-world empirical data.
In this paper we analyze the sequential movement of visitors, the spatial layout, and the
relationship between them in order to clarify the behavioral features of visitors in a large-
scale art museumThe Louvre Museum. We focus on visitors circulation from the entrance
to an exit as a whole mobility network rather than their movement in particular individual
rooms. The way of visiting exhibits is analyzed by means of the visitors length of stay and the
sequences in which they make their visits, because these determine the visitors perceptions
and attentions that shape their visiting experience (Bitgood, 2006). The length of stay might
be thought to be the key factor that determines the number of places visited and the sequence
in which they are visited, which results in a variety of different routes; the more time you are
given, the more opportunities you have, and vice versa. The question to be asked is whether
this hypothesis is actually true, and by its extension, how the length of stay and the sequence
of the places visited make visitors mobility style different, and how this dissimilarity is seen
Visitors behavior in The Louvre Museum analyzed 1115

in the museum. This understanding might be the key to improving the museum environment,
as well as to enhancing visitors experiences.
We employ a systematic observation method relying on Bluetooth proximity detection,
which makes it possible to produce large-scale datasets representing visitors sequential
movement with low spatial resolution. Large-scale datasets refers to the sample size we
used being much larger than those collected in art museums for previous studies [eg,almost
2000 in Melton (1935); 689 in Serrell (1998); 576 in Trndle etal (2012); and 50 in Sparacino
(2002)], although each of them contains different types of information with sufficient
resolution for their particular objectives and as good as human-based observation, GPS,
RFID, or ultra-wideband technology can achieve. Inour work we explore the global patterns
of visitors behaviors by increasing the quantity of the data, because when we increase the
scale of the data that we work with, we can do new things that werent possible when we just
worked with smaller amounts (Mayer-Schnberger and Cukier, 2013, page10).
Thus, we limit our research to dealing with visitors physical presence in and between
places, without questioning the introspective aspects (eg,learning process, making meaning
from the experience of the museum), which the previous studies tried to answer by small-scale
sampling [see Kirchberg and Trndle (2012) for a review]. However, the superimposition of
large amounts of data about individuals movements over time allows some patterns to appear
to be self-organizing in a bottomup way from seemingly chaotic, disordered, and crowded
movement. These results could shed light on the quality of visit conditions derived from
overcrowding, not only around the spots where the iconic art works are placed, but also the
spaces in the network between them that have dynamic visitor flow. Abetter understanding
of visiting features would help in designing more adequate spatial arrangements and give
insights to practitioners on how to manage visitor flow in a more efficient and dynamic way.

2 Visitors sequential movement and analysis framework

The use of large-scale datasets enables us to discover and analyze frequent patterns in human
activities. Such analyses have been conducted in the specific spatiotemporal limitations
derived from the limited measurement of mobile objects (Miller, 2005), in different contexts
and at various scales. These analyses have shed light on unknown aspects of human behavior
to discover patterns in human mobility (Gonzlez et al, 2008; Hoteit et al, 2014; Kung et
al, 2014), communication (Ratti et al, 2010; Sobolevsky et al, 2013), and urban activities
(Grauwin et al, 2014; Ratti et al, 2006; Pei et al, 2014) by studying cell-phone usage at the
regional scale. Other data like social media (Hawelka et al, 2014) or bank card transactions
(Sobolevsky et al, 2014) have also been used. In particular, the sequential patterns of tourists
at the local scale has been studied by looking at the number of locations visited, their order,
and the length of stay, obtained from GPS data (Shoval et al, 2013), and, for instance, some
aspects of customers purchasing behavior in a grocery store have been disclosed by analyzing
the customers path, length of stay, and the categories of products purchased through RFID
data (Hui et al, 2009).
Previous research (Yoshimura etal, 2012) proposed a Bluetooth-based data-collection
technique in alarge-scale art museum at the mesoscopic scale in order to classify visitors
behavior by their most-used paths and their relationship with the length of stay. Bluetooth
data collection is based on systematic observation which detects Bluetooth-activated mobile
devices, in the framework of unobtrusive measures, making use of the digital footprint
unconsciously left by visitors. Aconsiderable number of studies have employed this method
but not in the context of large-scale art museums. Examples include measuring the relationship
between peoples social networks (Eagle and Pentland, 2005; Paulos and Goodman, 2004),
analyzing mobility of pedestrians (Delafontaine etal, 2012; Kostakos etal, 2010; Versichele
etal, 2012), and estimating travel times (Barcel etal, 2010).
1116 Y Yoshimura, S Sobolevsky, C Ratti and coworkers

A Bluetooth proximity-detection approach to the analysis of visitor behavior in museums

has many advantages. Contrary to the granular mobile-phone tracking (Ratti etal, 2006), the
detecting scale using Bluetooth is much more fine grained. In addition, in contrast to RFID
tags (Hui etal, 2009; Kanda etal, 2007) and active mobile-phone tracking with or without
GPS (Asakura and Iryob, 2007), with Bluetooth previous registration is not required and it is
not necessary to attach any devices or tags. The fact that no prior participation or registration
is required enables a mass participation of subjects and the collection of an enormous amount
of data in the long term, unlike time constrained cases (McKercher etal, 2012; Shoval etal,
2013). Also, the unobtrusive nature of Bluetooth removes bias in the data, which could be
created if a subject is conscious of being tracked. Furthermore, Bluetooth proximity detection
succeeds inside buildings or in the proximity of tall structures, where GPS connectivity is
limited. All these advantages make this method adequate for detecting visitors sequential
movement between key places, without specifying their activities, attributes, or inner
thoughts, in a consistent way at the mesoscopic scale in a large-scale art museum.
We identify a visitors length of stay at a particular location as the indicator for
measuringtheir interest level at that exhibit by merely accounting for their presence without
questioning their inner thoughts. We estimate visitors routes between sensors and time at the
place from the collected data.
As our analysis and interpretation of the data were conducted within a specific
spatiotemporal framework, our approach has some limitations. Firstly the concept of trajectory
used in this paper is different from the one usually available when working with data collected
by GPS systems. This is because a Bluetooth proximity sensor just provides the time-stamped
sequence of individual transitions of a mobile device between nodes (eg,sequence of AB
D), while a GPS system can track all the movements of a device. However, the network of
rooms derived from the spatial layout of the museum determines the feasible routes, and
this enhances estimation of the paths used by visitors between sensors without observing
their exact trajectories and orientations per room (Delafontaine etal, 2012). Secondly, we
cannot deal directly with visitors introspective factors, their expectations, experiences, and
satisfactions (Pekarik etal, 1999). This results in excluding from our study research questions
about wayfinding, which refers to a visitors ability to find his or her way within a setting,
and orientation, which indicates an available knowledge in a setting through the use of the
hand-held maps and direction signs, because they consist of the complex interaction between
environmental cognition and the orientation devices. In addition, avisitors presence at a
specific place is not necessarily related to their time engaging with the exhibits, although
previous studies used this to measure visitor interest (Melton, 1935; Robinson, 1928).
Finally, our sample is possibly biased in two ways. First, the sample composition is affected
by the segments of the mobile-device holders and their decision to activate or not activate
the Bluetooth function. Although the latter requires calculating the sample representativeness
and is typically conducted by using a short-term manual counting method (Versichele etal,
2012), we employed a long-term (one-month) systematic comparison of the number of
devices detected at the entrance with the official museum head count and ticket sales. This
method provided us with more comprehensive information compared with previous research.

3 Concept definitions and data settings

In this section we define the locations of sensors used and the components of the dataset in
order to explore our method and data consistency. We collected our dataset during aspecific
period and processed it into a specific form required for the analysis.
Visitors behavior in The Louvre Museum analyzed 1117

3.1 Sensors settings in the museum and definition of node

Figure1 shows the location of seven sensors, deployed throughout the museum, covering
key places for detecting visitors. They are situated in one of the busiest trails, identified by
The Louvre Museum authorities, which lead visitors from the entrance to the Venus de Milo;
Entrance Hall(E), Gallery Daru(D), Venus de Milo(V), Salle des Caryatides(C), Great
Gallery(B), Victory of Samothrace(S), and Salle des Verres(G).
Each sensor defined a detection area, identified as a node, approximately 20m long and
7m wide. The area varied in size, depending on the museum settings and the location of the
sensor (eg,inside functional wooden boxes, desks, or in open space). However, all sensors
covered targeted areas along the paths to key iconic art works. Once a Bluetooth-activated
mobile device enters a detection area, the sensor receives the signal emitted by the mobile
device and the detection continues until the device leaves the area. The sensor registers the
time at which the signal from the mobile device first appears, called the checkin time, and
when the signal disappears, called the checkout time; the time difference between each
mobile devices check-in and check-out times can be calculated to define the length of stay
at the node. Similarly, by looking at the first check-in time and the last check-out time for
a mobile device over all nodes, provided that the first and last nodes correspond to an entry

Figure 1. [In color online.] Location of seven sensors E, D, V, C, B, S, and G, indicating their
approximate sensing range.
1118 Y Yoshimura, S Sobolevsky, C Ratti and coworkers

point and an exit from the museum, respectively, it is possible to calculate how long a visitor
stays in the museum. The series of check-in and check-out times registered for a mobile device
by all the sensors makes it possible to construct a visitors trajectory through the museum. In
addition to the length of the stay, the sensors time stamps allow calculation of the travel time
between nodes. The synchronization of all sensors makes it possible to perform fine-grained
time-series analysis. All this information can be achieved without invading visitor privacy,
because the SHA algorithm (Stallings, 2011, pages342361) is applied to each sensor where
the MAC ID is converted to a unique identifier (Sanfeliu et al, 2010).
3.2 Collected sample
We collected data over 24 days; from 30April to 9May 2010, 30June to 8July 2010, and
7August to 18August 2010. We selected data starting and finishing at node E in order to
measure the length of stay in the museum. Consequently, 24452 unique devices were chosen
to be analyzed for this study. Onaverage, 8.2% of visitors activated Bluetooth on their mobile
device while in The Louvre Museum (Yoshimura etal, 2012).
3.2.1 Data clean up
The data collection was performed at different periods by a different number of sensors. We
checked for possible synchronization issues arising from a lack of calibration, then adjusted
the data to remove any inconsistencies. Finally, we only used data from visitors who started
from node E and finished at node E in order to measure the complete length of the visit to
the museumsuch entries indicate that the visitor was correctly registered when he (orshe)
entered, moved around inside, and left the museum.
3.2.2 Data processing
Figure2 graphically shows the features of the logged data. It displays all entries in the database
for a visitor for one day. Each lettered circle symbolizes detection at the corresponding node.
It shows that this particular visitor made a sequential movement, ESDE, and stayed at
nodeE for 3min 10seconds, nodeS for 15min 20seconds, nodeD for 9min 34seconds
and, again, nodeE for 6min 3seconds. The travel times between corresponding nodes were:
12min 23seconds for ES, 8min 11seconds for SD, and 9min 34seconds for DE.
We built a database and designed a query engine to extract and transform the data
for the different stages of the analysis. Table1 shows an example of the components
of thetransformed dataset. There is one entry per visitor, and it includes the date of the
visit,thepath followedthrough the museum, the time of entry (check-in), time of exit (check-
out), and the total length of the visit to museum.

Nodes and movements

Minutes spent

Figure 2. [In color online.] Visualization of a relationship between the sequential movement and the
time of stay of a visitor.
 Visitors behavior in The Louvre Museum analyzed 1119

Table 1. Example of the dataset.

Reference Date Path Checkin time Checkout time Length of visit

Unique ID 30April 2010 ESDE 09:04:35 11:07:52 02:03:17

3.3 Partitioning of visitors

In order to find the characteristics, the typical patterns of visits, and other determinant
features of visitor behavior, we examined two extreme groups. Firstly, we sorted all the visits
of our sample (24452 visits) by their total time spent in the museum. By binning them into
deciles, we obtained equal-sized clusters of approximately 2446 visits. Referencing all visits,
those found in the first decile are called short visits and we refer to these visitors as short-
stay visitors. Similarly, we refer to the visits in the tenth decile as long visits and to these
visitors as long-stay visitors.

4 Results
In the following subsections, we present an overview of the statistical analysis built around
the previously described dataset. We discuss the path sequence length, which is the number
of nodes visited, including multiple visits executed without returning toE, the length of the
visitors paths, and the frequency of the appearance of each path. The distribution of the path
sequence length is also presented and analyzed. We reveal visiting patterns, and the similarity
and dissimilarity of the behaviors of the long-stay visitors and the short-stay visitors.
4.1 Basic statistics of visitors behavior
We analyzed all visitor data to capture the features of their behavior, focusing on the path
sequence length and its relationship with the length of stay in the museum.
Figure3(a) shows the distribution of the number of visits (yaxis) to the length of stay
in the museum binned for each hour (xaxis). Although the maximum length of stay is more
than 15hours, only 410visitors stayed for more than 8hours, which corresponds to 1.6% of
the total. Conversely, the minimum length of stay of less than 1hour was for only one visitor,
while more than 30% of visitors stayed for 12hours. These facts indicate that the extreme
visitors, whose length of stay is more than 8 hours or less than 1 hour, can be aggregated for
the statistical reliability without substantially affecting the time-sensitive behavioral analysis.
The distribution of the length of stay is positively skewed, with the majority of the visitors
staying for 46hours.
40 20
Percentage of total visitors

30 15

20 10

10 5

0 0
1 3 5 7 9 11 13 15 1 6 11 16 21 26
Length of visit to the museum (hours) Path sequence length
(a) (b)
Figure 3. (a)The distribution of visits against the length of stay in the museum. (b)The distribution
of the path sequence length.
 1120 Y Yoshimura, S Sobolevsky, C Ratti and coworkers

Next, we look at the distribution of the path sequence length (number of nodes) [see
figure3(b)]. Although the maximum length of the path sequence length is thirty nodes, the
percentage of visitors who visited more than fifteen nodes was only 0.5%. Ingeneral, this
plot shows a distribution slightly skewed to the right, but visitors who visited only one node
appear quite frequently, covering 15.2% of the total. Very few people visited two nodes (2.9%).
However, the length of the sequence by itself does not necessarily reveal the size of the
visitor mobility area, because a visitor could easily move between nearby nodes frequently
without radially expanding throughout the museum.
Figure4(a) represents the number of unique nodes visitors passed during their stay in the
museum. We can observe that visiting two nodes rarely happened, while visiting one or three
nodes have almost the same frequency. The most frequent number of unique nodes visited is
four or five nodes, while visiting all six nodes rarely happens. This indicates that in most of the
cases some factors prevent the exploration of all the nodes, while all nodes but one could be
explored much more often. In addition, figure4(b) reveals that the average number of unique
nodes visited against the duration of the visit is almost constant. The correlation coefficient
between these two variables (Spearmans correlation=0.072, pvalue<2.21016) indicates
that the unique number of visited nodes is independent of the duration of the visit to the
museum, and vice versa. Surprisingly, the long-stay visitors usually visit even fewer nodes
than the short-stay ones.

30
Average number of unique nodes visited

4.0
Percentage of total visitors

3.5
20

3.0

10
2.5

0 2.0
1 2 3 4 5 6 0 2 4 6 8
(a) Number of unique nodes visited (b) Length of visit to the museum (hours)

Figure4. (a)Distribution of the number of unique nodes visited other than nodeE. (b)The average
number of unique nodes visited against the duration of the visit.

Figure5(a) shows the frequency of visits for each node: 97% of all visitors passed
nodeS.Nodes D andB are visited frequently (nearly 80% for each). On the other hand,nodeG
is the most rarely visited, with just 30% of all visitors. Figure5(b) presents the attractivity of
the nodes depending on the duration of the visit. As we can see, for most nodes the probability
of visiting does not depend on the length of stay in the museum as the probability is nearly
constant for all nodes. NodeG behaves differently from the others, as its probability of
attracting visitors increases with the visitors length of stay in the museum. This shows that
short-stay visitors show a lower tendency to visit nodeG, while long-stay visitors seem more
attracted to visit this node (perhaps having more time to explore this part of the museum),
although its frequency does not surpass 40%, regardless of the visitor type.
We can observe the difference in the transition rates (probability of moving to the given
destination node right after visiting the given origin) from any node to nodeG for the two
types of visitors (see table2). All the transition rates increase as the visitors length of stay
increases; nodes D, C, andS show substantial increases (shown in bold in table2).
 Visitors behavior in The Louvre Museum analyzed 1121

1.0 1.0

Probability of a visitor visiting

0.8
Percentage of total visitors

0.4

0.6
0.6

a specific node
0.4

0.8
0.2
Node D Node V Node C
Node B Node S Node G
0.0 0.2
D V C B S G 1 2 3 4 5 6 7 8
Node Average length of visit to the museum (hours)
(a) (b)
Figure5. [In color online.] (a)The frequency of visits each node receives. (b)The frequency of
visiting different nodes at least once against the duration of stay.

Table 2. Two types of visitors transition rate from previous nodes to nodeG expressed as apercentage.
Bold type indicates a substantial increase.

Node before node G Short-stay visitors (%) Long-stay visitors (%) Difference (%)

D 4.00 7.17 3.17

V 1.38 3.17 1.79
C 4.86 9.91 5.05
B 5.60 6.30 0.70
S 2.53 5.69 3.16

4.2 Similarity of visitor behaviors

By looking at the path length of the visitors of different stay time we find another surprising
effect. Although the path length increases slightly with increased length of visit, the path
length of long-stay visitors is not substantially longer than that of the short-stay visitors. In
addition, the number of nodes that make up a visit is very similar.
Figure6(a) reveals that, while visitors tend to visit, on average, 4.3nodes whenvisitingthe
museum for 12 hours, they are likely to visit only 5.5nodes when they stay for 37 hours.

7 0.8
Average length of path sequence

Probability of a visitor visiting

6
0.6
1 node
5 2 nodes
0.4 3 nodes
multiple nodes

4 4+ nodes

0.2
3

2 0.0
0 2 4 6 8 0 2 3 4 5 6 7 8

(a) Length of visit to the museum (hours) (b) Length of visit to the museum (hours)

Figure6. [In color online.] (a)The average length of path sequence (yaxis) against the average length
of stay in the museum (xaxis). (b)The probability of a visitors path length being 1,2,3, or more
nodes by their length of stay in the museum (xaxis).
1122 Y Yoshimura, S Sobolevsky, C Ratti and coworkers

The longer length of stay is three times the shorter, but it results in an increase of only 28%
in the sequence length. In addition, the long-stay visitors (ie,910 hours) visited 6.6nodes
on average, which is even less than the 89 hour visitors. The path sequence length increases
as the duration of the visit increases, but the rate of change is not substantial (see table3)
especially if compared with the increases in visit times. Figure6(b) presents the probability of
visitors having a certain path length versus their length of stay in the museum. The probability
of visiting 1,2,3, or more nodes against the length of stay aggregated by each hour appears
almost flat, suggesting it is independent of the duration of the visit to the museum. We can
also observe this tendency by examining the frequently appearing paths of the short-stay and
long-stay visitors (table4).
Table 3. The average length of path sequence (number of nodes visited) per hour and its percentage
of increase.

Length of visit (hours) Path sequence length (nodes) Percentage increase

12 4.28 5.84
23 4.53 9.93
34 4.98 7.63
45 5.36 6.34
56 5.70 10.53
67 6.30 1.43
78 6.39 8.29
89 6.92 4.62

Table4. The probability of a visitor having a path length of 1,2,3, or more by the length of their stay
in the museum.

Length of visit (hours) Probability of path length

1 node 2 nodes 3 nodes >3 nodes

12 0.14 0.03 0.11 0.70

23 0.16 0.03 0.10 0.70
34 0.15 0.02 0.08 0.73
45 0.13 0.03 0.10 0.72
56 0.14 0.02 0.10 0.73
67 0.14 0.01 0.10 0.74
78 0.13 0.02 0.12 0.72
89 0.18 0.02 0.08 0.70
Average 0.14 0.02 0.10 0.72
Standard deviation 0.01 0.00 0.01 0.01

Table5 presents the top five most frequently appearing paths for both the short-stay and
long-stay visitors. We counted the number of paths which appear in both groups (visited at
least four nodes or visited less than four nodes) and divided by the total number of visitors
in the group (ie, 2445), in order to obtain the frequency of apath appearing. This reveals that
both groups have similar frequent path length; the short-stay paths are just slightly shorter
compared with the long-stay paths. For both groups, the first and second most frequently
Visitors behavior in The Louvre Museum analyzed 1123

Table5. Top five of the frequently appearing paths for paths of four nodes or more and for paths less
than four nodes for the long-stay and short-stay visitors.

Path of long-stay visitors Frequency (%) Path of short-stay visitors Frequency (%)

Visitors whose length of path is 4 nodes or more

EDSBDVCE 2.23 EDSBDVCE 8.26
EDSBDE 1.84 EDSBDE 6.89
EDSBDVE 1.50 EDSBVE 5.57
EDSBDSE 0.89 ESDVCE 4.67
EDSBDGE 0.78 ECVDSBE 3.71
Visitors whose length of path is less than 4 nodes
ESE 45.10 ESE 36.34
EDSBE 12.38 EDSBE 16.62
ESESE 7.64 EVDSE 4.51
EBE 5.19 EGSBE 4.12
EGSBE 4.28 ESBE 3.35

appearing paths for the long-stay and short-stay visitors are very similar, otherwise the
frequency of the group that visited more than four nodes is much lower than for those who
visited less than four nodes. The results show that the behavioral ways of short-stay and long-
stay visitors are not as significantly different as one might expect. Both types of visitors tend
to visit the same number of popular places but the long-stay visitors just tend to do so more
time extensively (spending longer studying the exhibits).
We examine in more detail the visitors whose path length is less than four nodes. Within
them, the most frequently appearing path for each category (ie, visited 1,2, or3 nodes)
coincides well between the groups of short-stay and long-stay visitors. Figures7(a), (b), (c)
present the probability of visiting 1node, 2nodes, or 3nodes, respectively. We can observe
that in each case only one path has a strong influence on the probability as a whole, especially
in figure7(a), where 89.4% of those visitors took the path ESE).
Similarly, visitors who followed the ESBE path, which is the most frequently
appearing path for those visiting two nodes (37.52%), added nodeB at the end of their visit,
while for EDSBE, the most frequently appearing path for those visiting three nodes
(38.94%), nodes D andB were added at the beginning and end of their visit, respectively.
There is no clear difference between long-stay and short-stay visitors to 1, 2, or3 nodes;
rather, their behavior seems very similar, other than the substantial difference in the length of
the visit to the museum.

5 Discussion
The previous sections revealed that many features of the behavior of the long-stay and
short-stay visitors, including the path sequence length and the unique nodes visited, do not
appear to be strikingly different between visits of different duration, and are sometimes even
independent or nearly independent of duration. In this section we show that visitors paths
and their variations are quite selective, with visitors mostly choosing the same paths in terms
of the path sequence length and sequential order although many other options exist. This
creates an uneven distribution of visitors among spaces, and is possibly one of the main
causes of high congestion and vacant spaces in the museum.
 1.0

Probability of a visitor 0.8

0.6 EDE EVE

ECE EBE
0.4
ESE

0.2

0.0
(a)
0.6
EDSE EVCE
EBSE ESDE
0.5
ESBE

0.4
Probability of a visitor

0.3

0.2

0.1

0.0
(b)
0.5

0.4

EDSBE EBSDE
Probability of a visitor

0.3
ESESE ESBDE

EGSBE
0.2

0.1

0.0
1 2 3 4 5 6 7 8 9 10
(c) Length of visit to the museum (hours)

Figure7. [In color online.] The probability that visitors take particular path lengths visiting (a)1node,
(b)2nodes, (c)3nodes, versus the length of their visit to the museum.
Visitors behavior in The Louvre Museum analyzed 1125

5.1 Uneven spatial distribution of visitors

The interplay between sensor locations and the spatial layout of the museum determines
thespecific and possible route(s) used by visitors. All sensors were placed logistically inthe
determinant positions for visitors route choice in the museum. Therefore, the transition
between two places makes it possible to estimate the determinant route that visitors take.
Thus, we can clarify the uneven spatial use of the museum accesses for visitors entry
and exit behaviors by analyzing the first two and last two locations, respectively, in their
sequence: 71.6% of visitors took ED, EB, or ES, meaning that they entered through the
Denon access, which indicates that only 28.3% used the Sully or Richelieu access (ie,EV,
EC, EG). 57.3% of visitors exited from Denon, 14.3% fewer than entered at Denon. This
technique enables us to determine the rooms visited without observing their exact trajectories.
Also, this indicates that we could speculate on the volume of visitors and their concentration
along specific paths without knowing the exact load per room.
In the previous section it was revealed that 13.5% of all visitors to the museum only visited
the Victory of Samothrace (nodeSone of the most iconic exhibits in the museum), not
visiting any of the other five nodes. Considering the museums spatial layout, these visitors
used the Mollien stairs, which connect the 16thcentury to 19thcentury Italian sculpture
rooms on the ground floor to the 19thcentury French painting room on the first floor, to visit
nodeS instead of using the Victory of Samothrace staircase where nodeD is located (see the
orange line in figure8).
From the spatial point of view this is an intriguing result because the shortest path from
the entrance (nodeE) to the Victory of Samothrace (nodeS) is the one which passes nodeD,
meaning that they turned to the left at the intersection between exhibit No.2 and No.4 (see
the orange dotted line in figure8). To use the route through the Mollien stairs signifies a
detour, both spatially and temporally, to reach nodeS.
Table6 reveals visitors route choice in more detail; from nodeE almost 40% of visitors
turned to the left to reach nodeD (ie,ED), while around 20% of visitors turned to the right
(ie,ES). Again, there is no significant difference between the behaviors of the long-stay
and short-stay visitors, meaning that both start their museum experience in a similar way. In
addition, since nodesD, V, B, and G are installed in some key points after exiting the Denon
Wing, all those visitors whose path was ESE tended to stay in a very confined area of the
Denon Wing during their visit. They just explore and stay in the small area during their visit,
and this tendency is stronger for the long-stay visitors than the short-stay ones (see table5).
On the other hand, the most frequently appearing path for both groups who visited at least
four nodes is EDSBDVCE, where the visitor visited the Gallery Daru, the Victory of
Samothrace, the Great Gallery, and the Venus de Milo (see the yellow line in figure8). This
path starts from a trail of ED and finishes with CE, indicating that the visitor entered the
museum at the Denon access, and exited from the Richelieu or Sully access. This suggests
that these visitors tend to explore the museum extensively through covering most of the iconic
exhibits rather than staying in only one part of the museum. In addition, the frequency of this
path for the short-stay visitors is much higher than that of the long-stay visitors. This could
indicate that the short-stay visitors might tend to select the most spatially optimized paths to
visit all the possible iconic exhibits within their limited available time in the museum.
We believe that short-stay visitors explore fewer of the popular places due to the limited
time that they have to spend in the museum. This is intuitive since a visitors movement
and their activities would be limited when the length of their visit to the museum is short.
Consequently, the trajectories of the long-stay visitors would be expected to be more
complex than those of the short-stay visitors, and vice versa. However, the results show that
the behavioral patterns of short-stay and long-stay visitors are not as significantly different as
 1126

Figure 8. [In color online.] (a) The map of the spatial layout of The Louvre Museum and the routes visitors used. (b)The transition percentages between
locations, showing only major links between each pair of nodes.
Y Yoshimura, S Sobolevsky, C Ratti and coworkers
Visitors behavior in The Louvre Museum analyzed 1127

Table6. Three types of visitor transition rates from nodeE to the subsequent node expressed as
apercentage.

Subsequent node from node E All visitors Short-stay visitors Long-stay visitors

D 43.32 42.41 40.34

V 11.25 12.80 11.38
C 9.59 10.02 11.32
B 6.80 9.20 7.51
S 21.53 21.02 20.82
G 7.51 4.54 8.63

one might expect. Both types of visitor tend to visit asimilar number of the popular rooms,
but the long-stay visitors tend to do so more time extensively.
The results imply that visitors trajectories seem to be quite limited in terms of the path
sequence length and its order, although there exist a number of possible routes including
repeating nodes. More generally, we might say thatand this partially agrees with Chois
(1999) statementthe more spaces available, the more the visitors path tends to be selective.
That is, when the number of the rooms with exhibits increases, visitors seem not to visit them
all, but visit a few of them selectively. But our findings tell us more; these limited paths and
their use are almost independent of the length of the visit to the museum, meaning that most
visitors, irrespective of whether their visit is short or long, tend to use the same trajectories.
We speculate that this similarity/dissimilarity of the patterns makes the distribution of
the quantity of visitors in the museum space uneven; for instance, the route EDSBD is
frequently observed, independent of the length of the visit, suggesting that there can be ahigh
concentration of visitors in those enclosed areas. Incontrast, some spaces can be found to be
quite vacant; the sequential pattern between nodeS and nodeG is rarely found, especially,
in the short-stay visits. This indicates that the topological proximity and the attractivity of
a node can be changed depending on the visitors length of stay (see figure5). It could be
thatnodeG, which tends to be visited when people have more time to explore the museum, is
not seen as a necessity or priority during the museum visit. Thus, the distribution of visitors
is uneven and the number of visits that each room receives varies.

6 Conclusion
In this study we examined visitors mobility styles and their respective spatial impacts by
analyzing large-scale datasets obtained through Bluetooth proximity detection in a bottom-up
methodology. This analysis and the results obtained give a great scientific advancement to
improving visiting conditions, which strongly affect the quality of avisitors experience in
the museum.
The results indicate that the behavior of short-stay and long-stay visitors is not as different
as one might expect. The path lengths grow at a much slower rate compared with increasing
duration of stay. Even more surprisingly, the number of unique nodes visited remains almost
constant, independent on the length of the visit. The correlation coefficient between these two
variables quantitatively indicates that the unique number of nodes visited is independent of
the duration of the visit to the museum, and vice versa. Both short-stay and long-stay groups
visit mostly the same number of sensor locations, while the long-stay visitors just tend to
do so more time extensively. Moreover, the probability of the appearance of visitors whose
path sequence length is small (<4nodes), is constant across all time divisions, meaning
that there always exists acertain category of visitors who do not try to explore museum
space extensively no matter how much time they have to do so. Also, we discovered that the
1128 Y Yoshimura, S Sobolevsky, C Ratti and coworkers

frequency of node visits per hour is almost constant and independent of the length of time
spent in the museum.
Conversely, we can point out key differences in visitors behavior within each of two
groupsthose who visited more than four nodes and those who visited fewer than four. The
average number of locations visited, for each of the groups, does not depend on the time
people have to spend in the museum (ie,it is independent of a visitor being classified as a
short-stay and long-stay visitor). For both short-stay and long-stay visitors the most frequently
occurring path in the group that visited at least four nodes is EDSBDVCE. We might
suggest that this path could be one of the most optimized paths, enabling visitors to explore
all the interesting places as quickly as possible. Alternatively, the group that visited just a few
nodes (less than four), which appears to be of relatively the same size among both short-stay
and long-stay visitors, might be interested in just a few of the iconic art works, or just not
motivated or informed enough to explore the bigger space.
All of this suggests that some routes used to explore the museum appear frequently for
both short-stay and long-stay visitors even though the latter might be expected to be much
more diverse in their choices given the longer time available. This implies that visitors
sequential movement in The Louvre Museum is quite limited in terms of path sequence
length and order, though there are a number of possible routes including repeating the same
nodes. We speculate that these similarities/dissimilarities could cause uneven distribution of
the number of visitors, resulting in congestion or sparcity in some museum spaces.
These findings present a significant advancement in describing patterns in visitors
activity and behavior in a museum, and might enable us to foresee visitor movement. This
also indicates the possibility of dynamically managing visitor flow and museum congestion,
taking into account time-related factors, and the possible advantages of the design of the
spatial arrangement. In addition, the transition rate and the probability of movement between
nodes makes it possible to foresee the specific quantity and flow of visitors at a certain time
and space, helping the development of more flexible and dynamic policies for space control.
For instance, the similarities/dissimilarities of both types of visitor, which were unknown
prior to this study, might make the practitioner reconsider the target of some management
techniques that should be applied carefully on the proper and segmented group types (Krebs
etal, 2007; Maddison and Foster, 2003). Also, a dynamic visitor-control system might be
developed, based on our findings, by using the audioguides to change suggested visitor
routes dynamically depending on the congestion level as calculated by the data gathered
from sensors installed throughout the museum.
Finally, these results might enable improvement in the quality of information that can be
provided to visitors at an adequate place and time in order to maximize their fulfillment of the
social and cultural experience, thereby optimizing the museum infrastructure.
Acknowledgements. We would like to thank the Studies and Research Department of The Louvre
Museum for their support, Kyuha Shim, Namju Lee, and Miriam Roure for the graphics materials, and
Alexander Amini and David Lee for text proofreading. We are indebted to many people for providing
an extremely stimulating research discussion. In particular, we would like to thank Jaume Barcel,
Catherine Guillou, Anne Krebs, Thomas Besancon, Marie-Anne Corniou, Michael Szell, Riccardo
Campari, Sebastian Grauwin, Markus Schlapfer, Chaogui Kaug, Tao Pei, Kael Greco, Dietmar
Offenhuber, Rodrigo de Oliveira, Kenichi Nakamura. We further thank the MIT SMART Program, the
Center for Complex Engineering System (CCES) at KACST and MIT CCES program, Volkswagen
Electronics Research Lab, BBVA, The Coca Cola Company, Ericsson, Expo 2015, Ferrovial, The
Regional Municipality of Wood Buffalo, and all the members of the MIT Senseable City Lab
Consortium. RS is supported by the James S. McDonnell Foundation. Of course, any shortcomings
are our sole responsibility.
Visitors behavior in The Louvre Museum analyzed 1129

References
Asakura Y, Iryob T, 2007, Analysis of tourist behaviour based on the tracking data collected using
a mobile communication instrument Transportation Research Part A: Policy and Practice
41684690
Barcel J, Montero L, Marqus L, Carmona C, 2010, Travel time forecasting and dynamic origin
destination estimation for freeways based on Bluetooth traffic monitoring Transportation
Research Record: Journal of the Transportation Research Board number 2175, 1927
Bitgood S, 2006, An analysis of visitor circulation: movement patterns and the general value
principle Curator 49 463475
Choi Y K, 1999, The morphology of exploration and encounter in museum layouts Environment
and PlanningB: Planning and Design 26 241250
Delafontaine M, Versichele M, Neutens T, Van de Weghe N, 2012, Analysing spatiotemporal
sequences in Bluetooth tracking data Applied Geography 34 659668
Eagle N, Pentland A, 2005, Reality mining: sensing complex social systems Personal and
Ubiquitous Computing 10 255268
Falk J H, Dierking L D, 1992 The Museum Experience (Walesback Books, Washington,DC)
Gonzlez M C, Hidalgo C A, Barabsi A L, 2008, Understanding individual human mobility
patterns Nature 453 779782
Grauwin S, Sobolevsky S, Moritz S, Gdor I, Ratti C, 2014, Towards a comparative science of
cities: using mobile traffic records in New York, London and Hong Kong, arXiv preprint
arXiv:1406.4400
Hamnett C, Shoval N, 2003, Museums as flagships of urban development in Cities and
Visitors: Regulating People, Markets, and City Space EdsLMHoffman, DJudd, SSFainstein
(Blackwell, Oxford) pp217236
Hawelka B, Sitko I, Beinat E, Sobolevsky S, Kazakopoulos P, Ratti C, 2014, Geo-located Twitter
as proxy for global mobility patterns Cartography and Geographic Information Science 41
260271
Hein G, 1998 Learning in the Museum (Routledge, London)
Hillier B, 1996 Space is the Machine: AConfigurational Theory of Architecture (Cambridge
University Press, Cambridge)
Hillier B, Hanson J, 1984 The Social Logic of Space (Cambridge University Press, Cambridge)
Hillier B, Tzortzi K, 2006, Space syntax: the language of museum space, in ACompanion to
Museum Studies Ed.SMacDonald (Blackwell, Oxford) pp282301
Hood M G, 1983, Staying away: Why people choose not to visit museums Museum News
61(4)5057
Hoteit S, Secci S, Sobolevsky S, Ratti C, Pujolle G, 2014, Estimating human trajectories and
hotspots through mobile phone data Computer Networks 64 296307
Hui S K, Bradlow E T, Fader P S, 2009, Testing behavioral hypotheses using an integrated model of
grocery store shopping path and purchase behaviour Journal of Consumer Research
36478493
Kanda T, Shiomi M, Perrin L, Nomura T, Ishiguro H, Hagita N, 2007, Analysis of people
trajectories with ubiquitous sensors in a science museum, in Proceedings 2007 IEEE
International Conference on Robotics and Automation (ICRA07) pp48464853
Kirchberg V, Trndle M, 2012, Experiencing exhibitions: areview of studies on visitor experiences
in museums Curator:The Museum Journal 55 435452
Klein H, 1993, Tracking visitor circulation in museum settings Environment and Behavior
25782800
Kostakos V, ONeill E, Penn A, Roussos G, Papadongonas D, 2010, Brief encounters: sensing,
modelling and visualizing urban mobility and copresence networks ACM Transactions on
ComputerHuman Interaction 17(1) 138
1130 Y Yoshimura, S Sobolevsky, C Ratti and coworkers

Krebs A, Petr C, Surbled C, 2007, La gestion de lhyper frquetation du patrimoine: dune

problmatique grandissante ses rponses indiffrencies et segmentes [Managing the
hyper-congestion of cultural heritage sites: from a growing problematic to its unspecialized
and segmented responses], in Communications of the 9th International Conference on Arts and
Culture Management University of Valencia, http://www.adeit.uv.es/aimac2007/index.php
Kung K S, Greco K, Sobolevsky S, Ratti C, 2014, Exploring universal patterns in human home/
work commuting from mobile phone data PLoS ONE 9(6): e96180
Laetsch W, Diamond J, Gottfried J L, Rosenfeld S, 1980, Children and family groups in science
centers Science and Children 17(6) 1417
McKercher B, Shoval N, Ng E, Birenboim A, 2012, First and repeat visitor behaviour: GPS
tracking and GIS analysis in Hong Kong Tourism Geographies 14 147161
Maddison D, Foster T, 2003, Valuing congestion costs in the British Museum Oxford Economic
Papers 55 173190
Mayer-Schnberger V, Cukier K, 2013 Big Data: ARevolution That Will Transform How We Live,
Work and Think (John Murray, London)
Melton A W, 1935 Problems of Installation in Museums of Art. American Association of Museums
Monograph New Series No. 14 (American Association of Museums, Washington,DC)
Miller H J, 2005, A measurement theory for time geography Geographical Analysis 37 1745
Parsons M, Loomis R, 1973 Visitor Traffic Patterns: Then and Now (Office of Museum Programs,
Smithsonian Institution, Washington,DC)
Paulos E, Goodman E, 2004, The familiar stranger: anxiety, comfort, and play in public places
Proceedings of the SIGCHI Conference on Human Factors in Computing Systems 6 223230
Pei T, Sobolevsky S, Ratti R, Shaw S L, Li T, Zhou C, 2014, A new insight into land use
classification based on aggregated mobile phone data International Journal of Geographical
Information Science (ahead-of-print), arXiv preprint arXiv: 1310.6129
Pekarik A J, Doering Z D, Karns D A, 1999, Exploring satisfying experiences in museums
Curator: The Museum Journal 42 152173
Ratti C, Pulselli R, Williams S, Frenchman D, 2006, Mobile landscapes: using location data from
cell phones for urban analysis Environment and PlanningB: Planning and Design 33 727748
Ratti C, Sobolevsky S, Calabrese F, Andris C, Reades J, Martino M, Claxton R, Strogatz S, 2010,
Redrawing the map of Great Britain from a network of human interactions PLoS ONE 5(12):
e14248. doi:10.1371/journal.pone.0014248
Robinson E S, 1928 The Behavior of the Museum Visitor. American Association of Museums
Monograph New Series No.5 (American Association of Museums, Washington,DC)
Sanfeliu A, Llcer M R, Gramunt M D, Punsola A, Yoshimura Y, 2010, Influence of the privacy
issue in the deployment and design of networking robots in European urban areas Advanced
Robotics 24 18731899
Schuster J M, 1995, The public interest in the art museums public, in Art in Museums Ed.S Pearce
(Athlone, London) pp109142
Serrell B, 1998 Paying Attention: Visitors and Museum Exhibitions (American Association of
Museums, Washington,DC)
Shoval N, McKercher B, Birenboim A, Ng E, 2013, The application of a sequence alignment
method to the creation of typologies of tourist activity in time and space Environment and
PlanningB: Planning and Design forthcoming doi:10.1068/b38065
Sobolevsky S, Szell M, Campari R, Couronn T, Smoreda Z, Ratti R, 2013, Delineating
geographical regions with networks of human interactions in an extensive set of countries
PloSONE 8(12) e81707
Sobolevsky S, Sitko I, Grauwin S, Tachet des Combes R, Hawelka B, Arias J M, Ratti C, 2014,
Mining urban performance: scale-independent classification of cities based on individual
economic transactions, arXiv preprint arXiv: 1405.4301
Sparacino F, 2002, The Museum Wearable: real-time sensor-driven understanding of visitorss
interests for personalized visually-augmented museum experiences in Proceedings of Museums
and the Web 2002 EdsDBearman, JTrant (Archives and Museum Informatics, Tronto)
Visitors behavior in The Louvre Museum analyzed 1131

Stallings W, 2011 Cryptography and Network Security: Principles and Practice 5thedition
(Prentice-Hall, Boston,MA)
Trndle M, Greenwood S, Kirchberg V, Tschacher W, 2014 An integrative and comprehensive
methodology for studying aesthetic experience in the field: merging movement tracking,
physiology, and pschological data Environment and Behavior 46 102135
Versichele M, Neutens T, Delafontaine M, Van de Weghe N, 2012, The use of Bluetooth for
analysing spatiotemporal dynamics of human movement at mass events: a case study of the
Ghent festivities Applied Geography 32 208220
Weiss R, Boutourline S, 1963, The communication value of exhibits Museum News November,
pp2327
Yoshimura Y, Girardin F, Carrascal J P, Ratti C, Blat J, 2012, New tools for studying visitor
behaviours in museums: Acase study at the Louvre, in Information and Communication
Technologies in Tourism 2012. Proceedings of the International Conference Helsingborg
(ENTER 2012) EdsMFuchs, FRicci, LCantoni (Springer, New York) pp391402