Visitor Dynamics in a Cultural Heritage Scenario

Salvatore Cuomo

, Pasquale De Michele

1,2

, Ardelio Galletti

, Francesco Pane

and Giovanni Ponti

Department of Mathematics and Applications, University of Naples “Federico II”, Naples, Italy

UTICT-HPC, ENEA Portici Research Center, Naples, Italy

University of Naples “Parthenope”, Naples, Italy

Keywords:

Computational Neural Models, Clustering, Data Mining, User Proﬁling, Cultural Heritage.

Abstract:

We propose a biologically inspired mathematical model to simulate the personalized interactions of users with

cultural heritage objects and spaces in the real case of an exhibition. The main idea is to measure the interests

of a spectator with respect to an artwork by means of a model able to describe the users behavioural dynamics.

In our approach, the user is assimilated to a computational neuron, and its interests are deduced by counting

potential spike trains, generated by external currents. As an effort, we relies on an huge amount of log ﬁles

that store visitors movements and interactions within a beautiful art exhibition named The Beauty or the Truth

located in Naples, Italy. The technological tools deployed within the exhibition aim to create a novel metaphor

stimulating user enjoyment and knowledge diffusion and the collected log ﬁles are useful data to analyse

how such technology an inﬂuence and modify user behaviours. We also performed an experimental analysis

exploiting clustering facilities to discover natural groups that reﬂect visiting styles. This is particularly suitable

to provide the tuning of a heuristic classiﬁer. The obtained results revealed to be particularly interesting also

to understand other important aspects hidden in the data and unattended in our ﬁrst analysis.

1 INTRODUCTION

In the cultural heritage area, the requirements of inno-

vative tools and methodologies to enhance the quality

of services and to develop smart applications is an in-

creasing requirement. Cultural heritage systems con-

tain a huge amount of interrelated data that are more

complex to classify and analyse. For example, consid-

ering an art exhibition, characterizing, studying, and

measuring the level of knowledge of a visitor with re-

spect to an artwork, and also the dynamics of social

interaction on a relationship network is an interesting

research scenario. To understand and analyse how art-

works observation can inﬂuence the social behaviours

is a very hard challenges. Indeed, semantic web ap-

proaches have been increasingly used to organize dif-

ferent art collections not only to infer information

about a cultural item, but also to browse, visualize,

and recommend objects across heterogeneous collec-

tions (Middleton et al., 2003). Other methods are

based on statistical analysis of user datasets in order

to identify common paths (i.e., patterns) in the avail-

able information. Here, the main difﬁculty is the man-

agement and retrieval of large databases as well as is-

sues of privacy and professional ethics (Kumar et al.,

2010). Finally, models of artiﬁcial neural networks,

typical of Artiﬁcial Intelligence ﬁeld are also adopted.

Unfortunately, these approaches seems to be, in gen-

eral, too restrictive in describing complex dynamics

of social behaviours and interactions in the Cultural

Heritage framework (Kleinberg, 2008). In this paper,

we refer to a computational neuroscience terminol-

ogy for which a cultural asset visitor is a neuron and

its interest is the electrical activity which has been

stimulated by appropriate currents. More speciﬁcally,

the dynamics of the information ﬂows, which are the

social knowledge, are characterized by neural inter-

actions in biological inspired neural networks. Rea-

soning by similarity, the users can be considered as

neurons in a network and their interests the morphol-

ogy; the common topics among users are the neuronal

synapses; the social knowledge is the electrical activ-

ity in terms of quantitative and qualitative neuronal

responses (spikes). In this context, several works pro-

posed a detailed comparative analysis in order to dis-

cover a reliable strategy to tune the model parameters.

In general, two different strategies can be adopted to

discover data groups: a Bayesian classiﬁer (Cuomo

337

Cuomo S., De Michele P., Galletti A., Pane F. and Ponti G..

Visitor Dynamics in a Cultural Heritage Scenario.

DOI: 10.5220/0005579603370343

In Proceedings of 4th International Conference on Data Management Technologies and Applications (KomIS-2015), pages 337-343

ISBN: 978-989-758-103-8

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

et al., 2014a) and an approach that ﬁnds data group-

ings in an unsupervised way (Cuomo et al., 2014b).

Such a strategy resorts to a clustering task employing

the well-known K-means algorithm (Jain and Dubes,

1988). Here we deal with the characterization of user

dynamics and behaviours starting from real datasets.

As a real scenario we have considered the art exhibi-

tion named the Beauty or the Truth located in Naples,

Italy, where new ICT tools and methodologies, pro-

ducing several users behavioural data, have been de-

ployed and currently are still active. Our aim is also to

classify visitors in the exhibit by using data collected

by the available technology. These data are used as

input current of the discussed model. The paper is or-

ganized as follows. In Section 2 we discuss the moti-

vation example. In Section 3 we describe the cultural

heritage information system. The Section 4 is devoted

to the experiments. Finally, the conclusions are drawn

in the Section 5.

2 MOTIVATION EXAMPLE

In order to better understand motivations behind this

work, it is important to deeply analyse the kind of

relations that exists between cultural spaces, people

and technological tools that nowadays are pervasive

in such environments. Accordingly, the behaviour of

a person/visitor, when immersed inside a space and

consequently among several objects, has to be anal-

ysed in order to design the most appropriate ICT ar-

chitecture and to establish the relationship between

people and technological tools that have to be non-

invasive. For this reason, it should be preferable to

provide cultural objects with the capability to interact

with people, environments, other objects and trans-

mitting the related knowledge to users through multi-

media facilities. In an intelligent cultural space, tech-

nologies must be able to connect the physical world

with the world of information in order to amplify the

knowledge but also and especially the fruition, involv-

ing the visitors as active players which offer the plea-

sure of perception and the charm of the discovery of

a new knowledge. In the follow, the architecture of

an Internet of Things (IoT) system, the technological

sensors immersed in the cultural environment and the

communication framework are presented. The sen-

sors aimed to transform cultural items in smart ob-

jects, that now are able to communicate with each

other, the visitors and the network; this acquired iden-

tity plays a crucial role for the smartness of a cul-

tural space. Accordingly, in order that this system can

perform its role and improve end-users cultural expe-

rience transferring knowledge and supporting them,

a mobile application has been designed; in this way

people have the opportunity to enjoy the cultural visit

and be more at ease simply using their own mobile

device. Furthermore, we present an interesting and

wide case study; it consists of a real art exhibition of

271 sculptures, divided into 7 thematic sections and

named “The Beauty or the Truth”

. This exhibition

shows, for the ﬁrst time in Italy, the Neapolitan sculp-

ture of the late nineteenth century and early twentieth

century, through the major sculptors of the time. The

sculptures are exhibited in the beautiful monumental

complex of San Domenico Maggiore, in the historical

centre of Naples. The proposed IoT system was en-

tirely deployed inside the exhibition, as illustrated in

Figure 1.

3 THE CULTURAL HERITAGE

INFORMATION SYSTEM

The overall data collected by the described ICT

framework will be used as the input of our compu-

tational model. In particular the LOG ﬁles are struc-

tured in order to store main informations about the

visitor behaviour in the exhibit. The following listing

shows the JSON schema diagram of a log ﬁle, charac-

terized by the fruition information w.r.t. the artworks.

1 {

2 " USER " : {

3 " SESSI O N " : [

4 {

5 " S T A R T _ S E S SI O N " : " 29 / 1 2/ 2 0 14 1 6 :1 5

:1 0 : 54 0 " ,

6 " IDU S E R " : " fdb 9 e ab 8 19 aa 3 f7 9 1 41 9 8 66

07 0 4 61 " ,

7 " PLATFOR M ": " A n d r o id " ,

8 " S T A R T _ L A N G " :" it ",

9 " IDEVENT O ": 1 0 ,

10 " R A N K _ E V E N T O ": 2 . 5 ,

11 " N U M B E R _ O P E R A " : " 2 7 1 " ,

12 " T R A N S A C T I O N ": [

13 {

14 " REQUE S T " :{

15 " R E Q U E ST_P A R A M E TE RS " : {

16 " C O D E C R I C K E T ": " [ 1 0 0 0 & Cr i c k e t 0

01 9 & ]" ,

17 " C ODEOPERA " : "1 2 8 ",

18 " DA TE " : " 2 9 / 1 2 /2 0 1 4 1 6 : 15 : 5 1: 8

18 " ,

19 " LANGUAG E ": " it "

20 }

21 },

22 " P A R A M E T ER S_ LO G " : {

23 " R ANK_TEXT " : 2. 5 ,

24 " R A N K _ A U D I O " :2 . 5 ,

http://www.ilbellooilvero.it

DATA2015-4thInternationalConferenceonDataManagementTechnologiesandApplications

338

Figure 1: Exhibit map.

25 " R A N K _ G A L L E R Y " : 2 . 5 ,

26 " ACT I O N " : [

27 {

28 " TY PE " : " A U DIO ",

29 " ID " :" 7 3 4" ,

30 " H O U R _ S T A R T " :" 2 9 /1 2 / 2 0 1 4 1 6:

16 : 0 0: 4 4 8" ,

31 " HOUR_EN D ": " 2 9/ 1 2 /2 0 1 4 1 6 : 16

:1 3 : 97 3 " ,

32 " TOT " :" 3 2 .9 2 2 "

33 },

34 {

35 " TY PE " : " I M AGE ",

36 " ID " : "/ M edia / 79 / 1 28 / I mag e / 20

14 -1 0 - 0 8 _ 1 7 -5 0 - 3 4 . j pg "

37 },

38 ...

39 ],

40 " IMG_SIZ E ": 3 ,

41 " A U D I O _ S I Z E " :2 ,

42 " H O U R _ OP ER A_STA R T ": " 2 9 /1 2 / 20 1 4

16 : 1 5: 5 1 :8 1 9 ",

43 " H O U R _ O PE RA _STOP " :" 2 9 / 12 / 2 01 4

16 : 1 6: 1 4 :1 5 "

44 }

45 },

46 . ..

47 }

We can observe, from JSON exam-

ple, that the visitor has viewed the image

2014-10-08 17-50-34.jpg associated to an

artwork. We notice that the exhibition The Beauty or

the Truth is still open and in this paper we analyse

over than 200 log ﬁles corresponding to the same

number of users that enjoy the available technological

instruments within such cultural space.

Respect to visitor classiﬁcation, we start from (Zan-

canaro et al., 2007) where personalized information

presentation in the context of mobile museum guides

are reported. In (Zancanaro et al., 2007) is shown that

visitor movements are compared to the behaviour

of four typical animals. In our work, we adapt this

classiﬁcation to ﬁnd how visitors interact with the

ICT technology and how a lot they are interested in

the exhibition. The visitor can be assimilated to:

• an ANT (A), if it tends to follow a speciﬁc path

in the exhibit and spends a lot of time using the

furnished technology;

• a FISH (F), if it moves around in the centre of the

room and usually avoids looking at media content

details;

• a BUTTERFLY (B), if it does not follow a speciﬁc

path but rather is guided by the physical orienta-

tion of the exhibits and stops frequently to look

for more media content;

• a GRASSHOPPER (G), if it seems to have a spe-

ciﬁc preference for some preselected artworks and

VisitorDynamicsinaCulturalHeritageScenario

339

spends a lot of time observing the related media

contents.

In Section 4, we will deﬁne a model that, starting from

data in the JSON ﬁle, is able to classify the user and

to predict its interest on an artwork or on the overall

exhibit.

4 USER BEHAVIOUR

REPRODUCTION AND DATA

MINING

The experiments described in this Section were car-

ried out from a dataset of 253 regular visitors, and

were performed on CRESCO HPC system (Bracco

et al., 2009), integrated into the ENEA-GRID infras-

tructure, and located in the ENEA Portici Research

Center.

. We have tracked the visitor behaviour by us-

ing a suitable Extrapolation Algorithm (EA) that has

the JSON ﬁle as input data. A typical EA output is

shown in the following:

IDUser : e7a5774700c1e88e1417618582735

# of artworks: 271

# of viewed artworks: 44

% of viewed artworks : 17.5%

...

-------------------------------------

i-th viewed artwork : 2

ID artwork : 128

Available audio (sec.) : 32.922

Listen audio (sec.) : 32.922

Available images : 3

Viewed images : 0

Available text : True

Viewed text : False

Interaction time (sec.) : 58.259

Path is followed : True

-------------------------------------

...

-------------------------------------

i-th viewed artwork : 6

ID artwork : 17

Available audio (sec.) : 85.141

Listen audio (sec.) : 85.141

Available images : 4

Viewed images : 2

Available text : True

Viewed text : True

Interaction time (sec.) : 103.141

Path is followed : False

-------------------------------------

Such ﬁles are particularly suitable to identify

users’ behaviour not only regarding their interactions

http://www.cresco.enea.it

with artworks, but also w.r.t. the whole artwork ex-

hibition. In fact, properly looking at the JSON ﬁles,

for each user it is possible to determine if the exhibi-

tion path followed, the sequence of visited sections,

the time spent to enjoy audio and images contents,

and if text information about a speciﬁc artwork are

visualized or not. It is easy to note that such a set of

information are useful to produce a detailed dataset,

which is most enriched w.r.t. the one exploited in an

our previous work.

This new dataset is suitable for the next step of

our analysis, which consists in executing an unsu-

pervised data mining algorithm in order to achieve

data groups that can reﬂect the user classiﬁcation de-

scribed in the previous paragraph. In this direction,

we propose a data structure containing not only user

interaction with the artwork, but also the indication of

how much he follows the path suggested by the exhi-

bition. The dataset structure in ARFF Weka format is

shown in the follow.

@RELATION ARTWORKS

@ATTRIBUTE audios NUMERIC [0..1]

@ATTRIBUTE images NUMERIC [0..1]

@ATTRIBUTE texts NUMERIC [0..1]

@ATTRIBUTE paths NUMERIC [0..1]

@ATTRIBUTE class {A,B,F,G}

@DATA

1,0.190476,0.190476,0.592593,G

0,0,0,0,F

...

0.84131,0.342857,0.114286,0.765432,?

...

0.573139,0.325581,0.697674,0.481481,G

The dataset contains, in its original version,

boolean values for some ﬁelds, such as paths and

text. However, it is well known in data mining liter-

ature that boolean ﬁelds are particularly critical for al-

gorithms, that may produce inaccurate results in such

cases. To avoid this problem, we propose a strategy

to transform binary ﬁelds into continuous ones. For

each binary ﬁeld, we take into account the ratio be-

tween the number of beneﬁted elements and the total

number of elements, in order to have a percentage of

beneﬁt. A different consideration should be done for

the paths ﬁeld: to obtain a percentage of following

path for a user, we counts the number of visited sec-

tions that are strictly consecutive, taking into account

only close consecutive section in an increasing order,

and perform the ratio of this count with the total num-

ber of sections.

Let we make an example. Assume that a user fol-

lows this path:

1 → 4 → 5 → 2 → 3 → 6 → 7

DATA2015-4thInternationalConferenceonDataManagementTechnologiesandApplications

340

The count here is 4, as we have 1 → 4 and 3 → 6

are consecutive but not strictly consecutive (the con-

tribute is 0.5 for each of these), while 4 → 5, 2 → 3

and 6 → 7 as strictly consecutive paths (the contribute

is 1 for each of these). Moreover, 5 → 2 are not con-

secutive (the contribute is −1), as the user goes back

to section 2 after visited the section 5). Note that the

count starts from 1. Hence, the percentage of fol-

lowed path is expressed as the ratio

count

#o f sections

= 0.571428

Starting from the data collected in the exhibit, we

classify the visiting behaviours by means of the math-

ematical model and some heuristics. More in detail,

we assume that

• Fs enjoy of almost the 7.5% of the overall media

contents and has a “small” number of spikes;

• Bs enjoy of at least the 50% of the overall media

contents, does not follow a speciﬁc path and has a

“medium” number of spikes uniformly distributed

w.r.t. the artworks.

• Gs enjoy of almost the 50% of the overall me-

dia contents, follows a speciﬁc path and has

a “medium” number of spikes focused only on

some artworks in the exhibit.

• As enjoy of at least the 70% of the overall media

contents, follows a speciﬁc path and has a “large”

number of spikes.

We exploit such a modiﬁed dataset for our exper-

iments. In a ﬁrst phase, we execute the Expectation-

Maximization (EM) algorithm to discover the best

number of clusters for our dataset. The algorithm pro-

duces K = 2 as the number of classes, which indicates

that only two of the four categories described before

are present in our data. With this input, we resorted

to the well-known K-means partitional clustering al-

gorithm (Jain and Dubes, 1988) and set the number of

classes to K = 2. Experiments underline that the two

categories in our data have been correctly identiﬁed

with an accuracy of almost 73%.

Investigating more deeply at the dataset, we have

known that they are data from the overall exhibition,

and the two behaviours present are F (i.e., ﬁsh), and

G (i.e., grasshopper). This means that users are not

typically interested in exploiting multimedia contents

in all the sections. In fact, in order to have a proof

for our intuition, we decided to take into account only

data from the ﬁrst two sections of the exhibition. We

discovered here all the four classes, and this indicates

that a user typically enjoys multimedia support only

in the ﬁrst phase of his visit.

In Table 1 we report the results of the clustering

(with K = 4) for the ﬁrst two sections of the exhibi-

tion. Note that Cluster0 corresponds to A, Cluster1

is F, Cluster2 represents G and Cluster3 is B, as this

is a typical majority voting based cluster assignment.

We achieved a slight improvement in terms of accu-

racy results (almost 81%). However, this clustering

session provides very interesting intuitions which can

be seen in the table. In fact, the behaviour labelled

with G is the most ambiguous one, as the grasshopper

is very difﬁcult to characterize. Moreover, regarding

A, these tuples are splitted mainly between two clus-

ters, i.e., Cluster0 and Cluster3, but the cluster la-

belling is negatively affected by G tuples. Finally, F

and B are well grouped in their clusters, with ∼ 85%

and ∼ 84% of accuracy, respectively.

Figure 2 shows the cluster assignments for the tu-

ples in the dataset. Tuples are coloured by the class

attribute, whereas on the axes there are class-ID and

cluster-ID.

Figure 2: Kmeans cluster assignment (K = 4).

We can see in the ﬁgure what saw before regard-

ing the difﬁculty of identifying with high accuracy the

behaviour of some classes, i.e., G (in cyan) and A (in

blue): they are mainly assigned to the right cluster,

but lots of the remaining instances are also present in

the other clusters.

Figure 3: Visitor Metamorphosis.

In Figure 3 we report the visiting styles in the sec-

tions 1 − 2 (red columns), the sections 1 − 4 (green

VisitorDynamicsinaCulturalHeritageScenario

341

Table 1: Results of the clustering for K = 4 for the ﬁrst two sections of the exhibition.

Animals Cluster0 Cluster1 Cluster2 Cluster3

A 14 11 4 22

B 4 0 1 27

F 7 72 4 2

G 16 9 30 30

columns) and sections 1 − 7 (blue columns) of the

cultural heritage event. For the ﬁrst 2 sections, we ob-

serve that only the ∼ 7.5% of the users are As, while

the ∼ 12.5% are classiﬁed as Bs. Moreover, the users

classiﬁed as Fs and Gs are the same amount of the

∼ 33%. The remaining ∼ 12.5% of users are not clas-

siﬁed. Furthermore, from the section 1 to the section

4 we note that all the users become Fs (∼ 63%) or

Gs (∼ 37%). Finally, by observing the blue columns

of the histogram in Figure 3, from the section 1 to the

section 7 (i.e., the entire exhibition) we note that there

is an adjustment of these metamorphosis of the users

in F (∼ 69.5%) and G (∼ 30.5%).

From these experiments we deduce that as the

time spent in the exhibit grows, visitors choose to not

use the available technology in an intensive way. We

deduce that if the event had lasted only 2 sections,

54% of users would continue to use the supplied tech-

nology, instead of the 30.5% on the entire exhibition.

5 CONCLUSIONS

In this paper we have described a framework that re-

ﬂects the computational methodology adopted to infer

information about visitors in a cultural heritage con-

text. Our challenge is to match, in a realistic way, the

biological morphology of a neuron and its behaviour

in this application scenario. In the model we propose,

the (R, C) couple represents the sensitivity of an user

respect to an artwork. Accordingly, we compared two

different strategies for tuning model parameters in or-

der to ﬁnd an accurate approach that is able to pro-

vide the best setting for the neuronal model. In this

respect, we shown experimental results for standard

Bayesian classiﬁer and a novel clustering methodol-

ogy to obtain starting groups from which these elec-

trical parameters can be tuned. From our experiments,

it has been highlighted that clustering task is able to

produce a more accurate setting.

Starting from the state-of-art about the Museum

visitors’ behaviour patterns, we have investigated how

the use of technological tools within cultural spaces

can affect visitors’ behaviour, causing behavioural

changes also during the same visit. In this particular

case, we have analysed such behaviour modiﬁcation,

introducing the concept of metamorphosis and show-

ing the analysis results in visitors’ styles.

An interesting observation and challenge for fu-

ture works is to adapt, in a smart way, this computa-

tional framework to many different application topics,

such as the context-aware proﬁling, feedback based

and/or recommendation systems.

ACKNOWLEDGEMENTS

Authors thank DATABENC, a High Technology Dis-

trict for Cultural Heritage management of Regione

Campania (Italy), and ENEA Portici Research Center,

UTICT-HPC Department, for supporting the paper.

REFERENCES

Bracco, G., Podda, S., Migliori, S., D’Angelo, P., Quintil-

iani, A., Giammattei, D., De Rosa, M., Pierattini, S.,

Furini, G., Guadagni, R., Simoni, F., Perozziello, A.,

De Gaetano, A., Pecoraro, S., Santoro, A., Scio’, C.,

Rocchi, A., Funel, A., Raia, S., Aprea, G., Ferrara, U.,

Novi, D., and Guarnieri, G. (2009). CRESCO HPC

System Integrated into ENEA-GRID Environment. In

Proc. of the Final Workshop of the Grid Projects of

the Italian National Operational Program 2000-2006

– Call 1575, pages 151–155.

Cuomo, S., De Michele, P., and M., P. (2014a). A biologi-

cally inspired model for describing the user behaviors

in a cultural heritage environment. In Proceedings

- Advanced Database Systems (SEBD), 22nd Italian

Symposium on, pages 292–302.

Cuomo, S., De Michele, P., Ponti, G., and M., P. (2014b).

A clustering-based approach for a ﬁnest biological

model generation describing visitor behaviours in a

cultural heritage scenario. In Proceedings - Data

Management Technologies and Applications (DATA),

3rd International Conference on, pages 427–433.

Jain, A. and Dubes, R. (1988). Algorithms for Clustering

Data. Prentice-Hall.

Kleinberg, J. (2008). The convergence of social and tech-

nological networks. Commun. ACM, 51(11):66–72.

Kumar, R., Novak, J., and Tomkins, A. (2010). Link Min-

ing: Models, Algorithms, and Applications, chapter

Structure and Evolution of Online Social Network,

pages 337–357. J. Am. Soc. Inf. Sci. Technol.

DATA2015-4thInternationalConferenceonDataManagementTechnologiesandApplications

342

Middleton, S., Shadbolt, N., and De Roure, D. (2003). Cap-

turing interest through inference and visualization:

Ontological user proﬁling in recommender systems.

In Proceedings - Knowledge Capture, 2Nd Interna-

tional Conference on, pages 62–69.

Zancanaro, M., Kuﬂik, T., Boger, Z., Goren-Bar, D., and

Goldwasser, D. (2007). Analyzing museum visitors’

behavior patterns. User Modeling 2007 - Lecture Note

in Computer Science, 4511:238–246.

VisitorDynamicsinaCulturalHeritageScenario

343