Improving Enjoyment of Cultural Heritage Through Recommender
Systems, Virtual Tour, and Digital Storytelling
M. Casillo
1 a
, F. Colace
2 b
, A. Lorusso
2 c
, D. Santaniello
1 d
and C. Valentino
2 e
1
DISPAC, University of Salerno, Giovanni Paolo II, 132 - 84084 Fisciano (SA), Italy
2
DIIN, University of Salerno, Giovanni Paolo II, 132 - 84084 Fisciano (SA), Italy
Keywords:
Recommender Systems, Virtual Tour, Situation Awareness, Ontology, Bayesian Network.
Abstract:
The integration of Information and Communication Technologies (ICT) within the world of cultural heritage
has the role of added value for its enhancement. In particular, improving the enjoyment of cultural Points of
Interest by suggesting personalized routes allows for better interaction between users and the cultural site. To
this end, this paper aims to introduce an architecture that, by employing Recommendation Systems integrated
with the Situation Awareness paradigm, allows for the identification of personalized paths for users through
the acquisition of data through smart sensors, which is then processed through the proposed approach, defined
as a Multilevel Graph (MuG) approach. This aims to filter through the data’s context and ontological lay-
ers to its processing through the Bayesian network, which is identified through structural learning algorithms
integrated with the domain’s semantic knowledge. The architecture also incorporates physical and virtual ex-
periences, exploiting the advantages of virtual tours and involving users more by employing digital storytelling
techniques. Testing of the proposed architecture based on the MuG approach took place through an offline
experiment aimed at evaluating the accuracy of the approach used and an online experiment to test the validity
of the designed architecture.
1 INTRODUCTION
Enhancing artistic and cultural heritage involves
strategies that enable both the preservation of the her-
itage and the improvement of visitor enjoyment (Bru-
mana et al., 2023).
This task requires the support of new technolo-
gies, particularly in Information and Communication
Technologies (ICT). In the case of preservation, these
technologies must provide the necessary tools for
monitoring. Through the data collected, it is essen-
tial to identify data processing strategies to prevent fu-
ture damage (Garc
´
ıa-Valldecabres et al., 2021). Con-
versely, in the case of improving the interaction be-
tween artistic and cultural heritage and users, ICTs
must provide the necessary data to enable users to
adapt the cultural experience and develop person-
alization strategies to tailor the cultural experience
a
https://orcid.org/0000-0003-0609-3781
b
https://orcid.org/0000-0003-2798-5834
c
https://orcid.org/0000-0002-0831-5694
d
https://orcid.org/0000-0002-5783-1847
e
https://orcid.org/0000-0001-9964-1104
to both the user and the environmental condition in
which the visit takes place (Ruotsalo et al., 2013).
To improve the enjoyment of cultural visitors, an-
alyzing and filtering data that will allow systems to
understand user preferences and adapt the cultural
experience accordingly is necessary. Specific tools
called Recommendation Systems (Colace et al., 2022;
Ricci et al., 2022), applied historically in tourism and
particularly useful in the cultural heritage sector, are
also exploited (Borr
`
as et al., 2014). Understanding
users’ preferences makes it possible to suggest the
most suitable points of interest (POIs) for the user. In
this field, various strategies allow Recommendation
Systems (RSs) to suggest the best content to cultural
users. Historically, the main strategies are Content-
Based, Collaborative Filtering, and Hybrid RSs (Ricci
et al., 2022). Over time, these have hybridized with
machine and deep learning techniques (Nikolakopou-
los et al., 2022; Zhang et al., 2022), but they preserve
the main features. Content-based aims to acquire data
about users and the products, usually called items to
be suggested. Thus, they aim to assess the affinity
between users and items (Musto et al., 2022). In con-
trast, Collaborative Filtering exploits interactions be-
Casillo, M., Colace, F., Lorusso, A., Santaniello, D. and Valentino, C.
Improving Enjoyment of Cultural Heritage Through Recommender Systems, Virtual Tour, and Digital Storytelling.
DOI: 10.5220/0013176000003905
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 14th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2025), pages 263-271
ISBN: 978-989-758-730-6; ISSN: 2184-4313
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
263

tween users and items, usually ratings, to understand
what users prefer to make predictions about items
with which it has not interacted (Koren et al., 2022).
Finally, hybrid approaches allow two or more strate-
gies to be integrated to overcome the problems of
individual recommendation approaches (Ricci et al.,
2022).
These tools integrate additional data to improve
their predictive ability in addition to recommendation
strategies. This is the case with Context-Aware Rec-
ommendation Systems (CARSs), which use Context
Awareness to adapt predictions to the conditions un-
der which the system processes data and suggests the
best POIs to users (Adomavicius et al., 2022; Co-
lace et al., 2023). The integration of Context Aware-
ness within RSs allows its acquisition in the percep-
tion phase, which directly links to the hint predic-
tion phase. However, to enable the Recommenda-
tion System to integrate into the Situation Awareness
paradigm, integrating an understanding phase before
the prediction phase is necessary (Endsley, 1995). It
is possible to incorporate ontological layers related to
domain knowledge to acquire knowledge from data
(Casillo et al., 2022; Musto et al., 2022).
This work aims to introduce an architecture for
enhancing artistic and cultural heritage by improving
the enjoyment of cultural users by defining a recom-
mendation system that integrates the Situation Aware-
ness paradigm within an RS. The recommendation
approach used takes advantage of a Multilevel Graph
approach that leverages three graph approaches: the
Context Dimension Tree (CDT) that enables context
management, a Domain Ontology that allows the ac-
quisition of knowledge about the data through seman-
tic data analysis, and finally, a Bayesian Network that
enables the prediction phase (Casillo et al., 2024).
The structure of the network is done through struc-
tural learning algorithms integrated with the filtering
phase through CDT and Ontology, which allow for a
more reliable construction of the Bayesian Network.
Coupled with the prediction phase via the Mul-
tilevel Graph (MuG) approach, it is necessary to in-
tegrate supporting techniques that aim to improve
the user’s cultural experience. To this end, the pro-
posed architecture leverages Digital Storytelling to
engage users through narratives centered on the path
suggested to users through the RS (Podara et al.,
2021; Selmanovi
´
c et al., 2020). In addition, the pro-
posed experiences will be hybrid through virtual tours
based on 360
◦
photos (Argyriou et al., 2020). Vir-
tual tours using 360
◦
photos allow users to create im-
mersive experiences of physical spaces by enabling
users to explore real environments through interactive
panoramic images. Virtual 360
◦
tours are created by
taking a series of panoramic photographs with a 360
◦
camera, thus requiring devices capable of capturing
360-degree images or video in all directions. Once the
360
◦
images have been captured, they must be pro-
cessed using software to create the interactive tour.
The resulting virtual environment can be freely ex-
plored through navigation points, defined as hotspots,
creating a continuous and fluid experience. The main
advantages of using 360
◦
Tours include providing an
immersive and realistic experience that relies on a
generally intuitive and easy-to-navigate interface. In
addition, integrating virtual tours allows for improved
accessibility of suggested POIs.
The paper is structured as follows: Section 2 de-
scribes the literature related to Recommender Sys-
tems and the use of Tour 360
◦
aimed at application in
the field of cultural heritage; Section 3 introduces the
proposed architecture and describes the proposed ap-
proach leveraged as a Recommender System that in-
tegrates Situation Awareness to suggest cultural paths
to users; Section 4 presents the experimental results
aimed at evaluating the introduced recommender ap-
proach and analyzing the effectiveness of the pro-
posed architecture through the development of a pro-
totype; finally, Section 5 describes the conclusion and
future work.
2 RELATED WORKS
Recommendation Systems employed in cultural her-
itage have a tradition stemming from using these tools
in tourism (Casillo et al., 2023).
(Cramer et al., 2008) investigate the impact of
transparency on user trust and acceptance of content-
based recommender systems in the context of cul-
tural heritage. The study explores whether providing
explanations for recommendations or showing how
confident the system is in its recommendations af-
fects users’ trust and acceptance. The authors exper-
imented with several groups of 60 participants, who
interacted with three versions of an artwork recom-
mendation system. This system proposes works based
on users’ ratings of other works of art. The three ver-
sions tested were a non-transparent system that did
not explain the recommendations, a transparent sys-
tem that explained why a recommendation was made,
and a system that showed how safe the system was
in giving the recommendation. The results show that
explaining the recommendations increased their ac-
ceptance but did not significantly improve confidence
in the system. Showing the system’s level of cer-
tainty influenced neither confidence nor acceptance.
The study derives some guidelines for the design of
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
264

recommendation systems in the cultural heritage do-
main.
In (De Gemmis et al., 2008), the authors explore
the integration of tags within a content-based recom-
mendation system. Combining static object descrip-
tions with dynamic user-generated tags aims to im-
prove content personalization, especially in the con-
text of cultural heritage. The study uses natural lan-
guage processing techniques to analyze and disam-
biguate tag meanings, transforming them into seman-
tic concepts that can be used to create more accurate
user profiles. These profiles are then used to improve
the recommendations provided to users. The paper
describes the recommendation process in three main
steps: content analysis, user profile learning, and rec-
ommendation generation. It focuses mainly on how
personal and social tags can improve the accuracy
of recommendations. Preliminary results show that
including tags increases recommendations’ accuracy
compared to using only static descriptions.
(Su et al., 2019) explore using artificial intel-
ligence based on Edge Computing technologies to
create an advanced recommendation system in the
context of cultural heritage. The proposed system
combines artificial intelligence techniques, Big Data,
and personalized recommendations to improve vis-
itors’ experience of museums, archaeological sites,
and other cultural spaces. The system is based on a
multilevel architecture that uses Big Data infrastruc-
ture to manage a large amount of heterogeneous data
from different sources, such as social media, digital
libraries, and environmental sensors. A crucial part
of the system is the Smart Search Museum mobile
application, which provides personalized suggestions
about museums and other cultural attractions based
on contextual recommendation techniques and arti-
ficial intelligence built into users’ devices. In addi-
tion, the authors describe how integrating different
data sources can provide personalized tourist routes
and recommendations of cultural objects, thereby op-
timizing the artistic experience.
In (Hong et al., 2017), the authors describe a so-
cial recommendation system for applications in cul-
tural heritage. It focuses on using digital technolo-
gies and social services to improve people’s interac-
tion with cultural spaces, make them dynamic, and
promote the discovery and sharing of new knowledge.
The proposed system uses social affinity among users
to provide personalized recommendations of artworks
and cultural content based on the characteristics of
the artworks and users’ experiences. An architecture
for managing group recommendations is also pro-
posed, addressing issues such as the scarcity of data
on user preferences and the sustainability of recom-
mendations in cultural contexts. It also describes how
using social networks and contextual information can
improve recommendations and presents an innovative
approach to calculating and exploiting social affinity
among users for group recommendations.
In addition to techniques for personalizing ser-
vices, 360
◦
virtual tours have become an innovative
tool for enjoying cultural heritage. Virtual tours can
improve accessibility to historic sites, allowing re-
mote and interactive explorations enhanced by mul-
timedia content such as text and video.
In (Valtolina et al., 2006), the authors describe
a system facilitating personalized access to cultural
heritage distributed across multiple museums. The
system meets the needs of two categories of users:
visitors, who can access information tailored to their
interests and interaction preferences, and domain ex-
perts, such as museum curators, who can create the-
matic pathways to provide a better understanding of
artifacts. The approach relies on a semantic represen-
tation of cultural heritage to build customizable vi-
sual interfaces called ”Virtual Wings” (VWs) that al-
low users to navigate through digital archives and the-
matic pathways, creating personalized virtual tours.
A practical example includes integrating customized
digital guides and 360
◦
panoramic images.
The authors of (Gunawan and Lesmana, 2023) de-
velop a 360
◦
virtual tour of the Dharma Rakhita Tem-
ple in Jamblang village as an educational tool to pro-
mote heritage knowledge. The temple, which is more
than 200 years old and rich in history, is little visited
due to its isolated location and poor tourism promo-
tion. The study proposes using 360
◦
images captured
with digital technologies to create a virtual tour ac-
cessible through web applications, allowing the pub-
lic to explore the temple remotely. The main goal is
to preserve the temple and use it as a source of cul-
tural learning by enhancing storytelling and historical
documentation through 360
◦
filming techniques.
3 THE PROPOSED
ARCHITECTURE
This section introduces the proposed architecture for
enhancing artistic and cultural heritage through im-
proving user enjoyment. The architecture aims to sug-
gest personalized paths for users to tailor the cultural
experience to visitors’ preferences. For added value,
the paths are integrated with digital storytelling tech-
niques that enhance engagement through storytelling
and exploit multimedia content. In addition, the pro-
posed experiences are hybrid as, depending on the
users’ needs and the specific context, suggested routes
Improving Enjoyment of Cultural Heritage Through Recommender Systems, Virtual Tour, and Digital Storytelling
265

integrate physical and virtual POI visits through 360
◦
tours.
The proposed architecture, illustrated in Figure
1, consists of four functional layers: the Acquisition
Layer, the Knowledge Base Layer, the Inference En-
gine Layer, and the Application Layer.
The acquisition layer consists of all the devices
that enable the acquisition of environmental data and
contextualize the suggestions that will be processed
during the data processing phase (Michalakis and
Caridakis, 2022). In this layer, through the Internet of
Things paradigm, it is possible to exploit sensors that,
once they have acquired the data, can communicate
it via the Internet using the MQTT protocol. Specif-
ically, the sensors used involve environmental moni-
toring. A weather station is installed to acquire tem-
perature, humidity, air quality, pressure, and atmo-
spheric precipitation data (Colace et al., 2018; Mitro
et al., 2022). In addition, user location data must
be acquired, and sensors must be installed to moni-
tor the level of crowding at each POI. Data acquired
through sensors should be integrated with API and
open-source data services that allow both the integra-
tion of the acquired data and the acquisition of data
beneficial for the architecture as multimedia content,
in addition to the content built specifically for the case
study and already used.
Once the data has been acquired, it must be stored
through the Knowledge-Base Layer. The raw data
collected via sensors and external services is initially
cleaned to detect possible missing or anomalous data
through a pre-processing module. Once cleansed, the
data are stored within the database, which must han-
dle structured and semi-structured data. In addition,
the pre-processing module must also work as a filter
for the data flow from the database to the Inference
Engine Layer to do the cleaning and preparation work
necessary for the actual processing of the data.
The third layer of the architecture, the Inference
Engine Layer, represents the beating heart of the ar-
chitecture. Here, data is processed to provide services
to users to improve their cultural experience. Four
modules are available here: the Experience Elabora-
tion Module, the Recommendation Module, the Con-
tent Selection Module, and the Digital Storytelling
Module.
The Recommendation Module leverages the Mul-
tilevel Graph (MuG) approach (Casillo et al., 2024),
which will be described in more detail later, and aims
to integrate the Situation Awareness paradigm within
the recommendation engine. For this purpose, it is
necessary to explain in detail how Endsley’s three lev-
els of awareness are obtained (Endsley, 1995).
The system achieves the perception stage from the
data acquisition stage to data storage, including the
pre-processing stages. Then, through the context fil-
tering that takes place through the Context Dimension
Tree (CDT) and the semantic filtering phase through
the Domain Ontology, the comprehension phase is
obtained (Kokar et al., 2009), which will allow the
predictions to be adapted both based on the ontology’s
domain knowledge and based on the environmental
conditions in which the system operates through the
CDT. In addition, the understanding phase permits the
integration of the application phase of the structural
learning algorithm that allows for the Bayesian Net-
work (BN) structure (Scanagatta et al., 2019). How-
ever, this is identified through the data, supplemented
by the filter represented by the Ontology and the con-
textual analysis performed. As a result, the BN en-
ables the prediction phase that aims to suggest cul-
tural paths to users based on gold preferences and
contextual conditions (Casillo et al., 2024; Scanagatta
et al., 2019). The schematic of how the MuG ap-
proach falls within the Situation Awareness paradigm
is described in Figure 2.
From the description given, it follows that the
Recommendation Module initially requires an offline
phase in which the BN is detected. In this phase,
the acquired data filtered through the CDT and Do-
main Ontology are exploited. This allows for identi-
fying constraints that force the identification of prob-
abilistic dependencies among the BN nodes or avoid
the identification of inconsistent dependencies. Such
constraints complement the applied structural learn-
ing algorithm and enable the identification of a more
reliable network structure. Once the offline phase is
completed, the network allows predictions to be ob-
tained during the online phase. In addition, new data
can periodically update the network.
In addition to the Recommendation Module, the
Inference Engine Layer exploits the Experience Elab-
oration Module, the Content Selection Module, and
the Digital Storytelling Module. Once the path to be
suggested to the user has been identified, the Experi-
ence Elaboration Module selects based on the context
related to POI crowding and the time available to the
user which POIs of the identified path are to present in
virtual form and which ones are not. Then, based on
this distinction, the Digital Storytelling Module com-
poses the narrative by assembling the available tex-
tual content and linking them appropriately accord-
ing to the strategy of non-linear Digital Storytelling.
Once the narrative has been identified, the Content
Selection Module identifies the multimedia content to
be provided to the users, which is distinguished by
whether the POI will be visited physically or virtu-
ally.
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
266

Figure 1: The proposed architecture designed to improve the cultural experience of users.
Figure 2: Scheme of the Multilevel Graph approach contex-
tualized in the Situation Awareness paradigm.
Finally, the architecture presents the Application
Layer, which provides services to the user. This layer
allows access to the custom path elaborated for the
user based on his preferences, integrating contextual
analysis and semantic analysis. In addition, the ser-
vice related to Digital Storytelling is also provided
with the addition of being able to visit the site through
the option of 360
◦
Tours. The system selects these via
the Experience Elaboration Module, but the user can
change how individual POIs or parts of the route are
visited. However, this requires modifying the history
provided and the multimedia content associated with
the POIs.
3.1 The MuG Approach as
Recommender System
To provide personalized pathways to users to enhance
their enjoyment of the arts and cultural heritage, it
is necessary to have a recommendation system ca-
pable of filtering and analyzing data to understand
their preferences. In addition, the recommendation
system will need to exploit context and leverage se-
mantic analysis to enable the system to gain situ-
ational awareness. To this end, the previously de-
scribed architecture exploits the Multilevel Graph ap-
proach based on three graphs: the Context Dimension
Tree (CDT), the Domain Ontology, and the Bayesian
Network (BN) that enables predictions.
The CDT is a tree depictable by a graph
G
CDT
=< N
CDT
, E
CDT
, r
CDT
>, (1)
where N
CDT
represents the nodes of the CDT, E
CDT
contains the edges of the graph, and r
CDT
is the root of
the tree (Bolchini et al., 2006). The nodes in N
CDT
are
divided into dimension nodes, concepts nodes, and
parameters. This structure allows context manage-
ment through the 5W+1H paradigm (Jia et al., 2016),
which represents the fundamental nodes among the
dimension nodes and is represented graphically with
black nodes. On the other hand, concept nodes spec-
ify the values assumed by a specific context domain
represented by the dimension nodes and are graphi-
cally represented through white nodes. Finally, pa-
rameters provide additional information to the con-
cept nodes and are represented graphically with trian-
gles. The CDT graph starts from the root r
CDT
, which
connects to the dimension and concept nodes. Finally,
the parameters are the children of the concept nodes.
The Ontology also consists of a graph structure
G
O
=

C, A, H, R
+
, R
−

, (2)
where C represents the concepts in the ontology to
which A attributes are associated. Then, it also in-
cludes the hierarchical type relations H and the de-
Improving Enjoyment of Cultural Heritage Through Recommender Systems, Virtual Tour, and Digital Storytelling
267

pendency and independence relations R
+
and R
−
, re-
spectively.
In particular, the CDT shares the graph nodes
with the domain ontology. Specifically, dimension
nodes are among the Ontology concepts, while con-
cept nodes are among the associated attributes. This
allows an interconnection between the two graphs and
joint filtering through context and semantic analysis.
The BN is also seen as a graph structure
G
BN
=< N
BN
, E
BN
>, (3)
where N
BN
contains the random variables of the BN
and the set E
BN
represents the edges that are the de-
pendency relations of the random variables. In addi-
tion, the random variables considered in the Bayesian
network (BN) also fall within the defined ontology
structure. The defined dependency and independence
relations can be identified among the random vari-
ables forced through the ontology and the indepen-
dence relations. This makes it possible to determine
the dependency R
(D)
constraint list and the indepen-
dence R
(I)
constraint list (Casillo et al., 2024).
Through the definition of the function g (h, π
h
),
which quantifies the connection between the node h
and the relatives π
h
within the graph, i.e., the depen-
dency relationships to be identified (Cooper and Her-
skovits, 1992), the goal for the construction of the
Bayesian network using the data is to maximize the
Bayesian network bound function
n
∏
h=1
max
π
h
P(π
s
i
−→ B
i
)g(h, π
h
) (4)
best suited to represent the data (Cooper and Her-
skovits, 1992), where n represents the number of ran-
dom variables. In addition, at the stage of identifying
the network B
i
, it is necessary to integrate the con-
straints R
(D)
and R
(I)
identified following the algo-
rithm 1.
4 EXPERIMENTAL PHASE
Once the architecture for enhancing artistic and cul-
tural heritage has been introduced based on the Multi-
level Graph (MuG) approach used as a recommender
system, this section presents the experimental phase
divided into two steps. In the first step, the accuracy
of the recommendation approach is evaluated by test-
ing how the classification using the Bayesian Network
(BN) succeeds in suggesting individual Points of In-
terest (POIs) and identified paths. In the second step,
the architecture’s ability to improve the enjoyment of
cultural aptitude is evaluated by exploiting an online
experiment and developing a prototype.
Data: N
BN
, R
(D)
, R
(I)
, u
Result: The structure of the BN
for i = 1, . . . , n do
Add the dependency relation of R
(D)
:
¯
π
i
=
n
(X
,
x
i
) ∈ R
(D)
o
;
Calculate P
old
= g (x
i
,
¯
π
i
);
status = True;
while status &
|
¯
π
i
|
< u do
Add the dependency to the node z;
Calculate P = g (x
i
,
¯
π
i
∪
{
z
}
);
if P > P
old
then
P
old
= P;
¯
π
i
=
¯
π
i
∪
{
z
}
;
else
status = False;
end
end
end
Algorithm 1: Multilevel Graph approach algorithm (Casillo
et al., 2024).
4.1 Accuracy Evaluation of MuG
The first step of the experimental phase aims to eval-
uate the proposed approach’s accuracy through data
collected about interactions between users and POIs
and between users and identified routes.
Specifically, two datasets were collected: the first
relating 81 POIs to 13 users for 347 identified inter-
actions. These interactions represent the ratings pro-
vided by users about the POIs based on the environ-
mental conditions in which the users were located.
Specifically, the 347 available ratings are divided into
five classes: 17 associated with the 1-star class, 67
with the 2-star class, 81 with the 3-star class, 96 with
the 4-star class, and 86 with the 5-star class. The
second dataset, conversely, contains interactions be-
tween 16 users and 27 possible paths for 140 avail-
able ratings. Again, the ratings are divided into five
classes: 4 associated with the 1-star class, 14 with the
2-star class, 35 with the 3-star class, 46 with the 4-
star class, and 41 with the 5-star class. The precision,
recall, and f-score metrics for each class and the sys-
tem’s overall accuracy were used to evaluate the pro-
posed approach’s accuracy (Gunawardana and Shani,
2015; Rainio et al., 2024). In addition, k-fold cross-
validation was used in the testing phase by dividing
the datasets into five parts (Koren et al., 2022).
In the case of the first dataset, Figure 3a reports the
confusion matrix obtained, while Figure 4a reports
the precision, recall, and f-score results for each class
considered. It can be seen from the figure how, for
each class, the system achieves at least 80% precision,
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
268

(a) Confusion matrix obtained on the first dataset of 347
ratings related to 13 users and 81 POIs.
(b) Confusion matrix obtained on the second dataset of
140 ratings related to 16 users and 27 routes.
Figure 3: Confusion matrices related to the two considered datasets.
(a) Precision, recall, f-score on the first dataset of 347
ratings related to 13 users and 81 POIs.
(b) Precision, recall, f-score on the second dataset of 140
ratings related to 16 users and 27 routes.
Figure 4: Precision, recall, f-score related to the two considered datasets.
which becomes higher than 85% in the case of the
2stars, 3stars, and 5stars classes. For recall, at least
80% is achieved for all classes except for the 1-star
class, which contains the least available ratings. The
results for the f-score, which is the harmonic mean
between precision and recall, achieve at least 80%
for each class considered. The accuracy obtained is
86.17%, which is satisfactory despite the imbalance
of available ratings in the various classes.
Instead, in the case of the second dataset, Figure
3b reports the confusion matrix obtained, while Fig-
ure 4b reports the precision, recall, and f-score results
for each class considered. The figure shows that the
system overcomes the 86% precision for classes 3-
stars, 4-stars, and 5-stars. For recall, the system over-
comes the 83% for the 3-star class, the 85% for the
4-star class, and the 92% for the 5-star class. The ac-
curacy obtained is 86.43%, which is adequate despite
the more significant imbalance of available ratings in
the various classes than in the previous dataset.
4.2 Online Experiment
Once the accuracy of the recommendation system had
been assessed, it was necessary to develop a prototype
for evaluating the effectiveness of the proposed ar-
chitecture in enhancing artistic and cultural heritage.
The development of the prototype also required the
development of virtual tours based on 360
◦
photos.
The first step of elaborating the virtual tour involved
collecting 360
◦
images by preliminarily defining the
hotspots associated with the POIs. Then, the photos
are processed through software that allows the cre-
ation of the virtual tour and, thus, implements the
hotspots that enable the user to move from one point
to another on the tour. In addition, the prototype re-
quired the installation of a weather station and sensors
to assess the crowding of individual POIs. The appli-
cation enabling user interaction was also developed.
The experimental phase involved 41 users who,
after using the services described in Section 3, eval-
uated their cultural experience by answering a ques-
tionnaire consisting of five sections:
• Section A aimed to assess the suggested route
Improving Enjoyment of Cultural Heritage Through Recommender Systems, Virtual Tour, and Digital Storytelling
269

Figure 5: Results of the questionnaire related to the online experiment.
to understand whether the system could meet the
users’ preferences;
• Section B focuses on the Virtual Tour and its abil-
ity to engage the user through the virtual experi-
ence while considering the ease of interaction;
• Section C requires evaluating the digital sto-
rytelling techniques applied, which will assess
whether the user felt engaged in the narrative pre-
sented;
• Section D aims to evaluate the hybrid experience,
divided between POIs visited physically and POIs
visited virtually overall;
• Finally, Section E seeks to assess the users’ cul-
tural experience overall.
Each section could be evaluated based on 5 possible
responses: Totally Agree, Agree, Neutral, Disagree,
and Totally Disagree.
It can be seen from Figure 5 that the results ob-
tained are more than satisfactory as each evaluated
section achieves excellent levels of evaluation.
5 CONCLUSIONS AND FUTURE
WORKS
This work introduces an architecture that aims to en-
hance the artistic and cultural patrimony by using a
recommendation system capable of integrating Situ-
ation Awareness, virtual tours based on 360
◦
photos,
and Digital Storytelling techniques. The design and
implementation of the prototype require defining four
layers that, starting from data acquisition and storage,
enable the elaboration of services to be provided to
users to improve the enjoyment of artistic and cultural
heritage. During the experimental phase, the accuracy
of the usage recommendation approach based on the
Multilevel Graph approach through two datasets was
evaluated on the one hand, and the ability of the pro-
posed architecture to improve the user experience on
the other hand. Possible future developments involve
increasing the data collected so that a more meaning-
ful sample of data can be used to evaluate the accu-
racy of the recommendation system. In addition, they
want to improve the prototype further and continue
the testing phase through the online experiment.
REFERENCES
Adomavicius, G., Bauman, K., Tuzhilin, A., and Unger, M.
(2022). Context-Aware Recommender Systems: From
Foundations to Recent Developments. Springer New
York, NY.
Argyriou, L., Economou, D., and Bouki, V. (2020). Design
methodology for 360° immersive video applications:
the case study of a cultural heritage virtual tour. Pers.
Ubiquitous Comput., 24(6):843 – 859.
Bolchini, C., Curino, C., Schreiber, F., and Tanca, L.
(2006). Context integration for mobile data tailoring.
In MDM’06, page 48 – 55.
Borr
`
as, J., Moreno, A., and Valls, A. (2014). Intelligent
tourism recommender systems: A survey. Expert Syst.
Appl., 41(16):7370 – 7389.
Brumana, R., Quilici, S., Oliva, L., Previtali, M., Gabriele,
M., and Stanga, C. (2023). Multi-sensor hr mass data
models toward multi-temporal-layered digital twins:
Maintenance, design and xr informed tour of the
multi-stratified appian way (paaa). Sensors, 23(20).
Casillo, M., Colace, F., Gupta, B. B., Lorusso, A.,
Marongiu, F., Santaniello, D., and Valentino, C.
(2022). A situation awareness approach for smart
home management. In ISMODE 2021, page 260 –
265.
Casillo, M., Colace, F., Gupta, B. B., Lorusso, A., San-
taniello, D., and Valentino, C. (2023). The role of
AI in improving interaction with cultural heritage: An
overview.
Casillo, M., Colace, F., Lorusso, A., Santaniello, D., and
Valentino, C. (2024). A multilevel graph approach for
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
270

iot-based complex scenario management through sit-
uation awareness and semantic approaches. J. Reliab.
Intell. Environ.
Colace, F., Conte, D., De Santo, M., Lombardi, M., San-
taniello, D., and Valentino, C. (2022). A content-
based recommendation approach based on singular
value decomposition. Connect. Sci., 34(1):2158 –
2176.
Colace, F., D’Arienzo, M. P., Lorusso, A., Lombardi, M.,
Santaniello, D., and Valentino, C. (2023). A novel
context aware paths recommendation approach for the
cultural heritage enhancement. In Proceedings - 2023
IEEE International Conference on Smart Computing,
SMARTCOMP 2023, page 273 – 278.
Colace, F., Lombardi, M., Pascale, F., and Santaniello, D.
(2018). A multi-level approach for forecasting critical
events in smart cities. In DMSVIVA 2018, page 31 –
35.
Cooper, G. F. and Herskovits, E. (1992). A bayesian method
for the induction of probabilistic networks from data.
Mach. Learn., 9(4):309 – 347.
Cramer, H., Evers, V., Ramlal, S., Van Someren, M., Rut-
ledge, L., Stash, N., Aroyo, L., and Wielinga, B.
(2008). The effects of transparency on trust in and
acceptance of a content-based art recommender. User
Model. User-Adapt. Interact., 18(5):455 – 496.
De Gemmis, M., Lops, P., Semeraro, G., and Basile, P.
(2008). Integrating tags in a semantic content-based
recommender. In RecSys’08, page 163 – 170.
Endsley, M. (1995). Toward a theory of situation awareness
in dynamic systems. Hum. Factors, 37(1):32 – 64.
Garc
´
ıa-Valldecabres, J., Galiano-Garrig
´
os, A., Meseguer,
L. C., and Gonz
´
alez, M. C. L. (2021). Hbim work
methodology applied to preventive maintenance: A
state-of-the-art review. In WIT Trans. Built Environ.,
volume 205, page 157 – 169.
Gunawan, E. S. and Lesmana, C. (2023). Developing 360
degree virtual tour of dharma rakhita temple as a cul-
tural learning source. In ICE-SMARTec 2023, page
151 – 154.
Gunawardana, A. and Shani, G. (2015). Evaluating recom-
mender systems. Springer New York, NY.
Hong, M., Jung, J. J., Piccialli, F., and Chianese, A. (2017).
Social recommendation service for cultural heritage.
Pers. Ubiquitous Comput., 21(2):191 – 201.
Jia, C., Cai, Y., Yu, Y. T., and Tse, T. (2016). 5w+1h pat-
tern: A perspective of systematic mapping studies and
a case study on cloud software testing. J. Syst. Softw.,
116:206 – 219.
Kokar, M. M., Matheus, C. J., and Baclawski, K. (2009).
Ontology-based situation awareness. Inf. Fusion,
10(1):83 – 98.
Koren, Y., Rendle, S., and Bell, R. (2022). Advances in
Collaborative Filtering. Springer New York, NY.
Michalakis, K. and Caridakis, G. (2022). Context awareness
in cultural heritage applications: A survey. J. Comput.
Cult. Herit., 15(2).
Mitro, N., Krommyda, M., and Amditis, A. (2022). Smart
tags: Iot sensors for monitoring the micro-climate of
cultural heritage monuments. Appl. Sci., 12(5).
Musto, C., de Gemmis, M., Lops, P., Narducci, F., and
Semeraro, G. (2022). Semantics and Content-Based
Recommendations. Springer New York, NY.
Nikolakopoulos, A. N., Ning, X., Desrosiers, C., and
Karypis, G. (2022). Trust Your Neighbors: A Compre-
hensive Survey of Neighborhood-Based Methods for
Recommender Systems. Springer New York, NY.
Podara, A., Giomelakis, D., Nicolaou, C., Matsiola, M., and
Kotsakis, R. (2021). Digital storytelling in cultural
heritage: Audience engagement in the interactive doc-
umentary new life. Sustainability, 13(3):1 – 22.
Rainio, O., Teuho, J., and Kl
´
en, R. (2024). Evaluation
metrics and statistical tests for machine learning. Sci.
Rep., 14(1).
Ricci, F., Rokach, L., and Shapira, B. (2022). Recommender
Systems: Techniques, Applications, and Challenges.
Springer New York, NY.
Ruotsalo, T., Haav, K., Stoyanov, A., Roche, S., Fani, E.,
Deliai, R., M
¨
akel
¨
a, E., Kauppinen, T., and Hyv
¨
onen,
E. (2013). Smartmuseum: A mobile recommender
system for the web of data. J. Web Semant., 20:50
– 67.
Scanagatta, M., Salmer
´
on, A., and Stella, F. (2019). A sur-
vey on bayesian network structure learning from data.
Prog. Artif. Intell., 8(4):425 – 439.
Selmanovi
´
c, E., Rizvic, S., Harvey, C., Boskovic, D., Hu-
lusic, V., Chahin, M., and Sljivo, S. (2020). Improv-
ing accessibility to intangible cultural heritage preser-
vation using virtual reality. J. Comput. Cult. Herit.,
13(2).
Su, X., Sperli, G., Moscato, V., Picariello, A., Esposito, C.,
and Choi, C. (2019). An edge intelligence empow-
ered recommender system enabling cultural heritage
applications. IEEE Trans. Ind. Inform., 15(7):4266 –
4275.
Valtolina, S., Mazzoleni, P., Franzoni, S., and Bertino, E.
(2006). A semantic approach to build personalized
interfaces in the cultural heritage domain. In Proceed-
ings of the Workshop on Advanced Visual Interfaces,
volume 2006, page 306 – 309.
Zhang, S., Tay, Y., Yao, L., Sun, A., and Zhang, C. (2022).
Deep Learning for Recommender Systems. Springer
New York, NY.
Improving Enjoyment of Cultural Heritage Through Recommender Systems, Virtual Tour, and Digital Storytelling
271