Creating a Personalized Recommendation Framework in Smart
Shopping by Using IoT Devices
Noura Abdaoui
1,2 a
, Ismahene Hadj Khalifa
3
and Sami Faiz
2,3
1
ENSI, University of Manouba, Tunisia
2
LTSIRS Laboratory, Tunisia
3
ISAMM, University of Manouba, Tunisia
Keywords: Ubiquitous Recommender System, Personalized Recommendation, IoT, Fog Architecture, Big Data, Context.
Abstract: Personalization and recommendation are two important prerequisites that must be incorporated in the Iot
environment where smart devices data are generated anywhere and anytime. Both prerequisites are essential
to produce a higher satisfaction level of ubiquitous recommender system which matches the preferences of
the user. Is the time to improve the quality of traditional ubiquitous recommender system which failed to
exploit dynamic and heterogeneous big data in delivering personalized recommendation. In this paper, we
create a framework of personalized recommendations in Smart shopping where Iot devices are connected. We
proposed a Fog computing architecture to solve the ubiquitous recommendations issues related to Iot
challenges. The given model is a multi-layer fog structure which aims to use the multi sources big data in
order to propose personalized offers according to the users’ profiles and analyze their feedbacks to improve
their experiences.
1 INTRODUCTION
Ubiquitous recommender systems assist the mobile
user by providing personalized recommendations of
items and services in ubiquitous environment where
context is the most important aspect (Mettouris and
al., 2014). Nowadays, devices used in ubiquitous
environment are connected to sensors coming under
Iot. Most of Iot applications are connected to cloud
computing. Sensors and other devices provide huge
amount of data in Iot applications. They generate Big
Data that will be processed and analyzed for suitable
and reactive actions. Data collected by sensors must
be analyzed in the cloud, which requires a high
bandwidth for the network used. Thus, these issues
can be solved by using fog computing (M. Chiang and
T. Zhang, 2016). This new technology was invented
by Cisco in 2012 as three-tier “Mobile (Iot)-Fog-
Cloud” system. It widens the cloud to bring the
environment close to the devices that interact with Iot
data (such as user devices).
This paper is an enhanced work since the
publication of (Abdaoui and al., 2018). Over there, it
a
https://orcid.org/0000-0002-7181-0401
has proposed a basic solution to send personalized ads
in ubiquitous environment after detecting the
customer’s localization. The motivation of this paper
is how to integrate the classic ubiquitous
recommendations in three-tier fog architecture. We
believe that our approach can recommend
personalized real time services while keeping features
of fog architecture as well as enhancing traditional
recommender approaches in ubiquitous computing.
In brief, the paper is organized as follows. Section 2
discusses the related studies of ubiquitous computing
and recommender systems related to Iot challenges,
and then outlines the use of fog computing. Section 3
presents the proposed architecture in details. In
section 4, we investigate the use of fog computing
architecture in the ubiquitous recommender system.
Section 5 describes the system’s implementation and
highlights the results. In Section 6, we wrap up the
paper with conclusions and horizons of work that
would improve the suggested ubiquitous fog-based
recommender system.
200
Abdaoui, N., Khalifa, I. and Faiz, S.
Creating a Personalized Recommendation Framework in Smart Shopping by Using IoT Devices.
DOI: 10.5220/0011969400003482
In Proceedings of the 8th International Conference on Internet of Things, Big Data and Security (IoTBDS 2023), pages 200-207
ISBN: 978-989-758-643-9; ISSN: 2184-4976
Copyright
c
 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

2 RELATED WORK
In this section we discuss the background and related
studies of ubiquitous computing in the field of
recommender systems. Then, we present the main
reasons behind fog computing’s appearance with
examples of existing recommender systems.
2.1 Ubiquitous Computing
The term ubiquitous computing, was first conceived
by Mark Weiser. It is an extended sort of
computational technology in the form of a
microprocessor, found in every object (Mettouris and
al., 2014). Recommendation in ubiquitous computing
is a service which uses contextual parameters in order
to provide more personalized recommendations to
users. Yet, there are various challenges to the
ubiquitous recommender systems. These limitations
could be either technological or related to storage.
While the first includes wireless technology problems
and energy concerns, the latter is related to context-
awareness, tacking user’s intentions and privacy
concerns.
2.2 Fog Computing Solution
To resolve issues mentioned previously, fog
computing technology has been set forth. Fog
computing is the outcome of the new requirements
that have been arisen to meet the needs of the ubiquity
of devices and the interest for faster management of
networks and services (F. Bonomi and al. 2014). Fog
computing applications processed different axes such
as healthcare (Waraich, A. 2019) and smart cities
(Aloqaily, M. and al., 2019). A number of fog
computing challenges exists such as the
simplification of mobility, reinforcement of privacy,
low latency, real-time interaction, low energy
consumption and the network bandwidth for real-life
applications in different sectors.
2.3 Recommender System
Certainly, recommender system RS proposes items
that users may prefer. Several approaches have been
used in the make of RS. The collaborative filtering CF
technique that recommends items based on similarity
measures between users and items, as defined by
(Prateek Parhi and al., 2017). (Asiri S. and al., 2016),
exploits the CF method to mold an Iot trust and
reputation model that investigates trust and reputation
among Iot nodes. Within the same framework,
Ubiquitous Context Aware Recommender Systems
for Ubiquitous Learning (UbiCARsUL) is proposed
by (Sukayna T. and al., 2015). The trust-aware
recommender system (TARS) is therefore introduced
to enhance the CF technique in Iot environment by
helping users finding reliable services (Weiwei Yuan
and al, 2013). But the recommender searching
mechanism of TARS is still always at its beginning.
It is not easy to find the most reliable
recommendations for the active users in the scale-free
network. In content-based filtering CBF technique, as
presented by (Umair Javed and al., 2021), the
algorithm recommends items and their similars that
were liked in the past. SOMAR by (Zanda et al.,
2012) suggests activities on Facebook from sensor
data. Similarly, (Koubai, N. and al., 2019) proposes a
recommender system to smart restaurants.
Knowledge-based technique suggests products based
on inferences about user’s needs and preferences. It is
based on the identified relationship between a user’s
needs and a possible recommendation. Another
aspect like Context-aware is also used in
recommendation. It uses the user’s context as time
and place to define the kind of recommendations.
(Hassani, A., and al., 2018) set forth Context-as-a-
Service (CoaaS) recommender system that uses an Iot
context service to provide contextual information in
smart shopping. In recent years, some attempts have
been made to integrate the real-world contexts and
emotions in the music recommenders like (Willian
Assuncao and al., 2022). Otherwise, hybrid
recommender technique as defined by (Erion Çano,
Maurizio Morisio, 2017) is the one that combines
multiple recommendation techniques together to
produce the output with higher accuracy. Besides,
(Twardowski, B. and Ryzko, D. (2015)) build an RS
that uses data from both mobile devices and other Iot
devices. Moreover, IBRS is an interest-based
recommender system proposed by (Punam Bedi and
Pooja Vashisth, 2015). It combines a hybrid RS
approach with automated argumentation-based
reasoning between cognitive agents. Machine
learning algorithms are used also in recommender
systems by many researchers namely (Sewak, M. and
Singh, S. 2016), (Abdaoui N. and al., 2017) and
(Ayata, D. and al., 2018).
2.4 The Cold Start Problem
The cold start problem occurs when the recommender
system is unable to form any relation between users
and items for which it has insufficient data due to
many reasons. Low interaction between users and
items as well as when a new user enrolls in the system
and the recommender is required to offer
Creating a Personalized Recommendation Framework in Smart Shopping by Using IoT Devices
201

recommendations for a set length of time without
depending on the user’s previous interactions.
According to (Dina Nawara and Rasha Kashef, 2020),
to provide the new user with reliable
recommendations, a content-based RS should have the
access to a sufficient number of user’s records that
allow it to determine the user’s preferences. Yet, user
might not receive accurate recommendations because
he has very few records. Moreover, although
recommending a new user’s top popular offers might
increase the user’s purchase likelihood as it could
decrease personalization. While, collaborative
approach can help improve personalization, the
recommendations’ precision might be rather weak.
In our work, we propose a hybrid algorithm that
combines the two algorithms: collaborative and
content-based algorithm in order to solve the cold
start problem for a new mobile user hence improve
the personalized recommendation. We design the
recommender model for the group of users sharing
similar characteristics, then we use this model to
predict the new user’s recommendations.
3 UBIQUITOUS FOG BASED
RECOMMENDER SYSTEM
ARCHITECTURE
While cloud computing has been widely used in
shopping center, it cannot cope with the challenges
arising in many Internet of Things scenarios, such as
network bandwidth constraints and constrained
devices. Fog architecture is a promising and effective
solution to these challenges to serving mobile users
and Iot devices by proactively catching and
processing the required data. Also, we propose hybrid
algorithm that combines the two algorithms:
collaborative and content-based algorithm in order to
solve the cold start problem for new mobile user and
improve the personalized recommendation. In the
proposed architecture, each mobile user is connected
with different fog nodes in different floors through
wireless access technologies WIFI. The fog server
can be interconnected by wireless communication
technologies. Each fog server is linked to the cloud
by IP core network. This architecture provides
efficient data processing and storing services. Each
fog node represents Ubiquitous fog-based RS that has
a number of mobile users’ interfaces connected to
both layers: things layer and the cloud layer. The
system uses virtualized machines with multiple VMs
running under a highly capable hypervisor. That
hypervisor includes real-time enhancements. The
Ubiquitous fog-based RS provides personalized and
localized recommendations to mobile user in real
time. There is a basic Linux host operating system, as
there are different modules of recommendation and
extensions for real-time operations and enhanced
security. Many resident modules are parts of the
Ubiquitous fog-based RS, including data
management, recommendations, results display and
identification with active mobile user. Therefore,
hybrid algorithms are implemented. The goal is to
learn contents from user’s profile. Then, according to
the required information and his contextual
information interacting with fog nodes, we provide
personalized recommendations in real time. More
details about the Ubiquitous fog-based RS are
discussed in the next section.
4 UBIQUITOUS FOG-BASED RS
MODEL
As shows Figure 1, there are four principal
components in our proposed system: mobile user
profile, Ubiquitous fog user interface, Process of
ubiquitous fog recommendations and fog data
processing.
4.1 Mobile User Profile
The user’s profile can be extracted from many sources
namely:
Inscription: through a registration form; preferences;
ratings on the items consulted and demographic
attributes. Demographic data can be used to calculate
recommendations for new mobile users. It used to
solve the cold start problem.
Consultation: “even without log-in”: analysis of
mobile user’s behavior implicitly. We call these
behaviors “traces of use” such as “copying/pasting”,
searching for a product on a page and navigation
indicators, such as frequency and duration of
browsing, number of clicks and mouse hovers on a
page or links, scrolling, etc.
Context: is the integration of contextual information
(location, time, physical environment, …) for the
generation of dynamic and personalized visit
itineraries.
IoTBDS 2023 - 8th International Conference on Internet of Things, Big Data and Security
202

Figure 1: Ubiquitous fog-based RS Model.
4.2 Ubiquitous Fog User Interface
It supports interactions between the potential user and
the system of recommendations that include many
functions. The Ubiquitous fog user’s interface
displays the recommended item to active user and the
reasons for the recommendation to explain why that
item has been selected to the potential user. It will
accompany the mobile user during all phases of his/
her visit to a shopping center: “before/during/after”.
He starts by researching brands and items to collect
information and make choices. The system matches
the itinerary with services and personalized content.
4.3 Process of Ubiquitous and Fog
Personalized Recommendations
In this module, we aim at providing personalized
recommendations to the users. The inputs of the
process are user, item, user-item preference,
contextual information, fog server id, workload
capability, storage capability and processing
capability. The output in most cases is directly shown
to the potential user by the Ubiquitous fog user’s
interface. The hybrid algorithms used in this module
combines Collaborative Filtering with Content based
filtering algorithms.
4.3.1 Collaborative Filtering
A collaborative filtering system collects and analyzes
a user's behavior based on a user's preferences given
in the form of feedback, ratings, and other
interactions. More specifically, a user-item rating
matrix of preferences for items by users is
constructed. From this matrix, user’s matches are
made on the basis of finding similar preferences and
interests by calculating similarities between user’s
profiles. The Equation (1) expresses how to calculate
the similarity between two users, where Vaj expresses
an assessment made by the active user a on a product
j, V
ij
the one made by the user u
i
, vvi the average of
the ratings of the user u
i
and vv
a
the average of the
valuations of the active user.
𝑆𝑖𝑚(𝑢

,𝑢

)=

𝑣

−𝑣𝑣



(𝑣

−𝑣𝑣

)


(𝑣

−𝑣𝑣

)


(𝑣

−𝑣𝑣

)

(1)
After matching between any products and users,
we try to find the k user (neighbors) with the highest
coefficient of resemblance to the active user. This
prediction is obtained from the weighted sum of the
valuations of other users using the following Equation
(2)
𝑃
(
𝑢

,𝑢

)
=𝑟𝑟


𝑟
,
−𝑟𝑟


∈
𝑆𝑖𝑚𝑢

−𝑢


∑
𝑆𝑖𝑚(
∈
𝑢

−𝑢

)
(2)
Then, there is the prediction of the valuation that the
active user 𝑃
(
𝑢

,𝑢

)
u
a
would assign to the item i. The
average rating of the user is rr and is the similarity
between the active user and the user 𝑆𝑖𝑚(𝑢

−𝑢

)i.
K is the subset of similar k-users while r
u,i
is the rating
of the neighbor u to item i .
4.3.2 Content-Based Filtering
In Ubiquitous fog-based RS, similar items are
proposed as part of the recommendations. This
technique has an advantage of providing
recommendations to the user with an item which has
not been rated yet. It provides a potential user’s
independence: by focusing only upon the dynamic
user’s ratings. It also provides transparency by
posting the characteristics of an item explicitly from
the list of recommendations.
4.3.3 Hybrid Filtering
Content-based filtering technique does not involve
the opinions of all users when recommending items
are consequently limited to making recommendations
that are in the range of a user's likes. However,
collaborative filtering cannot provide predictions to
items that have not yet been rated, commonly known
as the cold start problem. Therefore, hybrid filtering
techniques overcome these limitations and use a
combination of content-based and collaborative
techniques to improve performance. The idea is that
the resultant algorithm will provide more accurate
and effective recommendations than any single
Creating a Personalized Recommendation Framework in Smart Shopping by Using IoT Devices
203

algorithm. Given a new user, we do not intend to find
a single similar user, but we look for a group of users
with similar characteristics instead. Then, we do not
directly recommend to new users the offers that have
been bought by similar groups. Would rather, we use
a content-based method to build the recommender’s
model for the group and use the model to predict the
new user’s recommendations.
4.3.4 Personalized Recommendations
The filtering rules are applied. The result of a
personalized recommendation list is displayed to
Ubiquitous fog user’s interface taking into
consideration not only the user’s advanced and basic
profile, but also her/his customized preferences and
feedbacks. Assume that there is a set of instances X :
{X1,X2… Xn}and a binary preference function
pref(i,j,fh)>0 means i is preferred to j by user u in
Floor Fh pref(i,j,fh)= 0 means that neither of the two
items is preferred, and pref(i,j,fh)< 0 means j is
preferred to j by user u in Fh. shown in Equation 3.
Nu a set of users with similar preferences to those of
target user,𝑵
𝒖
𝑰,𝑱,𝒇
is the set of neighbors of u who rated
items i and j in the same fog server floor F.
𝒑𝒓𝒆𝒇
(
𝒊,
𝒋
,𝑭
𝒉
)
=
∑
𝒔𝒊𝒎
𝒖,𝒗,𝑭
𝒗∈𝑵
𝒖
𝑰,𝑱,𝒇
𝒓
𝒗,𝒊,𝑭
−𝒓
𝒗,𝑱,𝑭

∑
𝒔𝒊𝒎
𝒖,𝒗,𝑭
𝒗∈𝑵
𝒖
𝑰,𝑱,𝒇
(3)
𝑠𝑖𝑚
,,
is the similarity between u and v at the same
level of fog server F ; and denotes user’s ratings of
items i and j at the same level of fog server
F.𝑠𝑖𝑚
,,
𝑟
,,
𝑟
,,
. The process can be summarized
as following:
1. If a user exists, acquire input from the content-
based model
2. Produce a relevance score for all users and all
products.
3. Execute the collaborative filtering to generate
predicted ratings for all users.
4. Filter predictions.
5. Affect weighting factors for the content-based
model and collaborative filtering model to
maximize performance.
6. Turn back predictions in a descending order,
relevant items are at the top.
7. If user does not exist, the system insists to enter
user’s details such as age, sexe,etc.
8. A user’s information goes through a popularity-
based algorithm. This model finds the most
popular products by considering the average
rating for products and the maximum number of
ratings per product. A list of similar products is
then sorted in a descending order, filtered out and
represented as recommendations to the user.
9. Providing feedback on the quality of a
recommendation list in the form of ratings. This
feedback is saved for readjusting the weights and
recalibrating the recommender model.
To solve the cold start problem, we design the
user’s personal preference model that takes as an
input the historical data of a single user and as in
output the user’s contextual preference. For the newly
registered users, the system has no historical data.
Therefore, a clustering block is designed using 𝐾-
means and Density-based spatial clustering of
applications with noise (DBSCAN) (A. Smiti, Z. and
Elouedi, 2012) algorithms. These algorithms are used
easily with any data type. We employ the Elbow
method and the Silhouette coefficient in (James, G.
and al., 2013) to get the optimal number of clusters.
For each cluster, it collects the historical data of all
users belonging to that group. Next, the classification
module is used to learn about a group-level contextual
preference model. Having a clustering model and
group-level contextual preference models for groups
in hand, as soon as new user registers to the system.
Finally, the group-level contextual preference model
will be used as the user’s personal preference model.
The model will be in use until the system collects
enough historical data from the new user to build a
pure personal contextual preference model.
4.4 Fog Data Processing
In this module, raw datasets are filtered, converted
and stored into various databases. The aim of data
filtering is to extract useful information to the
recommender system. We categorize six databases
that describe a way of identifying the dataset. As
shows figure 1:
• Item Profile: this database contains item
attributes, such as item ID, description and
category, and virtual content size.
• User’s profile: this database contains user’s
attributes, such as age and gender.
• User-Item Preference: this database contains
user-item preference information that has been
converted from raw data. A preference could
either be explicit or implicit.
• Fog Server List: fog server ID, workload
capability, processing capability, storage
capability and power usage.
• Contextual Information: this database contains
contextual information such as user’s
IoTBDS 2023 - 8th International Conference on Internet of Things, Big Data and Security
204

localization, time, and network bandwidth
information.
• Recommendation Output: this database contains
the final output. It stores the output generated
during the process of creating a recommendation,
such as similarity matrix, training data and
testing data.
The basic idea is to maximize the use of a fog
server closer to the user. We try to process and store
data as close as possible to the user who is connected
to the target fog server. If the target fog server cannot
provide such a service, it will send a request to
neighboring fog servers. If neighboring fog servers
cannot provide the requested service, then the request
will be sent to the cloud.
5 EXPERIMENTS
In the following section, we implement the
Ubiquitous fog recommender system in smart
shopping by deploying five fog servers. We
performance a set of experiment based on a real-
world dataset.
5.1 Integration of Ubiquitous
Fog-based RS in Smart Shopping
Ubiquitous fog recommender system is evaluated
using a dataset containing ID address, destination IP
address, connected fog server IP address, and the
localization of the mobile user. We treat each ID
address as a user, because it is a unique identifier of
his Mobile in our fog environment. An item could
refer to Idsensor connected to a product or a web site
visited by a person. In our use case, we have used
many types of nodes in the three floors. Static nodes
such as beacons attached to different products and fog
nodes in the different floors. Also, we use mobile
nodes such as mobile phone. Each mobile user is
identified by his mobile’s API. Mobile user sensor is
detected by mobile devices which contact the fog
node with the proximity information. We have
deployed also the fog-based hybrid recommender
system on each fog server and an Alibaba cloud
server. The mobile user can access to the Internet and
use smart devices by connecting to the corresponding
fog server. The dataset is obtained from the deployed
fog servers.
We have used different platforms to simulate fog
servers. A typical platform is Windows10 OS with
Intel Core i5 CPU@2.7 GHz and 16 GB memory. All
algorithms were implemented using python with the
following installed libraries: NumPy, pandas,
matplotlib, scikit-learn, nltk, scipy. Anaconda has
been used for python package management and
deployment. We collect 200 records of different
users’ interaction with different items in the mall. All
data are obtained from the deployed five simulated
fog servers. Due to the small amount of data set, we
use all the data and split data into 80% for training
purpose, 20% as test data set.
5.2 Evaluation Data Utility vs RMSE
and MAE
In the following experiments, each floor which
includes fog server is the target level. Any user
connected to fog servers deployed on these floors is
considered in location. In the experiments, we vary
the weight value w
j
and attempt identifying the best
value of each parameter to obtain the most accurate
result. We use two popular evaluation methods for
recommendation, mean absolute error (MAE) and
root-mean-square error (RMSE), to justify our quality
of the prediction. RMSE (Equation 4) penalizes large
errors by amplifying the differences between the
predicted preferences items and the real ones:
𝑹𝑴𝑺𝑬=

∑
(𝒕𝒆𝒔𝒕 − 𝒓𝒔𝒍)
𝟐
|
𝒕𝒆𝒔𝒕
|
(4)
MAE (Equation 5) is the average absolute deviation
of the predicted ratings from the real ratings of items:
𝑴𝑨𝑬=
∑|
𝒕𝒆𝒔𝒕 − 𝒓𝒔𝒍
|
𝒕𝒆𝒔𝒕
(5)
We set up a decay factor for the weight parameter wj
to observe its impact on the prediction accuracy and
data utility. Here, the decay factor has been set up as
(
𝒘𝒋
𝒏
) for present location, the weight of the third level
of location is(
𝒘𝒋
𝒏
)
𝟑
. wj is the weight at level j that
controls prediction at each location level and affects
the final prediction. The weights satisfy the following
constraints: wj ϵ [0,1], w1+w2…+wj=1. In figure 2,
wj is fixed as 0.7, n varies from 1 to 4. This value has
also been mapped to the amount of privacy levels,
thus there are 4 privacy levels from ϵ [0, β] based on
location. In Figure 2, we observe that all three
measurements show a similar trend. The value of data
utility decreases from 1.3 to 0. 5. The RMSE and
MAE value both decreases. RMSE decreases from
3.5 to 0. 77. MAE decreases from 1.7 to 0. 5. If wj is
higher, both MAE and RMSE results are better. If n
is higher, both evaluation results are better as well.
However, the trend becomes weaker.
Creating a Personalized Recommendation Framework in Smart Shopping by Using IoT Devices
205

Figure 2: Data utility vs RMSE and MAE.
5.3 Evaluation Values of Precision and
Recall
Here, also each floor which includes fog server is the
target level.
Wk
is the weight value of the target level.
As a result, we need to be aware of the impact of
recommended content size on the server. We use the
recall and precision method to measure our result.
Both methods (as defined in (6) and (7)) are broadly
used in evaluating information retrieval and statistical
classification. In general, precision represents the
prediction accuracy, while recall represents the
prediction scale. Ideally, both values would be high.
𝑃=
𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑐𝑜𝑟𝑟𝑒𝑐𝑡 𝑖𝑡𝑒𝑚𝑠 𝑟𝑒𝑐𝑜𝑚𝑚𝑒𝑛𝑑𝑒𝑑
𝑇𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑖𝑡𝑒𝑚𝑠 𝑟𝑒𝑐𝑜𝑚𝑚𝑒𝑛𝑑𝑒𝑑
(6)
𝑅=
𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑖𝑡𝑒𝑚𝑠 𝑟𝑒𝑐𝑜𝑚𝑚𝑒𝑛𝑑𝑒𝑑
𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑟𝑒𝑙𝑒𝑣𝑎𝑛𝑡𝑖𝑡𝑒𝑚𝑠 𝑟𝑒𝑐𝑜𝑚𝑚𝑒𝑛𝑑𝑒𝑑
(7)
We vary the weight value w
k
from 0 to 1 with the
step size of 0,2 to observe the impact of result
accuracy. We also modify the number of prediction
items from 4 to 10 to observe the impact of result
accuracy. In Figure 3, each band represents the
change of the w
k
from 0 to 1. The height of each point
on a given band is the evaluation value of recall. The
length between two points on a given band represents
the difference between two results. Various bands
represent different numbers of predictions,
corresponding to items recommended to a fog server.
A higher value of the prediction number means
requiring more storage space on a fog server. Figure
3 shows that the more items are predicted, the higher
the recall value is.
In Figure 4, the height of each point on a band
represents the evaluation value of precision. We also
observe that the more items are predicted, the lower
the precision value. If w
k
is 0.3, we obtain most
accurate result on each band. So, if w
k
is 0,3, we
obtain the best evaluation results for both precision
and recall. The prediction number does not impact the
trend pattern of the evaluation value.
Figure 3: Evaluation value of recall.
However, the more items are predicted, the worse the
predicted results are, and the higher the recall is.
Figure 4: Evaluation value of precision.
The obtained results prove that the efficiency of
Ubiquitous fog-based RS and its sample algorithms
are feasible and can run independently from the cloud
server. The system helps fog servers choose the most
frequently requested content to purchase in order to
save bandwidth, storage resources and used network
resources. It provides also much more accurate
recommendation results for certain items based on
fog server location.
6 CONCLUSIONS AND FUTURE
LINES
The system is designed to address the problem of
information overloading ubiquitous environment, and
thus to become a tool of fog computing optimization.
We have reviewed the state of the art of ubiquitous
recommender system in several sectors then the
reason behind fog computing integration. The
proposed system of recommendation improves the
user’s experience in the smart shopping center where
IoTBDS 2023 - 8th International Conference on Internet of Things, Big Data and Security
206

we have used Iot devices. Experimental results
demonstrate that Ubiquitous fog-based RS provides
highly accurate and personalized recommendations to
mobile users. It considers the fog server as well as
contextual data of mobile user. Furthermore, it
incorporates feedbacks collected from mobile users.
Adding to that, it improves customers’ experiences
stored in the server and anticipates new users’ needs.
In future research, we intend to extend our proposal
to areas with deep learning algorithms and
reinforcement learning which can be used to improve
the current research and overcome limitations.
REFERENCES
Abdaoui Noura, Ismahène Hadj Khalifa, Sami Faïz (2017).
Process of Personalizing the Ubiquitous
Advertisements. International Journal of Knowledge
Society Research (IJKSR), 8 (2), pp.13-35.
Abdaoui Noura, Ismahène Hadj Khalifa, Sami Faïz (2018).
Improved Indoor Ge-olocalization System Between
Bluetooth Low Energy Technology of iBeacons and
Wi-Fi Network for Better Iot Systems Accuracy, in
Knowledge_Intensive Economies and opportunities for
Social,Organizational, and Technological Growth.
International Journal of Smart Education and Urban
Society (IJSEUS), IGI-global, chapter 9.
A. Smiti, Z. Elouedi (2012), Dbscan-gm: An improved
clustering method based on gaussian means and dbscan
techniques. 2012 IEEE 16th International Conference
on Intelligent Engineering Systems (INES), pp. 573–
578.
Aloqaily, M. , Al Ridhawi, I. , Salameh, H. B. , & Jararweh,
h. (2019). Data and service management in densely
crowded environments: Challenges, opportunities, and
recent developments. IEEE Communications
Magazine, 57 (4), 81–87.
Asiri, S. and Miri, A. (2016). An iot trust and reputation
model based on recommender systems. In Privacy,
Security and Trust (PST), 14th Annual Conference, pp.
561–568. IEEE.
Ayata, D., Yaslan, Y. and Kamasak, M.E. (2018). Emotion
based music recommendation system using wearable
physiological sensors. IEEE Transactions on Consumer
Electronics, 64, 196– 203.
Dina Nawara, Rasha Kashef (2020). IoT based
Recommendation Systems _An Overview. IEEE
International IOT, Electronics and Mechatronics
Conference (IEMTRONICS). pp 1-7.
Erion Çano, Maurizio Morisio (2017). Hybrid
recommender systems: A systematic literature review.
Intelligent Data Analysis. Volume 21. Issue 6.pp 1487–
1524.
F. Bonomi, R. Milito, P. Natarajan, and J. Zhu (2014). Fog
computing: A platform for internet of things and
analytics. Big Data and Internet of Things: A Roadmap
for Smart Environments. pp. 169–186.
Hassani, A., Haghighi, P.D., Ling, S., et al. (2018).
Querying iot services: A smart carpark recommender
use case. In Internet of Things (WF-IoT), 2018 IEEE 4th
World Forum on, pp. 619– 624.
James, G.; Witten, D.; Hastie, T.; Tibshirani, R (2013). An
Introduction to Statistical Learning; Springer: Berlin,
Germany; Volume 112.
J Li, Z He, Y Cui, C Wang, C Chen, C Yu, Min Zhang,
Yiqun Liu, Shaoping Ma. (2022). Towards Ubiquitous
Personalized Music Recommendation with Smart
Bracelets. Proceedings of the ACM on Interactive,
Mobile, Wearable and Ubiquitous Technologies. pp.1–
34
Koubai, N. and Bouyakoub, F. (2019). Myrestaurant: A
smart restaurant with a recommendation system.
International Journal of Computing and Digital
Systems, 8, 143–156.
M. Chiang and T. Zhang (2016). Fog and iot: An overview
of research opportunities. IEEE Internet of Things
Journal, vol. 3, no. 6, pp. 854–864.
Mettouris Christos, and George A. Papadopoulos (2014).
"Ubiquitous recommender systems." Computing 96.3.
223-257.
Prateek Parhi, Ashish Pal, Manuj Aggarwal (2017). A
survey of methods of collaborative filtering techniques.
2017 International Conference on Inventive Systems
and Control (ICISC).
Punam Bedi & Pooja Vashisth (2015). Argumentation-
enabled interestbased personalised recommender
system. Journal of Experimental & Theoretical
Artificial Intelligence,Volume 27, pp199-226. Issue 2
Sewak, M. and Singh, S. (2016). Iot and distributed
machine learning powered optimal state recommender
solution. In Internet of Things and Applications (IOTA),
International Conference on, pp. 101–106. IEEE.
Twardowski, B. and Ryzko, D. (2015). Iot and context-
aware mobile recommendations using multi-agent
systems. In 2015 IEEE/WIC/ACM International
Conference on Web Intelligence and Intelligent Agent
Technology (WI-IAT), pp. 33–40. IEEE.
Umair Javed, K. Shaukat, I. Hameed, Farhat Iqbal, Talha
Mahboob Alam, S. Luo (2021). A Review of Content-
Based and Context Based Recommendation Systems.
Int. J. Emerg. Technol. Learn.
Waraich, A. (2019). Remote health monitoring of elderly
through wearable sensors. Multimedia Tools and
Applications, 1–26.
Weiwei Yuan, Donghai Guan, Lei Shu & Jianwei Niu
(2013). Recommender Searching Mechanism for Trust
Aware Recommender Systems in Internet of Things, in
Automatika, Volume 54, - Issue 4.
Willian Assuncao, Lara S. G. Piccolo, and Luciana A. M.
Zaina. (2022). Considering emotions and contextual
factors in music recommendation: a systematic
literature review. Multim. Tools Appl. 81, 6, 8367--
8407.
Zanda, Andrea, Santiago Eibe, and Ernestina Menasalvas
(2012). SOMAR: A social mobile activity
recommender. Expert Systems with Applications 39.9.
8423-8429.
Creating a Personalized Recommendation Framework in Smart Shopping by Using IoT Devices
207