Recommendation Systems: A Deep Learning Oriented Perspective

Igor Luiz Lampa

, Vitoria Zanon Gomes

and Geraldo Francisco Doneg

a Zafalon

Department of Computer Science and Statistics, Universidade Estadual Paulista (UNESP),

Rua Crist

ao Colombo, 2265, Jardim Nazareth, S

ao Jos

e do Rio Preto - SP, 15054-000, Brazil

Keywords:

Recommendation Systems, Deep Learning, Collaborative Filtering, Content-Based, Hybrid Approach.

Abstract:

The massive use of the digital platforms has provided an exponential increase at the amount of data consumed

and daily generated. Thus, there is a data overload which directly affects the consume experience of digital

products, whether at ﬁnd a news, consume an e-commerce product or to choose a movie in a streaming

platform. In this context, emerge the recommendation systems, which have the ﬁnality of provide an efﬁcient

way to comprehend the user predilections and to recommend direct items. Thus, this work brings the classical

concepts and techniques already used, as well as analyzes their use along with deep learning, which through

evaluated results has a grater capability to obtain implicit relationships between users and items, providing

recommendations with better quality and accuracy. Furthermore, considering the review of the literature and

analysis provided, an architectural model for recommendation system based on deep learning is proposed,

which is deﬁned as a hybrid system.

1 INTRODUCTION

The high volume of digital data currently available,

whether these generated on social media, e-commerce

platforms, search, news agencies and other applica-

tions, caused an overload of information. In this con-

text, users when consuming services available on the

web, are faced with an excess of options, which in

many cases are divergent to their interests and not use-

ful for their personal proﬁle. So it is necessary to em-

ploy some methodology that is efﬁcient to provide the

information tailored to the proﬁle and interests of the

person requesting it.

According to Sarker and Matin (2021), with the

increase in web development, around 2.5 quintillion

of bytes are produced daily. In this sense, it is stated

this high amount of data harms the taking of de-

cisions. Thus, this situation has contributed to the

development of a segment of computing called rec-

ommendation systems, which using algorithmic ap-

proaches enables the customization of content tar-

geted at users and aligned with their expectations.

The main objective of a recommendation system

is to indicate to its users what they are most willing to

be interested in, providing an personalized experience

https://orcid.org/0009-0005-2099-9020

https://orcid.org/0000-0003-4176-566X

https://orcid.org/0000-0003-2384-011X

and avoiding excessive and unnecessary information

(Negi and Patil, 2021). Recommendation systems is

vital to improve access to information in order to sup-

port the decision-making process (Zhang et al., 2019;

Petter. and Jablonski., 2023).

The recommendation system primarily ability is

to understand users behaviors and habits in relation to

the items they are interacting to (Zhou, 2020). That

is, recognizing which trends of user preferences and

then indicate the items that match their expectations.

As speciﬁed by Da’u and Salim (2019), the devel-

opment of systems recommendation systems based on

deep learning has become a growing trend in present.

This condition is justiﬁed by the capacity of this new

technique to provide better representation learning of

the interaction between users and items, when com-

pared to traditional methods previously established in

literature. In this scenario, the hybrid use of tradi-

tional methods combined with deep learning, proved

to be innovative in discovering non-linear and im-

plicit relationships between users and items. There-

fore, high-quality and non-trivial recommendations

are generated.

For this reason, considering the literature review

and analysis provided in this work, we propose an

architectural model by applying innovative machine

learning techniques associated with the traditional

recommendation systems, with the aim of developing

682

Lampa, I., Gomes, V. and Zafalon, G.

Recommendation Systems: A Deep Learning Oriented Perspective.

DOI: 10.5220/0012622700003690

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 26th International Conference on Enterprise Information Systems (ICEIS 2024) - Volume 1, pages 682-689

ISBN: 978-989-758-692-7; ISSN: 2184-4992

a hybrid approach. In this scenario, we seek to re-

duce individual deﬁciencies of these methods while

improving the quality of the results by using them

together. So, this work focuses on providing a op-

timized strategy, which favors the recommendation

process in terms of quality and error reduction.

This work is organized as follows: in section 2,

we present a review concerning recommendation sys-

tems and deep learning; in section 3, related works are

presented; in section 4, a hybrid architectural model

is proposed supported by the review; and, ﬁnally, in

section 5, the conclusions are showed.

2 RECOMMENDATION

SYSTEMS AND DEEP

LEARNING

Traditional recommendation systems are classiﬁed

into three groups: content based, collaborative ﬁlter-

ing and hybrid approaches. These classiﬁcation are

explained at sections below.

2.1 Content Based

The content based approach directly depends on

meta-data regarding to users and items, that is, it re-

quires detailed information related to the user proﬁle,

as well as the attributes that describe the item. In the

case of users, data such as age, gender, geolocation,

among others can be considered registration data. For

items, assuming the example of a movie, they can

be observing the attributes of gender, title, summary,

classiﬁcation tags, running time, year of release, ac-

tors and other data deemed necessary to describe them

(Sarker and Matin, 2021). Therefore, this data is used

at comparisons between the user proﬁle and the items,

to obtain the recommendations (Bhanuse and Mal,

2021).

To establish the user’s preference proﬁle, it is nec-

essary to use of some learning technique, that is, al-

gorithms that reveal the relationship implicit in the di-

rect interaction between users and items. In this sense,

artiﬁcial neural networks, Support Vector Machine –

SVM and Bayesian classiﬁers are frequently used to

obtain the user proﬁle (Da’u and Salim, 2019).

One of the positive factors of content-based rec-

ommendation is user independence, since indications

are carried out in isolation, that is, disconnected from

the preferred proﬁle of other users. Another factor is

transparency, denoted by the clarity of the process of

obtaining items recommended. Also it is able to avoid

the absence of ﬁrst-rater problem, because when new

items are included and it is capable to make recom-

mendations, even without prior assessments of them.

On the other hand, one of the negative factors is

the occurrence of obviousness in recommendations

and this condition can result in predictable indica-

tions. According to Saat et al. (2018), this phe-

nomenon is classiﬁed as the ﬁlter bubble problem

recommendation, due to the mechanism of this strat-

egy being speciﬁc per user and not consider the rest

of the community’s preferences, that is, a cycle of

recommendations self-referenced by the user’s own

predilections. This fact along the time provide a re-

duction at variations and diversity that are considered

fundamental for the recommendations, moreover this

point was target of critics, because it can make user to

lose opportunities and newer possibilities (Grossetti

et al., 2019).

Although the content based approach is capable of

handling the insertion of new items, when new users

are included it becomes an obstacle. This condition

occurs due to the lack of interaction history of new

users with the available items, which would be used

for training and learning your preferences.

As described by Bhanuse and Mal (2021), there

are two types of content-based approaches:

• Case Dependent Reasoning Method: in this

model, items that are highly related are associ-

ated with those that have already been previously

appreciated by the user. There is a tendency to

increasing the quality of indications as the user

interacts with the items available, as your history

will be increased and new connections will be able

to be established.

• Attribute Dependent Method: Recommenda-

tions are made based on the list of item attributes

and its adherence in relation to the user proﬁle.

In this sense, there is no dependence in relation

to user interaction with items, therefore the inclu-

sion of new users do not create difﬁculties, as the

recommendations will use your proﬁle as a ba-

sis previously declared. For example, if a user is

registered at a computer science area of interest,

then the recommendation algorithm will use this

explicitly deﬁned attribute to make indications.

2.2 Collaborarive Filtering

Collaborative Filtering is one of the most widely used

methods (Da’u and Salim, 2019) and as opposed to

content based, this approach does not depend of the

attributes of users and items, but only of the rela-

tionship between them. In general, the mechanism

is based on extracting a user’s interaction history with

Recommendation Systems: A Deep Learning Oriented Perspective

683

the items and establish comparisons with the histori-

cal records of interaction of the other users with the

items, with the purpose of obtaining new classiﬁca-

tions (Sarker and Matin, 2021).

The fundamental concept employed by this ap-

proach is to consider the assumption that similar users

have similar interests (Bhanuse and Mal, 2021). The

recommendation is based on the assessment of the be-

havior of the community of users with the items, or

even the links between the items themselves (Da’u

and Salim, 2019). One of the most used algorithms

and considered the best calculation methods for this

purpose is called cosine similarity (Guo and Liu,

2019) .

In real system applications, a lack of data about

users or items is common, either due to their non-

existence or difﬁculty in processing to obtain it (Peng

et al., 2020). Therefore, the ability of this method

not to depend on these attributes for its operation, is

considered a positive advantage. Furthermore, it has

the ability to provide indications with diversity, since

does not consider the user exclusively and encom-

passes the analysis of all user-item relationships.

One of the factors that degrade the performance

of this approach is called cold start, which refers to

the condition in which there is a new item or new

user. In this context, there is an absence of a user

or item history, which causes a lack of resources to

recognize preferences. Therefore, the system will be

unable to generate assertive indications, a fact that

will reduce system performance of recommendation.

(Fayyaz et al., 2020; Liu et al., 2020)

Another factor that negatively impacts this method

is the scarcity of data, which occurs when the vol-

ume of interaction between user and items is low.

That condition generates a sparse user-item matrix

(Isinkaye et al., 2015), which causes the inability to

locate similar neighbors and therefore a recommen-

dation process with low assertiveness.

The constant increase in the volume of data used

in the recommendation is also considered a problem

for this method. The scalability problem, is related

to the difﬁculty that the method at deal with constant

data increase. A technique called Singular Value De-

composition - SVD which is based on dimensional-

ity reduction, also others strategies based on cluster-

ing process can be applied to mitigate this problem

(Fayyaz et al., 2020).

The collaborative ﬁltering strategy is subdivided

into two segments as follows:

• Memory Based:

Memory based is also categorized into two types,

user based and item based. In the ﬁrst approach,

the recommendation process focuses on under-

standing similar interests among users (Bhanuse

and Mal, 2021). In practical terms, items are rec-

ommended to the target user considering that they

have been rated or acquired by users who resem-

ble him. At the item based method, the objective

is to establish the relationship between the items

themselves and the interest in them, that is, based

on the item a particular user chose to make recom-

mendations for items that are similar to the same

(Anil et al., 2018).

In item and user-based techniques, there are fac-

tors that inﬂuence the choose between them.

Firstly, the similarity of items is considered with

greater importance stability, that is, the relation-

ship established between them is very likely to

stay. However, when considering the similarity

between users it is the opposite, since that users’

interests evolve and transform over time, as a re-

sult, new similarity calculations will be required

more frequently. Furthermore, another positive

factor of the item based technique is that, in gen-

eral, there are more users than items, therefore,

the item-item matrix is smaller in dimensions than

the user-user, that is, it can mean a competitive ad-

vantage in environments with resource limitation,

whether in terms of time or even hardware.

• Model Based:

On the other hand, model based is a category of

collaborative ﬁltering algorithms, that use statis-

tical and machine learning techniques to perform

model training and classiﬁcations. In this context,

the core of the strategy is detect how likely a given

user is to rate an item according to your interest or

not, based on the classiﬁcations previously carried

out (Bhanuse and Mal, 2021). Matrix factoriza-

tion is a classic algorithm of this area, whose ba-

sic function is based on a sparse user-item matrix,

produce probabilistic values for unﬁlled gaps, that

is, items without prior evaluation, which will be

used for recommendations.

In the meantime and given the scientiﬁc contri-

butions in the area of machine learning and pri-

marily advances at the sub-area of deep learning,

provided a fundamental ally for recommendation

systems. Since the use of models, such as auto-

encoders, Convolution Neural Network - CNN,

Recurrent Neural Network – RNN, among others,

aim to improve the performance of traditional rec-

ommendation algorithms by providing greater ef-

ﬁciency and lower error rate.

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

684

2.3 Hybrid Approach

Hybrid recommendation systems arise from the com-

bination of positive aspects of one or more recom-

mendation system strategies, with the aim of provid-

ing an optimized methodology, but also to mitigate

individual deﬁciencies inherent to them when used

in isolation (Da’u and Salim, 2019). The hybridiza-

tion of the recommendation process can be achieved

by combination of different techniques, for example,

there is the classic union of algorithms content based

with collaborative ﬁltering, which has the function of

providing greater accuracy of the prediction results.

Several current researches reveal the importance

and increase of quality of results obtained when rec-

ommendation strategies are used together. In this con-

text, one of the fastest growing areas in joint applica-

tion with traditional recommendation systems algo-

rithms is called deep learning (Zhang et al., 2019),

whose deﬁnition and use cases will be discussed in

the following section of this work.

2.4 Deep Learning

This concept is considered the new generation of ar-

tiﬁcial neural networks, which traditionally has been

one of the pillars of artiﬁcial intelligence and machine

learning (Da’u and Salim, 2019). The objective of its

application is to improve the representation of learn-

ing through multiple layers and stages of data pro-

cessing, being able to learn various levels of repre-

sentations and data abstraction (Zhang et al., 2019),

therefore, discover implicit relationships when mov-

ing between layers.

Currently, the application of deep learning is

growing and expanding, standing out in satisfactory

results in areas such as computer vision, natural lan-

guage processing, image processing and in recom-

mendation (Xu et al., 2021). As Da’u and Salim

(2019) highlighted, since the ﬁrst publications in-

volving deep learning with recommendation systems,

there was an accelerated growth of related studies and

most has been published in the last six years.

In the meantime, there is a massive demand from

both academia and in industry to use deep learning for

a wide range of applications, with in order to take ad-

vantage of their intrinsic ability to deal with complex

tasks processes and obtain improved results. Further-

more, this methodology surprises not only due to the

increase in performance in several aspects in these ar-

eas as well as its ability to learn representations from

the beginning of application, that is, from zero stage

(Zhang et al., 2019).

The effectiveness demonstrated in several scien-

tiﬁc studies regarding the application of this approach

with recommendation systems, enabled the develop-

ment and progress in the use of this method. In this

context, several algorithms that combine these two

approaches, created an important scientiﬁc ﬁeld with

a range of emerging applications capable to produce

results with greater performance than compared to

traditional algorithms (Peng et al., 2020; Sarker and

Matin, 2021).

3 RELATED WORKS

Some of the state of the art researches that consider

the recommendation systems based on deep learning

are presented as follows.

Zhou (2020) proposed a recommendation system

approach with the objective of reducing the problem

of high computational cost inherent to traditional al-

gorithms when new items or users are included. In

this sense, was used deep neural networks to perform

similar content search for new users or new products,

and then dynamically include them in the original sys-

tem. Through this strategy, it reduced the number of

recalculations that would be necessary and in compar-

ison to other algorithms, there were better results for

the mean squared error parameters (RMSE) and mean

absolute error (MAE).

A collaborative ﬁltering methodology applied

with deep learning was proposed by Negi and Patil

(2021), whose objective is to unite these two areas

and obtain better results in recommendations. Two

models were built, one based on stacked autoencoder

– AE, with the function of identifying multiple com-

pressed representations of the same data, and another

based on Restricted Boltzmann Machine – RBM,

used in the stage of building recommendations.

Sarker and Matin (2021) deﬁned a recommenda-

tion system hybrid based on matrix factorization and

deep neural networks, in addition used auxiliary in-

formation about users and items. The proposed model

aims to obtain internal and implicit relationships be-

tween users and items, for this it uses factorization

matrix and the Multilayer Perceptron – MLP model.

4 HYBRID MODEL OF

RECOMMENDATION SYSTEM

This section aims to describe the proposed model for

the system of recommendation. In this sense, it is ex-

plored the concepts of the hybrid model proposed for

recommendation based on collaborative ﬁltering and

Recommendation Systems: A Deep Learning Oriented Perspective

685

also content-based, both supported by deep learning.

According to Dellal-Hedjazi and Alimazighi

(2020) the most of the works that use recommenda-

tion systems based on deep learning uses collabora-

tive ﬁltering strategy and these approaches provide

high diversiﬁcation are capable to deal with the con-

stant evolution of volume of data. However, they are

impacted by the cold start problem and data scarcity.

On the other hand, although the content-based ap-

proach is less explored through the implementation

of algorithms based on deep learning, it is essential

to address the problem of cold start and data scarcity.

In that sense, it is a technique capable of providing

predilections, through descriptive attributes individ-

uals, regardless of the historical interaction relation-

ships between users and items.

In this context, the basic idea is to use a hybrid

model that is capable to unify the qualities of col-

laborative ﬁltering strategies with the content-based

type, and that is implemented with fundamental deep

learning algorithms. Furthermore, this is conﬁrmed

by Huang et al. (2019), that the hybrid recommenda-

tion algorithms are still little explored by deep learn-

ing approaches, and therefore it is an area with scope

to be explored for innovative contributions.

The architecture of the proposed is organized

into two main modules, namely Content-Based Deep

Learning and Collaborative-Filtering Deep Learning.

The ﬁrst uses as input data the user attributes and

items, and from them generates a model that is used

to generation of recommendations. The second uses

historical data of the relationship between users and

items to train the model and generate preferences. Fi-

nally, the results are combined and a recommendation

list is presented.

4.1 Content-Based Deep Learning

Module

This module has an architecture based on the recom-

mendation system developed by Dellal-Hedjazi and

Alimazighi (2020) and has two concepts which are

the demographic and content-based approach. That

is, consider both the attributes that describe and qual-

ify the items as well as those that are referring to

users, in order to be the basis for training and gen-

eration of recommendations. In Figure 1 are showed

the sequence of steps that are performed to obtain the

recommendation.

The steps that make up this module are described

below:

Pre-Processing:

Phase of receiving raw input data, in which the at-

tributes of interest are selected. Additionally, a data

Figure 1: Content-based Deep Learning Module.

cleaning process is applied for removing discrepant

(outliers) or unstructured data, and ﬁlling of missing

data. There is data coding, a stage in which the con-

version of data from the original format to digital for-

mat so that it can be used in mathematical calcula-

tions. Finally, the application of normalization called

Min-Max to reduce the data range between 0 and 1,

this operation being carried out with the in order to

facilitate the learning process of the neural network.

Learning Module:

One of the most suitable architectures for classiﬁca-

tion problems based on the attributes and characteris-

tics of the input data is called Multilayer Perceptron

– MLP, which is a deep neural network. In this case,

it was dimensioned to be composed of an input layer,

twelve hidden layers and a about to leave.

The input layer will be fed with the characteristics

of the items, for example, in the MovieLens database

the attributes of title, year of release and gender, and

also user attributes, as age and occupation. Regarding

the intermediate part, it will be composed of twelve

hidden layers, which may be resized to adapt the net-

work in relation to the results obtained and thus allow

their optimization. In observation, the choice of num-

ber of hidden layers is a relevant decision point for

the algorithm, since which can generate two types of

problems called overﬁtting and missﬁtting, the ﬁrst

occurs when the number of hidden layers is high and

this generates a very large in relation to the input data,

that is, the model in practice becomes specialized and

is not able to generalize to new evaluated data (Sabiri.

et al., 2022). The second problem is related to the in-

ability of the model to adhere to the behavior of input

data and becomes too much generic.

Finally, the output layer is responsible for gener-

ating the result of the recommendation, whose neuron

with the best probability is considered the network

prediction.

Furthermore, it has a variable number of neurons,

grouped functionally and fully connected. Basically,

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

686

the learning process starts with the pre-processed

data, of which 80% will be for training, 10% vali-

dation and 10% testing, so learning occurs by passing

data through the deep neural network and calculating

the error between the expected result and the found

result. In this way, at each interaction the error is min-

imized through backpropagation and the weights for

each parameter matrix are saved.

Recommendation Phase:

Step that considers the previously trained and learned

model, thus, loads the saved weights and uses them

under the new pre-processed data. The result obtained

is a list in descending order containing the largest val-

ues that were predicted.

4.2 Collaborative Filtering Deep

Learning Module

The main objective of this module is to enable the ex-

traction of predilections that are intrinsic to the his-

tory of user and item relationships. According to lit-

erary review research, there is a model that stands out

for its capacity extract the relationship between users

and items through multiple perspectives and that has

proven efﬁciency in obtaining results with greater per-

formance and quality. Therefore, this model proposed

by Huang et al. (2019) is called Deep Matrix Factor-

ization Learning – DMFL, and is used as a basis for

reference so that it can be adapted and coupled to the

hybrid algorithm proposed in this work. This module

is visually represented at Figure 2.

Figure 2: Collaborative Filtering Deep Learning Module.

The basic operating structure is segmented into

two stages, namely, feature learning and preference

generation, which are described in follow:

Feature Learning:

Composed of two parallel neural networks that are re-

sponsible for extract the latent feature vectors of items

and users. In this context, Multilayer Perceptron –

MLP structures are used with the aim of extract deep

and hidden characteristics regarding users and items.

Due to the characteristics of the items that will be

recommended are considered stable because they do

not change frequently, a method is established static

for training and extracting them. Therefore, the la-

tent vector data of item characteristics are extracted

directly from the static descriptive data of the items.

The characteristics are considered: id, name, year,

gender and other data that can describe it and that are

available in the MovieLens database.

On the other hand, data relating to user char-

acteristics can change more frequently according to

changes in preferences. Therefore, this method is a

process of dynamic learning which considers both ba-

sic and static data, for example, age, gender, occupa-

tion and region, as well as using the history of inter-

actions with items. Through this history, the vectors

of characteristics of the k items that the user “liked”

most recently and thus allows consider the natural dy-

namics of changes in user preferences.

The user’s preference history vector is deﬁned

through the equation 1, whose parameter k represents

the number of items that the user recently liked and y

is the preferred history item of user u

∑

t=1

(1)

Furthermore, in order to obtain the ﬁnal vector of

user characteristics, the vectors x

and y

, which refer

respectively to the basic data vector and static images,

and the preference history vector, are inserted into the

input layer of a Multilayer Perceptron. This ﬁnal vec-

tor is described by equation 2:

= f (W

: x

] + b

) (2)

Where f is the activation function of the neural

network layer, W

represents the weights and b

is the

bias of the network layer.

Generation of Preferences:

The result of the step described previously, which are

the characteristics of items and users, is used as in-

put to the algorithm responsible for generating user

preferences. This process is based on the concept of

simultaneously combining several models, so that it is

possible to extract characteristics of data from multi-

ple perspectives. In this way, in addition of being able

to combine the advantages of these different models,

also minimizes their deﬁciencies, therefore it makes

it possible to provide results with greater precision.

In this context, there are three sub-modules that

run in parallel and their results are merged to obtain

user preferences. The modules are called SDAE-FM,

Recommendation Systems: A Deep Learning Oriented Perspective

687

Figure 3: DMFL Architecture.

Deep Neural Network – DNN, Metric Learning as

presented in Figure 3.

The sub-modules of the preference generation

process are detailed below:

• SDAE-FM:

Based on two Matrix Factorization strategies

– FM and Stacked Denoising AutoEncoder –

SDAE, which together are responsible for reduc-

ing the dimension of features and extracting deep

latent features. In this way, it is used to learn the

importance of each characteristic, as well as the

relationship that occurs between them and, conse-

quently, obtains the user’s preferences.

• Deep Neural Network:

The main objective of this module is to explore

the relationships deep and non-linear relationships

arising from the relationship between items and

users, and ﬁnally, generate user preferences with

greater precision. Consider a neural network with

four layers, with the calculation in each of them

being carried out as equation 3:

(l+1)

= f (W

+ b

) (3)

Where a represents the input, W is the weights,

b is the bias, with all parameters referring to the

l − th layer. The function f corresponds to the

function of activation, in which in this case the

Rectiﬁed Linear Units – ReLUs method was used.

In observation, at this stage the use of the Batch

normalization with the aim of standardizing in-

puts at each layer of the neural network and pre-

vent the model overﬁtting problem.

• Metric Learning: At this stage the main objective

is to consider the users and items, and also mea-

sure them from a distance perspective. Uses two

parallel convolutional neural network models with

different parameters. The computational process

of these networks is represented by equation 4:

= f (x

∗ k

+ b

) (4)

Where x

refers to the input data, k

represents the

t- th ﬁlter, b

is the bias, f is the ReLUs activation

function and, ﬁnally, the symbol ∗ is the convolu-

tion.

In summary, each of the sub-modules aims to ob-

tain user preference for through speciﬁc perspectives:

linearity, non-linearity and distance. Thus, at the end

of executing the three steps there are three user pref-

erence vectors which will be processed through the

sigmoid operation. In this context, when executing In

this operation, the main objective is to obtain a com-

bination of results in order to effectively optimize rec-

ommendation accuracy. This operation is deﬁned by

equation 5:

i j

= sigmoid(y

i j

+ y

i j

+ y

i j

) (5)

In summary, the individual results of each of the

modules will be uniﬁed and then the recommendation

will be presented to the user. This process of joining

the results will initially be the simple union of the ﬁ-

nal prediction vectors individual issues arising from

each of the modules, however, it is a decision point

for the algorithm design and that it can be evolved

and used some speciﬁc operation when it is subjected

to evaluation criteria and tests.

5 CONCLUSIONS

The amount of resources that are offered and con-

sumed digitally through the Internet is constantly in-

creasing, involving information from work, leisure,

studies or simply for communication. In this con-

text, it is noted an overload of information that is con-

sidered harmful in relation to experience of users of

these resources and of the enterprises that are not able

to target the content that really matters to their cus-

tomers. In this way, it was veriﬁed how fundamental

it is to use recommendation to support the resolution

of the problem of information overload and thus em-

ploy useful strategies for it. Furthermore, it was found

that the use of lonely classical recommendation tech-

niques is not such efﬁcient when compared to its hy-

brid use or supported by other methodologies.

The study and analysis of approaches that adopt

deep learning in conjunction with traditional rec-

ommendation system techniques demonstrate greater

quality and accuracy than classical algorithms. In

addition, they provide the discovery of implicit rela-

tionships between users, items and user-items, there-

fore, providing better understanding of user prefer-

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

688

ences and then making predictions with greater efﬁ-

ciency and degree of assertiveness.

Finally, this work was conducted with the objec-

tive of provide improvements and optimizations to

recommendation systems based on development of

the proposed hybrid architectural model, considering

both aspects of content-based strategy and collabora-

tive ﬁltering supported by deep learning.

ACKNOWLEDGEMENTS

The authors would like to thank Coordenac¸

ao de

Aperfeic¸oamento de Pessoal de N

ıvel Superior -

Brasil (CAPES), under grant 88887.686064/2022-00,

and S

ao Paulo Research Foundation (FAPESP), under

grant 2020/08615-8, for the partial ﬁnancial support.

REFERENCES

Anil, D., Vembar, A., Hiriyannaiah, S., G.M., S., and Srini-

vasa, K. (2018). Performance analysis of deep learn-

ing architectures for recommendation systems. In

2018 IEEE 25th International Conference on High

Performance Computing Workshops (HiPCW), pages

129–136.

Bhanuse, R. and Mal, S. (2021). A systematic review:

Deep learning based e-learning recommendation sys-

tem. pages 190–197.

Da’u, A. and Salim, N. (2019). Recommendation system

based on deep learning methods: a systematic review

and new directions. Artiﬁcial Intelligence Review,

53:2709–2748.

Dellal-Hedjazi, B. and Alimazighi, Z. (2020). Deep learn-

ing for recommendation systems. pages 90–97.

Fayyaz, Z., Ebrahimian, M., Nawara, D., Ibrahim, A., and

Kashef, R. (2020). Recommendation systems: Algo-

rithms, challenges, metrics, and business opportuni-

ties. Applied Sciences, 10(21).

Grossetti, Q., du Mouza, C., and Travers, N. (2019).

Community-based recommendations on twitter:

Avoiding the ﬁlter bubble. In Cheng, R., Mamoulis,

N., Sun, Y., and Huang, X., editors, Web Information

Systems Engineering – WISE 2019, pages 212–227,

Cham. Springer International Publishing.

Guo, W.-w. and Liu, F. (2019). Research on collabora-

tive ﬁltering personalized recommendation algorithm

based on deep learning optimization. In 2019 Inter-

national Conference on Robots & Intelligent System

(ICRIS), pages 90–93.

Huang, Z., Yu, C., Ni, J., Liu, H., Zeng, C., and Tang, Y.

(2019). An efﬁcient hybrid recommendation model

with deep neural networks. IEEE Access, 7:137900–

137912.

Isinkaye, F., Folajimi, Y., and Ojokoh, B. (2015). Recom-

mendation systems: Principles, methods and evalua-

tion. Egyptian Informatics Journal, 16(3):261–273.

Liu, F., Song, Y.-j., Han, Y.-h., Wu, M.-y., Dong, W., Cai,

P., Long, C.-h., and Wang, B.-q. (2020). Research on

personalized recommendation model and algorithm in

deep learning mode. In 2020 International Confer-

ence on Robots & Intelligent System (ICRIS), pages

116–119.

Negi, R. and Patil, A. B. (2021). Deep collaborative ﬁlter-

ing based recommendation system. In 12th Interna-

tional Conference on Computing Communication and

Networking Technologies, ICCCNT 2021, Kharagpur,

India, July 6-8, 2021, pages 1–5. IEEE.

Peng, M., Zhang, J., Wen, S., and Liu, C. H. (2020). Com-

paring recommendation algorithms in session-based

e-commerce sites. In 2020 International Conference

on Intelligent Computing, Automation and Systems

(ICICAS), pages 377–380.

Petter., S. and Jablonski., S. (2023). Recommender systems

in business process management: A systematic litera-

ture review. In Proceedings of the 25th International

Conference on Enterprise Information Systems - Vol-

ume 2: ICEIS, pages 431–442. INSTICC, SciTePress.

Saat, N., Mohd Noah, S. A., and Mohd, M. (2018). Towards

serendipity for content–based recommender systems.

International Journal on Advanced Science, Engi-

neering and Information Technology, 8:1762.

Sabiri., B., El Asri., B., and Rhanoui., M. (2022). Mech-

anism of overﬁtting avoidance techniques for train-

ing deep neural networks. In Proceedings of the

24th International Conference on Enterprise Informa-

tion Systems - Volume 1: ICEIS, pages 418–427. IN-

STICC, SciTePress.

Sarker, M. R. I. and Matin, A. (2021). A hybrid col-

laborative recommendation system based on matrix

factorization and deep neural network. In 2021

International Conference on Information and Com-

munication Technology for Sustainable Development

(ICICT4SD), pages 371–374.

Xu, G., Cai, J., Jiang, S., Cai, J., Ouyang, M., and Li, J.

(2021). Physical examination package intelligent rec-

ommendation system based on deep learning. In 2021

6th International Conference on Computational Intel-

ligence and Applications (ICCIA), pages 21–25.

Zhang, S., Yao, L., Sun, A., and Tay, Y. (2019). Deep learn-

ing based recommender system: A survey and new

perspectives. ACM Computing Surveys, 52(1):1–38.

Zhou, Y. (2020). A dynamically adding information rec-

ommendation system based on deep neural networks.

In 2020 IEEE International Conference on Artiﬁcial

Intelligence and Information Systems (ICAIIS), pages

1–4.

Recommendation Systems: A Deep Learning Oriented Perspective

689