https://doi.org/10.31449/inf.v42i3.

0000 Informatica 45 page 501–yyy 1

A Deep Learning Model for Context Understanding in

Recommendation Systems

Ngo Le Huy Hien

Leeds Beckett University, Leeds, United Kingdom
E-mail: [email protected]

Luu Van Huy, Hoang Huu Manh, Nguyen Van Hieu*

The University of Danang - University of Science and Technology, Danang, Vietnam
E-mail: [email protected],[email protected], *[email protected]

Keywords: Recommendation System; Context Understanding; Convolutional Neural Network; Ma-

trix Factorization; Deep Learning; Text Processing

Received: November 2, 2022

Due to the robust growth in the amount of data and Internet users, there has been a sig-
nificant rise in information overload, hindering timely access to user demand. While in-
formation retrieval systems, such as Google, Bing, and Altavista have partially addressed
this challenge, prioritization and personalization of information have yet to be fully im-
plemented. Therefore, recommendation systems are developed to resolve the issue by
filtering and segmenting important information from an enormous volume of data based
on different criteria such as preferences, interests, and user behaviors. By collecting data
on users’ interests and purchased products, the system can predict whether a particular
user would enjoy an item, thus delivering an appropriate suggestion strategy. However,
the increased number of Internet users and items has resulted in sparseness in increas-
ingly vast datasets, reducing the performance of recommendation algorithms. Therefore,
this study developed a model integrating Convolutional Neural Network (CNN) and Ma-
trix Factorization (MF) to add extra product and user information, extract contexts,
and add bias to the observed ratings in the training process, attempting to enhance
the recommendation accuracy and context understanding. This approach can take ad-
vantage of CNN to efficiently capture an image’s or document’s local features, with the
combination of MF to create relationships between 2 main entities, users and items. The
results are promising and can act as a reference for developing context understanding
in recommendation systems, and future work may focus on optimizing the performance
and developing more text-processing techniques.
Povzetek:

1 Introduction sic to listen to, online news to read, or any-

thing else based on the desired industry [3]. Rec-
Recommendation systems (also known as recom- ommendation systems are substantially beneficial
mender systems [1]) are algorithms designed to when a person has to pick an item from an over-
deliver suggestions for the most pertinent items whelming number of options provided by a ser-
to a certain user by filtering out information from vice [4]. Netflix [5, 6] and Amazon [7], for ex-
a pool of data using various factors [2]. Nor- ample, employ recommendation systems to assist
mally, the recommendations pertain to different their consumers to choose a suitable product or
decision-making processes, including what movies movie. The recommendation system handles a
to watch, books to read, products to buy, mu- huge amount of data by filtering the most signif-
2 Informatica 45 page 501–yyy Hieu N.V. et al.

icant information from data given by a user and filtering approaches in various methods. It bal-
other criteria that correlate to their interests and ances out the corresponding deficiencies of differ-
preferences [3]. It determines the match between ent filtering techniques while using their respec-
the user and the item, then infers the similarities tive strengths. The methods can be weighted,
among them for suggestions [4]. switching, cascade, mixed, feature-combination,
Recommendation systems have been proven to feature-augmented, or meta-level hybrid depend-
provide decent benefits to both users and supplied ing on the operations of the combined techniques
services. They were characterized from the stand- [17].
point of E-commerce as a tool that assists users in However, the aforementioned filtering tech-
searching through a source of data associated with niques retain a few drawbacks, notwithstanding
users’ preferences [8]. Especially, under a com- their success. Overspecialization, limited content
plex and large accumulation of information, rec- analysis, and data scarcity are a few issues with
ommendation systems might showcase their ad- content-based filtering algorithms. In addition,
vantage to enhance the quality of decision-making cold-start, scalability, and sparsity issues remain
strategies [9]. This utility may result in decreas- to exist in collaborative techniques, reducing the
ing transaction costs associated with locating and effectiveness of recommendations [18]. It can be
selecting products in the E-commerce sector [10]. seen that the common problem with such filter-
Even in several companies, an efficient recommen- ing techniques is data sparsity. It is because of
dation system can generate colossal revenue, and the explosive growth in the number of users and
serve as a means to differ considerably from their items in the fast-growing service market, which
rivals [11]. increased the sparseness of product review data
It is prevailing to apply recommendation sys- from users [19]. This sparseness diminishes the
tems when having insufficient personal knowledge prediction accuracy of traditional filtering tech-
or expertise with the alternatives since the sys- niques [20].
tems may support and enrich the social process of In order to address the above data sparseness
making decisions based on the [9]. For instance, limitation, in this paper, different factors have
recommender systems are utilized in scientific li- been added to the recommendation system such
braries to assist users by enabling them to go be- as user information, user interactions, and prod-
yond catalog searches [3]. Therefore, these types uct description documents instead of only using
of systems can address the information overload- review data, attempting to enhance the accuracy
ing issue, which is commonly encountered in re- of the system. Moreover, traditional information
cent years [12], by operating accurate and efficient retrieval methods mostly use the bag-of-words
recommendation algorithms to deliver individual- model, which ignores the context information of
ized, distinctive service and content suggestions the text document [21]. To address this, the study
[13]. proposed a model to apply a Convolutional Neural
There are several recent techniques have been Network (CNN) in the recommendation system
developed for constructing recommendation sys- to better understand the text document. Owing
tems, including collaborative filtering, content- to the fact that CNN can efficiently capture lo-
based filtering, and hybrid filtering [14]. The most cal features of documents or images through local
developed and widely used technique is collabo- receptive fields, shared weights, and pooling [22].
rative filtering, which finds users who own sim- However, since CNN is primarily used in classifi-
ilar preferences and utilizes their views to sug- cation problems, this study proposed an approach
gest to another user [15]. Contrarily, the content- to integrate it into Matrix Factorization (MF) to
based approach links user attributes to content define relationships between users and items. The
resources. It hence often disregards inputs from combination makes it possible to take full advan-
other users and delivers recommendations solely tage of both CNN and MF [23]. Inspired by the
based on the information provided by the user work of Donghyun and colleagues [24], this study
[16]. Notwithstanding, hybrid filtering can im- aims to enhance the model by adding bias for
prove the effectiveness and accuracy of recom- the training more objectively; and supplement-
mendation systems, by combining two or more ing extra information from description documents
A Deep Learning Model for Context Understanding in . . . Informatica 45 page 501–yyy 3

of both users and items. The research outcomes suggestions. The user then receives suggestions
are promising and can be used as a reference for for unseen items that received favorable reviews
further developing context understanding in rec- from others in the neighborhood [26]. The sug-
ommendation systems. gestions can be in a form of recommendations or
predictions. A recommendation is a list of the top
items that the user would enjoy the best, whereas
2 Literature Review a prediction is an estimated favorable score of an
item for the target user [27].
2.1 The Development of
Recommendation Systems In contrast, content-based filtering links user
characteristics to the attributes of items. It hence
For a system to deliver its customers reliable and often disregards inputs from other users and de-
helpful recommendations, the usage of accurate livers recommendations solely based on the infor-
and efficient recommendation algorithms is essen- mation provided by the user [16]. This filtering
tial. Therefore, it is critical to clarify the advan- technique is significant when the suggested docu-
tages and limitations of various recommendation ments can be metadata-represented, which could
approaches. There are several recent techniques be books, news, and web pages. Content-based
for constructing recommendation systems, which filtering extracts characteristics from the content
are content-based filtering, collaborative filtering, of items previously rated by different users and
and hybrid filtering, as depicted in Figure 2.1 [14]. then merges them into a training set. From there,
the system recommends items that are greatly re-
lated to a user’s favorability to them. The tech-
nique can deliver recommendations even when a
user never offered ratings before [28]. As a re-
sult, users may receive suggestions without dis-
closing their profiles, ensuring their privacy. Fur-
thermore, content-based filtering could handle cir-
cumstances in which different users might not
have identical items, but only similar items that
shared common characteristics [29].
Nevertheless, by integrating two or more fil-
tering algorithms diversely, hybrid filtering can
increase the efficacy and accuracy of recommen-
dation systems. It compensates for the inade-
quacies of various filtering systems while maxi-
Figure 2.1: Different recommendation filtering mizing their unique strengths [17]. Depending
techniques. on the operations of the combined approaches,
the methods can be weighted, switching, cascade,
First of all, collaborative filtering is a technique mixed, feature-combination, feature-augmented,
to find users who own similar preferences and uti- or meta-level hybrid. Collaborative filtering and
lizes their views to suggest to another user. It has content-based filtering approaches can be used
become the most developed and widely used fil- differently, before being combined. Thereafter a
tering technique in recommendation systems [15]. unified model was formed that encompasses both
Collaborative filtering is prominent when the con- content-based and collaborative filtering capabil-
tent cannot be accurately and simply represented ities. Consequently, the data sparsity and cold-
by metadata, like music and movies [25]. This start issues could be solved by merging item rat-
technique aims to build a database of user pref- ings, characteristics, and demographic informa-
erences for things called a user-item matrix. By tion [30].
comparing the commonalities between users’ pro- Overall, recommendation systems have gained
files, it connects people with shared interests and considerable interest since their initial introduc-
preferences in a so-called neighborhood to provide tion and have been widely utilized in various sec-
4 Informatica 45 page 501–yyy Hieu N.V. et al.

tors, including e-commerce [8], e-library [31], e- TikTok, one of the most popular and rapidly
tourism [32], education [33], news [34], informa- expanding social media networks in the world,
tion retrieval, and digital content services [35]. has its secret strength as a unique recommenda-
Table 1 indicates the eminent applications of rec- tion system for discovering and distributing con-
ommendation systems in different domains. tent [50]. TikTok blends videos from newbies
and celebrities in the ‘For You’ feed, rewards
Item Type Recommendation Systems high-quality creative content based on page views,
E-commerce Amazon [7], eBay [36], and encourages emerging users to share videos
Products Shopify, Flipkart [37] with other viewers. Therefore, every user has
Videos Netflix [5], YouTube [38], Dai- the opportunity to become famous on the plat-
lymotion, Hulu [39], Movie- form, regardless of their fanbase or level of pop-
Lens, Nanocrowd, Jinni [40] ularity. High-quality creative work may be easily
Online News Google News, Yahoo! News, shared thanks to TikTok’s recommendation sys-
BBC, New York Times [41], tem, which regularly suggests videos to individu-
Findory [42], Digg, Zite [43] als with similar interests [51].
Music Spotify, Apple Music, Ama- It can be seen that recommendation systems
zon Music, Soundcloud, Pan- have been applied in numerous domains and have
dora, Mufin [44] helped businesses not only generate colossal rev-
Social Net- Facebook, TikTok, Twitter, enue but also serve as a means to differ consider-
working LinkedIn, Instagram [45] ably from their competitors.
Contents

3 Methodology
Table 1: Current eminent recommendation sys-
tems in different domains This study aims to develop a model integrating
Convolutional Neural Network (CNN) and Matrix
Leading e-commerce company Amazon applies Factorization (MF) to add extra product and user
a collaborative filtering technique to address scal- information and extract contexts before training,
ability challenges by offline generating a table of attempting to enhance the recommendation accu-
related items using an item-to-item matrix [7]. To racy. In this section, the architecture of CNN and
enhance suggestion quality, it employs topic di- MF is briefly presented.
versity algorithms. Following that, the algorithm
suggests items that are comparable online based
3.1 Convolutional Neural Network
on the customers’ past purchases [46]. Thanks to
this, items that are not among the shop’s 100,000 Convolutional Neural Network (CNN/ ConvNet
best-selling items have helped Amazon gain 20% - proposed by Fukushima Kunihiko) is a vari-
to 40% of sales [47]. ant of a feedforward neural network. Convo-
Netflix Recommendation Engine uses algo- lutional Neural Networks represent significant
rithms that filter its contents using each user’s progress and influence in the development of Deep
unique profile. The system uses 1,300 clusters Learning [52]. Many CNN variations, including
based on user choices to filter over 3,000 titles at VGGNet, MobileNet, Inceptions, ResNet, Reg-
once [48]. Cinematch, a proprietary recommen- Net, DenseNet, and EfficientNet have been devel-
dation system used by Netflix, has a root mean oped robustly. These variants emphasize different
squared error (RMSE) of 0.9525. In 2009, Net- facets of accuracy, efficiency, and scalability. The
flix held a competition called ’Netflix Prize’, at- field of computer vision is mostly dominated by
tempting to produce a recommender system that ConvNets models [53].
outperformed its algorithm, with a million-dollar The organization of the visual cortex and the
prize for the winner [6]. For that reason, 60% of human brain’s neural network both had an influ-
Netflix’s DVDs are rented thanks to recommen- ence on CNN’s architecture [54]. Individual neu-
dation algorithms, and 47% of North Americans rons can only respond to stimuli in the restricted
prefer Netflix with a retention rate of 93%. [49] visual field region known as the Receptive Field.
A Deep Learning Model for Context Understanding in . . . Informatica 45 page 501–yyy 5

A succession of similar fields that overlap encom- While U represents the correlation between users,
pass the entire visual field [55]. There are four V represents the relationship between items, as
main types of layers for a convolutional neural presented in Figure 3.2.
network: the convolutional layer (to extract lo-
cal features), the pooling layer (representing data
of the previous layer in a more concise form, i.e.,
select only the typical features with the highest
scores through activation functions), the ReLU
correction layer and the fully-connected layer [56],
as indicated in Figure 3.1.

Figure 3.2: The concept of matrix factorization.

The concept of latent features that reflect the

relationship between objects and users is funda-
mental in Matrix Factorization for Recommenda-
Figure 3.1: The Architecture of CNN. [57] tion Systems. For example, in a movie recommen-
dation system, the latent features can be crimi-
As shown in Figure 3.1, a CNN normally con-
nal, political, action, comedy, etc.; may also be
sists of two main components:
a combination of these features or anything that
1. Hidden layers or feature extraction layers: in may not need to be named [61]. Each item can
this component, the network will perform a bring some latent features to some extent corre-
series of convolution and pooling computa- sponding to the coefficients in its vector v. The
tions to detect features. For example, if an higher the coefficient corresponds the higher pos-
image of a zebra is inputted, in this compo- sibility of having that feature. Similarly, each user
nent, the network will recognize its stripes, will also tend to prefer certain latent features de-
two ears, and four legs. scribed by the coefficients in its vector u. The
higher the coefficient, the more likely users prefer
2. Classification: in this component, a class the movies with that latent feature. The value of
with full associations will act as a classifier the expression uv will be high if the correspond-
of previously extracted features. ing components of v and u are both high. This
The CNN model in natural language process- means that the item has latent features that the
ing often considers the local context aspect of the user likes, thus the system recommends this item
corpus [58]. These contexts are extracted through to that user.
filters or the kernel and aggregated at the pooling
layer [59]. However, since the CNN model is of- Assume that there are N users and M items,
ten used for classification problems, it is challeng- with a user-item rating matrix R, in which R
ing to apply CNN directly to the recommendation ∈ RN M . In Matrix Factorization, latent mod-
system. els of user i and item j can be represented as k-
dimensional models, ui ∈ Rk and vj ∈ Rk . The
observed rating rij of user i on item j is calcu-
3.2 Matrix Factorization
lated by the inner product of respective latent
Matrix Factorization is a commonly used collab- models of user i and item j. A common ap-
orative filtering method in recommendation sys- proach to training latent models is minimizing a
tems proposed by Simon Funk [60]. Matrix Fac- loss function L, which comprises sum-of-squared-
torization decomposes the performance evalua- error terms among the observed ratings and the
tion matrix into a product of 2 matrices U and V. predicted ratings. Therefore, the loss function in
6 Informatica 45 page 501–yyy Hieu N.V. et al.

this situation can be expressed as: input documents. The difference between those
N X
M latent feature vectors with matrix U and V is the
X 2 integration between CNN and MF in fully analyz-
L= Iij (rij − uTi vj ) +
i j ing descriptive documents and evaluation data.
N
X M
X
2
+ λu ||ui || + λv ||vj ||2 (1) 4.2 Adding Bias
i j
As mentioned in Section 3.2, the observed rating
where Iij is an indicator function that becomes 1 ri j of user i on item j is calculated by the inner-
if user i rated item j and equals 0 if not. product of respective latent models of user i and
item j, which can be indicated as:
4 Proposed Model
rij ≈ r̂ij = uTi vj (2)
4.1 General Architecture
However, to avoid overfitting issues, this study
adds bias to the observed rating:

r̂ij = uTi vj + di + bj (3)

in which:

– di is a coefficient representing the pleasant-

ness of user i. The higher the coefficient, the
better user i tends to rate the products.

– bj is a coefficient illustrating product quality,

the higher the coefficient. The more users
Figure 4.1: General Architecture. tend to rate that product better.

As depicted in Figure 4.1, MF (Matrix Factor- 4.3 Loss Function

ization, in the green box) is the decomposition of
the observed rating matrix R of user-item into two From there, the loss function now can be depicted
matrices with lower weights. Matrix U represents as:
the relationship between users, while matrix V N X
M
represents the correlation between the items. The
X Iij
L(U, V, W ) = (rij − r̂ij )2
model aims to add product features to the rec- 2
i j
ommendation system. CNN in natural language N
processing often considers the local context aspect λU X
+ ||vj − cnn(W v, Xv j )||2
of the text. Therefore, CNN is used to extract 2
j
features with local contexts of the user and item |wuk |
description sets and then add the information to λW u X
+ ||wuk ||2
matrix U (matrix containing vectors describing 2
k
characteristics of the user, such as age, gender, |wvn |
and occupation) and V (matrix containing vec- λW v X
+ ||wvn ||2 (4)
tors describing features of the item) respectively. 2 n
This technique can complement and clarify the
properties of the vectors in matrix U and V. The loss function is minimal when the deriva-
In Figure 4.1, Xu and Xv act as the set of doc- tive of the above equation is 0. The loss function
uments describing the user and item respectively, uses coordinate descent to find the function that
and Wu and Wv are the weights of the CNN model updates u and v. This optimizes having to iter-
for the user and item correspondingly. The out- ate over and over one variable while correcting the
puts of the CNN are latent feature vectors of those others.
A Deep Learning Model for Context Understanding in . . . Informatica 45 page 501–yyy 7

Assuming Wu , Wj , and V (or U) are constants,

the above equation becomes a quadratic function
with respect to U (or V). Therefore:
−1
ui ← (V IiV T + λu Ik ) (V Ri + λu nn(W u, Xui ))
di ← (rij − uTi vj − bj )
−1
vj ← (U Ij U T + λV Ik ) (U Rj + λv cnn(W v, Xv j )) Figure 5.1: Sample movie description text.
bj ← (rij − uTi vj − dj )
(5) From the description text of the movies, la-
Wu and Wj will be updated through the back- tent features were extracted to add to the training
propagation of the CNN. model. The input text set of movie descriptions
has been through different preprocessing steps, as
5 Experiment and Results shown in Figure 5.2, starting with cleaning to re-
move the noise in the text like HTML tags. The
5.1 Dataset next step is word splitting, meaning splitting the
sentences into single words. Those words were
This research utilizes Movielens 1M [62], a user’s
then normalized to the same font and type. And
movie review dataset, which contains 6000 users
finally, stopwords will be eliminated, which are
and 4000 movies. It was released in 2003 with a
words that appear frequently but contain trivial
rating rate of 4.6%. This dataset includes:
meanings, such as ‘is’, ‘that’, or ‘this’ in English.
– Movie information: id, movie name, genre,
release year;

– User information: gender, age, occupation;

– List of user reviews corresponding to movies

( 1 million samples).

The training was conducted on Google Colab

with the configuration specified in Table 2.

Type Specifications
CPU Intel(R) Xeon(R)
CPU @ 2.20GHz
Number of CPUs 2
RAM 12.0 GB
Memory 108.0 GB [44]
GPU Nvidia Tesla K80

Table 2: Device Specification.

Figure 5.2: Text Pre-processing Process.

5.2 Dataset Pre-processing

The input of the model is the item description 5.3 Training
document set. Particularly in this experiment, it
contains 4000 movie description texts correspond- The dataset was divided into 3 subsets, which
ing to 4000 movies in the dataset. A sample of the are training, validation, and testing sets. Cor-
movie description text is presented in Figure 5.1. responding to each user, the number of user re-
views will be divided by the ratio of 80% for the
8 Informatica 45 page 501–yyy Hieu N.V. et al.

training set, 10% for the test set, and 10% for the
validation set.

λU λV Dimension Train. Val. Test.

Loss Loss Loss
10 40 500 0.76 0.88 0.88
10 60 500 0.77 0.88 0.88
10 50 50 0.78 0.89 0.88
10 10 50 0.7 0.90 0.90
100 10 100 0.87 0.90 0.90
50 100 100 0.88 0.91 0.91

Table 3: Loss results in different hyperparame-

ters.
Figure 5.3: Plotting RMSE results.
From Table 3, it can be seen that the ratio be-
tween λU and λV significantly affects the results.
Evaluation Training Validation Testing
If λU is much larger than λV , meaning a higher
metric
priority is given to learning the parameters of U,
RMSE 0.76695 0.88974 0.88563
a good result could not be attained. While the
MSE 0.58821 0.79163 0.78435
goal of the problem is to use data from the item,
it is better to give preference to λV , making it
slightly higher than λU , to obtain a better result. Table 4: Result Evaluation in different metrics.

5.4 Evaluation
It can be seen that the model obtained a very
To evaluate the model’s general performance, this good RMSE of 0.89 in the testing set, which
study uses Root-mean-square error (RMSE) and means the model can relatively predict favorable
mean-square error (MSE), which represent the movies of users accurately. To determine how the
dispersion of the predicted data relative to the results correlate with the user amount, a compar-
actual data. ison of RMSE with different numbers of users is
r Pn presented in Table 5.
2
i (r̂i − ri ) )
RM SE = (6)
n
No. Train. Val. Test. Exec. Train.
N of RMSE RMSE RMSE time time
1 X
M SE = (rbi − ri )2 (7) users (s ) (s)
N
i 1000 0.87865 0.91478 0.90093 0.0062 110
The RMSE function evaluates the results after 2000 0.87205 0.91791 0.93004 0.0052 75
each iteration for all 3 training, validation, and 3000 0.87168 0.91896 0.92671 0.0053 91
testing sets. The model training process was re- 4000 0.86955 0.91383 0.92973 0.005 159
peated for about 100-200 iterations until the loss 5000 0.87865 0.91478 0.90093 0.0062 110
function gives the smallest value on the validat-
Table 5: Comparison of the RMSE with different
ing and testing sets. RMSE results of the model
numbers of users.
on the training, validating, and testing sets are
illustrated in Figure 5.4.
As can be seen from Figure 5.3, in the 8th itera- Table 5 demonstrates when increasing the num-
tion, the results began to deteriorate, and the val- ber of users in the dataset, from 1000 to 5000, the
idation RMSE increased while the training RMSE accuracy increases, but with a longer convergence
continued to be overfitting. Therefore, the result time. Therefore, in order to produce appropriate
was obtained in the 8th iteration. The evaluation recommendations, recommendation system appli-
of results for the entire data is shown in Table 4. cations need to employ a large dataset.
A Deep Learning Model for Context Understanding in . . . Informatica 45 page 501–yyy 9

5.5 Utilizing the training results means the model can relatively predict favorable
movies of users accurately. Testing on different
The results obtained after training the model are
amounts of users reveals that the more users,
2 matrices U and V. An evaluation matrix Y[i,j]
the higher the accuracy, but the longer the
can be generated as:
convergence time. Future research may aim
Y [i, j] = U [i] ∗ V [j]T (8) to overcome the scant user information (e.g.,
hobbies, location, marital status) by looking
in which: for a large dataset with more user information,
- i: i-th user including more features in the user description
- j: j-th item documents, leading to a higher impact on the
prediction. Moreover, the proposed model could
be developed further by swapping out Matrix
Factorization with more efficient techniques, such
as singular value decomposition (SVD).

Acknowledgement
This research was funded and implemented for the
Rising-Star project of University of Science and
Figure 5.4: Using the training results for creating Technology, The University of Danang, Vietnam.
recommendations.

As depicted in Figure 5.5, the evaluation ma- References

trix can be applied in the recommendation system
for further usage, which outputs a list of recom- [1] L. Lü, M. Medo, C. H. Yeung, Y.C. Zhang,
mended movies for the ith user. Z.K. Zhang and T. Zhou (2012) Recom-
mender systems, Physics reports , vol. 519,
no. 1, pp. 1-49.
6 Conclusions and Future Work
[2] H. Ko, S. Lee, Y. Park, A. Choi (2022) A
In this research, a deep learning model for rec- survey of recommendation systems: recom-
ommendation systems is proposed by integrat- mendation models, techniques, and applica-
ing Convolutional Neural Network and Matrix tion fields, Electronics, vol. 11, no. 1, p. 141.
Factorization to add extra information and ex-
tract contexts before training, attempting to en- [3] F. O. Isinkaye, Y. O. Folajimi and B. A.
hance recommendation accuracy and context un- Ojokoh (2015) Recommendation systems:
derstanding. Despite substantial previous efforts Principles, methods and evaluation, Egyp-
[21, 63, 64], this study adds additional informa- tian informatics journal, vol. 16, no. 3, pp.
tion on both user and item description documents 261-273.
and applied Convolutional Neural Networks to
[4] G. Shani and A. Gunawardana (2011) Rec-
efficiently capture their local features. Further-
ommender systems handbook, Recommender
more, this research adds bias to the observed
systems handbook, pp. 257-297.
ratings to avoid overfitting issues and uses Ma-
trix Factorization to create relationships between [5] Gomez-Uribe, C. A. and N. Hunt (2015) The
users and items. The proposed model can be netflix recommender system: Algorithms,
further used as a benchmark for developing con- business value, and innovation, ACM Trans-
text comprehension in recommendation systems, actions on Management Information Sys-
hence delivering more relevant recommendations tems (TMIS), vol. 6, no. 4, pp. 1-19.
for users.
It is observed that the model obtained a very [6] J. Bennett, S. Lanning (2007) The netflix
good RMSE of 0.89 in the testing set, which prize, in KDD cup and workshop, p.35.
10 Informatica 45 page 501–yyy Hieu N.V. et al.

[7] B. Smith, G. Linden (2017) Two decades ing and Innovative Technologies in Engineer-
of recommender systems at Amazon. com, ing (ICACITE), pp. 1360-1365.
IEEE internet computing, , vol. 21, no. 3,
pp. 12-18. [17] P. B. Thorat, R. M. Goudar and S. Barve
(2015) Survey on collaborative filtering,
[8] S. S. Li, E. Karahanna (2015) Online rec- content-based filtering and hybrid recom-
ommendation systems in a B2C E-commerce mendation system, International Journal of
context: a review and future directions, Computer Applications, vol. 110, no. 4, pp.
Journal of the association for information 31-36.
systems, vol. 16, no. 2, p. 2.
[18] G. Suganeshwari and S. P. S. Ibrahim (2016)
[9] K. Al Fararni, B. Aghoutane, J. Riffi, A.
A survey on collaborative filtering based rec-
Sabri and A. Yahyaouy (2020) Comparative
ommendation system 3rd international sym-
study on approaches of recommendation sys-
posium on big data and cloud computing
tems, Embedded Systems and Artificial Intel-
challenges , pp. 503-518.
ligence, pp. 753-764.

[10] C.-S. Juan-Pedro, I. Ramos-de-Luna, E. [19] N. L. H. Hien, T. Q. Tien and N. V.

Carvajal-Trujillo and Á. F. Villarejo-Ramos Hieu (2020) Web crawler: Design and im-
(2020) Online recommendation systems: plementation for extracting article-like con-
Factors influencing use in e-commerce, Sus- tents,Cybernetics and Physics, vol. 9, no. 3,
tainability Publisher, pp. 01-15. pp. 144-151.

[11] L. Ebrahimi, V. R. Mirabi, M. H. Ranjbar [20] P. Kumar and R. S. Thakur (2018) Recom-
and E. H. Pour (2019) A customer loyalty mendation system techniques and related is-
model for e-commerce recommendation sys- sues: a survey International Journal of In-
tems, Journal of Information Knowledge formation Technology, vol. 10, no. 4, pp. 495-
Management , vol. 18, no. 3 pp. 12-18. 501.

[12] Z. Wang, X. Yu, N. Feng and Z. Wang (2014) [21] S. Bhattacharya and L. Ankit (2019) Movie
An improved collaborative movie recommen- recommendation system using bag of words
dation system using computational intelli- and scikit-learn, Int J Eng Appl Sci Technol,
gence, Journal of Visual Languages Com- vol. 4, pp. 526-528.
puting, vol. 25, no. 6, pp. 667-675.
[22] M. Sheikh Fathollahi and F. Razzazi(2021)
[13] M. Robillard, R. Walker and T. Zimmer- Music similarity measurement and recom-
mann (2009) Recommendation systems for mendation system using convolutional neu-
software engineering, IEEE software, vol. 27, ral networks,International Journal of Multi-
no. 4, pp. 80-86. media Information Retrieval, vol. 10, no. 1,
pp. 43-53.
[14] L. Shah, H. Gaudani and P. Balani (2016)
Survey on recommendation system, Inter-
[23] A. F. Agarap(2017) An architecture com-
national Journal of Computer Applications,
bining convolutional neural network (CNN)
vol. 137, no. 7, pp. 43-49.
and support vector machine (SVM) for
[15] S. K. Raghuwanshi and R. K. Pateriya(2019) image classification,arXiv preprint, p.
Collaborative filtering techniques in recom- arXiv:1712.03541.
mendation systems, Data, Engineering and
Applications, pp. 11-21. [24] D. Kim, C. Park, J. Oh, S. Lee and H.
Yu (2016) Convolutional matrix factoriza-
[16] S. Eliyas and P. Ranjana (2022) Recom- tion for document context-aware recommen-
mendation Systems: Content-Based Filter- dation,Proceedings of the 10th ACM Confer-
ing vs Collaborative Filtering, 2022 2nd In- ence on Recommender Systems ” pp. 233-
ternational Conference on Advance Comput- 240.
A Deep Learning Model for Context Understanding in . . . Informatica 45 page 501–yyy 11

[25] B. S. Neysiani, N. Soltani, R. Mofidi and [33] A. C. Rivera, M. Tapia-Leon and S.

M. H. Nadimi-Shahraki (2019) Improve per- Lujan-Mora (2018) Recommendation sys-
formance of association rule-based collabora- tems in education: a systematic mapping
tive filtering recommendation systems using study,International Conference on Informa-
genetic algorithm,International Journal of tion Technology Systems, pp. 937-947.
Information Technology and Computer Sci-
ence, vol. 11, no. 2, pp. 48-55. [34] C. Feng, M. Khan, A. U. Rahman and A. Ah-
mad (2020) News recommendation systems-
[26] Z. Cui, X. Xu, X. U. E. Fei, X. Cai, Y. Cao, accomplishments, challenges future direc-
W. Zhang and J. Chen(2020), Personalized tions,it IEEE Access , vol. 8, pp. 16702-
recommendation system based on collabora- 16725.
tive filtering for IoT scenarios,IEEE Trans- [35] M. C. Kim and C. Chen (2015) A sci-
actions on Services Computing, vol. 13, no. entometric review of emerging trends and
4, pp. 685-695. new developments in recommendation sys-
tems,Scientometrics, vol. 1, pp. 239-263.
[27] W. Zhang and J. Wang(2018), Content-
bootstrapped collaborative filtering for med- [36] G. Wei, Q. Wu and M. Zhou (2021) A hybrid
ical article recommendations,2018 IEEE In- probabilistic multiobjective evolutionary al-
ternational Conference on Bioinformatics, gorithm for commercial recommendation sys-
pp. 1184-1188. tems, IEEE Transactions on Computational
Social Systems, vol. 8, no. 3, pp. 589-598.
[28] U. Javed, K. Shaukat, I. A. Hameed, F.
Iqbal, T. M. Alam and S. Luo (2021), A re- [37] V. T. R. M. Vivek, C. Saravanan and K.
view of content-based and context-based rec- V. Kumar(2022) A Novel Technique for User
ommendation systems,International Jour- Decision Prediction and Assistance Using
nal of Emerging Technologies in Learning Machine Learning and NLP: A Model to
(iJET), vol. 16, no. 3, pp. 274-306. Transform the E-commerce System, Big data
management in Sensing: Applications in AI
[29] J. Son and S. B. Kim (2017) Content-based and IoT, p. 61.
filtering for recommendation systems using
multiattribute networks,Expert Systems with [38] J. Davidson, B. Liebald, J. Liu, P. Nandy,
Applications, vol. 89, pp. 404-41. T. V. Vleet, U. Gargi and S. Gupta (2010)
The YouTube video recommendation sys-
[30] A. A. Kardan and M. Ebrahimi(2013) A tem,Proceedings of the fourth ACM confer-
novel approach to hybrid recommendation ence on Recommender systems, pp. 293-296.
systems based on association rules mining
[39] W. Bellante, R. Vilardi and D. Rossi(2013)
for content recommendation in asynchronous
On Netflix catalog dynamics and caching
discussion groups, Information Sciences, vol.
performance, /it 2013 IEEE 18th Interna-
219, pp. 93-110.
tional Workshop on Computer Aided Model-
[31] C. Tsai and M. Chen(2008) Using adap- ing and Design of Communication Links and
tive resonance theory and data-mining tech- Networks (CAMAD), pp. 89-93.
niques for materials recommendation based [40] X. H. Pham, T. N. Luong and J. J. Jung
on the e-library environment,The Electronic (2013) An black-box testing approach on user
Library, vol. 26, no. 3, pp. 287-302. modeling in practical movie recommendation
systems, International Conference on Com-
[32] R. Khan, H. Ur, C. K. Lim, M. F. Ahmed, putational Collective Intelligence, pp. 72-79.
K. L. Tan and M. B. Mokhtar(2021) Sys-
tematic review of contextual suggestion and [41] S. Raza and C. Ding (2021) News recom-
recommendation systems for sustainable e- mender system: a review of recent progress,
tourism,Sustainability, vol. 13, no. 15, p. challenges, and opportunities,Artificial Intel-
8141. ligence Review, pp. 1-52.
12 Informatica 45 page 501–yyy Hieu N.V. et al.

[42] F. Carmagnola, F. Vernero and P. [51] Z. Chen and C. Shi (2022) Analysis of Algo-
Grillo(2009) Sonars: A social networks- rithm Recommendation Mechanism of Tik-
based algorithm for social recommender Tok, International Journal of Education and
systems,International Conference on User Humanities, vol. 4, no. 1, pp. 12-14.
Modeling, Adaptation, and Personalization,
pp. 223-234. [52] S. Albawi, T. A. Mohammed and S. Al-Zawi
(2017) Understanding of a convolutional neu-
[43] W. Zou (2018) Design and application of ral network, international conference on en-
incremental music recommendation system gineering and technology, pp. 1-6.
based on Slope one algorithm,Wireless Per-
sonal Communications, vol. 102, no. 4, pp. [53] J. Chai, H. Zeng, A. Li and E. W. Ngai
2785-2795. (2021) Deep learning in computer vision: A
critical review of emerging techniques and
[44] W. Strank (2021) Analyzing Networks of application scenarios, Machine Learning with
Musical Context in the Digital Age1, 119: Applications, vol. 6, p. 100134.
The Oxford Handbook of Music and Adver-
tising. pp.6080-6088 [54] N. L. H. Hien, L. V. Huy and N. V.
Hieu (2021) Artwork style transfer model us-
[45] J. Sanz-Cruzado Puig(2021) Contact recom- ing deep learning approach, Cybernetics and
mendation in social networks: algorithmic Physics, vol. 10, no. 3, pp. 127-137.
models,diversity and network evolution,pp.
[55] L. Liu, F.-X. Wu, Y.-P. Wang and J. Wang
519-569
(2020) Multi-receptive-field CNN for seman-
tic segmentation of medical images, IEEE
[46] L. Lü, M. Medo, C. H. Yeung, Y.-C. Zhang
Journal of Biomedical and Health Informat-
and a. T. Z. Zi-Ke Zhang (2012) Recom-
ics, vol. 24, no. 11, pp. 3215-322.
mender systems, Physics reports, vol. 519,
no. 1, pp. 1-49. [56] N. L. H. Hien and A.-L. Kor (year) Analysis
and Prediction Model of Fuel Consumption
[47] E. Brynjolfsson, Y. Hu and M. D. Smith and Carbon Dioxide Emissions of Light-Duty
(2003) Consumer surplus in the digital econ- Vehicles, Applied Sciences, vol. 12, no. 2, p.
omy: Estimating the value of increased prod- 803.
uct variety at online booksellers, Manage-
ment science, vol. 49, no. 11, pp. 1580-1596. [57] Prabhu (2018) Understanding of Convo-
lutional Neural Network (CNN) — Deep
[48] X. Amatriain and J. Basilico (2015) Recom- Learning, medium [Online]. Available:
mender systems in industry: A netflix case https://medium.com/@RaghavPrabhu/
study, Recommender systems handbook, pp. understanding-of-convolutional-neural-network-cnn
385-419. [Accessed 29 10 2022].

[49] H. Verma (2022) Netflix Recommenda- [58] N. L. H. Hien, H. M. Hoang, N. V. Hieu

tion Engine - How Netflix uses Big data and N. V. Tien (2021) Keyphrase Extraction
and Analytics to Recommend you your Model: A New Design and Application on
Favourite Shows, startuptalky [Online]. Tourism Information, Informatica, vol. 45,
Available: https://startuptalky.com/ no. 4.
netflix-recommendation-engine/ [Ac-
cessed 29 10 2022]. [59] P. J. L. a. G. W. Li (2018) Application of
convolutional neural network in natural lan-
[50] M. Zhang and Y. Liu (2021) A commen- guage processing, 15th International Com-
tary of TikTok recommendation algorithms puter Conference on Wavelet Active Media
in MIT Technology Review 2021, Fundamen- Technology and Information Processing (IC-
tal Research, vol. 1, no. 6, pp. 846-847. CWAMTIP), pp. 120-122
A Deep Learning Model for Context Understanding in . . . Informatica 45 page 501–yyy 13

[60] Y. R. B. a. C. V. Koren (2009) Matrix factor-

ization techniques for recommender systems,
Computer, vol. 42, no. 8, pp. 30-37.

[61] H.-J. Xue, X. Dai, J. Zhang, S. Huang and J.

Chen (2017) Deep matrix factorization mod-
els for recommender systems, IJCAI, vol. 17,
pp. 3203-3209.

[62] grouplens (2003) MovieLens 1M Dataset

[Online]. Available: https://grouplens.
org/datasets/movielens/1m/ [Accessed 29
10 2022].

[63] H. Liu, C. Zheng, D. Li, X. Shen, K. Lin, J.

Wang, Z. Zhang, Z. Zhang and N. N. Xiong
(2021) EDMF: Efficient deep matrix factor-
ization with review feature learning for in-
dustrial recommender system, IEEE Trans-
actions on Industrial Informatics, vol. 18, no.
7, pp. 4361-4371.

[64] H. Wu, Z. Zhang, K. Yue, B. Zhang, J.

He and L. Sun(2018) Dual-regularized ma-
trix factorization with deep neural networks
for recommender systems, Knowledge-Based
Systems, vol. 145, pp. 46-58.