Ain Shams Engineering Journal 15 (2024) 102263

Contents lists available at ScienceDirect

Ain Shams Engineering Journal

journal homepage: www.sciencedirect.com

Knowledge graph-based recommendation system enhanced by neural

collaborative filtering and knowledge graph embedding
Zeinab Shokrzadeh a, Mohammad-Reza Feizi-Derakhshi b,⇑, Mohammad-Ali Balafar b,
Jamshid Bagherzadeh Mohasefi c
a
Department of Computer Engineering, Urmia Branch, Islamic Azad University, Urmia, Iran
b
Department of Computer Engineering, University of Tabriz, Tabriz, Iran
c
Department of Computer Engineering, Urmia University, Urmia, Iran

a r t i c l e i n f o a b s t r a c t

Article history: Recommendation systems are an important and undeniable part of modern systems and applications.
Received 9 December 2022 Recommending items and users to the users that are likely to buy or interact with them is a modern solu-
Revised 30 January 2023 tion for AI-based applications. In this article, a novel architecture is used with the utilization of pre-
Accepted 25 March 2023
trained knowledge graph embeddings of different approaches. The proposed architecture consists of sev-
Available online 21 April 2023
eral stages that have various advantages. In the first step of the proposed method, a knowledge graph
from data is created, since multi-hop neighbors in this graph address the ambiguity and redundancy
Keywords:
problems. Then knowledge graph representation learning techniques are used to learn low-
Social tagging systems
Recommender systems
dimensional vector representations for knowledge graph components. In the following a neural collabo-
Collaborative filtering rative filtering framework is used which benefits from no extra weights on layers. It is only dependent on
Knowledge graph matrix operations. Learning over these operations uses the pre-trained embeddings, and fine-tune them.
Neural collaborative filtering Evaluation metrics show that the proposed method is superior in over other state-of-the-art approaches.
Knowledge graph representation learning According to the experimental results, the criteria of recall, precision, and F1-score have been improved,
on average by 3.87%, 2.42%, and 6.05%, respectively.
Ó 2023 The Authors. Published by Elsevier BV on behalf of Faculty of Engineering, Ain Shams University.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-
nd/4.0/).

1. Introduction according to these historical interactions, which is also known as

collaborative filtering [2]. After presenting the idea of collaborative
With the advancement of internet and communication tech- filtering, various approaches have been introduced for this
nologies, people are facing increase in data. Obtaining useful infor- purpose.
mation from these big data is one of the important challenges. To The main element of recommendation systems is the recom-
address this challenge, recommendation systems (RSs) are one of mendation algorithm, which are mostly grouped into collaborative
the most useful tools [1]. To have highly accurate recommendation filtering (CF), content based, and hybrid recommendation systems
systems is an important challenge that has led researchers to con- [3]. Recommender systems, based on CF, model user interests
duct extensive studies in this area. So far, many solutions have according to the similarity of users or items from the interaction
been proposed to provide recommendation systems that can fol- data, while content-based systems utilizes content features of
low the recommendations according to users’ interests. This path item. CF-based recommender systems have been widely used
continues and requires more research and efforts. because they are effective for capturing user preferences and can
To create a personalized recommendation system, it is required be easily developed in multiple scenarios, without trying to extract
to employ the history data of users’ online behavior s such as rat- features in content-based recommender systems [4,5]. However,
ings, clicks, tags, and comments, and model users’ taste in items CF-based recommendations suffer from data fragmentation and
cold start problems [6]. To address these issues, hybrid recom-
mender systems have been proposed to equate interaction-level
⇑ Corresponding author.
similarity and content-level similarity [7]. In this process, various
E-mail addresses: [email protected] (Z. Shokrzadeh), [email protected].
ir (M.-R. Feizi-Derakhshi), [email protected] (M.-A. Balafar), J.bagherzade- types of side information have been investigated, such as item
[email protected] (J. Bagherzadeh Mohasefi). attributes [2], item reviews, and social networks of users [8].

https://doi.org/10.1016/j.asej.2023.102263
2090-4479/Ó 2023 The Authors. Published by Elsevier BV on behalf of Faculty of Engineering, Ain Shams University.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Z. Shokrzadeh, M.-R. Feizi-Derakhshi, M.-A. Balafar et al. Ain Shams Engineering Journal 15 (2024) 102263

Recently, knowledge graphs have attracted increasing attention
Multiplying the embeddings of users and tags, and calculating
and become a popular means of collecting, displaying, and extract- the dot product of the output of multiplication and embedding
ing structured knowledge that consists of various entities, their of source in the neural network facilitates to enrich the repre-
attributes, and their inter-relationships [9]. Knowledge graphs sentations users, tags and sources. Subsequently, our system
have launched by Google in 2012 for the first time [10]. Since then, is capable of recommending various items such as user, tag
various efforts have been proposed to enhance and apply such and source just by changing the input vectors. For example, to
knowledge structure in various fields such as adaptive learning, recommend tags, the embeddings of users and sources are fed
multimedia e-learning, software development, social media analy- into the matrix multiplication operation and then the dot pro-
sis, multi-task feature learning, air traffic management, data ware- duct of the previous operation and tag vectors were done.
houses, etc. Knowledge graphs have also been used in various
recommendation systems such as travel attraction recommenda- The structure of the article is stated as follows: Section 2 review
tions [11], personalized tweet and follower recommendations the fundamental concepts of knowledge graph, knowledge graph
[12], differentiated fashion recommendations [13], mashup-based embeddings and its methods and the related works. In section 3
APIs recommendations [14], points of interest recommendations the research method is explained in detail. Section 4 is experimen-
[15,16], group recommendation [17], etc. tal evaluation. Finally, in section 5, conclusions are descripted.
In this paper, a knowledge graph-based recommendation sys-
tem was proposed enriched with knowledge graph representation
learning and neural collaborative filtering. The proposed recom- 2. Background
mendation system first presents data in the form of a knowledge
graph which is constructed based on the user-user, user-tag, In this section, the concepts construct in the background of the
user-source, and tag-source relationships. These components study are expressed. First, a brief explanation of knowledge graphs
engender multi-hop neighbors in this graph that can abolish the will be provided. Then, the knowledge graph embedding tech-
ambiguity and redundancy problems. Although there are a number niques employed in proposed solutions were mentioned.
of studies that have been conducted on representing data in form In the real world, taking the form of a graph in a number of dif-
of knowledge graph to recommend items. However other works ferent scenarios have received remarkable attention [18]. For
just used some part of data to construct their desired knowledge instance, graph representation serves as the underlying support
graph. Following this, the next step enriches the original knowl- for various crucial domains in the field of health and diseases like
edge graph using advanced knowledge graph embedding tech- drug analysis and disease diagnosis [19]. Moreover, graph struc-
niques (ConvE, ComplEx and TransE) to learn low-dimensional ture has become a significant strategy for many applications
vector representations of knowledge graph elements. The final including data mining, natural language processing, and
stage is a neural collaborative filtering framework that receives question-answering systems [20].
an advantage, because learning a neural network provides capabil- One of the particular and popular types of graph structure is a
ity of performing classification tasks without additional weights. knowledge graph in which the node and edge respectively demon-
Our neural framework just relies on matrix operations, uses the strate the entity and relation between them. Since a knowledge
pre-trained embeddings obtained from the knowledge graph, and graph is a structure used to represent diverse types of information
fine-tunes them. The major contributions of this paper can be sum- in the form of different types of entities such as concepts, situa-
marized as follows: tions, objects, events, and portrays different types of relations, it
is renowned as a semantic network [10]. In traditionally, a knowl-

To the best of our knowledge, this paper is the first attempt to edge graph is represented by a triplet (s, p, o) where s, o indicates
completely consider all entities and their full connections to the subject and predicate entities accordingly, and p is the relation
construct a knowledge graph. In other words, all types of rela- associating s and o. Additionally, labels and attributes as extra
tionships including user-user, user-tag, user-source and tag- information can be assigned to entities and relations[21].
source are observed in the created knowledge graph. The main The goal of knowledge graph analysis is to provide a quantita-
advantage of this knowledge graph is that it presents an oppor- tive understanding of complex structures. Drawbacks of conven-
tunity to alleviate the ambiguity and redundancy problems. tional approaches mainly include high computational cost and

A further advantage is that issues such as cold-start, ambiguity memory requirements associated with the high dimensions [22].
and redundancy can be addressed with the help of multi-hop In fact, a group of methods have become a bottleneck in massive
neighbors in our constructed knowledge graph. Also, because network analysis.
of the similar neighborhood sub-graphs, synonymous tags can To prevail over the mentioned obstacles, a lot of research efforts
be semantically considered close to each other. have been devoted to learning low-dimensional vector representa-

The existence of user-user, user-tag, user-source, and tag- tions for knowledge graph components. Specifically, knowledge
source interactions in our knowledge graph enables us to deal graph embedding transforms entities and relations into a low-
with the issue of cold start for a new tag or source by inferring dimensional space in which the knowledge graph information is
the connections. preserved [23]. Knowledge graph embedding techniques based

Another significant benefit is to apply entity embedding learn- on their solutions may comprise some functions including similar-
ing in order to enrich the representations of entities. For this ity function and loss function. The similarity function is responsi-
purpose, three different categories of existing knowledge graph ble for estimating the identity of entities. In actual fact, the loss
embedding techniques (ConvE, ComplEx and TransE) were used. function looks over the quality of the vector representations.

Further gain to consider is that this work presents a neural col- This function includes a scoring function that assigns a score to
laborative filtering framework that uses matrix operations and each graph triple [23].
fine-tuning pre-trained embeddings of users, tags and sources Knowledge graph embedding methods can be classified into
to recommend various items. It means that embeddings gener- three groups: neural network, semantic matching and translational
ated by using knowledge graph embedding methods are distance models [24]. Neural network techniques apply neural net-
injected as pre-trained vectors into the neural collaborative fil- work models to obtain better embeddings of entities and relations.
tering network which are considered to be the inputs. In the semantic matching methods, the scoring function is

2
Z. Shokrzadeh, M.-R. Feizi-Derakhshi, M.-A. Balafar et al. Ain Shams Engineering Journal 15 (2024) 102263

similarity-based that determines the plausibility of facts by match- 2.1. Related works
ing the latent semantics of entities and relations [24]. On the other
coins, translational distance models utilize distance-based scoring Today, anyone is in the ocean of information. Finding useful
functions to determine the plausibility of a fact as the distance information from big data is very difficult and time consuming.
between the two entities [25]. In this study, ConvE, ComplEx, and Recommendation systems are a good solution to find useful infor-
TransE graph knowledge embedding approaches will be employed. mation according to users’ interests. Usually, recommendation sys-
Although there are many cutting-edge approaches such as DKN tem is a collection of algorithms that discover data patterns from
[26] and KBGAN [27], to name a few, these three models were the accessible dataset by learning and computing preferences of
hand-picked for two fundamental reasons. Firstly, these models user [18]. Finally, based on the correlation between the interests
cover three different categories of graph representation learning and needs, the system provides the most relevant and useful
models. The selected methods belong to neural network, semantic results [18].
machine and translational groups. Secondly, to remove any biases A knowledge-based recommendation system offers items to the
in the results because of using trendy approaches, these renowned user based on scope knowledge about how the items satisfy the
techniques were employed. In other words, to imply that the inno- user preferences [30].
vation of our proposed method does not rely on the type or the Knowledge graphs have shown great potential in supporting
innovations used in their suggested structures, these knowledge recommendation systems [31], [32,33,34,35,36,44]. The main idea
graph representation learning models were used. In the following of using knowledge graphs in thus knowledge-intensive applica-
paragraphs, these methods will be described in more detail. tions is how to transform their heterogeneous and semi-
ConvE [25] uses entity and node embedding training with 2D structured data into user-item relations, and extract useful fea-
convolutional neural networks. This architecture is designed in a tures from the KG [37]. Based on how these works employ the
way to address knowledge graph embedding for different types KG information, they are grouped into three categories:
of graphs. Embedding drop-out is one of the novelties of this embedding-based methods, path-based methods, and unified
method used to spread the features among different embedding methods [37]. The embedding-based methods generally use the
values rather than putting everything under specific values and information from the KG directly to enrich the representation of
indices. Link prediction is one of the tasks that this approach used items or users [37]. As mentioned in the background, in order to
to train the embeddings, which is a very important task in various exploit the KG information, knowledge graph embedding(KGE)
use cases. In case of recommendation systems, predicting the pos- algorithms need to be used to encode the KG into low-rank embed-
sibility of a link between different nodes is more like the recom- ding. All KGE algorithms can be divided into two classes [38]:
mendation itself. The score function of this technique is translation distance methods, such as TransE [29], TransH [39],
summarized in Eq. (1): TransR [40], TransD [41], etc., and semantic matching methods,
such as DistMult [42].
  h   i  
f conv E ¼ f v ec g es ; r p  X W e0 ð1Þ In path-based approaches create a user-item graph and use the
entity connectivity patterns in the graph to make recommenda-
tions [43].
where g is a non-linear activation function,  is a linear convolution In [26] Wang et al. proposed a news recommendation system
operator, v ec shows 2D reshaping. Also, f is the logistic sigmoid called DKN. This method models the news by combining the tex-
function. es , eo and r p denotes the embeddings of the subject, pred- tual embedding of sentences learned via Kim CNN [45] and embed-
 
icate, and object. Finally, es and r p are 2D reshaping of es and r p , ding of entities in news content at the knowledge-level by TransD
respectively. [25].
ComplEx [28] uses complex valued embeddings to provide Huang et al. [46] represented a KSR framework for sequential
knowledge graph embedding. Using such values can handle binary recommendations. In KSR, they used a GRU network with a
relations in embedding vectors and symmetric and asymmetric knowledge-enhanced key-value memory network (KV-MN) to
relations can also be handled as well. With regard to other meth- model comprehensive user preference from sequential interaction
ods of knowledge graph embeddings, this approach is much sim- [49]. The network of GRU utilizes sequential user preferences as
pler because it only uses dot-product operations which makes it input, while the KV-MN module uses knowledge base information
a good candidate for fast knowledge graph embedding. The scoring that learned with TransE to model the user’s feature-level prefer-
function of this approach is defined as Eq. (2): ence [46].
  In [47] Zhang et al. represented CFKG. In this method, they cre-

f ComplEx ¼ Re r p ; es ; eo ð2Þ ate a user-item knowledge graph. In this graph, behaviors of user
like purchase and mention are regarded as one relation type
between entities, and multiple types of item side information as
here Re indicates taking the real part of a complex value.
review, brand, category, etc. are contained [41]. This model, defines
TransE [29] is a translational distance model that uses
a function to measure the distance between two items by learning
translation-based transformation for embedding knowledge graph
the embedding of items and relationships in the graph.
entities. This approach uses a better approach compared to other
Wang et al. [48] proposed SHINE, which presented the celebrity
non-deep learning-based approaches. TransE uses a multi-
recommendation as a task of predicting sentiment links between
relational approach to create embeddings. Multi-relational
entities in a graph [10].
approach means taking into account different entity types and
Yang et al. [49] represented the movie recommender system.
nodes with same regard to created unified embeddings. In this
They used a model based on GAN called KTGAN. In the first step,
method, the score function uses L1 or L2 norm to measure the sim-
KTGAN incorporating the Metapath2Vec model [50] on the KG of
ilarity between the embedding of the subject. In Eq. (3) shown the
movies by learns the knowledge embedding vector for movie vec-
scoring function. In this equation, the number of entities is indi-
tor, and the tag vector embedding by the Word2Vec model [51] on
cated by n.
attributes of movie.
  Wang et al. [63] to aggregate information from knowledge
f TransE ¼ es þ r p  eo n ð3Þ graph employed graph attention networks. They used high order
connectivity in knowledge graph.
3
Z. Shokrzadeh, M.-R. Feizi-Derakhshi, M.-A. Balafar et al. Ain Shams Engineering Journal 15 (2024) 102263

Most embedding-based methods construct KGs with various Table 2

kinds of side information of item to enrich the item representation, Some of the related articles studied.

and such information can be used to more accurately model the Ref KG Embedding-based method Type of recommend
user representation [10]. Some models construct user-item graphs p p
[11] Item recommender
p p
by introducing users to the graph, which can directly model the [35] service recommendation
p p
user tastes [10]. The core of embedding-based methods is entity [37]
p p
service recommendation
embedding, and some papers modify embedding with GAN [49] [26] News recommendation
p p
[46] Sequential recommender
or BEM [52] for better recommendation. The methods of p p
[48] Celebrity recommender
Embedding-based use the information in the graph structure p p
[49] Movie recommender
p p
intrinsically. [63] Item recommender
Recently, the producing of recommendations with knowledge
graphs as side information has attracted remarkable attention. In
fact, with accurate presentation the information such as profiles researchers found this weakness in traditional and modern solu-
of users, tags, items and relationships of them, helps extract useful tions and decided to look for a suitable solution. Therefore, they
knowledge regarding the target user’s interests. tried to use the strengths and weaknesses of previous studies
In addition to the advantage of interpretability and accuracy, and presented an efficient method to provide high accuracy in sug-
another advantage of recommendations based on KG is that this gestions. Table 2 shows a number of related articles studied.
kind of side information can be naturally incorporated into recom-
mendation systems for various applications. To demonstrate the
effectiveness of KG as side information, recommendation systems 3. Research method
based on KG have been evaluated on datasets under different
scripts [10]. These results are shown in Table 1. The aim of this research is to design a neural collaborative filter-
Knowledge graphs provide additional information to solve ing model that is able to recommend multi-items such as user, tags
problems when faced with collaborative and content-based filter- and sources. The overall structure of the proposed method is
ing methods. On the other hand, it is possible to increase the accu- shown in Fig. 1. The steps taken for this method includes:
racy and variety of recommendations in the system with the help
of the semantic relations of the KG [54]. 1- constructing knowledge graph from dataset
Usually, the combination of knowledge graphs with collabora- 2- pre-trained knowledge graph embeddings
tive filtering is one of the most common methods. In this article, 3- collaborative filtering
using the knowledge graph, the embedding technique, the neural 4- classification
networks and collaborative filtering engender a solution to
increase the accuracy of the recommendation system has been In the first step, our solution presents information in the form of
presented. a knowledge graph. The main advantage of this representation of
Studying and reviewing the research done in the field of recom- data is to help address the ambiguity and redundancy problems.
mender systems, some of which were briefly reviewed in this sec- In the next step, the model takes advantage of the knowledge
tion, it is concluded that most of these systems suffer from graph representation learning technique to learn representations
insufficient accuracy in providing recommendations. This low of graph entities. Subsequently, a neural collaborative-filtering
accuracy is caused by a misdiagnosis of communication between network were presented to fine-tune pre-trained vectors for a rec-
users, items and sources. In other words, if users’ interests can be ommendation task. The final step includes a neural collaborative-
correctly identified and explicit and implicit connections between filtering framework that trains a neural network for a classifica-
users and tags and sources can be recognized correctly, recommen- tion task without additional weights. In another scene, vector rep-
dations with higher accuracy will definitely be obtained. Review- resentations of items generated from knowledge graph
ing the methods and solutions presented in various articles, the embedding methods are used as input features of the neural col-
laborative network. In order to capture items’ representations,
our method is comprised of two different neural networks: one
for node representation, and the other one for learning a shared
Table 1 representation between items. Then, the task of recommendation
A collection of different datasets in different application scenarios [13].
is modeled as a classification problem. However, as far as the pro-
Scenario Dataset posed structure includes four different stages, these phases were
Movie MovieLens-1 M divided into two groups and their procedures were described in
MovieLens-20 M detail in the following two subsections. The first subsection pro-
DoubanMovie vides information about constructing a knowledge graph and
MovieLens-100 K
knowledge graph embedding. In contrast, the characteristic of
Book DoubanBook
Amazon-Book the neural network and the classification procedure are described
DBbook2014 in the second subsection.
Book-Crossing
IntentBooks
News Bing-News 3.1. Constructing knowledge graph and learning embeddings
Product Alibaba Taobao
Amazon Product data Modern deep-learning architectures require a self-supervised
POI Yelp challenge
training before fine-tuning the dataset. In order to perform this
CEM
Dianping-Food
task for a recommendation system, a knowledge graph-based node
Music Last.Fm embedding architecture was used to self-supervise the model on
KKBox this data. The data is first transformed into a knowledge graph
Social Platform Weibo structure containing three different types of nodes: Users, Tags,
MeetUP
and Sources. Relations between these entities are created by using
DBLp
the dataset relations which, are as follows:
4
Z. Shokrzadeh, M.-R. Feizi-Derakhshi, M.-A. Balafar et al. Ain Shams Engineering Journal 15 (2024) 102263

Fig. 1. Proposed architecture for source recommendation.

-User-User Relation: User A follows User B or vice versa. 3.2. Neural collaborative filtering and classification
-User-Tag-Source Relation: User A used Tag T as Keyword for
Source S. The next step of the model is a neural collaborative filtering
Putting all of these relations in a single graph shows a unified framework that treats recommending as a classification task. The
version of data in form of a graph with these relations. generated embeddings obtained from previous step work as
Organizing data as a user-tag-source knowledge graph provides weights for this neural network. That means that instead of train-
more contextual semantic information, which can play a signifi- ing the neural network with randomly initialized weights, our pro-
cant role in dealing with the polysemy problem. Moreover, the posed approach uses the pre-trained vectors obtained from
multi-hop neighbors in this knowledge graph not only alleviate knowledge graph embeddings.
cold start, but also, they help distinguish between ambiguous enti- Apart from pre-trained and extracting vectors from the pre-
ties. To show how the created knowledge graph is capable of solv- trained model, it is necessary to fine-tune it. For fine-tuning, a neu-
ing these issues, consider a record in Delicious dataset that User U1 ral collaborative filtering network is used by only focusing on these
uses Tag }apple}, which is a polysemous tag. The relation between vectors and fine-tuning them. In order to avoid extra parameters
Tag-Source in the knowledge graph contributes to understand and focus solely on the vectors that were pre-trained from knowl-
the differences between concepts. Sources with titles like ‘‘iPad”, edge graph embedding approaches, the architecture, shown in
‘‘iPhone”, ‘‘Mac”, ‘‘Installing”, ‘‘Ecamm” and ‘‘Android” indicate that Fig. 2, was employed. In the input layer, the user, tag and source
tag ‘‘Apple” is a technology company. Therefore, our system cannot are one-hot encoded. We utilized just the identity of a user, tag
confuse ‘‘Apple”, the brand, with ‘‘apple” fruit. Moreover, due to and source as input features and convert them to binarized sparse
the fact that, a part of graph relies on the user-user interaction that vectors with one-hot encoding. Employing this kind of representa-
comforts to make recommendations for new users who just follow tion for inputs help our model easily adapt to cope with the cold-
their friends and there are no histories about their tagging behav- start problem by using content features to represent users, tags
iors. Additionally, the existence of user-user, user-tag and tag- and sources.
source interactions in our knowledge graph enables to deal with The embedding layer is above the input layer. The embedding
the issue of cold start for new tag or source by inferring the layers include user, tag and source vectors obtained from the
connections. knowledge graph embedding technique. In the context of the
After constructing our desirable knowledge graph, the next step latent factor model, the obtained user (item) embedding can be
of our proposed method is to map the component of this graph into inferred as the latent vector. Afterwards, in this article, there is a
low-dimensional vectors. The contemporary achievements in the neural architecture that is termed as neural collaborative filtering
knowledge graph to represent a knowledge graph’s entities and layer. This framework maps the latent vectors to recommend a
relations as continuous vector spaces are knowledge graph embed- source. The first layer of this neural network is a multiplication
dings. The key property of knowledge graph embeddings is to cap- layer that executes element-wise multiplication on the user and
ture the semantics and inherent structures of entities and relations tag embeddings. This layer recognizes the latent structures of
[37]. To achieve the goal, three different knowledge graph embed- user–tag interactions. Subsequently, the next layer is the dot pro-
ding techniques were used, including ConvE [48], ComplEx [27], duct of user-tag matrix multiplication and source vector which,
and TransE [28]. The details of how these methods work were learns the interaction between user, tag and source latent features.
described in background section. The result of this layer connects to the last layer where the output
The key idea behind the second phase of Fig. 1 is to leverage the needs to be constrained in the range of [0, 1]. In order to easily
productive facts in the knowledge graph to enhance the represen- achieve this goal, the sigmoid function was used as the activation
tations of users, tags and sources. Also, these embeddings allow to function for the output layer.
find the top similar tags, users or sources. That means that this part As mentioned, the immediately following step is a binary classi-
of our model guarantees that entities of the same type to stay close fication which is the main task that is why embeddings are fined-
to each other in the embedding space. tuned. This classification is used to ensure that each User-Tag-
5
Z. Shokrzadeh, M.-R. Feizi-Derakhshi, M.-A. Balafar et al. Ain Shams Engineering Journal 15 (2024) 102263

Fig. 2. Architecture of neural network.

! !  
Source vector is fine-tuned in a way that: if a user tagged source S u  t ¼ ue1  t e1 ; ue2  t e2 ;    ; ued  ted ð6Þ
with tag T then the result is 1 and 0 otherwise. According to this
analogy, the pre-trained vectors are fine-tuned and can be put into Dot-product multiplication ensures that the resulting vector
a vector search database. from previous equation is combined with the source. This opera-
Training a binary classification on recommendation-based data tion ensures the presence of user, tag and source embeddings in
is also one of the important novelties of this work that is accom- our collaborative filtering model which engenders an accurate rec-
plished by creating negative data. In order to create data for each ommendation system. Eq. (7) shows it. From this equation, it is
of the three recommendation tasks (User recommendation, Tag clear that we combined all the elements of embedding vectors
recommendation, and Source recommendation) syntactic data cre- one by one.
ation was used by randomly sampling (uniform distribution) over
users, tags, and resources. For example, in the case of user recom-   X
d
! ! !
mendation, there is a user-tag-source triplet (U, T, S) that has hap- u  t  s ¼ uei tei sei ð7Þ
i¼0
pened before. User U has tagged source S with tag T. It is accepted
as positive sample, and negative samples were created by fixing S
A sigmoid activation is applied to make sure the range of output
and T and changing U with other users that are not in the dataset.
is in [0, 1]. Eq. (8) shows whole neural network matrix operations
The same applies to tag and source recommendations.
which is indicated by nn.
For more clarification, as far as, operations used in neural net-
work architecture as mathematical expressions were presented, !
  X
d
the following equations have presented. One-hot encoded vectors ! !!
nn u ; t ; s ¼ sigmoid uei t ei sei ; ½0; 1 ð8Þ
regarding a triplet, user, tag, and source are created from their
i¼0
respective ids. Some examples are shown in Eq. (4).
onehotencoder
ðuid ; tid ; sid Þ ! uoh ½0010    0n
toh ½0000    1m ð4Þ 4. Experimental evaluation
soh ½0001    0k
In this section, an overview of the dataset, which applied in the
where uoh , t oh , soh are one-hot vectors user, tag, and source. In this proposed method, will be provided. The dataset used in this
Equation n, m, k show the number of users, tags, and sources, research is real world. This dataset is widely experimented by
respectively. Each one-hot vector is the input of the proposed archi- social recommendation systems. In the following, the number of
tecture. In this architecture, three different dense layers are trans- baseline algorithms will be introduced to compare with the pro-
forming one-hot vectors into embeddings which, for all of the posed approach against them. Next, the used parameter settings
triplet items are in the same size d. Eq. (5) show these embeddings are described. Finally, the metrics hired to evaluate our proposed
that are also pre-trained from knowledge graph embedding. algorithm will be discussed.
!  
uoh ! u ! ue1 ;    ; ued
4.1. Dataset
!  
t oh ! t ! t e1 ;    ; t ed One of the most important parts of any recommendation sys-
tem is data collection. If it was done in the accurate and regular
!  
soh ! s ! se1 ;    ; sed ð5Þ way, data analysis can be done with great speed and accuracy
[55]. In this system, among the published valid datasets, the widely
! ! !
where u , t , s are embedding vectors and uei ; t ei andsei vector val- used Hetrec2011-Delicious-2 k dataset such as [5657] was
ues.. Vector multiplication between elements of user and tag vec- employed in the experiments, which contains 53,388 tags, 1867
tors are applied that Eq. (6) shows it. users, and 69,226 sources collected from Delicious.com and pub-
6
Z. Shokrzadeh, M.-R. Feizi-Derakhshi, M.-A. Balafar et al. Ain Shams Engineering Journal 15 (2024) 102263

Table 3 lished in [51]. In this dataset, users are connected in a social net-
User contacts. work created from Delicious interactions [58]. The sample data of
UserID UserID this relations shown in Table 3. In the data set each user has their
22 102 own tags, sources, and tag assignments [58]. Table 4 show the sam-
1432 54 ple data.
298 8 In this article, the trained and test data as 80% and 20% of the
. . ... . . ... total dataset were determined. Recommendations are produced
. . ... . . ...
based on the known information in the training set, and then to
evaluate the performance of the proposed system test set is used.
All experiments implemented an Intel(R) Core i7 computer with
Table 4
2.67 GHz CPU and 16.00 GB RAM. Google Colab has been used to
User–tag- source.
train the most of the models and the AmpliGraph library has been
UserID TagID SourceID used to implement the knowledge graph representation learning
22 7434 4 techniques.
1432 765 8128 In order to make the information of the dataset clearer, the
54,643 98,195 87
visualization of word cloud of tags and the knowledge graph are
. . ... . . ... . . ...
. . ... . . ... . . ... presented in Figs. 3, 4. In Fig. 3, each tag is picturized with its fre-
quency. The generated knowledge graph in Fig. 4 is plotted with
Gephi software. This graph includes 111,987 nodes and 746,404
edges. To have a great depth of knowledge of the dataset, the con-
tents of the tag data and construction graph are visualized.

4.2. Parameter settings

Parameters of the approach has been selected using the grid

search and the parameters that have been searched are learning
rate and the embedding size of the knowledge graph embedding
method (pre-trained vectors). Best parameters that have been
selected for this work are 0.001 for adam optimizer learning rate
and 128 for vector size.
In total, our approach has 14,334,336 parameters that are pre-
trained using knowledge graph embedding. All of these parameters
Fig. 3. Word cloud of tags. fine-tuned on the down-stream task and none of them freezed dur-
ing training.

Fig. 4. The constructed knowledge graph from dataset.

7
Z. Shokrzadeh, M.-R. Feizi-Derakhshi, M.-A. Balafar et al. Ain Shams Engineering Journal 15 (2024) 102263

4.3. Evaluation metrics integrated algorithm on the graphs and calculates the similarity
of the users by the graphs [60].
The evaluation metrics used in the literature are top-k preci-
FURG: Based on the user’s interest, a user similarity graph is
sion, recall, and F1-score. The same metrics were also used to make used to recommend a friend. First, in this method, LDA is
the results comparable to previous works. This metrics are the applied to identify topics of interest to users in order to reduce
most popular and powerful metrics to evaluate quality of recom- the problems of tag redundancy and data sparseness. Then, an
mender systems [53]. The acceptable range for all three criteria interest-based user similarity graph is constructed by various
is between 0 and 1. One of the challenges for researchers is to views of users’ favorite topics, shared collections, and co-
design approaches that can increase these values. These metrics annotated tags. Finally, it provides interest-based recommenda-
are described in Eqs. (9), (10), and (11) according to [59]. tions by graph mining [64].
jlr \ ltj
P¼ ð9Þ
jlrj
In the following paragraphs, the baselines for tag recommenda-
jlr \ ltj tion are listed:
R¼ ð10Þ
jltj

pR
CFA: It Uses a sparse auto-encoder to latent representations of

F1 ¼ 2  ð11Þ user profiles and applied user-based CF for recommendation
pþR
[56].
In Eqs. (9) and (10), lr is the recommended items list and lt is
DSPR: In this method, MLPs are used to extract user and item
the list of items being tested. representations. Here, MLPs with common parameters were
suitable for tag-based feature processing [57].
4.4. Baselines
DeepFM: factorization machines [65] and deep neural network
combined for feature learning. Then recommendations of the
To evaluate the efficiency of the proposed method for sources, tags are done [66].
users and, tags recommendations, the baseline algorithms were
PinSage: this technique deploys GraphSage [67] on industrial
divided into three groups. While one class includes previous stud- application. Then extracts node representations through non-
ies that recommend source, the other reveals the researches pro- spectral graph convolution [27].
posed models for user and tag recommendation. The following
models recommend source: 4.5. Evaluation results

CCS method: The basis of this method is the use of cosine sim- In this section, the efficiency of the proposed model will be
ilarity based on clustering. In other words, to model users and investigated on the three different recommendation items: source,
sources as attributes vector based on the hierarchical clustering user and tag according to recall, precision and F1-score metrics. It
method has been used. Then content-based filtering based on should be noted that top-k in evaluation metrics shows number of
cosine similarity has been used to provide recommendations relative items that are used for validation of approach. In addition
[60]. to this, the effect of the number of embedding dimensions will be

ACF method: In this method, the automatic encoder is used in examined on the performance. Then, in order to make more judg-
the content-based method. An automated encoder is used to ments about our solution, the clusters created by embeddings of
obtain summary introductions from user profiles along with users, tags, and sources are going to be visualized.
recommendations based on collaborative filtering. This method
has shown with different tests on collaborative filtering and 4.5.1. Source recommendation results
with different number of hidden layers that deeper architec- The three evaluation criteria, recall, precision and F1-score
tures can provide better performance if the depth of the neural show that the proposed method have been significantly improved,
network is properly set [56]. according to the known algorithms. The results of source recom-

CCF or CF based on clustering: This method is similar to CCS but mendation based on different criteria have shown in Table 5, 6
here CF method based on user is used to provide recommenda- and 7. According to the results provided in Tables 5-7, it can be
tions [61].

PMF method: This technique is also based on collaborative fil-
tering. It uses a user rating matrix. This method works with Table 5
the assumption that similar sets of items, if ranked by users, Source Recommendation Results (Recall in %).
means that those users have similar preferences [62]. Models R@5 R@15 R@30

KGAT: this is state-of-the-art knowledge-based model, per- CCF 0.439 1.051 1.499
forms knowledge-aware attentive graph convolution in KG for ACF 0.590 1.209 1.917
high-order modeling of relation [63]. CCS 0.938 2.271 3.774
PMF 1.302 2.851 4.988
KGAT 12.121 18.080 22.654
In the succeeding paragraphs, the baselines for user recommen-
Proposed method(TransE) 64 12.3062 18.1397 23.4488
dation are listed: 128 13.0062 18.7049 22.7944
256 11.1136 17.6240 23.7954

Tag-co: In this algorithm used the cosine similarity. It calculates Proposed method(ComplEx) 64 13.0350 19.1087 23.3158
user similarity based on tags [64]. 128 13.0450 19.1630 23.5312
256 13.2771 19.2323 23.8975

Item-co: This method similar to Tag_co, with the difference that Proposed method(ConvE) 64 12.0131 17.0943 22.0237
in this method calculates user similarity based on items [64]. 128 12.7751 18.3415 23.4005

Tri-graphs: This method provides a recommendation algorithm 256 11.8413 17.6472 21.8941
with diffusion based on bipartite graphs. Then it uses the

8
Z. Shokrzadeh, M.-R. Feizi-Derakhshi, M.-A. Balafar et al. Ain Shams Engineering Journal 15 (2024) 102263

Table 6 Table 9
Source Recommendation Results (Precision in %). User Recommendation Results (Precision in %).

Models P@5 P@15 P@30 Models P@5 P@10

CCF 0.913 0.757 0.597 Tag-co 10.3696 8.0450
ACF 1.120 0.909 0.736 Item-co 11.3658 8.0664
CCS 2.397 1.903 1.564 Tri-graphs 11.3658 9.6668
PMF 9.157 7.467 6.784 FRUG 14.5581 11.0927
KGAT 14.01 13.88 11.98 Proposed method(TransE) 64 15.0000 13.0012
Proposed method(TransE) 64 15.3333 14.0033 13.6533 128 15.0321 13.0021
128 16.6897 14.0011 13.7778 256 16.0210 14.1200
256 15.5556 14.7963 13.5852 Proposed method(ComplEx) 64 14.3333 13.0313
Proposed method(ComplEx) 64 15.6667 14.1111 13.3704 128 14.0100 13.0000
128 15.3514 14.7027 13.6396 256 14.0110 13.0100
256 16.7846 14.9487 13.6610 Proposed method(ConvE) 64 14.0010 13.701
Proposed method(ConvE) 64 16.0120 14.7667 13.5556 128 14.0110 13.012
128 15.6667 14.5556 13.2302 256 15.0010 14.0901
256 16.0000 14.6800 13.2500

Table 10
User Recommendation Results (F-Measure in %).
Table 7 Models F@5 F@10
Source Recommendation Results (F1 in %).
Tag-co 7.8497 8.8354
Models F1@5 F1@15 F1@30 Item-co 8.6039 8.8593
Tri-graphs 8.6741 10.6016
CCF 0.593 0.880 0.854
ACF 0.791 1.038 1.064 FRUG 10.2450 11.5488
CCS 1.349 2.070 2.205 Proposed method(TransE) 64 14.4246 15.5360
PMF 2.280 8.791 8.803 128 14.1059 15.4043
256 14.9105 16.3400
KGAT 12.998 15.704 15.672
Proposed method(ComplEx) 64 13.2435 15.3512
Proposed method(TransE) 64 13.6540 15.8053 17.2580
128 13.5311 15.1846
128 14.6195 16.3704 17.1746
256 13.2209 15.5731
256 12.9647 16.0868 17.2959
Proposed method(ComplEx) 64 14.2302 16.2340 16.9950 Proposed method([ConvE] 64 13.3389 15.8578
128 14.1045 16.6391 17.2693 128 13.2630 15.4468
256 14.8262 16.8221 17.3843 256 13.9602 16.1735
Proposed method(ConvE) 64 13.7272 15.8455 16.7819
128 14.0739 16.2307 16.9035
256 13.6101 16.0274 16.5090

Table 11
tag Recommendation Results (Recall in %).

concluded that presenting data in form of this knowledge graph, Models R@10 R@20
and using collaborative neural network make source recommenda- CFA 3.69 3.98
tion more accurate. Additionally, it is obviously that in all most DSPR 12.19 13.49
cases ComplEx of our approach using 256 embedding sizes pro- Deep FM 12.41 13.87
PinSage 12.12 13.25
vides best results on top-k precision, Recall, and F1-score.
Proposed method(TransE) 64 14.1884 16.0783
128 14.7153 17.0409
256 15.5551 16.0366
4.5.2. User recommendation results Proposed method(ComplEx) 64 14.3212 16.4979
Tables 8-10 illustrate the obtained results for user recommen- 128 15.9978 16.8458
256 14.9218 16.9050
dation. The results demonstrate that our solution could signifi-
Proposed method(ConvE) 64 15.7249 16.6398
cantly improve the performance in terms of recall, precision and 128 15.8115 16.5984
F1-score. 256 15.5551 16.6483

Table 8
User Recommendation Results (Recall in %).

Models R@5 R@10 There are three major reasons why our system can suggest
Tag-co 6.3152 9.7991
users in accurate manner. Firstly, our constructed knowledge
Item-co 6.9220 9.8252 graph relies on user-user, user-tag and tag-source interactions.
Tri-graphs 7.0133 11.7367 Secondly, knowledge graph embedding techniques infer and ana-
FURG 7.9035 12.0442 lyze these connections when they learn items’ representations.
Proposed method(TransE) 64 13.8918 19.2987
Finally, using the knowledge graph embedding vectors along with
128 13.2872 18.8954
256 13.9439 19.3883 neural collaborative filtering could provide better understandings
Proposed method(ComplEx) 64 12.3077 18.6760 of items specifically users. In other words, learning users’ features
128 13.0839 18.2517 along with sources and tags guaranties better user recommenda-
256 12.5152 19.3939 tion. Furthermore, it is clear that, user recommendation employing
Proposed method(ConvE) 64 12.7366 18.8205
128 12.5909 19.0027
TransE (with 256 embedding sizes) as graph representation learn-
256 13.0545 18.9798 ing, can engender better results on top-k precision, recall, and F1-
score.

9
Z. Shokrzadeh, M.-R. Feizi-Derakhshi, M.-A. Balafar et al. Ain Shams Engineering Journal 15 (2024) 102263

Table 12
Tag recommendation results (precision in %).

Models P@10 P@20

CFA 3.51 4.36
DSPR 12.12 14.69
Deep FM 11.90 12.71
PinSage 14.16 10.27
Proposed method([TransE] 64 16.3051 13.3051
128 15.3750 14.1406
256 15.0159 12.8730
Proposed method([ComplEx] 64 16.2648 14.2085
128 16.5500 14.2106
256 16.3651 13.1429
Proposed method([ConvE] 64 16.2899 13.2464
128 15.7532 13.6623
256 15.9322 13.5424

Fig. 5. User clusters: t-SNE dimensionality reduction / K-Means Clustering with 10

clusters.

Table 13
Tag recommendation results (f1 in %).

Models F1@10 F1@20

CFA 0.5865 0.5265
DSPR 2.5443 2.2772
Deep FM 1.0322 1.0101
PinSage 13.9365 13.981
Proposed method(TransE) 64 15.1733 14.5608
128 15.0379 14.8387
256 15.2807 14.2817
Proposed method(ComplEx) 64 15.2312 15.2679
128 16.2692 15.4164
256 15.6102 14.7884
Proposed method(ConvE) 64 16.0024 14.7504
128 15.7823 14.9879
256 15.7414 14.9356

4.5.3. Tag recommendation results

The tag recommendation results can be shown in Tables 11-13. Fig. 6. Tag clusters: t-SNE dimensionality reduction / K-Means Clustering with 10
It is obviously that in most cases ComplEx of our approach using clusters.
128 embedding sizes provides best results on top-k precision and
F1-score measures. TransE with 128 embedding sizes on top 20
recall measure, imply best performance.
The main reason that how our system can precisely make sug-
gestions is that the relation between tag-source in our created
graph contributes to understand the differences between concepts
and our system can better distinguish sources of ambiguous tags.
Moreover, the user-user interaction enables our system to make
recommendations for new users who just follow their friends
and there are no histories about their tagging behaviors.

4.5.4. Visualization task

In this section, the extracted embedding was used to cluster
users, tags and sources. For this goal, t-SNE dimensionality reduc-
tion and K-Means clustering method with 10 cluster on the embed- Fig. 7. Source clusters: t-SNE dimensionality reduction / K-Means Clustering with
dings obtained from ComplEx with size 64 were applied. Since 10 clusters.

lower dimensions cannot offer significant features than high

dimensional feature space, the size 64 for the embedding dimen-
sions was considered to evaluate the performance of the proposed precisely recognize users’ preferences and eventually do the spec-
method in a low dimensional feature vector. The results of these ified recommendation.
clusters shown in Figs. 5-7. The important goal of this investigation For more investigation, the community detection task on the
is to discover how to do clustering with created user, tag and knowledge graph were carried out. These communities are
source embeddings. The embeddings derived from our proposed detected using Louvain algorithm. The modularity value is 0.413.
learning model handles the problem of high dimensionality in an Normally, regardless of the semantic relationship, and only from
effective manner, hence, more accurate clusters are formed. A great the knowledge graph and communications of the nodes, communi-
clustering of tags leads to more precise recommendations. By look- ties have obtained, that is shown in Fig. 8. Also, Fig. 9 indicates the
ing at how users tag sources, the recommendation system is able to size of each community.
10
Z. Shokrzadeh, M.-R. Feizi-Derakhshi, M.-A. Balafar et al. Ain Shams Engineering Journal 15 (2024) 102263

Fig. 8. Communities detected from the knowledge graph.

Fig. 9. Size of communities.

Two main observations can be conducted from the figure. 5. Conclusion

Firstly, by applying the community detection algorithm on the
constructed knowledge graph, a group of tags, users, and sources In this paper, the researchers are motivated to formulate the
(bookmarks) that are similar to each other can be generated in recommendation task as a classification problem by constructing
communities. The second monitoring point is that the communi- a knowledge graph from user, tag, and source entities. The key idea
ties illustrate how ambiguity and redundancy can be addressed behind using embeddings is to accurately find similar items. Fun-
with the help of multi-hop neighbors in our constructed knowl- damentally, the hypothesis declares that a knowledge graph con-
edge graph. structed from different types of items aids in comprehending

11
Z. Shokrzadeh, M.-R. Feizi-Derakhshi, M.-A. Balafar et al. Ain Shams Engineering Journal 15 (2024) 102263

how traits are related to one another. Moreover, neural collabora- SIGKDD international conference on Knowledge discovery and data mining.
2009, June: p. 397–406.
tive filtering assets to learn accurate representations of items. To
[9] Chen X, Jia S, Xiang Y. A review: Knowledge reasoning over knowledge graph.
capture statistical dependencies and provide more contextual Expert Syst Appl 2020;141:112948.
semantic information, the problem was formulated in terms of [10] Guo Q, Zhuang F, Qin C, Zhu H, Xie X, Xiong H, et al. A survey on knowledge
knowledge graph construction. The connections in the knowledge graph-based recommender systems. IEEE Trans Knowl Data Eng 2020.
[11] Zhang W, Gu T, Sun W, Phatpicha Y, Chang L, Bin C. Travel attractions
graph contributes to understand the differences between concepts. recommendation with travel spatial-temporal knowledge graphs. In:
Moreover, a part of graph constructs based on the user-user inter- International conference of pioneering computer scientists, engineers and
action that make it possible to suggest tags and sources to new educators. Singapore: Springer; 2018. p. 213–26.
[12] Pla Karidi D, Stavrakas Y, Vassiliou Y. Tweet and followee personalized
users with no histories about their tagging behaviors. Additionally, recommendations based on knowledge graphs. J Ambient Intell Hum Comput
these interactions enable this model to deal with the cold-start 2018;9(6):2035–49.
problem for new tag or source. Broadly, the researchers argue [13] Yan C, Chen Y, Zhou L. Differentiated fashion recommendation using
knowledge graph and data augmentation. IEEE Access 2019;7:102239–48.
how the model can ease ambiguity and redundancy issues that [14] Wang X, Wu H, Hsu CH. Mashup-oriented API recommendation via random
arise as a result of the lack of contextual semantics for tag and walk on knowledge graph. IEEE Access 2018;7:7651–62.
source recommendation. [15] Christoforidis G, Kefalas P, Papadopoulos A, Manolopoulos Y. Recommendation
of points-of-interest using graph embeddings. In: In 2018 IEEE 5th
The proposed method uses simple matrix operations and learns International Conference on Data Science and Advanced Analytics
over them with no need for extra parameters on model Utilization (DSAA). IEEE; 2018. p. 31–40.
of pre-trained vectors from knowledge graph embedding ensures [16] Hu S, Tu Z, Wang Z, Xu X. A POI-sensitive knowledge graph-based service
recommendation method. In: In 2019 IEEE International Conference on
that the pre-trained vectors are more adapted to the structure of
Services Computing (SCC). IEEE; 2019. p. 197–201.
graph. These embeddings are semantic representations of the [17] Deng, W., Zhu, P., & Ma, J. (2018, June). Enhancing group recommendation by
graph for each node regardless of its type. The size of these vectors knowledge graph. In 22nd Pacific Asia Conference on Information Systems
is all the same for any entity type (User, Tag, and Source). (PACIS 2018) (p. 214).
[18] Cui P, Wang X, Pei J, Zhu W. A survey on network embedding. IEEE Trans
Additionally, the simplicity of our approach in combining pre- Knowl Data Eng 2018;31(5):833–52.
trained vectors in the architecture makes it much easier to use [19] Oramas S, Ostuni VC, Noia TD, Serra X, Sciascio ED. Sound and music
any other knowledge graph embedding and keep the training time recommendation with knowledge graphs. ACM Trans Intell Syst Technol (TIST)
2016;8(2):1–21.
much smaller. On the other hand, instead of using a very huge [20] Lu C, Laublet P, Stankovic M. Travel attractions recommendation with
model with complexities on different sides (from number of layers knowledge graphs. In: European Knowledge Acquisition
to types of layers), only a simple neural collaborative neural net- Workshop. Cham: Springer; 2016. p. 416–31.
[21] Hogan A, Blomqvist E, Cochez M, d’Amato C, Melo GD, Gutierrez C, et al.
work that includes vector operations on entity vectors is utilized Knowledge graphs. ACM Comput Surveys (CSUR) 2021;54(4):1–37.
which makes model to fine-tune embeddings for the downstream [22] Sun Z, Yang J, Zhang J, Bozzon A, Huang L-K, Xu C. Recurrent knowledge graph
task faster and with less data. All of these novelties in the method- embedding for effective recommendation. In: Proceedings of the 12th ACM
Conference on Recommender Systems. 2018, p. 297–305.
ology helps the approach to overcome much complex methods. [23] Fan H, Zhong Y, Zeng G, Ge C. Improving recommender system via knowledge
Binarization of model output and creating negative samples in graph based exploring user preference. Appl Intell 2022:1–13.
order to create a balanced dataset also makes outputs more robust [24] Makarov I, Kiselev D, Nikitinsky N, Subelj L. Survey on graph embeddings and
their applications to machine learning problems on graphs. PeerJ Comput. Sci.
to data and concept drift.
2021;7.
In the future, the researchers aim to formalize the problem as a [25] Dettmers T, Minervini P, Stenetorp P, Riedel S. Convolutional 2d knowledge
dynamic knowledge graph. They reckon that dynamic knowledge graph embeddings. In Proceedings of the AAAI Conference on Artificial
graph construction is a trendy solution that may raise interesting Intelligence (Vol. 32, No. 1) 2018, April.
[26] Wang H, Zhang F, Xie X, Guo M. Dkn: Deep knowledge-aware network for
questions. So, they are eager to model the sequential dependencies news recommendation. In: Proceedings of the 2018 World Wide Web
over the user-tag, user-source and source-tag interactions in a Conference, ser. WWW ’18. Republic and Canton of Geneva, Switzerland:
sequence. In other words, in the future work, they will investigate International World Wide Web Conferences Steering Committee, 2018, p.
1835–44.
a dynamic knowledge graph-based approach that recommends [27] Ying R, He R, Chen K, Eksombatchai P, Hamilton WL, Leskovec J. Graph
items in a sequential form of interactions. convolutional neural networks for web-scale recommender systems. In:
Proceedings of the 24th ACM SIGKDD international conference on
knowledge discovery & data mining. 2018, July: p. 974–83.
[28] Trouillon T, Welbl J, Riedel S, Gaussier É, Bouchard G. Complex embeddings for
Declaration of Competing Interest
simple link prediction. In: International conference on machine
learning. PMLR; 2016. p. 2071–80.
The authors declare that they have no known competing finan- [29] Bordes A, Usunier N, Garcia-Duran A, Weston J, Yakhnenko O. Translating
cial interests or personal relationships that could have appeared embeddings for modeling multi-relational data. Adv Neural Inf Proces Syst
2013;26.
to influence the work reported in this paper. [30] Shokeen J, Rana C. A study on features of social recommender systems. Artif
Intell Rev 2020;53(2):965–88.
[32] Dai Y, Wang S, Xiong NN, Guo W. A survey on knowledge graph embedding:
References Approaches, applications and benchmarks. Electronics 2020;9(5):750.
[33] Xu M. Understanding graph embedding methods and their applications. SIAM
Rev 2021;63(4):825–53.
[1] Ahmadian S, Ahmadian M, Jalili M. A deep learning based trust-and tag-aware
[34] Chen Y, Zhou M, Zheng Z, Chen D. Time-aware smart object recommendation
recommender system. Neurocomputing 2022;488:557–71.
in social internet of things. IEEE Internet Things J 2019;7(3):2014–27.
[2] Chen YC, Hui L, Thaipisutikul T. A collaborative filtering recommendation
[35] Wang H, Wang Z, Hu S, Xu X, Chen S, Tu Z. DUSKG: A fine-grained knowledge
system with dynamic time decay. J Supercomput 2021;77:244–62.
graph for effective personalized service recommendation. Futur Gener Comput
[3] Ni J, Huang Z, Cheng J, Gao S. An effective recommendation model based on
Syst 2019;100:600–17.
deep representation learning. Inf Sci 2021;542:324–42.
[36] Wang H, Zhang F, Zhao M, Li W, Xie X, Guo M. May). Multi-task feature
[4] Hou YE, Gu W, Dong W, Dang L. A Deep Reinforcement Learning Real-Time
learning for knowledge graph enhanced recommendation. In: In The world
Recommendation Model Based on Long and Short-Term Preference. Int J
wide web conference. p. 2000–10.
Comput Intell Syst 2023;16(1):4.
[37] Mezni H, Benslimane D, Bellatreche L. Context-aware service recommendation
[5] Sun Z, Guo Q, Yang J, Fang H, Guo G, Zhang J, et al. Research commentary on
based on knowledge graph embedding. IEEE Trans Knowl Data Eng 2021.
recommendations with side information: A survey and research directions.
[38] Wang Q, Mao Z, Wang B, Guo L. Knowledge graph embedding: A survey of
Electron Commer Res Appl 2019;37:100879.
approaches and applications. IEEE Trans Knowl Data Eng 2017;29
[6] Mısır M, Sebag M. Alors: An algorithm recommender system. Artif Intell
(12):2724–43.
2017;244:291–314.
[39] Wang Z, Zhang J, Feng J, Chen Z. Knowledge graph embedding by translating
[7] Wang R, Jiang Y, Lou J. ADCF: Attentive representation learning and deep
on hyperplanes. in Twenty-Eighth AAAI conference on artificial intelligence,
collaborative filtering model. Knowl-Based Syst 2021;227:107194.
2014.
[8] Jamali M, Ester M. Trustwalker: a random walk model for combining trust-
based and item-based recommendation. In: Proceedings of the 15th ACM

12
Z. Shokrzadeh, M.-R. Feizi-Derakhshi, M.-A. Balafar et al. Ain Shams Engineering Journal 15 (2024) 102263

[40] Lin Y, Liu Z, Sun M, Liu Y, Zhu X. February). In Twenty-ninth AAAI conference [64] Wu BX, Xiao J, Chen JM. Friend recommendation by user similarity graph
on artificial intelligence: Learning entity and relation embeddings for based on interest in social tagging systems. In International Conference on
knowledge graph completion; 2015. Intelligent Computing; 2015, August. p. 375–86. Springer, Cham.
[41] Ji G, He S, Xu L, Liu K, Zhao J. July). Knowledge graph embedding via dynamic [65] Rendle S. Factorization machines. In 2010 IEEE International conference on
mapping matrix. In: In Proceedings of the 53rd annual meeting of the data mining, 2010, December, p. 995–1000, IEEE.
association for computational linguistics and the 7th international joint [66] Guo H, Tang R, Ye Y, Li Z, He X. DeepFM: a factorization-machine based neural
conference on natural language processing (volume 1: Long papers). p. network for CTR prediction, 2017. arXiv preprint arXiv:1703.04247.
687–96. [67] Hamilton W, Ying Z, Leskovec J. Inductive representation learning on large
[42] Yang B, Yih WT, He X, Gao J, Deng L. Embedding entities and relations for graphs. In: Advances in neural information processing systems. p. 30.
learning and inference in knowledge bases; 2014. arXiv preprint
arXiv:1412.6575.
[43] Hu B, Shi C, Zhao WX, Yu PS. July). Leveraging meta-path-based context for
top-n recommendation with a neural co-attention model. In: Proceedings of Zeinab Shokrzadeh was born in Tabriz, Iran, in 1975
the 24th ACM SIGKDD international conference on knowledge discovery & and received her B.S. degree from Isfahan University in
data mining. p. 1531–40. computer engineering. She received her M.S. degree in
[44] Tiddi I, Schlobach S. Knowledge graphs as tools for explainable machine computer engineering from Islamic azad university
learning: A survey. Artif Intell 2022;302:103627. sience and research branch. Her candidate Ph.D. student
[45] Chen Y. Convolutional neural network for sentence classification. University of is in Islamic azad university Urmia branch from
Waterloo); 2015. Master’s thesis. Department of Computer Engineering. She is a lecturer
[46] Huang J, Zhao WX, Dou H, Wen JR, Chang EY. June). Improving sequential in Department of Software Engineering, shabestar
recommendation with knowledge-enhanced memory networks. In: In The University.
41st International ACM SIGIR Conference on Research & Development in
Information Retrieval. p. 505–14.
[47] Zhang Y, Ai Q, Chen X, Wang P. Learning over knowledge-base embeddings for
recommendation; 2018. arXiv preprint arXiv:1803.06540.
[48] Wang H, Zhang F, Hou M, Xie X, Guo M, Liu Q. February). Shine: Signed
heterogeneous information network embedding for sentiment link prediction.
In: In Proceedings of the eleventh ACM international conference on web search
and data mining. p. 592–600. Mohammad-Reza Feizi-Derakhshi received his B.S. in
[49] Yang D, Guo Z, Wang Z, Jiang J, Xiao Y, Wang W. In: A knowledge-enhanced Software Engineering from the University of Isfahan.
deep recommendation framework incorporating GAN-based models. IEEE; Then he received his M.S. and Ph.D. in Artificial Intelli-
2018. p. 1368–73. gence from the Iran University of Science and Technol-
[50] Dong Y, Chawla NV, Swami A. metapath2vec: Scalable representation learning ogy. He is currently a faculty member at the University
for heterogeneous networks. In Proceedings of the 23rd ACM SIGKDD of Tabriz. His research interests include: natural lan-
international conference on knowledge discovery and data mining. 2017, guage processing, optimization algorithms, social net-
August, pp. 135–144. work analysis and intelligent databases.
[51] Mikolov T, Chen K, Corrado G, Dean J. Efficient estimation of word
representations in vector space; 2013. arXiv preprint arXiv:1301.3781.
[52] Ye Y, Wang X, Yao J, Jia K, Zhou J, Xiao Y, et al. November). Bayes embedding
(bem) refining representation by integrating knowledge graphs and behavior-
specific networks. In: Proceedings of the 28th ACM international conference
on information and knowledge management. p. 679–88.
[53] Colombo-Mendoza LO, Valencia-García R, Rodríguez-González A, Alor-
Hernández G, Samper-Zapater JJ. RecomMetz: A context-aware knowledge-
based mobile recommender system for movie showtimes. Expert Syst Appl Mohammad-Ali Balafar is associate professor at uni-
2015;42(3):1202–22. versity of Tabriz. His research area includes AI and
[54] Chicaiza J, Valdiviezo-Diaz P. A comprehensive survey of knowledge graph- Computer Vision. He has published more than 30 jour-
based recommender systems: Technologies, development, and contributions. nal papers in respected journals.
Information 2021;12(6):232.
[55] Nikzad-Khasmakhi N, Balafar MA, Feizi-Derakhshi MR. The state-of-the-art in
expert recommendation systems. Eng Appl Artif Intel 2019;82:126–47.
[56] Zuo Y, Zeng J, Gong M, Jiao L. Tag-aware recommender systems based on deep
neural networks. Neurocomputing 2016;204:51–60.
[57] Xu Z, Chen C, Lukasiewicz T, Miao Y, Meng X. October). Tag-aware
personalized recommendation using a deep-semantic similarity model with
negative sampling. In: Proceedings of the 25th ACM International on
Conference on Information and Knowledge Management. p. 1921–4.
[58] Shokrzadeh Z, Feizi-Derakhshi M.-R., Balafar MA, Bagherzadeh-Mohasefi J.
Graph-Based Recommendation System Enhanced with Community Detection;
2022. arXiv preprint arXiv:2201.03622.
[59] Zhang J, Peng Q, Sun S, Liu C. Collaborative filtering recommendation Jamshid Bagherzadeh Mohasefi was born in Urmia,
algorithm based on user preference derived from item domain features. Iran, in 1974 and received his B.S. degree from Sharif
Physica A 2014;396:66–76. University in computer engineering. He received his M.
[60] Shepitsen A, Gemmell J, Mobasher B, Burke R. October). Personalized
S. degree in computer engineering from Sharif Univer-
recommendation in social tagging systems using hierarchical clustering. In:
sity, Tehran in 2005. His Ph.D. degree was from IITD. He
Proceedings of the 2008 ACM conference on Recommender systems. p.
is a faculty member at urmia univetsity in Department
259–66.
[61] Xu Z, Yuan D, Lukasiewicz T, Chen C, Miao Y, Xu G. Hybrid deep-semantic of Software Engineering urmia.
matrix factorization for tag-aware personalized recommendation. In ICASSP
2020-2020 IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP), p. 3442–6. IEEE; 2020, May.
[62] Mnih, A., & Salakhutdinov, R. R. (2007). Probabilistic matrix factorization.
Advances in neural information processing systems, 20.
[63] Wang X, He X, Cao Y, Liu M, Chua TS. KGAT: Knowledge graph attention
network for recommendation. In: Proceedings of the 25th ACM SIGKDD
international conference on knowledge discovery & data mining. p. 950–8.

13