Received 20 May 2022, accepted 16 July 2022, date of publication 26 July 2022, date of current version 4 August 2022.

Digital Object Identifier 10.1109/ACCESS.2022.3194063

Knowledge Graphs in Education and

Employability: A Survey on Applications
and Techniques
YOUSRA FETTACH 1 , MOUNIR GHOGHO 1,2 , AND
BOUALEM BENATALLAH3,4 , (Fellow, IEEE)
1 TIC Laboratory, College of Engineering & Architecture, International University of Rabat, Sale 11103, Morocco
2 Facultyof Engineering, University of Leeds, Leeds LS2 9JT, U.K.
3 School of Computing, Dublin City University, Dublin 9, D09 W6Y4 Ireland
4 School of Computer Science and Engineering, UNSW Sydney, Sydney, NSW 2052, Australia

Corresponding author: Yousra Fettach ([email protected])

ABSTRACT Studies on the relationship between education and employability are of paramount importance
for policy makers, training institutions, companies and students. The availability of large scale data on the
Internet, such as online job ads, has been leveraged to better investigate this relationship. Recent studies have
shown that representing this data using knowledge graphs can be useful in identifying the job market needs
and establishing better assessment methods. Such studies also highlight the skill mismatch between education
and the job market and help mitigate it. In this paper, we provide a comprehensive review of the applications
of knowledge graphs in education and employability. We review knowledge graph based frameworks that
have been proposed in previous work on different aspects of education and employability. The survey
introduces new taxonomies for these applications and their methodologies, and provides a thorough outlook
on promising research directions. We also present a use case to illustrate how knowledge graphs can be used
to investigate the relationship between education and the job market.

INDEX TERMS Educational technologies, employability, knowledge graphs, representation learning.

I. INTRODUCTION resumes, has been incorporated in order to examine and

Education has been an important tool for economic and social analyze the issue.
growth throughout the modern era. Higher education cre- The increased availability of large amount of data of differ-
dentials in particular are considered by many people to be ent types also impelled researchers to leverage semantic and
essential to enter the labor market. However, recent research comceptual representation. That led to the birth of knowledge
demonstrated that sometimes higher education falls short in graphs (KG). A KG is a structured representation of facts,
fulfilling the needs of the job market. Therefore, the issue of consisting of entities, relationships, and semantic descrip-
skill disparities between the job market and education has tions [1]. KGs are core component for several information
received much attention from policy makers, higher educa- systems which require access to structured knowledge [2].
tion institutions, employers as well as researchers. With the With the support of KGs as a representation tool for education
availability of massive amounts of structured and unstruc- and labor market entities, one can unravel correlations across
tured data engendered by the digital transformation, online these entities and construct semantic connections between
content such as job ads, massive open online courses and them in order to turn data into usable knowledge.
Previous work proposed numerous methods based on data
analysis to improve education by learning student behav-
The associate editor coordinating the review of this manuscript and ior [3] or building recommender systems for students [4]
approving it for publication was Ali Shariq Imran . for a better academic performance and education-to-job

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
80174 VOLUME 10, 2022
Y. Fettach et al.: Knowledge Graphs in Education and Employability: A Survey on Applications and Techniques

matching [5]. Several studies investigated ways to eval- one [16]–[18]. Several approaches were utilized to tackle
uate the job market using job representations [6], pre- this problem. One of the first studies in this realm was
senting taxonomies to help job seekers [7] and predicting to enable prerequisite prediction for within-university and
job market trends [8]. Furthermore, many large-scale KGs cross-university settings. The initial intuition of the work was
have been developed by various academic and corporate the fact that courses may be specific to one institution whereas
institutions [9], [10]. concepts are shared across different providers [19]. Some
This review is based on the analysis of research papers on of the approaches adopted supervised learning to perform
the use of knowledge graphs as a tool to improve education binary classification, using a set of features extracted via nat-
and employability for job seekers and employers. Based on ural language processing methods, in order to automatically
this analysis, we suggest a unified KG-based framework for identify prerequisites [20]. Pan et al. [17] have implemented
education and employability. The main contributions of this different classification methods using novel contextual and
paper are as follows: structural features extracted from MOOCs. Authors of [16]
• A comprehensive survey of KG applications in applied active learning based on graph-based features for
education. representing concept pairs. In [21], the authors claimed that
• A comprehensive survey of KG applications for the problem of determining an effective learning path from a
employability. corpus of documents depends on the accurate identification
• A unified KG-based framework for education and of concepts that are learning objectives and the identification
employability. of concepts that are prerequisites. The authors of [21] then
• An outlook on future directions. implemented a supervised machine learning approach for
• A case study for exploring the relationship between annotation based on multiple extracted local features and
education and the job market. contextual ones. Other studies focused on the relationship
between learning resources instead of concepts [22]. In [23],
the authors suggested an approach to acquire concepts maps
A. RELATED SURVEYS
from textbooks, where concepts are associated with prereq-
Previous survey papers on KGs mainly focus on tech-
uisite relationships.
nical aspects [1], KG applications in general [2] and
Various studies were established in order to assist job
their embeddings [11]. Surveys like [12] and [13] pro-
seekers for the purpose of finding a good fit [24], [25].
vide general overviews of research from the perspective
Graphs were first incorporated into these methodologies with
of job market evaluation using data analysis methods.
graph-based job recommender systems [26]. In [27], the
Other literature reviews highlighted the potential of KGs in
authors created a job transition network that presented the
education [14], [15]. Our survey combines all these aspects
talent flow between organizations, which was then used to
and dives deeper into the applications of KGs in education
build a talent circle detection model for talent exchange pre-
and the techniques behind them. Besides, our paper pro-
diction. Other studies presented methodologies that aimed to
vides a comprehensive review of KG-based methodologies
create new taxonomies or enhance existing ones for job-skill
for employability in addition to showcasing the role of KGs
matching [7].
in evaluating the link between education and the job market.
Overall, KGs started getting more attention in recent
years [28] and being incorporated in various tools for educa-
B. SURVEY ORGANIZATION tion and job market assessment. The next section represents
In Section II, we describe the background of KG method- a brief report on KG techniques and how they are applied in
ologies for education and employability. Section III dis- education and employability.
cusses the KG techniques in education and employability.
KG applications in education and employability are discussed III. KG TECHNIQUES IN EDUCATION AND
in Section IV and V, respectively. In Section VI, we elabo- EMPLOYABILITY
rate on the importance of building a unified framework for Knowledge representation went through many iterations
representing education and employability data in a single based on frame semantics, which is the idea that understand-
knowledge graph. This framework is illustrated using a use ing one word relies on the understanding of the knowledge
case in Section VII. Future research directions are discussed related to it, representing coherent structure of related con-
in VIII and conclusions are drawn in Section IX. cepts. The concretization of this concept in computational
linguistics can be found in the Semantic Web and KGs. The
II. BACKGROUND Semantic Web started by developing the concepts of the tech-
Studies took different directions in order to use semantic nical standards such as Uniform Resource Identifier (URI),
metadata to improve education. One of these directions is Resource Description Framework (RDF) and Web Ontology
exploring ways to automatically define prerequisite relation- Language (OWL). The concept of Linked Data started as
ships between concepts. From a pedagogical point of view, linking datasets to each other in the Semantic Web to make
a prerequisite is a dependency relation that states which them as a unified large KG. In 2012, the concept of KGs
concepts a student has to learn before moving to the next became more popular as Google launched it and utilized the

VOLUME 10, 2022 80175

 Y. Fettach et al.: Knowledge Graphs in Education and Employability: A Survey on Applications and Techniques

semantic knowledge in web search in order to identify and by jointly learning topology view, semantic view, job transi-
differentiate entities in addition to providing links to related tion balance view, and job transition duration view. In addi-
entities for the purpose of creating more of a connected and tion, they presented an encode-decode based fusion method to
uniform search results. obtain a unified representation from the multi-view represen-
There are many technical aspects in KGs [1]. In this section tations. JobBERT is the latest effort in this trajectory, where
we try to tackle the most prominent ones in the field of the authors of [39] presented a neural representation model
education and employability. for job titles, by augmenting a pre-trained language model
with co-occurrence information from skill labels extracted
A. KNOWLEDGE ACQUISITION from job vacancies.
In KGs, knowledge acquisition aims to extract knowledge Due to the literature gap in the area, using specific KG
from unstructured text or various other structured or semi embedding techniques in educational KG construction was
structured sources. It also aims at completing established KGs one of the direction [40] have undertaken. The authors pre-
and recognizing entities and relations. Knowledge acquisition sented a model for embedding learning of eduational KGs.
consists of many tasks including entity recognition and align- Notably, the displayed method jointly learns embeddings
ment and relation extraction [1]. These tasks have been tack- from pre-trained structural (i.e. TransE) and literal embed-
led using various methods. For example, in [29], entity recog- ding vectors (i.e. BERT). These are considered to be pro-
nition was carried out using a Gated Recurrent Unit (GRU) jections of KGs in the knowledge representation spaces e.g.
network, and relation extraction was performed using proba- Euclidean, non-Euclidean which is computationally efficient
bilistic methods. In [30], the authors adopted a concept prun- and representative but they suffer from lack of interpretability.
ing strategy that is based on a measure called SemRefD to This issue is remedied by rule based KG completion methods,
analyze references between concepts. This measure has been which take advantage of the inherent structure of the KGs in
utilized in KGs to assess the likelihood of one concept being order to find meaningful patterns.
a prerequisite of another. In order to evaluate the relation What follows is a review of recent KG applications in edu-
extraction, the algorithm XGBoost was used for binary clas- cation. Then, we present an analysis of KG based applications
sification. In several other techniques, entity extraction was for employability, highlighting several factors that we think
cast as a classification problem which was addressed using are critically important for a unified KG-based framework for
algorithms like Support Vector Machine (SVM) [31] and education and the labor market.
Conditional Random Field classifier (CRFClassifier) [32].
Other methodologies focused on entity extraction using IV. KG APPLICATIONS IN EDUCATION
semantic similarity [33] or Stack Long Short Term Memory Several studies have proposed methodologies to use KGs as
(Stack-LSTM) with Spacy1 [34]. In a similar way, some a tool to improve education and its outcomes. We classify
researchers used KGs for job market data representation. the reviewed studies into three main categories as shown
For example, the authors of [35] implemented entity linking in Figure 1.
using rule-based methods. Qin et al. [36] proposed a novel
approach for entity recognition by feeding the word vec- A. ASSISTED INSTRUCTION
tors into a bidirectional Long Short Term Memory and hav- Assisted instruction refers to approaches to support instruc-
ing the outputs fed into Conditional Random Fields layer tors and faculty members in different aspects of their tasks.
(LSTM-CRF). They also improved the reliability of the Several studies have addressed this area using KGs. In the
extracted skill entities using click through data as a way for next subsections, we describe two subcategories in which
entity filtering. Additionally, the relation extraction process assisted instruction is the main topic.
was done through discovering the hypernym-hyponym rela-
tions between skill entities. 1) INSTRUCTIONAL CONCEPTION
One of the main areas of educational research is instructional
B. REPRESENTATION LEARNING conception and its role in students learning. Instructional con-
Knowledge representation learning aims to map entities and ception is the creation of learning experiences and materials
relations of a KG onto a dense low-dimension feature space in a manner that results in the acquisition and application
that captures the information and the structure of the graph. of knowledge and skills. The authors of [29] conceptualized
This was mainly carried out for job market-related graph a system called KnowEDU that constructs a KG for edu-
applications. Studies such as [37] proposed representation cation automatically. The result was a KG with nodes that
learning algorithms such as Joint Bayesian Personalized represent instructional concepts that learners should master.
Ranking (Joint-BPR) and Joint-Margin In [38], the authors The main purpose of KnowEDU is to identify the prereq-
proposed new algorithms for job representation learning uisite relations between concepts in one course or subject,
using a directed graph as input. Other studies, such as [6], pro- so that inter-course and inter-subject relations can be fur-
posed a collective multi-view representation learning method ther explored. The instructional concepts are identified from
pedagogical data using recurrent neural networks (RNN).
1 https://spacy.io/universe/project/video-spacys-ner-model The links between the instructional concepts are established

80176 VOLUME 10, 2022

Y. Fettach et al.: Knowledge Graphs in Education and Employability: A Survey on Applications and Techniques

FIGURE 1. Taxonomy of KG applications in education.

through the probabilistic association rule mining algorithm corresponding relations, including 4 platforms, 604 univer-
by using students’ performance data. The same methodology sities, 18671 teachers and 9312 courses. [35] constructed a
was used for the automatic construction of an educational KG from video lectures by following a methodology that
KG from K-12 educational subjects [41]. In [42], the authors consists of transcript extraction from the video then applying
devised a procedure that enables the automatic construction named entity recognition, co-reference recognition and triple
of an educational KG using a model that combines Bidirec- extraction. The proposed methodology however has many
tional Encoder Representations from Transformers (BERT), limitations such as language restrictions, time restrictions and
Bidirectional LSTM (BiLSTM) and the CRF model. Their accent restrictions.
KG was built using two large datasets, specifically subject
teaching resources and an online encyclopedia. 2) KNOWLEDGE MANAGEMENT
[30] developed a strategy for identifying concepts’ prereq- Knowledge management using KGs explored tools such
uisites based on the semantic relations found in a KG. The as SoLeMiO [31] to integrate learning material in popular
process consists of two phases: i) a candidate prerequisite authoring software The study focused on semantic repre-
searching module that generates an initial list of candidates, sentation by exploiting information from open KGs through
which are concepts that are members of the same cate- expansion and filtering strategies in order to assist teachers
gory as the target concepts; the associations are established in the authoring process of learning material through semi-
using properties like Common Memberships (CM), Linked automatic annotations. The authors used the identified con-
Concepts (LC) in addition to novel pruning methods; and cepts to extract and retrieve related learning resources from an
ii) a prerequisite evaluation module, that determines pre- open corpus. Furthermore, [44] presented a KG-based design
requisite relationships between concepts using a machine of a scientific publication management model for scientific
learning approach that models the problem as a binary clas- metadata using data analysis technologies. For evaluation,
sification. In [43], the authors presented a process to sup- they implemented the use case of an entrepreneurship scien-
port domain experts in creating KGs by automating parts tific publication management prototype system, in order to
of their generation. They utilized a method that consisted encourage sustainable learning. Similar work tackled courses
of entity recognition, relation extraction and graph comple- allocation and management using KGs [45]. One of the lat-
tion using the T-Mitocar tool. The study resulted in creating est studies in this direction of research was ELODIE [46],
annotated educational KGs for personalized learning. In the a mechanism for knowledge management. The study pre-
same context, [33] adopted an acknowledged educational sented the task at hand as an information retrieval problem.
ontology to conceptualize and build a KG for Massive Open The authors proposed a faceted search interface enhanced
Online Courses. Entities represent 28591 instances with their by a natural language query capabilities. ELODIE organises

VOLUME 10, 2022 80177

 Y. Fettach et al.: Knowledge Graphs in Education and Employability: A Survey on Applications and Techniques

nodes and links of the KG by facets. Students can click on any system based on a multi-view KG. The study started by con-
provided option to formulate their queries. The user can then structing a KG that presents faculty information such as sci-
manipulate the dataset and visualizations according to their entific research papers, patents, funds, monographs, awards,
needs. The system presents a simple approach for engaging professional activities and teaching performance. Based on
the user in information retrieval and knowledge management. the KG embeddings of this data, an academic development
factor (ADF) was proposed for making predictions about
B. ASSISTED LEARNING faculty academic development.
Researchers argued that using KGs in assessing students
would improve educational outcomes. Based on our inves- V. KG APPLICATIONS FOR EMPLOYABILITY
tigation, we group these applications into two subcategories, In previous studies, KGs are used to represent and visualize
as described next. the entities and relationships related to social commerce,
e.g. [54]. Inductive logic programming techniques were used
• Personalized Learning Personalized learning in KGs to discover implicit information that can interpret the behav-
was presented as techniques to help students engage iors and intentions of the users. In this section, we explore
and understand their study material. [47] constructed a more studies that focused on analyzing the job market using
personalized learning system based on knowledge struc- knowledge graphs.
tural graphs. They then defined and implemented an
optimized learning path generator in order to provide A. JOB RECOMMENDATION
learners with personalized learning content. Interactive [38] tackled the job representation aspect in knowledge
applications using KGs is a relatively new direction to graphs. The authors developed a representation of job posi-
improve education. In [50], the authors addressed inter- tions consisting of job title and company pairs which captures
active visual navigation and inductive reasoning in KGs. the similarity and the ordering relations among job positions.
Visualization was also the topic of the study carried out In the context of job recommendation, [37] created three
by [49]. The work evaluated the utility of the educational types of information networks from the historical job data: job
KG by a visualization analysis in order to provide a transition network, job-skill network, and skill co-occurrence
better comprehension of its structure. network. The authors of [37] conceptualized a representation
• Question Answering [48] relied on crowdsourcing to learning model that uses the information from all three net-
enrich a math-related KG (MathGraph). The main pur- works. This study presented the pair wise ranking objective
pose of the KG was to help students solve mathematical which learns job and skill vector representations into a shared
exercises with mathematical derivations and calcula- latent space using three pre-processed graphs.
tions. The authors also presented an algorithm to align Recent studies started exploring the graph based recom-
a semantically parsed exercise to MathGraph in order to mendation systems. For example, [55] developed a recom-
automatically resolve the mathematical problem. mender system by measuring the skill importance of each
occupation for each country and feeding the best vector model
C. EDUCATIONAL ASSESSMENT to a graph database. [56] built a KG based recommender sys-
Educational assessment focuses on assessment of a learner’s tem for jobs called GUApp. The latter recommends relevant
knowledge and the level of skill they have acquired. Multiple jobs to the users. The system also integrates a domain-specific
studies tackle this aspect using KGs. [32] explored establish- KG that allows it to support natural dialogues with users
ing teacher support methodologies where the authors adopted about the discussed domain. This leads to an improvement
a KG-based approach to detect plagiarism. They focus on in the quality of recommendations results. Last but not least,
building a semantic KG from unstructured data. The built sys- [57] proposed an explainable job recommendation system,
tem implements entity recognition, followed by relationship which uses a KGs to model job-postings and user profiles into
extraction, then performs a co-reference resolution. It uses a unified framework. The system generates human-readable
semantic similarity and sentence replacement for purpose of explanation for each recommendation based on the graph
plagiarism detection. structure itself and a named entity classifier.
KGs also contributed to advancing the question generation
field especially in education. [51] implemented an end-to-end B. JOB-SKILL MATCHING AND TALENT INTELLIGENCE
approach to generate natural language questions from knowl- The latest studies in this direction is the Skills & Occupation
edge graphs. An entity is first selected as an answer. A struc- KG that was created in [58]. The KG was built by leverag-
tured triple-pattern query having this answer as its sole result ing existing skills and occupation taxonomies enriched with
is then generated. If a multiple-choice question is desired, the external job posting. [35] built a KG representing the artificial
approach selects alternative answer options. It also adopts a intelligence job market and the corresponding skills required.
template based method to verbalize the structured query and The KG was used to explore the relationship of different key
yield a natural language question. The authors based their skills of artificial intelligence.
approach on an earlier methodology that follows the same Talent intelligence exploration started with [6] presenting a
pattern [52]. [53] presented a university faculty evaluation data-driven approach for job title benchmarking. The authors

80178 VOLUME 10, 2022

Y. Fettach et al.: Knowledge Graphs in Education and Employability: A Survey on Applications and Techniques

TABLE 1. KG applications in education.

constructed a job title transition graph where nodes represent in mediating between the needs of industry and research.
job titles and links denote the correlations between jobs. They [63] emphasized the importance of soft business-related skills
purpose of the study was to provide an accurate guidance and competencies and the need for prior work-experiences.
and considerable convenience for both talent recruitment and Investigating the dynamics between education and the job
job seekers for position and salary calibration and prediction. market is a very relevant issue which has been investigated
Other like [36] constructed a KG of job skills through mining in many ways before. Addressing this issue using KGs is
the abundant historical recruitment data. The authors utilized an interesting research direction. We illustrate this using the
the constructed Skill-Graph for building a personalized ques- following case study.
tion recommendation algorithm for improving job interview
assessment. VII. CASE STUDY: DATA SCIENCE/AI
A summary of the studies is provided in the reference Here, we propose a framework to showcase the use of KG
tables 1 and 2. in the fusion of education and employability information.
The framework is based on the idea that education provides
students with certain skills and the job market needs certain
VI. TOWARDS A UNIFIED FRAMEWORK FOR EDUCATION skills to accomplish tasks in a certain job. This leads us to rely
AND EMPLOYABILITY on skills to build links between the job market and education.
Research has shown a lot of interest in diagnosing the issues Figure 3 illustrates a subgraph of the entire KG, built to
that education and especially higher education suffers from, describe the skills needed for a certain job and the courses that
specifically when considering the job market. Previous work offer to teach those skills. We choose as a case study about
has analyzed the relationship between academia and the job data science jobs. We aim to give a full picture of what skills
market to understand how they influence each other [59]. [60] one would need in order to compete for a certain job. The
examined the collaborations between stakeholders and higher schema should encompass all aspects of data science jobs,
education institutions to identify the lack of communication the skills needed for a given job and the courses that teach
between the job market and the youth. Other studies aimed these skills. The KG may be used as recommender system for
to determine the relationship between skill mismatch and students to identify the skills needed for the job they want, for
the regional unemployment. Researchers found that regions institutions to update their courses according to the job market
with higher skilled populations have lower unemployment evolution and eventually also to predict additional links.
rates [61]. [62] presented multiple data analyses and visu- We used a public dataset of data science/AI job offers
alizations to understand discrepancies and temporal delays collected from Indeed.2 Each line in the dataset contains
between evolving job market needs, course offerings, and the job title, the company that posted the job offer and the
science and technology developments. The findings of this
work suggest that educational efforts play an important role 2 www.indeed.com

VOLUME 10, 2022 80179

 Y. Fettach et al.: Knowledge Graphs in Education and Employability: A Survey on Applications and Techniques

TABLE 2. KG applications in employability.

FIGURE 2. Methodology for job-skill-course KG construction.

hard skills needed for the job. We define hard skills as the goals. The weight of the relationship between nodes is the
technical skills needed to complete a certain task, whereas number of co-occurrences of these entities. We normalize the
soft skills are abilities that are related to how an individual weights in order to have values between 0 and 1.
works and interacts with others. In our study, we focus solely For visualization purposes, we chose the top ten skills
on hard skills for simplicity reasons. We collected data about required for a data scientist (the edges with the highest
Massive Open Online Courses using a Web crawler where weights that are connected to the node ’data scientist’),
we gathered attributes such as the name of the course, the as shown in Figure 3 and the top two courses for each skill
university that offers the course, and the language in which (the edges with the highest weights that are connected to the
the course is taught. In addition, using the categories offered chosen skills).
by the Website and entity recognition, we created a list of
skills for each course, extracted from the course title, the VIII. FUTURE RESEARCH DIRECTIONS
description of the course and its category. A full view of the In this section, we discuss the limitations of the existing work
pipeline adopted in our framework is presented in Figure 2. and propose research directions for future work.
We created two weighted directed graphs: the first one • KG Construction and Data Quality Most of the exist-
has nodes as jobs and skill entities (Skill-Job KG) and ing work paid very little attention to discussing the
the second one has nodes representing skills and courses details of and the motivations behind the techniques used
(Skill-Course KG). In order to create the final KG, we rely on for entity recognition and relation extraction. It would
semantic matching for entity alignment between the Skill-Job be interesting to compare these techniques in order to
KG and Skill-Course KG for the skill entity. The KG is identify those that perform the best in the different tasks.
heterogeneous as it contains three types of nodes: jobs, skills Furthermore, to the best of our knowledge, there are
and courses. The job nodes are linked to certain skill nodes no methods to fully automate the process of KG con-
based on the skills needed for the job, and skill nodes are struction. Developing such methods would be a promis-
linked to certain course nodes based on the fact that these ing research direction, particularly when dealing with
courses offer teaching those skills as part of their course large-scale unstructured content.

80180 VOLUME 10, 2022

Y. Fettach et al.: Knowledge Graphs in Education and Employability: A Survey on Applications and Techniques

FIGURE 3. A snapshot of job-skill-course knowledge graph. The blue nodes represent the job titles, the pink
nodes represent the skills, and the green nodes represent the courses.

The methodologies reviewed in this paper did not con- limitation of traditional knowledge representation and
sider the use of standardized and appropriate data quality reasoning by considering the temporal dimension [1].
measures when constructing KGs. The data collection • KG-based recommendation KG-based recommender
process gave little consideration to the credibility of the systems attracted considerable interest in recent years.
generated information. An interesting research direction Unlike traditional recommender systems, KG-based rec-
would be to develop methods to evaluate the credibility ommender systems present a variety of entities and rela-
of the information generated from KGs. This could be tions linking users and items, and illustrate better the
carried out by building probabilistic KGs where each reasoning process. Generating recommendations with
relation/link is associated with a measure of its certainty. the KG as side information for job recommendation and
• Temporal KGs Temporal KGs are one of the new direc- curriculum design would benefit institutions, job seekers
tions that researchers have recently started to explore and employers. Advancements in this field focus on
especially in education. This need to expansion comes multi-task learning where transferring knowledge from
from the fact that static KGs do not give a full represen- KG-related tasks, such as entity representation, classi-
tation about real world data and are only considered to fication and resolution would enhance recommendation
be a static snapshot of multi-relational data. Studies such performance [67].
as [64] initiated research in this direction by construct-
ing a novel deep evolutionary knowledge network that
efficiently learns non-linearly evolving entity represen- IX. CONCLUSION
tations over time in a multi-relational setting. The use of knowledge graphs to improve education and
Work on incorporating temporal KGs in the job market is employability has received much attention in recent liter-
very limited. There has been some descriptive work done ature. This study examined the different applications of
on using time-series on clusters of skillsets for job mar- such knowledge graphs as well as the approaches taken to
ket forecasting [65], [66]. A dynamic KG, together with construct them. We detailed these studies by categorizing
learning algorithms capturing dynamics, can address the them into different aspects of education and employability.

VOLUME 10, 2022 80181

 Y. Fettach et al.: Knowledge Graphs in Education and Employability: A Survey on Applications and Techniques

To contribute to reducing the mismatch between higher edu- [21] S. Changuel, N. Labroche, and B. Bouchon-Meunier, ‘‘Resources sequenc-
cation and market need, we advocate the development and ing using automatic prerequisite-outcome annotation,’’ ACM Trans. Intell.
Syst. Technol., vol. 6, no. 1, pp. 1–30, Mar. 2015.
exploitation of knowledge graphs that jointly represent edu- [22] F. Gasparetti, C. De Medio, C. Limongelli, F. Sciarrone, and M. Temperini,
cation and job market data. As a step towards this, we intro- ‘‘Prerequisites between learning objects: Automatic extraction based on
duced a framework for building such knowledge graphs, and a machine learning approach,’’ Telematics Informat., vol. 35, no. 3,
pp. 595–610, Jun. 2018.
use the field of data science/AI as a case study. A number [23] S. Wang, A. Ororbia, Z. Wu, K. Williams, C. Liang, B. Pursel, and
of research avenues remain to be explored, and this article C. L. Giles, ‘‘Using prerequisites to extract concept maps fromTextbooks,’’
discussed a few. in Proc. 25th ACM Int. Conf. Inf. Knowl. Manage., Oct. 2016, pp. 317–326.
[24] C. Zhu, H. Zhu, H. Xiong, C. Ma, F. Xie, P. Ding, and P. Li, ‘‘Person-job fit:
Adapting the right talent for the right job with joint representation learn-
REFERENCES ing,’’ ACM Trans. Manage. Inf. Syst., vol. 9, no. 3, pp. 1–17, Sep. 2018.
[1] S. Ji, S. Pan, E. Cambria, P. Marttinen, and P. S. Yu, ‘‘A survey on [25] C. Qin, H. Zhu, T. Xu, C. Zhu, C. Ma, E. Chen, and H. Xiong,
knowledge graphs: Representation, acquisition and applications,’’ 2020, ‘‘An enhanced neural network approach to person-job fit in talent recruit-
arXiv:2002.00388. ment,’’ ACM Trans. Inf. Syst., vol. 38, no. 2, pp. 1–33, Mar. 2020.
[2] X. Zou, ‘‘A survey on application of knowledge graph,’’ J. Phys., Conf., [26] W. Shalaby, B. Alaila, M. Korayem, L. Pournajaf, K. AlJadda, S. Quinn,
vol. 1487, Mar. 2020, Art. no. 012016. and W. Zadrozny, ‘‘Help me find a job: A graph-based approach for job
[3] M. Hussain, W. Zhu, W. Zhang, S. M. R. Abidi, and S. Ali, ‘‘Using machine recommendation at scale,’’ in Proc. IEEE Int. Conf. Big Data (Big Data),
learning to predict Student difficulties from learning session data,’’ Artif. Dec. 2017, pp. 1544–1553.
Intell. Rev., vol. 52, no. 1, pp. 381–407, Jun. 2019. [27] H. Xu, Z. Yu, J. Yang, H. Xiong, and H. Zhu, ‘‘Talent circle detection in
[4] F. Bousbahi and H. Chorfi, ‘‘MOOC-Rec: A case based recommender job transition networks,’’ in Proc. 22nd ACM SIGKDD Int. Conf. Knowl.
system for MOOCs,’’ Proc. Social Behav. Sci., vol. 195, pp. 1813–1822, Discovery Data Mining, Aug. 2016, pp. 655–664.
Jul. 2015. [28] Y. Mao, ‘‘Summary and evaluation of the application of knowledge graphs
[5] A. Fedorets, F. Lottmann, and M. Stops, ‘‘Job matching in connected in education 2007–2020,’’ Discrete Dyn. Nature Soc., vol. 2021, pp. 1–10,
regional and occupational labour markets,’’ Regional Stud., vol. 53, no. 8, Sep. 2021.
pp. 1085–1098, Aug. 2019. [29] P. Chen, Y. Lu, V. W. Zheng, X. Chen, and B. Yang, ‘‘KnowEdu: A
[6] D. Zhang, J. Liu, H. Zhu, Y. Liu, L. Wang, P. Wang, and H. Xiong, system to construct knowledge graph for education,’’ IEEE Access, vol. 6,
‘‘Job2Vec: Job title benchmarking with collective multi-view representa- pp. 31553–31563, 2018.
tion learning,’’ in Proc. 28th ACM Int. Conf. Inf. Knowl. Manage., 2019, [30] R. Manrique, B. Pereira, and O. Mariño, ‘‘Exploring knowledge graphs for
pp. 2763–2771. the identification of concept prerequisites,’’ Smart Learn. Environments,
[7] I. Khaouja, G. Mezzour, K. M. Carley, and I. Kassou, ‘‘Building a soft skill vol. 6, no. 1, pp. 1–18, Dec. 2019.
taxonomy from job openings,’’ Social Netw. Anal. Mining, vol. 9, no. 1, [31] C. Grévisse, R. Manrique, O. Mariño, and S. Rothkugel, ‘‘Knowl-
pp. 1–19, Dec. 2019. edge graph-based teacher support for learning material authoring,’’ in
[8] N. Pontika, ‘‘Roles and jobs in the open research scholarly communications Proc. Colombian Conf. Comput. Cham, Switzerland: Springer, 2018,
environment: Analysing job descriptions to predict future trends,’’ LIBER pp. 177–191.
Quarterly, J. Assoc. Eur. Res. Libraries, vol. 29, no. 1, pp. 1–20, Jul. 2019. [32] K. Khadilkar, S. Kulkarni, and P. Bone, ‘‘Plagiarism detection using
[9] X. Dong, E. Gabrilovich, G. Heitz, W. Horn, N. Lao, K. Murphy, semantic knowledge graphs,’’ in Proc. 4th Int. Conf. Comput. Commun.
T. Strohmann, S. Sun, and W. Zhang, ‘‘Knowledge vault: A web-scale Control Autom. (ICCUBEA), Aug. 2018, pp. 1–6.
approach to probabilistic knowledge fusion,’’ in Proc. 20th ACM SIGKDD [33] F. Dang, J. Tang, and S. Li, ‘‘MOOC-KG: A MOOC knowledge graph
Int. Conf. Knowl. Discovery Data Mining, Aug. 2014, pp. 601–610. for cross-platform online learning resources,’’ in Proc. IEEE 9th Int. Conf.
[10] A. Hogan, E. Blomqvist, M. Cochez, C. d’Amato, G. de Melo, C. Gutierrez, Electron. Inf. Emergency Commun. (ICEIEC), Jul. 2019, pp. 1–8.
J. E. L. Gayo, S. Kirrane, S. Neumaier, A. Polleres, R. Navigli, [34] G. S. Shanmukhaa, S. K. Nandita, and M. V. K. Kiran, ‘‘Construction of
A.-C. N. Ngomo, S. M. Rashid, A. Rula, L. Schmelzeisen, J. Sequeda, knowledge graphs for video lectures,’’ in Proc. 6th Int. Conf. Adv. Comput.
S. Staab, and A. Zimmermann, ‘‘Knowledge graphs,’’ 2020, Commun. Syst. (ICACCS), Mar. 2020, pp. 127–131.
arXiv:2003.02320. [35] S. Jia, X. Liu, P. Zhao, C. Liu, L. Sun, and T. Peng, ‘‘Representation
[11] Y. Dai, S. Wang, N. N. Xiong, and W. Guo, ‘‘A survey on knowledge of job-skill in artificial intelligence with knowledge graph analysis,’’ in
graph embedding: Approaches, applications and benchmarks,’’ Electron- Proc. IEEE Symp. Product Compliance Eng. Asia (ISPCE-CN), Dec. 2018,
ics, vol. 9, no. 5, p. 750, May 2020. pp. 1–6.
[12] M. Papoutsoglou, A. Ampatzoglou, N. Mittas, and L. Angelis, ‘‘Extracting [36] C. Qin, H. Zhu, C. Zhu, T. Xu, F. Zhuang, C. Ma, J. Zhang, and H. Xiong,
knowledge from on-line sources for software engineering labor market: A ‘‘DuerQuiz: A personalized question recommender system for intelligent
mapping study,’’ IEEE Access, vol. 7, pp. 157595–157613, 2019. job interview,’’ in Proc. 25th ACM SIGKDD Int. Conf. Knowl. Discovery
[13] I. Khaouja, I. Kassou, and M. Ghogho, ‘‘A survey on skill identification Data Mining, Jul. 2019, pp. 2165–2173.
from online job ads,’’ IEEE Access, vol. 9, pp. 118134–118153, 2021. [37] V. S. Dave, B. Zhang, M. Al Hasan, K. AlJadda, and M. Korayem,
[14] M. Rizun, ‘‘Knowledge graph application in education: A literature ‘‘A combined representation learning approach for better job and skill
review,’’ Acta Universitatis Lodziensis. Folia Oeconomica, vol. 3, no. 342, recommendation,’’ in Proc. 27th ACM Int. Conf. Inf. Knowl. Manage.,
pp. 7–19, Aug. 2019. Oct. 2018, pp. 1997–2005.
[15] B. Abu-Salih, ‘‘Domain-specific knowledge graphs: A survey,’’ J. Netw. [38] H. Luo, S. Ma, A. J. B. Selvaraj, and Y. Sun, ‘‘Learning job representation
Comput. Appl., vol. 185, Jul. 2021, Art. no. 103076. using directed graph embedding,’’ in Proc. 1st Int. Workshop Deep Learn.
[16] C. Liang, J. Ye, Z. Wu, B. Pursel, and C. Giles, ‘‘Recovering concept Practice High-Dimensional Sparse Data, 2019, pp. 1–5.
prerequisite relations from university course dependencies,’’ in Proc. AAAI [39] J.-J. Decorte, J. Van Hautte, T. Demeester, and C. Develder, ‘‘JobBERT:
Conf. Artif. Intell., 2017, pp. 1–6. Understanding job titles through skills,’’ 2021, arXiv:2109.09605.
[17] L. Pan, C. Li, J. Li, and J. Tang, ‘‘Prerequisite relation learning for concepts [40] S. Yao, R. Wang, S. Sun, D. Bu, and J. Liu, ‘‘Joint embedding learning of
in moocs,’’ in Proc. 55th Annu. Meeting Assoc. Comput. Linguistics, vol. 1, educational knowledge graphs,’’ in Artificial Intelligence Supported Edu-
2017, pp. 1447–1456. cational Technologies. Cham, Switzerland: Springer, 2020, pp. 209–224.
[18] J. Gordon, L. Zhu, A. Galstyan, P. Natarajan, and G. Burns, ‘‘Modeling [41] P. Chen, Y. Lu, V. W. Zheng, X. Chen, and X. Li, ‘‘An automatic knowledge
concept dependencies in a scientific corpus,’’ in Proc. 54th Annu. Meeting graph construction system for K-12 education,’’ in Proc. 5th Annu. ACM
Assoc. Comput. Linguistics, 2016, pp. 866–875. Conf. Learn. at Scale, Jun. 2018, pp. 1–4.
[19] Y. Yang, H. Liu, J. Carbonell, and W. Ma, ‘‘Concept graph learning from [42] Y. Su and Y. Zhang, ‘‘Automatic construction of subject knowledge graph
educational data,’’ in Proc. 8th ACM Int. Conf. Web Search Data Mining, based on educational big data,’’ in Proc. 3rd Int. Conf. Big Data Educ.,
Feb. 2015, pp. 159–168. Apr. 2020, pp. 30–36.
[20] P. Talukdar and W. Cohen, ‘‘Crowdsourced comprehension: Predicting [43] R. Manrique, B. Pereira, and O. Marino, ‘‘Exploring knowledge graphs for
prerequisite structure in Wikipedia,’’ in Proc. 7th Workshop Building Educ. the identification of concept prerequisites,’’ Smart Learn. Environ., vol. 6,
Appl. Using NLP, pp. 307–315. 2012. pp. 1–18, Dec. 2019.

80182 VOLUME 10, 2022

Y. Fettach et al.: Knowledge Graphs in Education and Employability: A Survey on Applications and Techniques

[44] Y. Chi, Y. Qin, R. Song, and H. Xu, ‘‘Knowledge graph in smart education: [66] E. M. Sibarani and S. Scerri, ‘‘SCoDis: Job advert-derived time series
A case study of entrepreneurship scientific publication management,’’ for high-demand skillset discovery and prediction,’’ in Proc. Int. Conf.
Sustainability, vol. 10, no. 4, p. 995, Mar. 2018. Database Expert Syst. Appl. Springer, 2020, pp. 366–381.
[45] I. Aliyu, A. F. D. Kana, and S. Aliyu, ‘‘Development of knowledge graph [67] Q. Guo, F. Zhuang, C. Qin, H. Zhu, X. Xie, H. Xiong, and Q. He, ‘‘A survey
for university courses management,’’ Int. J. Educ. Manage. Eng., vol. 10, on knowledge graph-based recommender systems,’’ IEEE Trans. Knowl.
no. 2, p. 1, 2020. Data Eng., vol. 34, no. 8, pp. 3549–3568, Aug. 2020.
[46] D. R. De, M. Garofalo, D. Malandrino, M. A. Pellegrino, and A. Petta,
‘‘Education meets knowledge graphs for the knowledge management,’’
in Proc. Int. Conf. Methodologies Intell. Syst. Technol. Enhanced Learn.
Cham, Switzerland: Springer, 2020, pp. 272–280.
[47] X. Huang, ‘‘Study of personalized E-learning system based on knowledge
structural graph,’’ Proc. Eng., vol. 15, pp. 3366–3370, Jan. 2011.
[48] T. Zhao, C. Chai, Y. Luo, J. Feng, Y. Huang, S. Yang, H. Yuan, H. Li, K. Li, YOUSRA FETTACH received the M.Sc. degree
F. Zhu, and K. Pan, ‘‘Towards automatic mathematical exercise solving,’’ from ENSIAS, Rabat, Morocco, in 2019, with a
Data Sci. Eng., vol. 4, no. 3, pp. 179–192, Sep. 2019. major in data science and big data. She is currently
[49] K. Sun, Y. Liu, Z. Guo, and C. Wang, ‘‘EduVis: Visualization for education pursuing the Ph.D. degree with the International
knowledge graph based on web data,’’ in Proc. 9th Int. Symp. Vis. Inf.
University of Rabat. Her research interests include
Commun. Interact., Sep. 2016, pp. 138–139.
knowledge graphs and NLP. She was a recipient of
[50] V. Telnov and Y. Korovin, ‘‘Semantic web and interactive knowledge
graphs as an educational technology,’’ in Cloud Computing Security-
the UIR research scholarship, in 2020.
Concepts and Practice. London, U.K.: IntechOpen, 2020.
[51] D. Seyler, M. Yahya, and K. Berberich, ‘‘Knowledge questions from
knowledge graphs,’’ in Proc. ACM SIGIR Int. Conf. Theory Inf. Retr.,
Oct. 2017, pp. 11–18.
[52] D. Seyler, M. Yahya, and K. Berberich, ‘‘Generating quiz questions from
knowledge graphs,’’ in Proc. 24th Int. Conf. World Wide Web, May 2015,
pp. 113–114.
[53] Q. Lin, Y. Zhu, H. Lu, K. Shi, and Z. Niu, ‘‘Improving university faculty
evaluations via multi-view knowledge graph,’’ Future Gener. Comput. MOUNIR GHOGHO received the M.Sc. and
Syst., vol. 117, pp. 181–192, Apr. 2021. Ph.D. degrees from the National Polytechnic
[54] L. Yuan, Z. Huang, W. Zhao, and P. Stakhiyevich, ‘‘Interpreting and Institute of Toulouse, France, in 1993 and
predicting social commerce intention based on knowledge graph analysis,’’ 1997, respectively. He was an EPSRC Research
Electron. Commerce Res., vol. 20, no. 1, pp. 197–222, Mar. 2020. Fellow with the University of Strathclyde
[55] A. Giabelli, L. Malandri, F. Mercorio, M. Mezzanzanica, and A. Seveso, (Scotland), from September 1997 to November
‘‘Skills2Job: A recommender system that encodes job offer embed- 2001. In December 2001, he joined the School of
dings on graph databases,’’ Appl. Soft Comput., vol. 101, Mar. 2021, Electronic and Electrical Engineering, University
Art. no. 107049.
of Leeds, U.K., where he was promoted to a Full
[56] G. M. Biancofiore, T. Di Noia, E. Di Sciascio, F. Narducci, and P. Pastore,
Professor, in 2008. While still affiliated with the
‘‘Enhancing job recommendations with knowledge graphs,’’ MDPI, Basel,
Switzerland, Tech. Rep. 116-138, 2021.
University of Leeds, in 2010, he joined the International University of
[57] C. Upadhyay, H. Abu-Rasheed, C. Weber, and M. Fathi, ‘‘Explain- Rabat (Morocco), where he is currently the Dean of the College of Doctoral
able job-posting recommendations using knowledge graphs and named Studies and the Director of the ICT Research Laboratory (TICLab). He was
entity recognition,’’ in Proc. IEEE Int. Conf. Syst., Man, Cybern. (SMC), a recipient of the 2013 IBM Faculty Award, and a recipient of the 2000 U.K.
Oct. 2021, pp. 3291–3296. Royal Academy of Engineering Research Fellowship. He is the Co-Founder
[58] M. de Groot, J. Schutte, and D. Graus, ‘‘Job posting-enriched knowledge and the Co-Director of the CNRS-Associated International Research Lab-
graph for skills-based matching,’’ 2021, arXiv:2109.02554. oratory DataNet, in the field of big data and artificial intelligence. His
[59] S. Angioni, A. A. Salatino, F. Osborne, D. R. Recupero, and E. Motta, research interests include machine learning, signal processing, and wireless
‘‘Integrating knowledge graphs for analysing academia and industry communication, on which he has published over 300 papers in journals and
dynamics,’’ in Proc. ADBIS, TPDL EDA Common Workshops Doctoral conferences. He has coordinated around 20 research projects and supervised
Consortium. Cham, Switzerland: Springer, 2020, pp. 219–225. over 30 Ph.D. students in the U.K. and Morocco. In the past, he served as an
[60] I. Khaouja, I. Makdoun, G. Mezzour, I. Rahhal, H. Benchekroun, Associate Editor for many journals, including the IEEE Signal Processing
Y. El Hatib, and I. Kassou, ‘‘Social network analysis of job market stake- Magazine and the IEEE TRANSACTIONS ON SIGNAL PROCESSING.
holders in Morocco,’’ in Proc. Int. Conf. Social Comput., Behav.-Cultural
Modeling, Predict. Behav. Represent. Modeling Simulation, Washington,
DC, USA, 2017.
[61] A. Newell, ‘‘Skill mismatch and regional unemployment in Poland,’’
in The European Labour Market. Cham, Switzerland: Springer, 2006,
pp. 187–202.
[62] K. Börner, O. Scrivner, M. Gallant, S. Ma, X. Liu, K. Chewning, L. Wu, BOUALEM BENATALLAH (Fellow, IEEE) has
and J. A. Evans, ‘‘Skill discrepancies between research, education, and jobs
been a Full Professor in computing with Dublin
reveal the critical need to supply soft skills for the data economy,’’ Proc.
City University (DCU), Ireland, since January
Nat. Acad. Sci. USA, vol. 115, no. 50, pp. 12630–12637, Dec. 2018.
[63] J. Andrews and H. Higson, ‘‘Graduate employability, ‘Soft skills’ versus
2022. He has had over 21 years as a Senior Lec-
‘hard’ business knowledge: A European study,’’ Higher Educ. Eur., vol. 33, turer, an Associate Professor, a Professor, then a
no. 4, pp. 411–422, Dec. 2008. Scientia Professor with UNSW Sydney, Australia.
[64] R. Trivedi, H. Dai, Y. Wang, and L. Song, ‘‘Know-evolve: Deep temporal His research interests include developing funda-
reasoning for dynamic knowledge graphs,’’ in Proc. Int. Conf. Mach. mental concepts and techniques in web service
Learn., 2017, pp. 3462–3471. composition, services engineering, crowd sourc-
[65] E. M. Sibarani and S. Scerri, ‘‘Generating an evolving skills network from ing services, data curation, cognitive services, and
job adverts for high-demand skillset discovery,’’ in Proc. Int. Conf. Web business processes management.
Inf. Syst. Eng. Cham, Switzerland: Springer, 2020, pp. 441–457.

VOLUME 10, 2022 80183