JobIQ: Recommending Study Pathways Based on Career Choices

Tomas Trescak

, Laurence A. F. Park

and Mesut Kocyigit

School of Computer Data and Mathematical Sciences, Western Sydney University, NSW, Australia

Keywords:

Employability, Skill Frameworks, Curriculum Development, Curriculum Benchmarking.

Abstract:

Modern job markets often require an intricate combination of multi-disciplinary skills or specialist and techni-

cal knowledge, even for entry-level positions. Such requirements pose increased pressure on higher education

graduates entering the job market. This paper presents our JobIQ recommendation system helping prospective

students choose educational programs or electives based on their career preferences. While existing recom-

mendation solutions focus on internal institutional data, such as previous student experiences, JobIQ consid-

ers external data, recommending educational programs that best cover the knowledge and skills required by

selected job roles. To deliver such recommendations, we create and compare skill proﬁles from job advertise-

ments and educational subjects, aggregating them to skill proﬁles of job roles and educational programs. Using

skill proﬁles, we build formal models and algorithms for program recommendations. Finally, we suggest other

recommendations and benchmarking approaches, helping curriculum developers assess the job readiness of

program graduates. The video presenting the JobIQ system is available online

∗

1 INTRODUCTION

“Intelligent” technology, marking the fourth indus-

trial revolution, is disrupting world job markets (Xu

et al., 2018). Governments and organisations are

trying to analyse the impact of such disruptions,

analysing the employability proﬁles of 21

century

workers (Daly and Lewis, 2020). A popular way

of deﬁning such proﬁles is by listing the in-demand

skills for job roles. To deﬁne skill proﬁles, gov-

ernments and organisations use an ever-increasing

number of skill frameworks, such as SFIA

for sci-

ence, technology and business skills or a more generic

ESCO

framework from the European Union or the

Australian Skill Framework

from the Australian

Skills Commission.

However, research and experience show that this

approach is ﬂawed. First, the skills listed in these

frameworks are not always the ones that are required

for a job role, being wildly different depending on the

industry, company size, or location (Holmes, 2001).

Second, the skill proﬁles in skill frameworks are in-

https://orcid.org/0000-0002-2540-6002

https://orcid.org/0000-0003-0201-4409

∗

https://www.youtube.com/watch?v=LHTW5P1tNr0

https://sfia-online.org/

https://esco.ec.europa.eu

https://www.nationalskillscommission.gov.au

complete, often specifying only three to ﬁve skills for

complex roles, listing outdated or missing new tech-

nologies or approaches. Last, interviews with em-

ployers show that personal values such as honesty

and foundational knowledge are more critical during

the employee selection process than skills (Manyika

et al., 2017). Speciﬁcally, technical skills are often

rated very low in importance due to their short shelf

life (Collet et al., 2015b).

This paper proposes a different approach using Jo-

bIQ, our novel analytical and recommendation sys-

tem. Rather than trying to invent a new generic

framework, JobIQ uses “live” job market advertise-

ments to extract information from job advertisements,

such as the required hard skills (i.e. specialist skills),

soft skills (i.e. personal skills), and knowledge (i.e.

domain or technical). JobIQ processes job markets

daily, obtaining a large dataset of available positions

with information on employers, industries or salaries.

Traditionally, related research provided no access to

such datasets (Gupta et al., 2020). While due to

present an approach to building datasets allowing for

real-time analysis and projections on various aspects

of the job market, such as demand for skills and

knowledge. By understanding the requirements of

jobs across multiple sectors, we can reach a much

ﬁner granularity of skill and knowledge requirements,

Trescak, T., Park, L. and Kocyigit, M.

JobIQ: Recommending Study Pathways Based on Career Choices.

DOI: 10.5220/0011754000003470

In Proceedings of the 15th International Conference on Computer Supported Education (CSEDU 2023) - Volume 1, pages 137-145

ISBN: 978-989-758-641-5; ISSN: 2184-5026

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

137

building localised, industry or employer-speciﬁc “em-

ployability” proﬁles. Moreover, we can detect and

track emerging skills.

Complementary to “employability” proﬁles for

job roles, JobIQ uses the same skill and knowledge

extraction approach to analyse our undergraduate and

postgraduate programs, extracting skills and knowl-

edge proﬁles from our subjects, specialisations and

programs. JobIQ compares educational proﬁles with

employability proﬁles, providing recommendations

for students and curriculum developers. Prospec-

tive students can explore how various degrees work

towards their career goals. JobIQ also helps cur-

rent students select electives and proactively helps

them maintain their development to meet career goals.

Last, curriculum developers can benchmark their pro-

grams concerning the job roles they support and dis-

cover opportunities to introduce new knowledge or

challenges with outdated content.

Section 2 of the paper provides background infor-

mation on our research, further expanding on ideas

from the introduction. Section 3 of this paper explains

and evaluates our approach to skill extraction from

job advertisements and educational content. Section 4

presents formally deﬁnes the supporting structures

and recommendation algorithms in JobIQ. Last, in

Section 5 we discuss our approach and future work.

2 BACKGROUND

Modelling job roles using sets of required skills pro-

vides the opportunity to understand the retiring and

emerging qualities of 21

-century job markets (Apps,

1988), (Peetz, 2019). Increasingly, we see the no-

tion of employability, deﬁning competitive employ-

ment proﬁles (Collet et al., 2015a), (Hinchliffe and

Jolly, 2011). Consequently, world governments and

organisations analyse and model employability per-

spectives using skill proﬁles to predict the growth of

individual job roles and impose demands on immi-

gration, educational institutions, or ﬁnancing schemes

(Peetz, 2019).

We can use several generic or specialised skill

frameworks to specify job role proﬁles. For exam-

ple, the Skills for Information Age framework (SFIA)

deﬁnes technical and business roles using 147 skills,

and each is further decomposed into seven areas of

responsibility. The European Union, Australian Pub-

lic Service, USA National Initiative for Cybersecurity

Education (NICE) and other organisations and gov-

ernments used SFIA to model job role skill proﬁles.

Similarly, the European Union created the Euro-

pean Skills/Competences, Qualiﬁcations and Occu-

pations framework (ESCO), deﬁning over 3000 oc-

cupations using 13500 skills/competencies and 11500

qualiﬁcations. More recently, in a less gargantuan ef-

fort, the Australian National Skills Commission pub-

lished the second version of the Australian Skills

Classiﬁcation (ASC), which deﬁnes 1100 occupations

mapped to ANZSCO

proﬁles using 2000 skills.

Following the governmental efforts requiring

higher educations institutions to prepare “job-ready”

graduates (Daly and Lewis, 2020), inspired by (Her-

bert et al., 2013), we intended to use skill frameworks

(i.e. SFIA, ASC or ESCO) to deﬁne graduate pro-

ﬁles for in-demand jobs in local job markets. Sub-

sequently, we planned to redesign and improve our

undergraduate programs, matching graduate proﬁles

to job market expectations and hoping to obtain infor-

mation about the entry-level job roles’ skill require-

ments. Finally, we aimed to validate our changes by

estimating how well our redesigned program deliv-

ers the skills required by targeted job roles and subse-

quent careers.

Unfortunately, we were not able to fully appreci-

ate skill frameworks for such a purpose. The SFIA

framework deﬁnes job roles with very few skills (i.e.

usually using three to ﬁve skills), serving only a lim-

ited understanding of the complex professional re-

quirements. Contrary to SFIA, ESCO and ASC deﬁne

numerous essential and optional skills, competencies

and knowledge. Yet, we found the deﬁnitions some-

what ﬂawed, strangely speciﬁc or lacking.

For example, the Software Developer occupation

lists “create ﬂowchart diagram” or “perform scien-

tiﬁc research” as essential skills, yet “use object-

oriented programming” or “develop creative ideas”

only as optional. We would expect this to be the

other way around. Moreover, optional knowledge

lists many historical or scientiﬁc languages such as

Erlang, Cobol, Haskell or Smalltalk, but lacks modern

in-demand languages such as GoLang, Kotlin, Rust or

Julia. We understand that it is difﬁcult to list all the

currently used technologies as that list would be very

long. In this case, ASC has a much better approach;

instead of listing individual programming languages,

it deﬁnes and maps a single technological skill, Soft-

ware development and programming languages, al-

beit losing the opportunity to specify the exact re-

quired technologies.

Moreover, in stark contrast to governmental ef-

forts that increasingly use and depend on skill frame-

https://www.abs.gov.au/statistics/classiﬁcationsanzs

co-australian-and-new-zealand-standard-classiﬁcation-

occupations/latest-release

http://data.europa.eu/esco/occupation/f2b15a0e-e65a-

438a-affb-29b9d50b77d1

CSEDU 2023 - 15th International Conference on Computer Supported Education

138

works, (Manyika et al., 2017) surveyed numerous

CEOs and members of senior and middle manage-

ment, as well as reviewed multiple works dealing

with employment and “employability”, discovering

that such a simplistic view is not possible. The skill

requirements for the same role vary across industries

or the use of underlying skill frameworks. Employ-

ers often rated technical or digital skills towards the

bottom, primarily due to their “short shelf life”. Sim-

ilarly, (Holmes, 2001) and (Collet et al., 2015b) ex-

pose the ﬂaws in trying to deﬁne the optimal graduate

skill sets (i.e. graduate proﬁles) due to a large pool of

diverse graduates trying to match the requirements of

individual employers from varied backgrounds with

unique values.

Additionally, (Holmes, 2001) argues that the con-

cept of graduate employability cannot be deﬁned by

the acquisition of measurable skills due to the unmea-

surable complexity of personal qualities. Instead of

skills, (Manyika et al., 2017), (Collet et al., 2015b)

or (Hinchliffe and Jolly, 2011) note the importance

of identity, values and abilities as a driving force of

graduate employment. For example, it is much more

important to be trustworthy and reliable with good

communication abilities than to have speciﬁc techni-

cal skills.

Such ﬂaws of the skill-driven approach are further

ampliﬁed by using context-agnostic skill frameworks

designed for organisations of any size, sector or in-

dustry. Role proﬁles are rarely updated and become

“stale” ’ or obsolete. However, governments aim to

help drive policy, recruitment and training by provid-

ing a data-driven classiﬁcation of skills using such

skill frameworks. Considering the previously men-

tioned research stressing the different, ever-changing

needs of individual businesses and personalised pref-

erences, such efforts have a questionable effect and

informative value for employers and potential em-

ployees (Manyika et al., 2017) (Council et al., 2012).

Consequently, in this work, we present means of

building datasets that allow for the automated ex-

traction of skills and knowledge from job advertise-

ments. Using this data, we generate highly granular

skill and knowledge proﬁles, understanding nuances

in requirements across different industries, regions or

employers. Consequently, using the same extraction

approach, we analyse skill and knowledge acquisition

in educational activities, providing recommendations

to students about subjects or courses that best cover

the requirements of particular job roles.

3 SKILL EXTRACTION

Students attend university and select speciﬁc subjects

to learn skills, knowledge, technologies and abilities,

developing their intelligence and increasing their em-

ployment chances (for simplicity, in the rest of this

paper, we write only “skills” instead of “skills, knowl-

edge, technologies and abilities”). Employers seek

out people with skills to take on speciﬁc roles. Identi-

fying the skills required for a job or gained from tak-

ing a subject allows us to optimise the coverage of

job roles available after completing a degree. There-

fore, skill extraction is essential to a subject/role rec-

ommender system.

Our approach relies on understanding the skill re-

quirements of individual jobs based on job advertise-

ments’ descriptions. Then, we can build role skill pro-

ﬁles for each job role, aggregating the skills from re-

lated job ads. Optimally, job advertisements would

use a robust skill framework and explicitly list the

required skills, knowledge and technologies. Educa-

tional institutions would use the same skill framework

to specify their teaching skills and compare the de-

mand and supply. Unfortunately, this is not the case.

As a result, we depend on skill extraction from the

description of job advertisements and related meta-

data. Since our institution does not explicitly specify

the skills covered in educational content, we also used

skill extraction to create the list of skills for subjects.

Please note that we only used the automated approach

to prepare recommendations for subject coordinators,

who further redacted and curated the list.

To automatically extract skills from text,

(Kivim

aki et al., 2013) used a graph-based approach,

mapping the text to Wikipedia articles which map

further to LinkedIn

skills. Unfortunately, we did

not ﬁnd a way how to use their method with our

target ACS skill framework or any other framework.

Moreover, the author’s approach worked well for

extracting skills from scientiﬁc articles with match-

ing inputs in Wikipedia, less so with advertised job

ads. More recently, the team at LinkedIn (Gupta

et al., 2020) presented their (Bhola et al., 2020)

system, which uses salience and a market-aware skill

extraction system. Skills extracted by this system

are not only those found in the description of the job

advertisement but also those generally required by

related job-role.

Moreover, the system also ﬁlters out the skills for

which there is a supply on the market. While this sys-

tem would possibly be a good match for our purposes,

the underlying data structures used for training are not

available outside LinkedIn. Unfortunately, this is the

https://linkedin.com

JobIQ: Recommending Study Pathways Based on Career Choices

139

trend of most of the related works in the area, with

only a minimal number of works releasing their data

(Bhola et al., 2020) (Zhang et al., 2022) and none,

apart from (Zhang et al., 2022), releasing their anno-

tation guidelines.

Other approaches (Smith, 2021) or (Zhang et al.,

2022) use natural language processing to automati-

cally build the skill database with high accuracy in

detecting skills and knowledge. While this approach

is interesting by having the possibility to see emerg-

ing skills, it does not ﬁt our purpose and serves only

complementary functionality to do so. The main rea-

son is that we need to explain to our users what it

means to have a speciﬁc skill and how to obtain it. Us-

ing automatically extracted emerging skills makes ex-

planations very difﬁcult while using established skill

frameworks, we can provide them.

As a result, our requirement is for a hybrid system

that detects skills from a selected (interchangeable)

skill framework. Using simple text matching is im-

possible as most skill frameworks skills are deﬁned

using multiple words, which can be expressed in var-

ious ways (for example, ASC skill “Provide technical

support for computer network issues”). As a result,

we devised a skill-matching strategy based on match-

ing sentence embeddings.

3.1 Skill Extraction Using Sentence

Embeddings

The goal of skill extraction is to determine whether a

given text contains a notion of skill from a speciﬁc

skill framework. In our case, we are trying to ex-

tract skills from job and subject descriptions. Note

that descriptions are written in natural language, often

specifying activities related to applying a skill rather

than providing the name of the skill. This prohibits

us from quickly identifying the skills. Initial exper-

iments using keyword extraction provided poor re-

sults. Therefore, we needed to understand the mean-

ing of sentences, not the words themselves. The re-

cent advancement in natural language processing us-

ing word embeddings (Jatnika et al., 2019) and deep

transformer networks (Kenton and Toutanova, 2019)

provided us with vector representations of sentences

that capture the meaning of the sentences rather than

just the words. By converting the sentences from a job

description and a skill description into a set of vectors,

we can identify how similar each skill is to a particu-

lar role. The vectors are created so that the similarity

of two sentences is computed using the cosine of the

angle between the two vectors. We use the sentence

embedding approach to extract skills belonging to a

speciﬁc skill framework from arbitrary descriptions

(e.g. job or subject description). Formally, we deﬁne

skills framework as:

Deﬁnition 1. A skill framework S is a set of skills or

other discrete competencies, where each skill is rep-

resented in natural language.

Deﬁnition 2. The skill-matching function M(t, d) →

s, given a text t and a description d of the skill, re-

turns a Skill Match Strength s, represented by a value

from interval [0, 1] with 0 representing no match, val-

ues between 0 and 1 a partial match and 1 representing

a complete match.

In our case, the skill-matching function M uses

the sentence embedding approach. It takes the text of

the job advertisement and, for a skill represented by

its title and description, estimates whether that skill

is mentioned (i.e. embedded) in the job description.

This skill is more probably mentioned when the value

of skill match strength is closer to 1.

Deﬁnition 3. The skill extraction function

E(S, M, s, t) → S

′

given:

• S is a Skill Framework

• M is a Skill-Matching Function

• s is a minimum value of Skill Match Strength

• t provided text t (e.g. job-description)

extracts a subset of skills S

′

⊆ S from a pro-

vided text t (e.g. job-description) whose skill match

strength is bigger or equal to provided s.

In our case, we take a job or subject description,

and the Skill Extraction Function extracts all skills

from the ASC skill framework most probably con-

tained in the job description. Through experimental

evaluation, we discovered the value 0.45 for s being

the optimal skill match strength for extraction using

our approach.

3.2 Evaluating Skill Extraction

The Australian Skill Classiﬁcation (ASC) dataset de-

ﬁnes over 2000 skills assigned to every role in the

ANZSCO dataset. For each ANZSCO role, ASC pro-

vides a list of skills, core competencies (e.g. numer-

acy, literacy) and used technologies.

For example, a Web Developer contains 32 skills,

such as

• Design websites or applications

• Update website content or

• Test software performance

Furthermore, the Web Designer role uses 12 tech-

nologies such as

CSEDU 2023 - 15th International Conference on Computer Supported Education

140

• Software development and programming lan-

guages,

• Graphics or photo imaging software, and

• Social media and web publishing software.

We verify the validity of our approach by

analysing how well our algorithm extracts skills from

job advertisements related to roles speciﬁed in the

ASC dataset. We also verify whether our approach

can discover new, feasible skills not mentioned in the

ASC dataset.

Our dataset has 41,273 job advertisements down-

loaded from Australian job websites. First, we classi-

ﬁed each job advertisement with one of the ANZSCO

roles, discovering over 700 ANZSCO roles. Second,

from job descriptions, we extracted the skills and used

technologies. Figure 1 summarises the results of skill

extraction, analysing the coverage of ASC skills. We

see that we extracted 100% of the skills from ASC in

almost 8% of job roles. Overall, for 73% of job roles,

we extracted at least 50% of the skills deﬁned in the

ASC dataset.

Furthermore, we can see a correlation between the

number of job advertisements and the accuracy of our

extraction. The curve is not monotonic, declining to

the group of job advertisements with 100% skill cov-

erage. The reason is that in this group, there are ad-

vertisements for expert jobs in medicine, which pro-

vide exhaustive descriptions of activities, allowing us

to extract all of the ASC skills. Job advertisements in

other areas were less speciﬁc. We found no correla-

tion with the average length of the description of the

job advertisement.

Figure 1: Completion Criteria of Mathematics Major.

The roles with high coverage of skills contained

mainly medical roles such as “Resident Medical Ofﬁ-

cer” or “Registered Nurse (Aged Care)” with a de-

tailed description of responsibilities. On the other

hand, low-coverage roles, such as the “Web Devel-

oper” role, often described only the company culture

and left responsibilities as assumed. For example,

some of the skills from the ASC framework not de-

tected in the “Web Developer” job ads were:

• Troubleshoot issues with computer applications

or systems

• Develop diagrams or ﬂow charts of system opera-

tion

• Prepare graphics or other visual representations of

information

Such skills are usually assumed and only rarely

speciﬁed in job advertisements. Consequently, in Jo-

bIQ, we use the ASC skills as a baseline skill-set for

every role, extending them with detected skills. But,

the JobIQ approach becomes helpful when detecting

ASC skills not covered by the ASC framework role

proﬁles. JobIQ allows for a high granularity of anal-

ysis, building personalised skill proﬁles by locality,

employer or industry.

Figure 2: The emergence of skill framework roles in skill

proﬁles.

Figure 2 depicts the number of emergent skills.

This time we only considered skills that appear in

at least 10% of job advertisements. We see that the

ASC coverage dropped signiﬁcantly, to an average of

10.8%, but detecting three times more skills than in

the ASC dataset on average. For example, the system

has detected 13 emerging skills for the

“Software and Applications Programmers” contained

in at least 10% of the job advertisements, including:

• Support individuals with diverse needs to under-

stand, access and utilise information or services

• Deliver culturally appropriate programs, policies

or services

• Maintain a working understanding of the cultural,

diversity and accessibility needs of others and

how this applies to the role

• Evaluate projects to determine compliance with

technical speciﬁcations

• Gather information to provide services to clients

JobIQ: Recommending Study Pathways Based on Career Choices

141

• Provide technical support for software mainte-

nance or use

Furthermore, we used our approach to extract ASC

skills from subjects being offered at our institution.

The extraction coverage was very similar to the exper-

imental data. We also conﬁrmed that the skill detec-

tion failed when the subject description lacked infor-

mation about related activities or outcomes. Updating

the descriptions to include such information improved

our algorithm’s accuracy and coverage. This proved a

good strategy for analysing our subject catalogue, al-

lowing us to consult subject coordinators about pos-

sibly lacking descriptions of their subjects and pro-

viding recommendations for improvement. These

improvements improve student understanding of the

skills they develop during the offered subject or pro-

gram.

4 RECOMMENDING STUDY

PROGRAMS TO PROSPECTIVE

STUDENTS

In the previous sections, we presented our approach

to extracting skills from job advertisements and ed-

ucational subjects. In this section, we demonstrate

how we can build a recommender system to in-

form prospective students who study programs de-

liver skills, knowledge and technology most related

to their career choices. Our approach is novel as ex-

isting methods focus on recommendations based only

on student data, not considering external sources. For

example, (Chaturapruek et al., 2018) provides recom-

mendations based on historic student preferences, or

(Farzan and Brusilovsky, 2006) considers student rat-

ings. More recently, researchers employed neural net-

works to recommend courses and subject sequences

based on enrolment data (Pardos et al., 2019) or grade

prediction (Ren et al., 2019). To our knowledge, none

of the existing systems considers the job market pref-

erences.

We need a Skill Framework and a set of job roles

with their skill proﬁles to build such a recommender

system. We also need a dataset of job advertisements,

where each ad speciﬁes the description and the job

role required.

Similarly to a skill proﬁle of a job role, we can

use the skill extraction function to extract all the skills

from a job advertisement description to build a job

advertisement skill proﬁle.

Deﬁnition 4. Considering a skill framework S and a

job advertisement a, a job advertisement skill proﬁle

′

⊆ S is a subset of skills from S contained within the

job advertisement a.

Such a proﬁle contains the set of skills that are

found relevant only to the given advertisement. We

can then ﬁlter and aggregate all the job advertisement

skill proﬁles (e.g. by industry, employer) to build

a highly granular job role proﬁle that considers the

proportion of job advertisements requiring each skill.

Formally we deﬁne:

Deﬁnition 5. Considering a skill framework S with

n roles, a job role and a set of job advertisement

skill proﬁles, we deﬁne a job role proﬁle as a vec-

tor (c

, c

. . . c

) where each vector element c

where

k ≤ n represents the ratio of job ads containing the

skills s

and a total number of advertisements for this

role.

In other words, for each skill from the skill frame-

work, the job role skill proﬁle deﬁnes how many job

advertisements require this skill in proportion to all

the available job advertisements for that given role.

This approach allows us to model demand for partic-

ular skills.

For example, consider a skill framework with

three skills, two job roles r

, r

, two job advertise-

ments for role r

and three for role r

with job adver-

tisement skill proﬁles from Table 1. Then, the job role

proﬁles would be those found in Table 2.

Table 1: Job advertisement skill proﬁles for roles r

and r

Ad # s

Role r

1 0.6 0.0 0.2

2 0.4 0.0 0.0

Role r

1 0.0 0.7 0.0

2 0.0 0.8 1.0

3 0.0 0.2 0.4

Table 2: Job role skill proﬁles for roles r

and r

with skill

match strength threshold 0.

Role r

1.0 0.0 0.5

Role r

0.0 1.0 0.66

We can further specify that we only consider skills

if their skill match strength is higher than a thresh-

old value. Table 3 depicts how the role skill proﬁle

changes when considering a threshold value. For ex-

ample, skill s

drops the value from 1 to 0.66 as only

two advertisements match this skill with strength 0.45

or above.

With job role proﬁles, we can assess which skills

are desired for a given job role. What we aim to do,

CSEDU 2023 - 15th International Conference on Computer Supported Education

142

Table 3: Job role skill proﬁles for roles r

and r

with skill

match strength threshold 0.45.

Role r

0.5 0.0 0.5

Role r

0.0 0.66 0.33

is to recommend the educational program which will

deliver those skills. To achieve this goal, similar to the

job advertisement skill proﬁle, we deﬁne the subject

skill proﬁle and a pathway skill proﬁle:

Deﬁnition 6. Considering a skill framework with n

roles, we deﬁne a subject skill proﬁle as a vector

, c

. . . c

) where each vector element c

where k ≤

n represents either the skill match strength of skill s

or zero, if that skill match strength is below the thresh-

old value.

In other words, the subject skill proﬁle provides

which skills are delivered in a subject, along with a

probability (i.e. strength) of their detection in a de-

scription of the subject.

Deﬁnition 7. Considering a skill framework S with n

roles and an educational pathway with l subjects, we

deﬁne pathway skill proﬁle as a vector (d

, d

. . . d

)

where each vector element d

where k ≤ n represents

a maximum value of a c

from all of the l subject skill

proﬁles.

In other words, we build a pathway skill proﬁle by

combining all the skills delivered in all the subjects

in the pathways and remembering only the maximum

match strength value for each detected skill. By doing

so, we understand what skills are most probably deliv-

ered in the given pathway of an educational program.

For example, Table 4 depicts the pathway with only

two subjects and their related subject skill proﬁles:

Table 4: Example pathway with two subject skill proﬁles.

Subject

0.7 0.0 0.5

Subject

0.0 0.6 0.33

Taking the maximum value from each subject skill

proﬁle, the resulting pathway skill proﬁle would be:

(0.7, 0.6, 0.5)

We are almost ready to recommend educational

programs for a deﬁned job role. First, we generate

all possible pathways in an educational program. We

can only consider representative pathways through

the program if there are too many combinations. Sec-

ond, we assess how well a particular pathway matches

the skill requirements of a given role. We can do this

by multiplying the values of a role skill proﬁle with

a pathway skill proﬁle and then summing up all the

elements.

∑

, c

. . . c

) × (d

, d

. . . d

)

Please remember that the role skill proﬁle deﬁnes

how important the skill s

is for a given role, and the

pathway skill proﬁle deﬁnes how strongly skill s

covered in the pathway. We obtain a weighted value

of skill importance by multiplying these two vectors.

For example, if a skill is not essential, i.e. its role

proﬁle value is 0, and no matter how well this skill is

covered in a pathway, it will not affect the ﬁnal value.

Third, we process all the pathways and remember

only the value of the best matching pathway as a pro-

gram representative. With this approach, for a given

job role, we can order all the programs based on the

value of the best matching pathway and recommend

the programs that deliver most of the desired skills for

a given job.

The JobIQ recommender system pre-computes the

recommendation data that is then delivered in real-

time through the web interface depicted in Figure 3.

We evaluated the performance of the recommenda-

tion by assessing which programs are recommended

for given job roles. The recommendations for more

technical programs performed as expected, recom-

mending our IT programs for ICT roles and busi-

ness programs (accounting) for business (accounting

roles). More interesting recommendations appeared

when assessing other roles, such as school teachers

ranking our arts or criminology programs high.

4.1 Other Recommendations

Following a similar approach, we were able to deliver

other recommendations that our students desired.

• The recommendation of electives helps current

students ﬁnd electives that best match their career

choices. Through careful selection of often inter-

disciplinary electives, students can maximise their

skill and knowledge uptake. The recommendation

uses the same approach as program recommenda-

tions, using subject skill proﬁles instead of path-

way skill proﬁle.

• The recommendation of careers and individual

jobs based on educational programs is an inverse

recommendation strategy to program recommen-

dation. This strategy is popular with current stu-

dents who look for job opportunities towards the

end of their program.

• Benchmarking and comparison of programs allow

prospective students to compare the different op-

tions that programs provide concerning opportu-

nities in current job markets.

JobIQ: Recommending Study Pathways Based on Career Choices

143

Figure 3: JobIQ system interface.

Considering that our system processes daily op-

portunities on job markets, we built a proactive job,

and career-monitoring system that overlooks student

performance, provides updates on skill development

related to current opportunities and ﬁnds alternative

pathways in case a student needs more practice or

fails a subject.

5 CONCLUSIONS AND FUTURE

WORK

Our approach delivers a novel strategy for recom-

mending educational programs or electives based on

career choices. For this purpose, we designed a skill

extraction system that can compare the skills cover-

age based on supply (i.e. education) and demand (i.e.

job markets). Overall, the accuracy of the automated

extraction depends on the quality of the description.

We can build highly accurate proﬁles only with a suf-

ﬁcient number of jobs in the dataset. As a side effect,

our approach is helping curriculum designers to write

better subject descriptions mentioning the skills and

capabilities covered during the subject delivery, pro-

viding more information to students. We developed a

recommender system that allows review and assigns

skills to subjects, helping to order and assign correct

skills to knowledge subjects. We will evaluate this ap-

proach and further optimise our skill extraction strat-

egy as part of our future work.

We are currently working on enabling JobIQ to

help with life-long learning. Since the system under-

stands its users’ capabilities, skills and knowledge, it

can monitor for career opportunities and up-skilling.

For example, it can provide notiﬁcations of opportu-

nities to take a speciﬁc course to achieve the skill set

necessary for a new, possibly more exciting career.

REFERENCES

Apps, J. W. (1988). Higher Education in a Learning Soci-

ety. Meeting New Demands for Education and Train-

ing. ERIC.

Bhola, A., Halder, K., Prasad, A., and Kan, M.-Y. (2020).

Retrieving skills from job descriptions: A language

model based extreme multi-label classiﬁcation frame-

work. In Proceedings of the 28th International Con-

ference on Computational Linguistics, pages 5832–

5842.

Chaturapruek, S., Dee, T. S., Johari, R., Kizilcec, R. F., and

CSEDU 2023 - 15th International Conference on Computer Supported Education

144

Stevens, M. L. (2018). How a data-driven course plan-

ning tool affects college students’ gpa: evidence from

two ﬁeld experiments. In Proceedings of the ﬁfth an-

nual ACM conference on learning at scale, pages 1–

10.

Collet, C., Hine, D., and Du Plessis, K. (2015a). Employa-

bility skills: perspectives from a knowledge-intensive

industry. Education+ Training.

Collet, C., Hine, D., and Plessis, K. d. (2015b). Employa-

bility skills: perspectives from a knowledge-intensive

industry. Education + Training, 57(5):532–559.

Council, N. R. et al. (2012). Education for life and work:

Developing transferable knowledge and skills in the

21st century. National Academies Press.

Daly, A. and Lewis, P. (2020). The proposed job-ready

graduate package: A misguided arrow missing its

target. Australian Journal of Labour Economics,

23(2):231–251.

Farzan, R. and Brusilovsky, P. (2006). Social navigation

support in a course recommendation system. In In-

ternational conference on adaptive hypermedia and

adaptive web-based systems, pages 91–100. Springer.

Gupta, R., Liu, Y., Shah, M., Rajan, S., Tang, J., Prakash,

B. A., Shi, B., Yang, J., Guo, F., and He, Q. (2020).

Salience and Market-aware Skill Extraction for Job

Targeting. Proceedings of the 26th ACM SIGKDD

International Conference on Knowledge Discovery &

Data Mining, pages 2871–2879.

Herbert, N., Lewis, I., and De Salas, K. (2013). Career out-

comes and sﬁa as tools to design ict curriculum. In

ACIS 2013: Information systems: Transforming the

Future: Proceedings of the 24th Australasian Confer-

ence on Information Systems, pages 1–11. RMIT Uni-

versity.

Hinchliffe, G. W. and Jolly, A. (2011). Graduate iden-

tity and employability. British Educational Research

Journal, 37(4):563–584.

Holmes, L. (2001). Reconsidering Graduate Employability:

The ’graduate identity’ approach. Quality in Higher

Education, 7(2):111–119.

Jatnika, D., Bijaksana, M. A., and Suryani, A. A. (2019).

Word2vec model analysis for semantic similarities in

english words. Procedia Computer Science, 157:160–

167.

Kenton, J. D. M.-W. C. and Toutanova, L. K. (2019). Bert:

Pre-training of deep bidirectional transformers for lan-

guage understanding. In Proceedings of NAACL-HLT,

pages 4171–4186.

Kivim

aki, I., Panchenko, A., Dessy, A., Verdegem, D.,

Francq, P., Bersini, H., and Saerens, M. (2013). A

graph-based approach to skill extraction from text.

In Proceedings of TextGraphs-8 graph-based methods

for natural language processing, pages 79–87.

Manyika, J., Lund, S., Chui, M., Bughin, J., Woetzel, J., Ba-

tra, P., Ko, R., and Sanghvi, S. (2017). Jobs lost, jobs

gained: Workforce transitions in a time of automation.

McKinsey Global Institute, 150.

Pardos, Z. A., Fan, Z., and Jiang, W. (2019). Connec-

tionist recommendation in the wild: on the utility and

scrutability of neural networks for personalized course

guidance. User modeling and user-adapted interac-

tion, 29(2):487–525.

Peetz, D. (2019). The realities and futures of work. ANU

Press.

Ren, Z., Ning, X., Lan, A. S., and Rangwala, H. (2019).

Grade prediction based on cumulative knowledge and

co-taken courses. International Educational Data

Mining Society.

Smith, E. W. (2021). Skill Extraction for Domain-Speciﬁc

Text Retrieval in a Job-Matching Platform. pages

116–128.

Xu, M., David, J. M., Kim, S. H., et al. (2018). The fourth

industrial revolution: Opportunities and challenges.

International journal of ﬁnancial research, 9(2):90–

95.

Zhang, M., Jensen, K., Sonniks, S., and Plank, B. (2022).

SkillSpan: Hard and Soft Skill Extraction from En-

glish Job Postings. Proceedings of the 2022 Confer-

ence of the North American Chapter of the Associa-

tion for Computational Linguistics: Human Language

Technologies, pages 4962–4984.

JobIQ: Recommending Study Pathways Based on Career Choices

145