The labor market is constantly evolving. Occupations are changing, being added, or disappearing to fit the needs of today's market. In recent years the pace of this change has accelerated, due to factors such as globalization, digitization, and the shift to working from home. Diferent factors are relevant when selecting employment, e.g., cultural fit, compensation, provided degree of freedom. To successfully fulfill an occupation the gap between required (by the job) and possessed (by the job seeker) skills needs to be as small as possible. Decreasing this skill-gap improves the fit between a job candidate and occupation. In this paper we propose a custom-built Skills & Occupation Knowledge Graph (KG) that fits the above described dynamic nature of the labor market, by leveraging existing skills and occupation taxonomies enriched with external job posting data. We leverage this KG and explore several applications for skillsbased matching of jobs to job seekers. First, we study link prediction as a means to quantify relevance of skills to occupations, which can help in prioritizing learning and development of employees. Next, we study node similarity methods and shortest path algorithms for career pathfinding. Finally, we leverage a term weighting method for identifying which skills are most “distinctive” for diferent (types of) occupations.
• Computing methodologies → Ontology engineering; • Theory of computation → Graph algorithms analysis; • Information systems → Content analysis and feature selection. labor market, skill matching, knowledge graphs
In recent years the number of people that change their job is
increasing [
In addition, with demand of skills changing over time, having
the correct skills for specific occupations is more crucial than ever.
The increasing amount of digitization has made computer skills
more valuable [
The volatility in the labor market results in a change of
occupations with new required skills, and being able to keep up with
the latest developments is a challenge. To find relevant vacancies
and job postings, individuals can use external services to match
their skills with their desired work. In 2019, employment
agencies were responsible for fulfilling 10% of the available jobs in the
Netherlands [
As explained above, in recent years the labor market has become
more competitive, and requirements more dynamic. As a result of
this, there is a rising interest in skill-based matching of candidates
to jobs [
To facilitate candidate to job posting matching, it is important to know which skills are relevant, in demand, and in supply. Here, the need for a flexible data representation for skills arises. This representation should facilitate various tasks, such as a skills similarity metric to be able to quantify likeliness between skills, skills-tooccupation similarity metrics, to help people navigate the labor market and find new occupations, and understanding which skills relate to which occupations to inform which skills are needed for desired occupations. And since relations between skills and occupations are not static and need robust and accurate updating methods to ensure the information does not get outdated.
In this paper we address the task of skills and occupation graph construction which we describe in Section 2, and apply this data representation to the following set of use-cases: link prediction for identifying novel skills-occupation relations in Section 3, skillsbased occupational similarity for career pathfinding in Section 4, and identifying distinctive skills per occupational group for learning & development in Section 5. 2
Our Skills & Occupational KG is based on existing structured data, more specifically, we combine the ISCO (occupations) and ESCO (skills) taxonomies (bottom row in Figure 1). Next, we enrich this existing data with information from noisy, unstructured job postings (top row in Figure 1) to ensure our KG represents the current state of the labor market.
Job Postings
The first step involves constructing a shared Skills & Occupational
Knowledge Graph, through combining the existing ISCO and ESCO
taxonomies.
2.1.1 ISCO (occupations). The International Standard
Classification of Occupations (ISCO) is ordered as a taxonomy of
occupational groups with four granularity levels across ten diferent major
groups. An occupation is defined as “a set of jobs whose main tasks
and duties are characterized by a high degree of similarity”, where
a job is defined as “a set of tasks and duties performed, or meant to
be performed, by one person, including for an employer or in
selfemployment.” [
Professional Technicians and associate professionals Clerical support workers Service and sales workers Skilled agricultural, forestry and fishery workers Craft and related trades workers Plant and machine operators, and assemblers Elementary occupations
Armed forces occupations
2.1.2 ESCO (skills). We define our initial high-level occupation
groups by using the ISCO standard. For skills, we turn to The
European Skills, Competences, Qualifications and Occupations (ESCO)
taxonomy [
The ESCO covers 13,485 skills, connected to 2,942 occupations (in 27 languages).
We link our ISCO occupations to ESCO by using the direct links that are defined between ISCO level 4 groups (most finegrained/lowest level of the taxonomy) and ESCO concepts, in the ESCO. These links between ESCO and ISCO are not (necessarily) 1-to-1, as multiple ESCO occupations can be linked to a single (level 4) ISCO group.
In Figure 2 we illustrate this connection between ISCO and ESCO. ESCO occupations are shown in blue, with ISCO occupation groups in purple. In addition to the ESCO occupations shown in the image, ESCO also defines skills (not shown), e.g., the ESCO occupation “Cattle breeder,” has skills linked to them such as “feed livestock” and “assist animal birth.” 2.2
Now that we have our high-level KG structure based on ISCO and ESCO, which defines occupations and skills as nodes, and edges as links between ESCO and ISCO objects, we turn to job posting data to account for the dynamic nature of associations between skills and occupations, as described in Section 1. To make sure our KG reflects the current status of the labor market, we use information from job postings to enrich the structure of our KG. More specifically, we create additional edges by identifying and extracting ESCO skills for each job posting’s ISCO occupation group, and assign weights to edges by relying on co-occurrence statistics of skills and occupations.
This second step of our process revolves around extracting skills
from job postings. We describe our job posting dataset in Section
2.2.2, our approach for skill extraction in Section 2.2.2, and how we
match extracted skills to ESCO skills in 2.2.3.
2.2.1 Vacancy data. Our vacancy dataset consists of sample of
600,000 Dutch vacancies collected by Jobdigger [
Posting
Normalized
candidate skil
n-gram of
candidate skil
by selecting a uniform distribution of ISCO level 1 occupations, to
make sure our set covers the entire breadth of the labor market.
Prior to sampling our set at the ISCO level 1, the initial dataset was
cleaned by discarding low quality and noisy job postings, such as
postings that represented multiple occupations, or job postings that
contained a low number of sentences. Here, we treat vacancy data
as a proxy for the demand in the job market. By doing so, internal
promotions and career paths and informal channels are not taken
into account.
2.2.2 Skill Extraction. For skill extraction we rely on the
industrystandard Textkernel Extract [
Our final KG, resulting from the process shown in Figure 1 and described in the previous section, consists of 1,220 nodes, of which 983 represent (ESCO) skills, and 237 (ISCO) occupations. These nodes are connected through 3, 910 edges, with an average node degree of 6.4.
This KG is a subset of the full ESCO (13.485 skills), and ISCO (436 occupations) taxonomies. There are several reasons why our KG is a subset and does not span the entirety of the ISCO and ESCO taxonomies.
First, it is conceivable that not all ISCO occupations are in current demand, e.g., we found that there were no vacancies for ISCO occupation code 8111: “mining-plant operators,” which is not surprising with currently no mines in operation in The Netherlands. Next, it is likely we are dealing with coverage issues, from (i) the likely incomplete coverage of the TextKernel Extract method we use for skill extraction, and (ii) our skills matching methodology further reducing the number of identified skills. As the focus of this paper is on downstream applications, we consider matching out of scope, and rely on our naive but solid character -grams-based method. 3
One of the challenges of modeling skills and occupations is the dynamic nature of the labor market. In this section we explore our first down-stream application of our data-driven dynamically constructed Skills & Occupation Knowledge Graph: matching occupations to skills. We focus on discovering novel connections between skills and occupations through leveraging the structure of our knowledge graph enriched with job posting data.
More specifically, in this section we compare link prediction algorithms, to quantify the relatedness between a skill and occupation node, in order to discover novel connections between skills and occupations, not present in our initial KG. We describe our two link prediction methods in the following sections, the first, Preferential Attachment, is described in Section 3.2.1, next, Node2Vec is described in Section 3.2.2 3.1
We employ link prediction to estimate the relatedness between skills and occupation nodes. To evaluate and reliably compare diferent methods, we first split our KG into train, test, and validation sets. More specifically, we sample 55% of all edges for training the link prediction algorithms (where applicable), leaving leave 30% for testing, and 15% for validation. For each existing pair of occupation and skills node — which we consider a positive sample in our train, test and validation sets — we randomly generate a negative sample (i.e., a pair of skills and occupation nodes that do not exist in our KG). An overview of the number of edges in each set is shown in Table 2.
3.2.1 Method 1: Preferential Atachment (PA). The first link
prediction method is preferential attachment [
A higher score here corresponds to a larger probability the nodes are connected. The intuition behind this is that if both nodes have a high amount of neighbors the nodes might function as a hub. Most graphs have the property that hubs have a higher chance to be connected.
To compute all scores, we represent our KG as a matrix, where
each node is represented as a row and a column. Note that this
matrix is symmetric since the value for row and column is equal
to the value at row and column . At the intersecting cell of two
nodes, we store the preferential attachment. We normalize this
matrix by dividing each score by the maximum Closeness score, to
ensure that each value is between 0 and 1. We consider the resulting
normalized Closeness score as the probability the corresponding
nodes are related.
3.2.2 Method 2: Node2Vec (N2V). The second link prediction method
we use is the Node2Vec algorithm [
These parameters were selected after a grid search on a large number of possible combinations of parameters. 3.3
When the number of positive and negative edges in the test set
is equal, PA outperforms the more complex N2V method, with an
f1-score for the positive class of 0.78 against 0.65. In most realistic
situations however, we may want to explore how a node can be
linked to any other node, making the number of comparisons, or
edges to predict 1-to-(N-1), i.e., for each node we compare each other
node (excluding self). To approximate this real world performance
the ratio of negative to positive edges should reflect these more
realistic proportions. To do so we compute F1-score at increasing
ratios of positive-to-negative edges, ranging from 1 (as shown in
Table 3) to 7. Results are shown in Figure 4. The figure shows that
up to ratio of 3:1, N2V is on par with PA, but as ratios increase,
N2V outperforms PA, suggesting N2V is better suited for most real
world situations.
1.0
0.8
s
s
a
l
c
ive0.6
t
s
o
p
e
h
tf
o
re0.4
o
c
S
1
F
0.2
Now that it has been established that N2V is more suitable for our
task, we aim to employ this algorithm to predict the relationships
between occupations and skills. When doing so we need to realize
that the graph which we use as input is imperfect in terms of
correctness and completeness [
Looking at the false positives of the algorithm, skills that are — according to our dataset — incorrectly linked to occupations can be identified. For KG completion, we aim to identify those skills that are not linked to occupations, but should be. Table 4 shows a random sample of False Positives: it reinforces our intuition that link prediction can be employed for KG completion, as some of the predicted edges make sense, e.g., the skill: “preparing materials for dental procedures” is shown as a relevant skill for the occupation: “dentist.” By consulting domain experts, skills can be eficiently added to enrich the current graph.
To further explore these intuitions, in Figure 5 we show the edges to skill nodes predicted by N2V, for the node representing ISCO code 2611: “Lawyers.” The y-axis shows skills edges, and the x-axes show the link prediction probabilities, for all predictions with a probability>0.5 (i.e., positive predictions by the method). The green bars denote True Positives (i.e., correctly predicted edges between the skill and occupation), and blue bars depict False Positives (skills that are predicted to have an edge with the occupation, but do not exist in our KG). The figure shows “education law” and “investigation research methods” as newly identified skills for lawyers, not found in the original ESCO taxonomy nor in co-occurrences in job postings. 4
According to recent data (2019) 1.1 million people switched
occupation in the Netherlands [
much from the other. Consequently, occupations that share a large number of skills should be easier to transfer between. In this chapter we focus on leveraging skills for better informing transitions between occupations. More specifically, we aim to leverage the KG structure for matching occupations with occupations, to identify how an individual can change jobs in the most optimal way.
To determine the feasibility of an occupation transfer, we propose to model the distance between occupations with Jaccard distance. We compute Jaccard distances between occupations by representing each occupation as the set of its required skills (which we extract from our KG), and computing the overlap between two sets of skills. See Figure for an illustration 6.
In our KG a total of 120, 952 links can be made between pairs of skills and pairs of occupations. From these pairs 89.3% is between skills and 10.7% between occupations. To gain insight in the overall similarity of skills and occupations, we study the distribution of jaccard distances in Figure 7.
Looking at the distribution of Jaccard distance one can see that on average, skills are more similar to one another than occupations. This becomes apparent when looking at the mean value of both distributions: for occupations the mean is 0.96, and for skills around A
C 0.5 D 0.5 B
E F 0.88. Over 99% of occupations have a Jaccard distance between 0.8 and 1, meaning that occupations require distinct skillsets. Both distributions are skewed to the left, meaning that the mean (average of the observations) is left of the mode (most observed value).
In the distribution we see a number of spikes, which can be explained by the prevalence of some fractions over others, e.g., if half of the neighbors are shared, the Jaccard distance will be 12 , which can be achieved in a number of diferent ways. Other spikes occur at additional common fractions such as 23 and 34 .
In Table 5 we show a description of the distance distributions. For both skills and occupations the minimum distance is 0, meaning that a skill is shared by every occupation where the skill is connected to or that two occupations share every skill. An example is “Food service counter attendants” and “Hotel receptionists,” both share the same skillset and thus have a Jaccard distance of 0. Skills with a distance of 0 are for example “Lop trees” and “Pruning techniques.” count mean std min 25% 50% 75% max 107959 0.825 0.163 0.000 0.800 0.875 0.923 0.985 Skill The highest distance found in the dataset is 0.993, this corresponds with the occupations “Electronics engineers” and “Policy administration professionals.” They share at least one skill but are — next to the shared skill — completely diferent. The common skill in this example is “perform project management.” 4.2
With the distances between each occupation and between skills, we can proceed to identify the most eficient transition between every pair of occupations. This is done by assigning the Jaccard distance scores as edge weights between nodes in our graph, to enable computational methods for finding the most eficient path between a start node (the current occupation) and an end node (the desired occupation). We show an example of such a transition in Figure 8: here we set a threshold for the maximum possible distance at 0.8. This threshold was determined to be optimal based on eyeballing and comparing a diferent cutof points. If two occupations are further apart than 0.8 we consider the step too large.
W
Z 0.2 0.4
X 0.9 Y 0.6 0.5
In this example we start at node and want to go to node .
We are not able to directly transition between and because
the occupations are not similar enough (0.9 > 0.8).
4.2.1 Method. Finding the most eficient path in an undirected
weighted graph can be done by applying shortest path algorithms.
For this paper we turn to Dijkstra’s algorithm [
We show a real world example in Figure 9. Due to the COVID-19 pandemic a lot of people find themselves out of a job, especially individuals that work in restaurants. Using the described model we can calculate which occupation has the smallest distance to the occupation: “cook.” Dijkstra’s algorithm yields “bakers, pastrycooks and confectionery makers” as most feasible transition. Next to fine-grained analysis of occupations and skills, gaining macro-level insights is an important task for monitoring and understanding the labor market. The ISCO taxonomy provides multiple levels of granularity, which allows us to aggregate the information contained in our KG at diferent levels, too. In this section we explore a method for identifying the most relevant skills occupations (ISCO level 4) and aggregation of occupations (ISCO level 1-3). More specifically, we match skills to occupations at an aggregated level.
As we’ve seen in the previous section, diferent occupations may share skills. Several skills, such as teamwork, are commonly required for a large number of occupations, which can be considered generic or sector-independent skills. At the other side, we may have highly specialized skills, that are only required for specific occupations or occupation groups. Whether a skill is specifically or generically important can be quantified in diferent ways. For a skill to be specific to an occupation or occupational group, we define two criteria: • A skill needs to be frequently required within its context (occupation or occupation group).
• A skill needs to be characteristic for its context. 5.1
The two criteria described above fit naturally to the Term
Frequency–Inverse Document Frequency (TF-IDF) weighting scheme
for terms [
We “transplant” this TF-IDF weighting scheme from terms in documents to skills associated to occupations. TF-IDF consists of two parts: Term Frequency (TF) is the frequency of a word (skill) used in a given document (observed with an occupation), Inverse Document Frequency (IDF) is a way to discount highly common terms, i.e., it is high when a word (skill) appears in a smaller number of documents (observed with a low number of occupations). Common terms (skills) will thus yield a lower IDF score.
For our TF-IDF-based model, we consider skills identified in job postings terms, and documents can be modeled as a collection of job postings belonging to an ISCO group. The counts of skills, which model term frequency, correspond to the number of times a skill is found in a job posting associated to a certain ISCO code. (2) 5.2
5.2.1 Level 1 ISCO groups. The resulting score provides us with skills that are common for a given occupation (group) but uncommon in all other occupation(s) (groups). Table 6 shows the top 5 skills for the level 1 ISCO groups.
In this table Microsoft Ofice appears both in the Managers and Clerical support workers groups. For this skill to score high in multiple contexts (occupation groups) the frequencies need to be substantial in both, to be able to compensate for the IDF component of the metric. In the Managers group, Microsoft Ofice has a TF of 9% and in Clerical support workers a TF of 5%. 5.2.2 Multiple ISCO levels. ISCO level 1 helps us to understand which skills are relevant for the least granular level; to deepen our understanding we look at the development of multiple layers of ISCO group 2 in Figure 10. Here, we show the 3 most relevant skills for several ISCO levels of the “Professionals” ISCO group.
We notice the following: First, communication-related skills appear in multiple forms across occupation groups. The terms communication, communication sciences, communication studies, ICT communication protocols, manage online communications and communication disorders seem to be closely related. Because these skills are defined as distinct skills, each skill receives its own ranking. This concept can appear multiple times.
Next, “Nursing professionals” and “Nursing and midwifery professionals” share the same set of relevant skills, which are highly similar to those of their parent group “Health professionals”. Skills that appear in those groups are the most frequent skills in the parent group.
Finally, the further down the figure we go, the more specialized the skills appear to be, and more specialized skills, such as “dental Professionals (2) Network marketing Manage online communications
Communication
Health professionals (22) Coordinate care Have computer literacy Citizen involvement in healthcare
Teaching professionals (23) Communication sciences Communication studies
ICT communications protocols Nursing and midwifery professionals (222) Coordinate care Have computer literacy Solve problems in healthcare
Other health professionals (226) Radiofarmaceutica Work analytical y Analytical chemistry
Other teaching professionals (235) Communication Communication disorders Microsoft Visio Nursing professionals
(2221) Coordinate care Have computer literacy Solve problems in healthcare
Dentists (2261) Dental studies Lead the dental team Handle payments in dentistry
Pharmacists (2262) Radiopharmaceuticals Work analytical y Analytical chemistry
Special needs teachers
(2352) Education law Communication disorders Pedagogy studies,” are more commonly observed in level 4 ISCO groups. A possible explanation for this is that specialized skills do only appear at specialized occupations. 6
In recent years the labor market has changed drastically. This is mostly due to increased globalization, a growing working population and disappearing jobs due to digitalization. The COVID-19 pandemic has accelerated this change. This paper aims to explore algorithmic and data-driven methods for exploring and improving the fit between job seekers and vacancies by modeling skills and occupation data in a knowledge graph. Modeling and leveraging relationships between occupations and skills can provide insights for job seekers with existing skill sets.
After constructing our knowledge graph by relying on the existing ISCO and ESCO taxonomies for occupations and skills, we enrich our KG by relying on job posting data.
We explore our final KG using three diferent applications.
First, we study link prediction methods for quantifying the relatedness between skills and occupations in Section 3. We compare and evaluate two diferent link prediction methods, and find that “Node2Vec” performs best. Next to quantifying relatedness between occupations and skills for, e.g., ranking skills for an occupation or using as edge weights in our KG, we explore Node2Vec for identifying skills-to-occupation links that are not present in the original KG.
Next, in Section 4 we explore our KG for finding eficient job transitions. When an individual is searching for a job, knowing which occupations can help the search process. In our next application we explore shortest path finding algorithms for identifying potential careerpath prediction. We use a skills-based Jaccard similarity metric to model distance between occupations. Furthermore, we show examples of job transitions and study properties of our KG by analyzing the distribution of distances between skills and occupations. 1 2 3 4 5
Finally, in Section 5 we study a method to determine which skills are most relevant to diferent levels of aggregated occupations, using the ISCO taxonomy. The skill relevance to an ISCO (group) is calculated by taking the frequency of the skills being required for an ISCO (group) with the uniqueness of the skill in the overall ISCO taxonomy. Here, the uniqueness is high if a skill occurs more often in one group compared to the other groups. The metric that reflects this intuition is called “TF-IDF.” By doing so we construct a birds-eye view of the labor market.
The findings from the three sections described above are all variations to the same theme, of finding or enabling the perfect fit between a job seeker and a vacancy, by leveraging skills. 7
In this paper we present diferent KG-driven applications for
skillsbased job matching. In principle, the methods presented are
dataagnostic, as long as similar concepts (occupations and skills) and
data (job postings with identified occupations and skills) are
available. More specifically, we leverage the ISCO and the ESCO
taxonomies, which are available in a large number of languages, and
are considered standards that are freely available. Other frameworks
could be used as well, where ESCO is widely used in Europe, the
O*NET framework [
The outcome of any research is heavily dependent on the
available data. In the case of this research this data is preprocessed
in a number of steps, one of which is the skill matching step
described in Section 2.2.3. We opted for a naive character -gram
based method for matching surface forms found in a job posting
with skill names in ESCO. Obviously, more refined methods can be
employed, e.g., by considering additional representation of the skill
in the job posting (contextual words, occupation), and at the same
time additional context at the side of the KG (e.g., skill descriptions,
associated occupations, etc.). In general, this problem of matching
can be considered an entity linking task, which is considered out of
scope for this application paper. Having a flawed knowledge graph
as a result of sub-optimal prepossessing does not invalidate the
methods used. Whichever approach is used to create a knowledge
graph, the outcome will never be perfect [
Finally, two out of three applications of our KG are not validated
empirically: for both our shortest path finder (Section 4) and
identifying the most relevant skill per ISCO group (Section 5, we focused
on the analysis and interpretation of results, omitting a more formal
evaluation methodology. For future research it would be interesting
to benchmark the current against diferent career path prediction
models. Validating if, e.g., the discovered paths between
occupations indeed are practically the shortest one, requires additional
data. Unfortunately, no such data was available at the time of
writing. One place to acquire such data, is, e.g., by collecting data of
historic career paths. However, collecting such data and composing
was determined out of scope for this work. The same arose for the
method for quantifying the relevance of skills per ISCO group; these
aggregated insights were dificult to validate. We could imagine
involving human expert annotators to annotate which skills they
deem (most) relevant to a certain ISCO (group). However, similarly
to the above, collecting and analyzing such data did not fit in the
scope of our present work. In summary, our paper revolves around
studying algorithmic methods that aim to help both jobseekers and
recruiters find a better match between individuals and occupations,
we consider studies with actual end-users out of scope [
A special thanks to the thesis supervisors for the project Niels van Weeren and Prof. Aske Plaat as well as everybody at Randstad involved in this project.