JOB MATCHING USING ARTIFICIAL INTELLIGENCE

International Journal on Soft Computing, Artificial Intelligence and Applications (IJSCAI),
Vol.11, No.3, August 2022
DOI : 10.5121/ijscai.2021.11301 1
JOB MATCHING USING ARTIFICIAL INTELLIGENCE
Sukit Kitichalermkiat1
, Vishakha Singh2
, Suntaree Chaowiang2
, Sumet
Tangprasert2
, Pornthep Chiraprawattrakun2
and Nopbhorn Leeprechanon1
1
Department of Electrical and Computer Engineering, Thammasat University, Thailand
2
Depth First Co. Ltd., Bangkok, Thailand
ABSTRACT
The job market has expanded exponentially in the past few years. With many recruiters and candidates, it
is not an easy task to match a perfect candidate with a perfect job. The recruiter targets candidates with
the required skill sets mentioned in the job descriptions, while candidates target their dream jobs. The
search frictions and skills mismatch are persistent problems. In this paper, we build a model that would
match companies with candidates with the right skills and workers with the right company. We have
further developed an algorithm to investigate people’s hiring history for better results.
KEYWORDS
Job matching, Recruiter, TSIC, TSCO, Machine Learning
1. INTRODUCTION
Classifications are not effortless, but the complexity of document/text classification increases
two-fold. Generally, the classification approach of text/image/video consists of defining the
training data into different classes or categories for training and then using the trained model for
classifying the new unseen data into their specified category/class. The data has also increased in
the past decade, making it even harder to classify by a human. The job sector has seen
tremendous growth in the past few years, and as we see currently with covid 19, the job postings
and job searches are only done through the internet.
When it comes to job title, it can sometimes be very vague. For example, a computer engineer
does the same job as a web developer. Python programmer does the same job as a python
developer, which is quite similar to a python software engineer, but one may be faster and
stronger than the other. Therefore, when a recruiter uploads a job posting, it would want to see all
the available candidates that possess the skills that are needed for the role. The same applies to
the candidates as well. Also, many job titles are written differently but hold the same
requirements. It may make the perfect candidate miss that because of the misunderstanding of the
job title. Therefore, it is of utmost importance that we have a system that can classify these titles
based on their requirements which can benefit the company and the candidate in the long run. It
would also be a fairer platform where the candidates will be shortlisted purely based on their skill
set.
In Thailand, jobs/occupations are divided into various categories based on the job's skills and
requirements. They are regarded through the Thailand Standard Classification of Occupations
(TSCO) and the Thailand Standard Industrial Classification (TSIC). In this paper, we propose a
job matching classification system that would use the job description alongside TSCO and TSIC
to classify the incoming jobs into their specific category. To do the classification job, we used the

2
Stochastic Gradient Descent, a supervised learning algorithm in our research. Furthermore, we
have developed our algorithm, Hiring History, which will review a person's previous hiring
details.
2. LITERATURE
2.1. Review
Rodrigues and Chavez [1] in their paper described job matching to be a way to control exactly
how a job register and a job candidate are correctly paired together. Their paper extracted the
attributes from the resumes and then used WEKA for data mining, they then adopted clustering
algorithm to match the profile of the job seekers against the job requirements given by the
respective companies.
Almalis Et al. [2], proposed a content-based recommendation algorithm called FoDRA (Four
Dimensions Recommendation Algorithm) where they use a structured form of the job and
candidate’s profile that were produced from the job description and CVs to assess the
appropriateness of a job seeker for a particular job position.
Harris [3], in his paper evaluated three approaches to find best candidates to match a set of job
skills. He used crowdworkers in a gamified environment, information retrieval-based search
methods and a text-mining approach that used feature and elements from the IR-based search
engine. He found that the crowdsourcing environment provided the best results for the technical
job postings and the crowd and text-mining both performed equally well for the non-technical job
postings.
Chalidabhongse, Jirapokakul and Chutivisarn [4] proposed a decision support system called Job
Application Support System to facilitate the recruitment process where they focused on the part
where the applicants have to fill out application forms and the screening process.
Mishra, Rodrigues and Portillo [5] in their paper “An AI Based Talent Acquisition and
Benchmarking for Job” proposed a methodology to solve problem of selecting best CV from a
pool of CVs by matching the skill graph generated from CV and Job Post. Their approach is to
understand the business aspect to explain why these kinds of problem generate and how one can
solve it using natural language processing and machine learning techniques.
Koh and Chew [6] in their paper proposed an intelligent job matching with self-learning
recommendation engine for the self-operation of resume matching/ranking. Their parameters
include domain of job, job title, position, knowledge, experience, location, salary and other. Their
engine is going to extract the data from ontology to ensure the data stability.
Lee, Kim and Na [7], in “Artificial Intelligence based Career Matching” developed a method for
career matching amidst university students and companies by the name of Artificial Intelligence
based Design platform (AID). They analysed the results from the model with statistical methods
like least squares, Pearson correlation, Manhattan distance. In their experimentation they found
that their model/methods gave them zero miss-matching between student’s skills and company’s
need on the other hand statistical method gave 30% miss-matching.

3
2.2. Natural Language Processing
We as a human species mainly communicate with each other via text or speech. We see texts
wherever we go from road signs, news outlets, emails, messages, to menus and instructions, that
is naturally how we communicate around the world. Natural language processing or NLP can be
defined as a branch of Artificial intelligence that gives computers the ability to understand the
text and speech like the humans do. In other words, it is used to describe the way of a software’s
automatic manipulation of these natural languages (speech or text) [8] [9].
In the recent times. NLP has grown popular and can be seen everywhere, from the translation
software, GPS systems to chat bots. It’s still emerging area and is one of the most complicated
areas to tackle in the AI world, as the data has increased tremendously and the fact that human
language is filled with ambiguities. But today’s technology and high computational machines
analyses more data (language based) that humans in an unbiased way and they work tirelessly
[10].
NLP can further be divided into different tasks ranging from speech recognition, word sense
disambiguation named, entity recognition, sentiment analysis to natural language generation. In
this paper, we are attempting to solve the challenges faced by the job sector by applying AI to it.
2.3. Stochastic Gradient Descent (SGD)
Gradient Descent is one of the most popular optimization techniques in Machine learning and
Deep Learning and can be used with most of the learning algorithms. The gradient or slope is the
rate of change of a function and is typically used for functions that has several inputs but a single
output.
For our work, we have chosen Stochastic Gradient Descent (SGD), as it is one of the most
efficient approaches to fit linear classifiers and regressors under covex loss functions such as
linear Support Vector Machines and Logistic Regression. The word ‘stochastic’ refers to a
system that is related with random probability, that is, this algorithm selects samples at random
and not the whole data for training for each iteration [11]. It has gotten more famous in the recent
times due to the large-scale learning and is often used for machine learning problems that are
faced in text classification and natural language processing (NLP). SGD is easy to implement and
is efficient but can be sensitive to feature scaling.
2.4. TSCO
TSCO stands for Thailand Standard Classification of Occupations (TSCO-2001), it is inspired
from ISCO (International Standard Classification of Occupations) where according to the skill
level occupations (Jobs) are divided into various categories ranging from low to high and armed
forces [12]. They are further divided into Major groups where there are total of 9 categories, from
Managers, Technicians and Associate professionals to Elementary Occupations and Armed
Forces Occupations. Influenced by ISCO, Thailand in 2001 made their own Standard
Classification and termed it TSCO, which is what we have used in our paper.
2.5. TSIC
Similar to TSCO, there is another set of industrial classification known as TSIC (Thailand
Standard Industrial Classification). Inspired from ISIC (International Standard Industrial
Classification of all Economic Activities) which is a United Nations system for classifying

4
different economic data. Various kinds of economic activity in the fields of production,
employment, gross domestic product and other statistical areas are classified here [13]. TSIC
consists of 21 sections, 88 divisions, 243 groups, 440 class and 1089 industries in total. Ranging
from agriculture, forestry, construction to manufacturing, mining etc.
In our paper we have made use of TSCO and TSIC for the classification of different
occupations/jobs.
3. METHODOLOGY
The goal of our work is to get a classifier that receives a job position which would be in a form of
a sentence, written in natural language (English and Thai) and the return the TSCO code for that
job position.
Our work is carried out using the Python language and the implementation of the trained
classifiers is done in Scikit Learn library, and classification is done using SGD classifiers.
3.1. Data Processing
The classifier we have used in our work belongs to supervised learning, and they are trained on
the data that we have collected from Department of Employment/Labour and are manually
classified. The input data in our work is job descriptions and job titles with their respective
TSCO. The aim is for the model to learn the TSCO for different jobs and when model is fed with
unseen data, they map the TSCO to their categories.
The input data looks like the following data shown in the table 1, which is provided to the model
in a tab separated file:

5
Table 1. Input dataset for the model
Position Classification
ผู้ปฏิบัติงานอาชีพในกลุ่มนี้รวมถึงนักบริหารของหน่วยงานรัฐบาลซึ่งมิได้จัดป
ระเภทไว้ในที่อื่นกลุ่มอาชีพในหมู่นี้มีดังนี้
1120
ควบคุมการรับ–
จ่ายเงินการบัญชีและงบดุลของพรรคตามที่กฎหมายหรือตามที่คณะกรรมกา
รบริหารพรรคกาหนด
1141
ปฏิบัติงานหลักมูลฐานเช่นเดียวกันกับกานันแต่ปกครองบรรดาราษฎรที่อยู่ใน
เขตหมู่บ้านและดาเนินการในเรื่องต่างๆตามที่กานันมอบหมาย
1130
ผู้ปฏิบัติงานอาชีพในกลุ่มนี้รวมถึงผู้จัดการด้านภัตตาคารและโรงแรมซึ่งมิได้
จัดประเภทไว้ในที่อื่นอาชีพในหน่วยนี้มีดังนี้
1225
ผู้ปฏิบัติงานอาชีพในกลุ่มนี้รวมถึงผู้จัดการด้านการขนส่งการสื่อสารและคมน
าคมซึ่งมิได้จัดประเภทไว้ในที่อื่น
1226
หรือสถานประกอบการอื่นๆ 1233
ใจว่าได้รับมอบตามกาหนดเวลา 1235
ลอบตัดต้นไม้ 1311
ผู้ปฏิบัติงานอาชีพในกลุ่มนี้รวมถึงผู้ประกอบวิชาชีพด้านคอมพิวเตอร ์ซึ่งมิได้
จัดประเภทไว้ในที่อื่นกลุ่มอาชีพในหมู่นี้มีดังนี้
2139
The first column refers to the <job description> and the second column defines the respective
<job position> (TSCO code). Since the classifier is a supervised learning algorithm, we are
training them using manually classified data.
We then generated dataframe for the given input file that has the job description and the TSCO
code. After which splitting is done for training (train(x)) and testing (test(y)). Normalization of
the dataframe is done according to our defined criteria, the results from the above process gives
us the X_train, X_test and Y_train, Y_test sets. So, this step splits our input data into training
(80%) and testing set (20%).
3.2. Training and Tuning our Model
SGD classifiers consists of parameters and hyper-parameters. Parameters are the values that
corresponds to the mathematical model, that are adjusted after the training while, hyper-
parameters are values that are relate to the way of training and are adjusted using the selected part
of the training set. Learning rate, momentum, decay and nesterov are the hyperparameters that
can be optimized in SGD.
To train and adjust the parameters is feasible to use a fit function, in our work we have made use
of tools provided by Scikit learn known as Pipeline and GridSearchCV with the aim to make
extensive search to achieve the hyperparameters needed to optimize the results [14]. Below table
1 and 2 shows the configuration of SGD classifier and GridSearchCV used in our paper.

6
Table 2. Parameters of SGD Classifier
Parameters Definition Value
Leaning rate To control the way an algorithm learns and to help
the performance through tuning
0.01
loss To evaluating how well the algorithm is modelling
the given data
‘hinge
’
fit_intercept Whether the intercept should be estimated or not. True
max_iter The max number of passes over the training data. 1000
l1_ratio It is the Elastic Net Mixing parameter. with 0 <=
l1_ratio <= 1. l1_ratio=0 corresponds to L2
penalty, l1_ratio=1 to L1.
0.15
Penalty It is the total of absolute values of weights. ‘l2’
Shuffle To determine if the training data should be shuffled
or not.
True
power_t To inverse scale the learning rate. 0.5
validation_fracti
on
The amount of data set aside for validation. 0.1
class_weight Weights that are associated with the given classes. ‘balan
ced’
Tol The stopping criterion. 1e-3
epsilon It is used when there is a selection of specific
action based on the Q values we already have.
0.1
Table 3. Parameters of GridSearchCV
Parameters Definition Value
n_jobs To determine the number of jobs that run
in parallel
2
cv To determine the cross-validation splitting
strategy
5
iid To assume if the data is identically
distributed across the folds or not.
false
3.3. Our Algorithm (Hiring History Algorithm)
We have developed a model that after the job matching model is done will look into a person’s
hiring history for better prospect. This algorithm improves the matching score and can be used
for further personalization and the idea is adopted from Discounted Cumulative Gain. Our
algorithm can be expressed as below:

7
Rating (R)
Eq. 1
R = Rating
i = the ith
job in the candidate’s work history
t = Time (how much time the candidate worked)
N = total number of jobs in the candidate’s work history
Fig 1. Job matching API to find job
Fig 2. Job matching API to find candidate
Figure 1 and 2 shows the way the information is carried out in our model. The algorithm is tested
on the data gathered from the Department of Employment. The model is already in use
throughout Thailand, the steps and working of the algorithm are as follows:
1. Firstly, the algorithm will look into the candidate’s current and previous job profile, i.e.,
where they are currently working, where in the past they have worked and what kind of job
positions were they holding at that time. Fig 3 shows layout of it.

8
Fig 3. Job history of a candidate A
2. After that, we will calculate the rating for each job to evaluate which job industry is better
suited for that particular candidate.
3. Higher rating will be given to the current job position and the score will keep on getting
lower with the past jobs. Fig 4 predicts the rating for respective job position.
Fig 4. Rating for the job best suitable for candidate A.
4. Finally, a smart job center (SJC) will filter and sort the jobs in its database according to the
suggested TSCO and TSIC.
5. With this, our algorithm is able to determine which sector of occupation is best for the
given candidate.
4. RESULTS AND CONFUSION MATRIX
As we are aware, the NLP is of the most complex of AI problems. In this paper, we attempted to
tackle the job matching problem with Stochastic gradient descent and we received the overall
accuracy of 99.89%. Our result also gives macro average and weighted average, as macro
average gives each prediction equivalent weight when it is calculating but it may sometime occur
that our data might be imbalanced and we would like to give importance and value to some
prediction more, there we need weighted average. So, our results compromises of all the given
accuracies.
Fig 5. Accuracies of the job matching model

9
Table 4. Testing result of the model
Position Classification
เจ้าหน้าที่ดูแลเอกสารด้านกฎหมาย M72
นักศึกษาฝึกงาน E-Commerce M742
เจ้าหน้าที่ Online Marketing S9610
Supervisor Laboratory ประจาโรงพยาบาลศรีระยอง M72
หัวหน้าแผนก QA G471
Creative Content & Copy Writer M742
เจ้าหน้าที่แมสเซนเจอร ์ M72
The prediction results from our model on an unseen data looks like above data represented in the
table 4. The left side represents the job title and the right side depicts the TSCO. We provided the
model with only job titles and got their respective TSCO in return.
Figure 6 shows the confusion matrix where various job positions were classifies into their
respective categories. A part of our confusion matrix for job matching model can be seen in fig 6
below. The horizontal line is our True Class whereas the vertical line is our Predicted Class, and
it is done in order to visualize our classification algorithm’s performance. As we can see from
figure below, our algorithm is able to predict the true class with good accuracy,
Fig 6. Confusion matrix
4.1. Hiring History Algorithm
The algorithm, test on the data from Department of Employment show how well it will work for
the future usage. Hiring History algorithm gives the accuracy of 73.81% and precision of
77.02%. The TSCO code is the result that we receive from our above job matching model, and
the TSIC code is the result from our hiring history algorithm. In Figure 7, it shows that when a
person looking for a job in Human Resources field, the algorithm will return all jobs with TSCO
1232, which is Personnel and Industrial Relations field, and TSIC 46599 (Machine & Computer
trading), 47412 (Video Games and Software) and 47112 (Discount Store and Hyper market)
respectively according to the person’s working history.

10
Fig 7. Testing result of Hiring History algo
Similarly, confusion matrix for the above test data is given in fig 8 where, the predicted class and
true class matrix is defined. The red diagonal numbers refer to the number of correct predictions
whereas the yellow highlighted number explains which of true class were predicted wrong.
Fig 8. Confusion matrix of Hiring History

11
5. CONCLUSIONS
In this paper, we have used a job matching model based on stochastic gradient descent to solve
the problem occurring in the job sector. We have further developed an algorithm to further
enhance our model. The algorithm takes a review at the people’s previous hiring setting, and
provides the best suitable candidate for a particular job profile. The model is currently being used
internally by the Department of Employment, Thailand.
6. LIMITATIONS AND FUTURE WORK
First limitations we faced was due to the nature of dataset necessary for this experiment. For
training and testing we needed personal information of people in regard to their work which
makes it difficult to get vast dataset. Along with that we can never be sure of the information
received from candidates, if the satisfaction level stated by them is true or otherwise. Another
limitation is that sometimes our model will suggest a former job that might not be available
anymore.
Overall limitation of our work would only be the dataset. Our model can be used in other
countries if the department that is using this have a right to access their people’s personal data.
Future work of our model will depend on the suggestions we will receive by the Department of
Employment, Thailand. We would like to add more attributes that is responsible in a hiring
process. We will try to gather the data from different companies and will look into their hiring
process, their requirements and other internal managements and will add those to our existing
model.
ACKNOWLEDGEMENTS
Financial support granted by Depth First Co, Ltd. (Thailand) is greatly appreciated and
acknowledged.
REFERENCES
[1] Rodriguez, L. G., & Chavez, E. P. (2019). Feature selection for job matching application using profile
matching model. 2019 IEEE 4th International Conference on Computer and Communication Systems
(ICCCS). https://doi.org/10.1109/ccoms.2019.8821682
[2] Almalis, N. D., Tsihrintzis, G. A., Karagiannis, N., & Strati, A. D. (2015). FoDRA — A new content-
based job recommendation algorithm for job seeking and recruiting. 2015 6th International
Conference on Information, Intelligence, Systems and Applications (IISA).
https://doi.org/10.1109/iisa.2015.7388018
[3] Harris, C. G. (2017). Finding the best job applicants for a job posting: A comparison of human
resources search strategies. 2017 IEEE International Conference on Data Mining Workshops
(ICDMW). https://doi.org/10.1109/icdmw.2017.31
[4] Chalidabhongse, J., Jirapokakul, N., & Chutivisarn, R. (2006). Facilitating job recruitment process
through job application support system. 2006 IEEE International Conference on Management of
Innovation and Technology. https://doi.org/10.1109/icmit.2006.262244
[5] Mishra, R., Rodriguez, R., & Potillo, V. (2020). An AI Based Talent Acquisition and Benchmarking
for Job.
[6] Koh, M. F., & Chew, Y. C. (2015). Intelligent job matching with self-learning recommendation
engine. Procedia Manufacturing, 3, 1959-1965. https://doi.org/10.1016/j.promfg.2015.07.241
[7] Lee, D., Kim, M., & Na, I. (2018). Artificial Intelligence based career matching. Journal of
Intelligent & Fuzzy Systems, 35(6), 6061-6070. https://doi.org/10.3233/jifs-169846
[8] What is natural language processing? (2019, August 7). Machine Learning Mastery.
https://machinelearningmastery.com/natural-language-processing/

12
[9] What is natural language processing? (2020, July 2). IBM - United States.
https://www.ibm.com/cloud/learn/natural-language-processing
[10] What is natural language processing? (n.d.). https://www.sas.com/en_us/insights/analytics/what-is-
natural-language-processing-nlp.html
[11] ML | Stochastic gradient descent (SGD). (2021, September 13). GeeksforGeeks.
https://www.geeksforgeeks.org/ml-stochastic-gradient-descent-sgd/
[12] International standard classification of occupations (ISCO). (n.d.). ILOSTAT.
https://ilostat.ilo.org/resources/concepts-and-definitions/classification-occupation/
[13] Thailand Standard Industrial Classification: TSIC-2009. Employment Promotion Division
Department of Employment Ministry of Labour MAY 22, 2014.
[14] Scikit-learn. Retrieved December 7, 2021, from https://scikit-
learn.org/stable/modules/classes.html#module-sklearn.linear_model

JOB MATCHING USING ARTIFICIAL INTELLIGENCE

More Related Content

Similar to JOB MATCHING USING ARTIFICIAL INTELLIGENCE (20)

Recently uploaded (20)

JOB MATCHING USING ARTIFICIAL INTELLIGENCE