2023 3rd International Conference on Pervasive Computing and Social Networking (ICPCSN)

Resume Classification using Support Vector

Machine
Gopal Kamineni Kandula Akhil Sai Dr. G. Siva Nageswara Rao
Department of CSE Department of CSE Department of CSE
Koneru Lakshmaiah Education Koneru Lakshmaiah Education Koneru Lakshmaiah Education
Foundation Foundation Foundation
2023 3rd International Conference on Pervasive Computing and Social Networking (ICPCSN) | 979-8-3503-2284-2/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICPCSN58827.2023.00021

Vaddeswaram, India Vaddeswaram, India Vaddeswaram, India

gopalkamineni2001@gmail.com 190030731cse@gmail.com sivanags@kluniversity.in

Abstract—Recruitment is one of the company's most important

and crucial tasks. A company needs to select the best person
for the job who has all the skills to fulfil the duties. A resume is
a reflection of a person's abilities. These days, companies
receive an enormous number of applications for a job.
Traditional recruiting methods such as analyzing and class
manually analysing and classifying resumes are becoming
consuming a lot of time. So, there is a need for automation of
the resume analysis and classification phase. The automation
should work in such a way that it will take resumes, job roles,
and job descriptions as input and classify whether the resumes
are suitable for the job or not. This can be done with the help
of text analysis and classification algorithms such as the
Support Vector Machine.

Keywords—Resume classification, Support vector machine,

Text analysis, Cosine similarity, Classification algorithms.

I. INTRODUCTION
In recent years, companies have received an enormous
number of applications from candidates for job openings.
And the HR department needs to pick the best one out of
those applications. It’s very difficult for the hiring
department to look into each resume, analyse them based on
their education, skills, and certifications, and shortlist them
manually, and this makes the hiring process slow down. As
the corporate world runs on tight deadlines, a delay in the
addition of team members during the hiring process could
lead to heavy losses. Furthermore, the candidate who applied
for the job must wait a long time for a response from the Fig 1: Basic Flow of the Model
hiring department. Another problem with manual processes Although many models are proposed for the automation
is that they require a lot of manpower, which can sometimes of the resume screening process, they have some limitations.
lead to human errors. And thereby, the manual analysis and We are planning to build a hybrid model based on those
shortlisting of resumes are highly inefficient. models, removing all those limitations, and getting high
Also, in a recent study, it was reported that in the accuracy.
upcoming years, there will be 75–100 applicants for every There are two main modules in this system; they are: -
job opening. And also, many companies are opening their
jobs to remote workers, which means the applicant need not 1.) Text Analysis: Making the unstructured resumes into
stay in the country of the job opening. And this will further a structured format and breaking the job description
increase the number of applicants. Which makes the manual into vectors for training the classifier.
screening of resumes more inefficient. 2.) Resume Classification: The classifier will use the job
To overcome the above-mentioned limitations, this description as a parameter and classify whether the
research study presents an approach that will automate the resumes are suitable for the job or not.
tasks of resume analysis and classification. This is an
intelligent system that can automate the tasks of resume
analysis and classification. We need the system to take the
resumes as input and classify them according to whether they II. RELATED WORK
are suitable for the job or not. The parameters for the There have been many approaches and techniques
classification will be education, skills, certifications, and proposed for the automation of resume analysis and
experience. which, in return, will save a lot of time in the classification. In this context, some approaches attempt to
hiring process. classify the resumes using cosine similarity as their basis for

979-8-3503-2284-2/23/$31.00 ©2023 IEEE 91

DOI 10.1109/ICPCSN58827.2023.00021
Authorized licensed use limited to: VIT University. Downloaded on February 07,2025 at 17:59:28 UTC from IEEE Xplore. Restrictions apply.
resume classification; another approaches the problem using 7.) Creation of Training and Validation data: splitting
the K-Means algorithm, in which they classify the resumes the data into train and test.
into K-Sets based on the similarities between them; and
And the data is tested onto different classifiers such as
another approach suggests profiles to the job recruiter based
KNN, GB, BAG, SVM, and XGB. And XGB classifier
on the similarities between the job description and LinkedIn
performed best among all.
profile.
Now I am going to list all of the research papers I have As this model is highly data-centric, we can make use of
gone through and mention their limitations of them and how the data analysation methods for building our model.
we used the insights from them to create a hybrid model. Ref. [4] Proposed a model in which job recruiters can
post job openings and applicants can apply for the openings.
Ref. [1] Proposes a system called JRC that helps to
At first, the resumes pass through an NLP for text analysis
recommend jobs to job seekers and is built around the
and get broken down into the form of vectors. Later, the
coupling of two modules, namely the conceptual
similarity between the job description and the CV is scored
classification module and the job posting classification
based on the formulae of statistical concept relatedness, the
module. The conceptual classification module consists of an
same as in [1]. This formula considers a variety of factors,
integrated knowledge base, and the job post classification
including job experience, skills, experience, and cosine
module consists of job postings and required skills. And it
similarity. This model was able to achieve an accuracy of
uses the statistical concept of relatedness. This is one of the
97%.
most unique and solid approaches, and it achieved an
accuracy of 97%. The limitation of this model is that it can’t The limitation of this model is that it is highly
classify the resumes for the job opening that is not in its concentrated on the formulae and doesn’t use any classifiers
database. for classification purposes.
Ref. [2] Proposes a system that contains four main Ref. [5] Proposed a system using KNN and cosine
phases. They are: similarity. Here, the system takes resumes as input and then
classifies them into K-Groups based on their similarity with
1.) Extracting Plain Text: Extracts the contents of the
the job profile. And this uses an open tool called "genism,"
resume and job description using the Apache Tika
which gives a summary of the text. Later, this summary is
server.
used for training the KNN classifier. This model was able to
2.) Pre-processing: cleaning and tokenizing the input place all the resumes into their respective groups. This
files. system had an accuracy of 98%.
3.) Feature Representation: Representing the word Ref. [6] Unlike other models that use Random Forest and
feature vectors using pre-trained glove-100- K-NN, this model used the linear support vector machine
dimensional word vectors. algorithm. Like every other model, this one also deals with
the formatting of the text first. The resumes and job
4.) Classification: It feeds the vectors into CNN for descriptions will be passed through the NTLK toolkit and get
classification, which predicts where the input file features extracted using TF-IDF. It analyses the text and
belongs. gathers insights from it. And the mappings are passed
This model’s limitation is that it is overly dependent on through the classifiers. Linear SVM got an accuracy of
the job description, and JD will not always be available for 0.7853, which is the best among all the other classifiers such
job openings. as logistic regression (0.6240), Naive Bayes (0.4439), and
random forest (0.3899).
Ref. [3] Proposes a system for automatic segment
extraction from resumes. This model is mainly based on data The only limitation of this model is its accuracy, and
extraction and broadly consists of seven phases. They are: although 0.785 is good accuracy in real life, this could lead
to some very big problems.
1.) Data Collection: Collection of resumes
Ref. [7] This system used domain adaptation. Here, a list
2.) Data labelling: involves Assigning labels to the data of classifiers is trained on a large dataset of job description
points so that the information can be used for the snippets. And the structured data for training the classifiers
training of supervised machine learning models. comes from the text classification system, which uses
3.) Data Pre-Processing: Converting raw data into techniques such as TF-IDF, lemmatization, and many more.
usable data with the help of NLP. And these classifiers were later used for classifying the
resumes. Character-level CNN showed better performance
4.) Feature Extraction: Extraction and analysis of compared to all other classifiers.
features from the text to collect meaningful insights
This contains many sub-sections, such as Word Ref. [8] This paper is not directly related to the resume
Count, Length of the Text, and "Average Word classification model, but it gives us valuable insight into
Length. different clustering algorithms, such as how they function
and their time and space complexity. Their shapes and
5.) Feature Analysis: Extracted features from the capabilities Their limitations and many more
previous step are analysed.
6.) Feature Transformation: Normalization of highly III. PROPOSED MODEL
skewed features to make sure that the data is ready. This section presents an overview of the proposed
system’s architecture and discusses its main modules. As

92

Authorized licensed use limited to: VIT University. Downloaded on February 07,2025 at 17:59:28 UTC from IEEE Xplore. Restrictions apply.
shown in Figure 2, the proposed system comprises several
modules that are organized as follows:

Fig 2: System Design of the Proposed Model

There are two basic prerequisites for the model. They are: -
1.) Resumes from the applicants
2.) Job Description.
First of all, the application will perform pre-processing on
the job description. Next, different types of features are
extracted and analyzed to collect meaningful insights from
the job description, such as requirements such as experience,
education, and certification, and the description is
summarized into key points. This is done with the help of
sub-modules such as Word Count, Length of Text, Average
Word Length, Numeric, Stop Words, and Special Fig 3: A Sample Resume
Characters. And this data, along with already existing
corpus data, is used to train the SVM classifier. The SVM
will be trained around a few main parameters, such as
education, skills, work experience, and certifications.
Now, we can move on to the second module, which deals
with resumes. In general, all resumes are unstructured, and
we must convert them to a structured format for further Fig 4: After removing unnecessary data from the resume
classification. This consists of a few steps. They are:
1.) First, we clean up the resumes.
2.) Remove unnecessary items such as hobbies, extra-
circular activities, and a motto; these are not
required for the classification. Fig 5: Data from the resume converted into a data frame
3.) Now, we will be using data frames for storing Now that we have converted the unstructured resume data
structured data. (.pdf) to a structured format (data frame), we are ready for
4.) We will be using 1 data frame for every resume, the classification part.
which consists of the candidate's name, education, In this step, each resume is classified as to whether it is
skills, and work experience. suitable for the job or not with the help of the classifier
5.) And we do the same for all the resumes. trained in the first module. The classifier will use different
All the steps taken together will allow us to perform text parameters from the JD and existing corpus to make the
analysis on the resumes and gather insights from them. classification. And once all resumes are classified, we will
Firstly, the application takes the resume as input, and then have a list of resumes that are suitable for the job
we pass it through the function, which helps us with the first description.
two of the above-mentioned five steps. They are cleaning Unlike the other approaches, which train the model once and
the resumes and removing unnecessary fields. We have use it, this approach will update its corpus for every job
done this with the help of libraries such as PyPDF2, and classification, helping us to get better accuracy every time.
text-extracted string.
IV. DETAILED OVERVIEW OF THE MODEL
The most important task of this model is converting
unstructured data from job descriptions and training the
classifier.
Now, we will see the conversion of job descriptions and
the collection of insights from them.

93

Authorized licensed use limited to: VIT University. Downloaded on February 07,2025 at 17:59:28 UTC from IEEE Xplore. Restrictions apply.
 subskill, it reduces the score of the resume, thereby making it
unsuitable for the job. Let’s look into it with the help of an
example. Consider an applicant who listed "React" on his
resume as a skill. And a person who learned React should be
familiar with other skills such as HTML, CSS, and
Javascript. Let’s look at a JD that has HTML, CSS, and JS as
its required skills. When the classifier is not able to identify
the subskills, it makes the applicant ineligible for the job.
This brings the accuracy of the model down. This is where
corpus comes into play; it checks the react skills in corpus
and gets all the subskills of the skills, which will help us
increase the accuracy of the model.

V. EXPERIMENTAL EVALUATION
This section shows the evaluation of the model. Here we
will compare the score of our model with manual models and
see how it performed compared to others.
Table 1 score comparison between manual and model
Difference
Resume Manual Model’s
Job title (Manual- Result
index score score
Models)
Perfect
R1 0.65 0.65 0.00
match
Under
R2 0.70 0.63 0.07
qualified
Java
Full Perfect
R3 0.44 0.44 0.00
Stack match
Dev Over
R4 0.80 0.95 -0.15
qualified
Fig 5: Flow chart for the Structurisation of Job Description R5 0.50 0.38 0.12
Under
qualified
The data stored in the data frame will be compared with Perfect
R6 0.82 0.82 0.00
the data in the corpus, and they both together will be used for Data match
the training of the SVM model and updating the corpus. analyst Perfect
R7 0.57 0.57 0.00
match
The classification will consist of parameters such as Over
R8 0.63 0.72 -0.09
"education required, "skill required, "work experience" (if qualified
mandatory), and "projects. As shown in Table 1, we have taken two job postings and
Now, we will see the conversion of unstructured data eight resumes. The first job post is for a Java Full Stack
from resumes to a structured format. This is much more Developer, and it has the following requirements: 2 years of
complex when compared to the job description. Because experience in web development using Java, front-end
resumes will be in different formats, colours, and types. And languages such as HTML, CSS, and JavaScript, frameworks
the model needs to convert them correctly. If not, it will lead such as React, Django, A strong grip over databases such as
to a lot of errors, resulting in a decrease in accuracy. MySql and MongoDB and certification in Red Hat will be
helpful. And Job Post 1 has four resumes with it, namely R1,
Now, each resume will be classified with the help of the R2, R3, and R4. As we can see from the resumes R1 and R3,
support vector machine. The SVM will classify the resumes the difference is 0, so it’s a perfect match. In resume R2, the
into two categories: - difference is 0.07, which is because the manual score is
higher than the model score, so it's underqualified. For
1.) Eligible for the job.
resume R4, the difference is -0.15 because the model score is
2.) Not eligible for the job. higher than the manual score, so it's overqualified.
The eligibility of a candidate for the job will be based on Now we look at Job Post 2, which is Data Analyst. Its
whether the skills mentioned in the resume match the job requirements are 1 year of experience, a strong grasp of
description and the corpus; if any mandatory work languages such as Python, and tools such as Tableau, Weka,
experience is mentioned in the job description, the classifier and PowerBI. Certification in Hadoop will be an added
will also check it in the resume and classify them advantage. And we have taken four resumes, namely R5, R6,
accordingly. The model adds additional scores to the R7, and R8. As we can see for the resumes R6 and R7, the
resumes if they consist of any certificates or certifications difference is 0 because the manual score and model score is
that match the job description. equal, and thereby it’s a perfect match. For resume R5, the
difference is 0.12 because the manual score is greater than
Need for Corpus: The corpus is what makes the the model score and it is underqualified. For me 8, the
difference between this model and all other existing models. difference is -0.09 because the model score is higher than the
This helps us increase the accuracy of the system. Most of manual score, so it’s overqualified.
the time, the applicant will list a high-level skill that consists
of many subskills, and when the job description mentions a
subskill as a requirement and the classifier can’t find that

94

Authorized licensed use limited to: VIT University. Downloaded on February 07,2025 at 17:59:28 UTC from IEEE Xplore. Restrictions apply.
 Table 2 Comparison with other approaches communication skills, and a Certificate in Human Resources
Model Management from York University or SHRM certification,
Resume Manual Cosine
Job title
index score
KNN
Similarity
Score which will be an added advantage. And it has three resumes,
score
namely R4, R5, and R6. As shown in Table 2, the manual
Marketing R1 0.56 0.20 0.37 0.51 scores and model scores are almost the same. And KNN
Manager
R2 0.89 0.56 0.60 0.78 performed the second best, and cosine similarity performed
the least.
R3 0.40 0.20 0.30 0.45
Human R4 0.61 0.55 0.50 0.65 The third job posting is Software Engineer, and its
Resources R5 0.46 0.35 0.20 0.46 requirements are B. Tech. or B.S. A degree in computer
Operator R6 0.53 0.41 0.35 0.54
Software R7 0.35 0.20 0.20 0.35
science, a good understanding of C/C++ and Java, Python
Engineer R8 0.70 0.61 0.70 0.75 full-stack development, and AWS certification will be an
R9 0.20 0.20 0.25 0.25
added advantage. It has three resumes, namely R7, R8, and
Now we compare the results of the model with the manual R9. Similar to the results of the previous 2 job posts, the
score and other models such as the KNN score and cosine scores of our model and manual are almost equal. This time,
similarity. cosine similarity came in at the second and KNN came in at
the third.
KNN stands for k-nearest neighbours algorithm; it is a And the results of the 3 job postings show that the model is
supervised learning technique used to make classifications. not only working well but also performing better compared
Using this algorithm, we can categorize objects into K
to the other available models and algorithms.
different groups. For the resume classification application,
The reason why our model performed better than the cosine
we will be choosing K as 2. So, there are two categories: one
is suitable for the job, and another is not suitable for the job. similarity is that the cosine similarity just looks at the
And we will pass the resumes into the model so the similarity between the job description and the resumes and
algorithm can determine whether they are suitable for the job gives out a score; there is much more to look at than that.
description or not. For instance, a candidate might have all the skills and
certifications mentioned in the job description but not satisfy
the experience requirement. As most of the matter is similar,
even though the requirement is not satisfied, he still gets a
score.
The reason why our model performed better than the KNN
algorithm is that the KNN algorithm doesn’t use its previous
This is the Euclidean distance formula used in the KNN classification insights to classify the resumes. For this
algorithm; it helps us determine whether the resume is reason, every classification is a new classification. Because
suitable for the job or not. of this, the KNN algorithm is also called a lazy learning
Cosine Similarity Cosine similarity is used to measure algorithm. It uses the same old data for all the resumes
similarities between two documents or any two sequences in because it is one-time trained data for all the classifications.
general. This is one of the most important concepts in data But our model overcame this with the help of corpus. Which
science. The application takes two files as input; the files helps the model have high accuracy.
might be of any format (.docx,.txt,.xlsx,.csv). And analyzes
the number of similarities between them and gives us a score. VI. RESULTS AND DISCUSSION
We have passed a list of resumes into the application along
with a job description. And they ate the results of it.

This is the mathematical formula for cosine similarity.

Using this, we can analyze the files and get their similarity.
We have taken three job postings and nine resumes and
compared the scores of our model with those of other
available models. The first job post is marketing manager,
and its requirements are 2 years of experience in
management, an MBA degree, an 8.0+ CGPA throughout
the education process, and good communication skills. And
it has three resumes, namely R1, R2, and R3. As we can see,
the manually generated-score and the model-generated score
are almost equal. And the other two models have a high
difference, with cosine similarity performing the second best
and KNN performing the last.
The second job post is that of a human resource operator. Its
requirements are 3+ years of experience in the hiring
department of a multinational company, great
Fig 6: Output from the Application

95

Authorized licensed use limited to: VIT University. Downloaded on February 07,2025 at 17:59:28 UTC from IEEE Xplore. Restrictions apply.
These are the people who have the highest model score model on a day-to-day basis, thereby ensuring that the
between their job description and their resume, i.e., people model's accuracy won’t decrease at any point in the future.
who have skills that are very close to their job description.
Now let’s investigate the accuracy, precision, and other REFERENCES
metrics. [1] A. Zaroor, M. Maree and M. Sabha, "JRC: A Job Post and Resume
These are the accuracy scores in detail, which further prove Classification System for Online Recruitment," 2017 IEEE 29th
that our model is a success and can work more accurately International Conference on Tools with Artificial Intelligence
(ICTAI), Boston, MA, USA, 2017, pp. 780-787, doi:
than any manual model. 10.1109/ICTAI.2017.00123.
[2] S. Nasser, C. Sreejith and M. Irshad, "Convolutional Neural Network
with Word Embedding Based Approach for Resume Classification,"
2018 International Conference on Emerging Trends and Innovations
In Engineering And Technological Research (ICETIETR),
Ernakulam, India, 2018, pp. 1-6, doi:
10.1109/ICETIETR.2018.8529097.
[3] B. Gunaseelan, S. Mandal and V. Rajagopalan, "Automatic Extraction
of Segments from Resumes using Machine Learning," 2020 IEEE
17th India Council International Conference (INDICON), New Delhi,
India, 2020, pp. 1-6, doi: 10.1109/INDICON49873.2020.9342596.
[4] Avisha Anand, Sandeep Dubey "CV Analysis Using Machine
Learning", Ijraset Journal For Research in Applied Science and
Engineering Technology (IJRASET), 2022, ISSN: 2321-9653, doi:
https://doi.org/10.22214/ijraset.2022.42295
[5] Rajath V, Riza Tanaz Fareed, Sharadadevi Kaganurmath, 2021,
Resume Classification and Ranking using KNN and Cosine
Similarity, INTERNATIONAL JOURNAL OF ENGINEERING
RESEARCH & TECHNOLOGY (IJERT) Volume 10, Issue 08
(August 2021)
[6] Pradeep Kumar Roy, Sarabjeet Singh Chowdhary, Rocky Bhatia, A
Machine Learning Approach for automation of Resume
Recommendation system, Procedia Computer Science, Volume 167,
2020, Pages 2318-2327, ISSN 1877-0509,
https://doi.org/10.1016/j.procs.2020.03.284.
[7] S. Ramraj, V. Sivakumar and K. Ramnath G., "Real-Time Resume
Classification System Using LinkedIn Profile Descriptions," 2020
International Conference on Computational Intelligence for Smart
Power System and Sustainable Energy (CISPSSE), Keonjhar, India,
2020, pp. 1-4, doi: 10.1109/CISPSSE49931.2020.9212209.
[8] K. Bindra and A. Mishra, "A detailed study of clustering algorithms,"
2017 6th International Conference on Reliability, Infocom
Technologies and Optimization (Trends and Future Directions)
(ICRITO), Noida, India, 2017, pp. 371-376, doi:
10.1109/ICRITO.2017.8342454.
[9] Z. Li, F. Kang, P. Yu and H. Shu, "A Review of Field Text Analysis,"
Fig 7: Accuracy precision and other metrics 2021 IEEE Conference on Telecommunications, Optics and
The accuracy of the model came out to be 99%. We have Computer Science (TOCS), Shenyang, China, 2021, pp. 52-58, doi:
also used the application on our peers’ resumes and 10.1109/TOCS53301.2021.9688944.
calculated their chances of getting a job manually. And the [10] Pal, Riya & Shaikh, Shahrukh & Satpute, Swaraj & Bhagwat,
Sumedha. (2022). Resume Classification using various Machine
results came out to be accurate and matched those of the Learning Algorithms. ITM Web of Conferences. 44. 03011.
manual screening. The application performed the job on the 10.1051/itmconf/20224403011.
10th attempt at manual screening. [11] Gopalakrishna, Suhas & Varadharajan, Vijayaraghavan. (2019).
Automated Tool for Resume Classification Using Semantic Analysis.
VII. CONCLUSION International Journal of Artificial Intelligence & Applications. 10. 11-
23. 10.5121/ijaia.2019.10102.
In conclusion, resume classification using support vector [12] ALI, Irfan et al. Resume Classification System using Natural
machines has been demonstrated to be an effective and Language Processing and Machine Learning Techniques. Mehran
efficient method for automating the resume screening University Research Journal of Engineering and Technology, [S.l.], v.
process. By training a model on a dataset of labelled 41, n. 1, p. 65 - 79, Jan. 2022. ISSN 2413-7219. doi:
resumes, the model can learn to accurately classify new http://dx.doi.org/10.22581/muet1982.2201.07
resumes as suitable for the job or not using fields such as [13] Takkar, Sakshi & Arora, Mohit & Chopra, Shivali. (2019). A Deep
Insight of Automatic Resume Classifiers For Skill Mapping By
education level, work experience, skills, and certifications. Recruiters. 1247.
This can save significant time and resources for HR
[14] S. Bharadwaj, R. Varun, P. S. Aditya, M. Nikhil and G. C. Babu,
departments and recruiters by reducing the amount of time "Resume Screening using NLP and LSTM," 2022 International
spent manually reviewing resumes. Conference on Inventive Computation Technologies (ICICT), Nepal,
2022, pp. 238-241, doi: 10.1109/ICICT54344.2022.9850889.
Most of the present-day resume classifiers are only one- [15] G. Sudha, S. K. K, S. J. S, N. D, S. S and K. T. G, "Personality
time trained, and thereby their accuracy will go down at Prediction Through CV Analysis using Machine Learning Algorithms
some future point because of the updated data. But our for Automated E-Recruitment Process," 2021 4th International
application will overcome this. We have implemented a Conference on Computing and Communications Technologies
corpus that will update the available data with every (ICCCT), Chennai, India, 2021, pp. 617-622, doi:
10.1109/ICCCT53315.2021.971178
classification that the application performs and train the

96

Authorized licensed use limited to: VIT University. Downloaded on February 07,2025 at 17:59:28 UTC from IEEE Xplore. Restrictions apply.