3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254 – 6529 Ed. 40 Vol. 11 N.º 1 Marzo - Junio 2022
251 https://doi.org/10.17993/3ctic.2022.111
Gasim Othman Alandjani
Computer Science and Engineering Department. Yanbu University College, (Saudi Arabia).
E-mail: alandjanig@rcyci.edu.sa ORCID: https://orcid.org/0000-0003-0321-7013
Recepción: 21/10/2021 Aceptación: 20/12/2021 Publicación: 29/03/2022
Citación sugerida:
Alandjani, G. (2022). Online fake job advertisement recognition and classication using machine learning. 3C TIC.
Cuadernos de desarrollo aplicados a las TIC, 11(1), 251-267. https://doi.org/10.17993/3ctic.2022.111.251-267
3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254 – 6529 Ed. 40 Vol. 11 N.º 1 Marzo - Junio 2022
252 https://doi.org/10.17993/3ctic.2022.111.251-267
Machine learning algorithms handle numerous forms of data in real-world intelligent systems. With the
advancement in technology and rigorous use of social media platforms, many job seekers and recruiters
are actively working online. However, due to data and privacy breaches, one can become the target
of perilous activates. The agencies and fraudsters entice the job seekers by using numerous methods,
sources coming from virtual job-supplying websites. We aim to reduce the quantity of such fake and
fraudulent attempts by providing predictions using Machine Learning. In our proposed approach,
multiple classication models are used for better detection. This paper also presents dierent classiers’
performance and compares results to enhance the results through various techniques for realistic results.
Machine learning, Supervised algorithms, Fake jobs detection, Classication.
3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254 – 6529 Ed. 40 Vol. 11 N.º 1 Marzo - Junio 2022
253 https://doi.org/10.17993/3ctic.2022.111.251-267
Every organization nowadays is the internet and social media dependent. Systems like enterprise
applications, management information systems, Information systems for Human Resources, and oce
automation applications are pivotal for running work. Creating an eective workforce recruitment
process is considered by employing online applications, as it is more convenient for applicants. The
majority of the human asset specialists and associations empower the online application framework
for the enlistment and choice cycle. It has many benets. Candidates can apply without the time and
transfer their educational program vitae for additional references. Managers additionally can channel
the applications rapidly and make waitlists within a brief period.
In this way, electronic enrollment makes human resource capacities fast. It gives an ideal chance to
online scammers to exploit their distress on these needy occasions when thousands and millions of
individuals seek jobs. Over time, there is an expansion in these fake job posts where ads appear to be very
ordinary, frequently these organizations will likewise have a site and will have an enlistment interaction
like dierent rms in the area (Ward, Gbadebo, & Baruah, 2015).
Online Recruitment Fraud (ORF) is becoming a severe issue in recent times. Due to hype in social media,
online job advertisements are growing rapidly, but with advantages, there are many scammers, fraud
employers scam them for money or taking personal information. Deceitful jobs ads can be posted using
a well-known organization for disregarding their validity (Ward et al., 2015). Detecting fake job posts has
taken consideration for acquiring an automatic tool, recognizing fake ads positions, and revealing them
to individuals to stay away from the application for such positions.
All Fraud jobs advertisements can be viewed as bogus data on the web and as a type of scam. Information
on the internet can be false, which is divided into misinformation and disinformation. If information is
3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254 – 6529 Ed. 40 Vol. 11 N.º 1 Marzo - Junio 2022
254 https://doi.org/10.17993/3ctic.2022.111.251-267
falsely created by misunderstanding or misconception, disinformation is purposefully made to cheat per
user (Kumar & Shah, 2018). Fake job ads are considered disinformation. Supervised and unsupervised
learning solves disinformation-related problems such as fake news and reviews.
Bondielli and Marcelloni (2019) suggested two approaches in their paper. The rst approach uses fact-
checking websites for source information validation, it is named knowledge-based detection and the
second approach uses the key attributes and extraction of essential features from source information.
Fake or bogus news datasets are created manually based on multiple resources that are:
Creation of Fact-checking websites such as FakeNewsNet Dataset (Murtagh, 1991).
Using document samples labeled dataset in Burfoot Satire News Dataset by Burfoot and Baldwin
Credbank Dataset by Mitra and Gilbert (2015) approach by dataset gathering by using expert
For classication, supervised and unsupervised, both algorithms can work. Random forest agave
learning-based approach where each classier comprehends numerous tree-like classiers applied to
various examples, and each tree votes in favor of the most tting class. Another helpful technique can be
boosting, which can work with multiple classiers for a single classier to improve classication results.
Extended innovation applies an algorithm for classifying the weighted adaptations of training data
and chooses the grouping of the more signicant voting classier. AdaBoost illustrates a procedure of
boosting, which delivers better eectiveness (Murtagh, 1991). Expanding algorithms implies tackling
issues with spam ltration viably. In addition, Gradient boosting is an extra boosting procedure for a
Classier dependent on the decision tree rule (Prentzas et al., 2019). It likewise limits the deciency of
Algorithms approaches that can distinguish fake advertisements in online media are the decision forest.
Models of a quick, controlled ensemble. The decision tree can be the best model assuming the need to
3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254 – 6529 Ed. 40 Vol. 11 N.º 1 Marzo - Junio 2022
255 https://doi.org/10.17993/3ctic.2022.111.251-267
anticipate a target for up to two tests. It is suggested to train and test dierent models by utilizing the
Tune Model Hyperparameters system. Alghamdi and Alharby (2019) provided a model for detecting
scam posts in online job ads systems. The authors had used the EMSCAD dataset on various machine-
learning algorithms.
The methodology is divided into 3 steps preprocessing, selection of features, and identifying scams by
the classier:
In step, one unwanted noise and tags are removed from the data and bringing into general text.
To reduce extraversion features that are not in use selective features are selected using a support
vector machine and random forest classier.
It is reported that the detect fake job posts classication accuracy showed 97.4%.
Rathi and Pareek (2013) implemented various data mining techniques to detect spam mail in conjunction
with analyzing various data mining approaches on the spam dataset to search for the best classier for
email characterization. Support vector machine was utilized to classify and investigate data. A Naïve
Bayes classier was utilized to locate a specic feature of a class that was irrelevant to the existence of
some other feature, analyze and clean data by breaking down the information, and eliminate immaterial
and repetitive features from the data feature selection methods were used. The outcomes showed that
well exactness of the classier Random Tree is 99.715% (Rathi & Pareek, 2013).
Van Huynh et al. (2020) put forward a method in which authors used deep neural networks retrained
models with text datasets. The classication was done on IT-related jobs. Models were text CNN,
BiGRU CNN, and Bi-GRU-LSTM CNN. The TextCNN model is fully connected and contains layers
of convolution and pooling (Mujtaba et al., 2021; Mujtaba & Ryu, 2020). The training was done using
layers (convolution and pooling). Softmax function was used in this model for classication with that
ensemble classier was used to get more accuracy. Reported accuracy was 66% from text CNN. Bi-
GRU- LSTM CNN 70% accuracy
3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254 – 6529 Ed. 40 Vol. 11 N.º 1 Marzo - Junio 2022
256 https://doi.org/10.17993/3ctic.2022.111.251-267
Zhang, Dong, and Philip (2020) presented a model, an automatic fake detector. Utilizing text processing
separates good and false news, containing articles and subjects. They had gathered a custom dataset of
information or articles using the Twitter account PolitiFact site. For the proposed GDU diusive unit
model custom dataset was used to train. As there are multiple sources of information simultaneously, this
prepared model has worked well.
What is the best suitable classication algorithm for detecting Fake job advertisements?
What are the appropriate and important features for fraudulent job detection?
This research aims at constructing a suitable model to detect fraudulent job advertisements, to protect
the expatriates from falling into the trap. This research falls under the category of an empirical study that
would be based on observation, testing, evaluations, and comparison of the applied algorithms.
The research understudy can be described as a three-tier approach starting with the dataset preprocessing,
feature selection, and classifying by applying dierent machine learning models and evaluating them.
Let us look at the research that has already been done in this eld of detecting fraudulent advertisements
or detection of spam emails etc., over a period. It is observed that many researchers have applied several
classication algorithms, including SVM, NB, MLP, KNN, ID3, J48, decision tree, etc., among which
SVM outperformed in many cases (Mitra & Gilbert, 2015). Considering this performance of SVM as
a parameter to be validated, this research focuses on applying SVM, multinomial NB, decision tree,
random forest, and K-nearest neighbor on the dataset and comparing their results (Islam et al., 2020).
3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254 – 6529 Ed. 40 Vol. 11 N.º 1 Marzo - Junio 2022
257 https://doi.org/10.17993/3ctic.2022.111.251-267
Figure 1. Work Methodology.
Source: own elaboration.
3.2.1. DATASET
This research works on a dataset from Kaggle to categorize a job advertisement as fraudulent or not
based on some attributes derived from the advertisements available on dierent sources. The data was
available in a CSV le having 17880 instances of jobs advertisements. Each advertisement is dened in
terms of attributes on which we are working, that data is then preprocessed and classied through several
algorithms. As the dataset has many missing values and anomalies, it needs a preprocessing step before
it can be used as an input to any classication algorithm.
The initial dataset had 17 attributes based on which this model would be predicting the status of an
advertisement. These 17 attributes include job id, title, location, department, salary range, company
prole, description, requirements, benets, and telecommuting, has the company logo, has questions,
employment type, required experience, required education, industry, and function. Each attribute
contains either object or integer data. The label is binary for the specic problem domain, i.e., 0 for
non-fraudulent and 1 for fraudulent.
The preprocessing phase starts after analyzing the dataset for missing values and some basic statistical
operations on the integer data. Our integer elds include job id, telecommuting, has the company logo,
has questions, and the nal label of being fraudulent or not. Figure 1 describes the number of missing
values in each eld; this description justies the deletion of job IDs and salary range containing the
3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254 – 6529 Ed. 40 Vol. 11 N.º 1 Marzo - Junio 2022
258 https://doi.org/10.17993/3ctic.2022.111.251-267
maximum missing values. The integer elds were ten checked for the correlation, and Figure 2 depicts
the correlation heat map.
Figure 2. Key attributes.
Source: own elaboration.
After doing the exploratory data analytics, the process calls for proper preprocessing, including removing
the missing values and stop words, deleting the irrelevant attributes that can be observed from the
correlation heat-map, and nally removing the extra space. Now, the dataset is ideal for transforming
into categorical encoding to achieve a feature vector. This feature vector would then be the nal and
transformed input to the classiers.
3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254 – 6529 Ed. 40 Vol. 11 N.º 1 Marzo - Junio 2022
259 https://doi.org/10.17993/3ctic.2022.111.251-267
Figure 3. Attributes heat map.
Source: own elaboration.
The proposed approach compares the performance of classiers on two dierent feature sets. The
rst feature set includes the processed data discussed above and the second feature set has the integer
attributes, benets, and location. In this research, several classiers are engaged, such as Naive Bayes
Classier, Decision Tree Classier, K- Nearest Neighbor, and Random Forest Classier, classifying job
posts as fake. Note that ‘fraudulent’ is the target class for the research under discussion. Moreover, the
feature sets on which the models are trained are mentioned in Table 1 below:
3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254 – 6529 Ed. 40 Vol. 11 N.º 1 Marzo - Junio 2022
260 https://doi.org/10.17993/3ctic.2022.111.251-267
Table 1. Feature Sets.
Feature set 1 Feature set 2
title, location, department, company
prole, description, requirements, benets,
telecommuting, has the company logo,
has questions, employment type, required
experience, required education, industry, and
title, telecommuting, has the company logo,
has questions, benets
Source: own elaboration.
For both the feature sets, the classiers are passed on to the training phase with 80 percent of the entire
dataset, the remaining 20 percent would be used for the prediction phase. Training the classiers for the
proposed approach starts with choosing the right and tuned parameters as default parameters do not
guarantee the best and promising results. After the prediction of the testing data, the model would be
then evaluated on metrics such as Accuracy, F-measure, and Cohen- Kappa score. They are keeping
the work on both the feature sets in parallel. The best classier would be chosen to have outstanding
performance among all the peer classiers for each feature set.
To evaluate the performance of any machine-learning model, evaluation metrics are used for this
purpose. Given metrics are considered for evaluating and identifying the subtle approach for solving
a problem. Accuracy metric aims to identify the true cases (predictions) from overall numbers to cases
given to test. Accuracy may not be the primary metric for checking the model’s performance as false cases
(prediction). If a false result is taken as true, one will become problematic. It is important to consider false
positive and false negative cases to requite the wrong classication. Precision checks the ratio of the right
identied positive case from the total positive results given by the classier. Recall presents the correct
results of positive cases divided by the number of cases relevant. F-measure is a metric, which is involved
in precision and recall, calculation is done by the harmonic mean of precision and recall.
3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254 – 6529 Ed. 40 Vol. 11 N.º 1 Marzo - Junio 2022
261 https://doi.org/10.17993/3ctic.2022.111.251-267
After getting the predictions from all ve classiers discussed in this research, their performance is
compared based on a couple of evaluation metrics to conclude the best classier for predicting fraudulent
job advertisements. Table 1 displays the comparative study of the classiers concerning evaluating
metrics for both feature sets.
Table 2. Comparative table of classiers performance.
Source: own elaboration.
Figure 4. Accuracy Metric Comparisons of algorithms.
Source: own elaboration.
3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254 – 6529 Ed. 40 Vol. 11 N.º 1 Marzo - Junio 2022
262 https://doi.org/10.17993/3ctic.2022.111.251-267
Figure 5. F1 Metric Comparison of algorithms.
Source: own elaboration.
Figure 6. Cohen-Kappa Score of algorithms.
Source: own elaboration.
3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254 – 6529 Ed. 40 Vol. 11 N.º 1 Marzo - Junio 2022
263 https://doi.org/10.17993/3ctic.2022.111.251-267
Figure 7. MSE calculations.
Source: own elaboration.
Platforms such as online job portals or social media for job advertisements are an exciting way of
attracting potential candidates on which many enterprise companies are dependent on the hiring process.
Fake jobs scam detection at an early stage can save a job seeker and make them only apply for legitimate
companies. For this purpose, various machine learning techniques were utilized in this paper. Specically,
supervised learning algorithms classiers were used for scam detection. This paper experimented with
dierent algorithms such as naïve Bayes, SVM, decision tree, random forest, and K-Nearest Neighbor.
It is reported that the K-NN classier gives a promising result for the value k=5 considering all the
evaluating metrics. On the other hand, Random Forest is built based on 500 estimators on which the
boosting is terminated. In the future, the proposed method can be used for mobile devices using energy-
ecient techniques (Mujtaba, Tahir, & Soomro, 2019; Mujtaba & Ryu, 2021).
3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254 – 6529 Ed. 40 Vol. 11 N.º 1 Marzo - Junio 2022
264 https://doi.org/10.17993/3ctic.2022.111.251-267
Alghamdi, B., & Alharby, F. (2019). An intelligent model for online recruitment fraud detection. Jour-
nal of Information Security, 10(03), 155. https://www.scirp.org/journal/paperinformation.aspx?pa-
Bondielli, A., & Marcelloni, F. (2019). A survey on fake news and rumour detection techniques.
Information Sciences, 497, 38-55. https://app.dimensions.ai/details/publication/pub.1114201506
Islam, M. R., Liu, S., Wang, X., & Xu, G. (2020). Deep learning for misinformation detection on on-
line social networks: a survey and new perspectives. Social Network Analysis and Mining, 10(1), 1-20.
Kumar, S., & Shah, N. (2018). False information on web and social media: A survey. arXiv preprint ar-
Mitra, T., & Gilbert, E. (2015). Credbank: A large-scale social media corpus with associated credibility
annotations. In Ninth international AAAI conference on web and social media. https://ojs.aaai.org/index.
Mujtaba, G., & Ryu, E. S. (2020). Client-driven personalized trailer framework using thumbnail con-
tainers. IEEE Access, 8, 60417-60427. https://ieeexplore.ieee.org/document/9046852
Mujtaba, G., & Ryu, E. S. (2021). Human Character-oriented Animated GIF Generation Framework.
In 2021 Mohammad Ali Jinnah University International Conference on Computing (MAJICC) (pp. 1-6). IEEE.
Mujtaba, G., Lee, S., Kim, J., & Ryu, E. S. (2021). Client-driven animated GIF generation fra-
mework using an acoustic feature. Multimedia Tools and Applications, 1-18. https://link.springer.
3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254 – 6529 Ed. 40 Vol. 11 N.º 1 Marzo - Junio 2022
265 https://doi.org/10.17993/3ctic.2022.111.251-267
Mujtaba, G., Tahir, M., & Soomro, M. H. (2019). Energy ecient data encryption techniques in
smartphones. Wireless Personal Communications, 106(4), 2023-2035. https://link.springer.com/arti-
Murtagh, F. (1991). Multilayer perceptrons for classication and regression. Neurocomputing, 2(5-6), 183-
197. https://pure.hud.ac.uk/en/publications/multilayer-perceptrons-for-classication-and-re-
Prentzas, N., Nicolaides, A., Kyriacou, E., Kakas, A., & Pattichis, C. (2019). Integrating ma-
chine learning with symbolic reasoning to build an explainable AI model for stroke prediction.
In 2019 IEEE 19th International Conference on Bioinformatics and Bioengineering (BIBE) (pp. 817-821).
IEEE. https://ieeexplore.ieee.org/document/8941679
Rathi, M., & Pareek, V. (2013). Spam mail detection through data mining-A comparative performan-
ce analysis. International Journal of Modern Education and Computer Science, 5(12), 31. https://www.
Van Huynh, T., Van Nguyen, K., Nguyen, N. L. T., & Nguyen, A. G. T. (2020). Job prediction: From
deep neural network models to applications. In 2020 RIVF International Conference on Computing and
Communication Technologies (RIVF) (pp. 1-6). IEEE. https://ieeexplore.ieee.org/document/9140760
Vidros, S., Kolias, C., Kambourakis, G., & Akoglu, L. (2017). Automatic detection of online re-
cruitment frauds: Characteristics, methods, and a public dataset. Future Internet, 9(1), 6. https://
Ward, A., Gbadebo, A., & Baruah, B. (2015). Using job advertisements to inform curricula design
for the key global technical challenges. In 2015 International Conference on Information Technology Based
Higher Education and Training (ITHET) (pp. 1-6). IEEE. https://ieeexplore.ieee.org/abstract/docu-
3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254 – 6529 Ed. 40 Vol. 11 N.º 1 Marzo - Junio 2022
266 https://doi.org/10.17993/3ctic.2022.111.251-267
Zhang, J., Dong, B., & Philip, S. Y. (2020). Fakedetector: Eective fake news detection with deep
diusive neural network. In 2020 IEEE 36th International Conference on Data Engineering (ICDE) (pp.
1826-1829). IEEE. https://arxiv.org/abs/1805.08751
3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254 – 6529 Ed. 40 Vol. 11 N.º 1 Marzo - Junio 2022
267 https://doi.org/10.17993/3ctic.2022.111.251-267