“eRReBIS” Business Intelligence based Intelligent Recommender System
for e-Recruitment Process
Siwar Ayadi
, Manel Bensassi
and Henda Ben Ghezala
RIADI Lab, National School of Computer Science, Manouba University, Tunisia
e-Recruitment, Business Intelligence, Content-based Recommendation, Similarity Measure, Prescriptive
Analysis, Machine Learning Algorithms.
Due to the continuous and growing spread of the corona virus worldwide, it is important, especially in the
business era, to develop accurate data driven decision-aided system to support business decision-makers in
processing, managing large amounts of information in the recruitment process. In this context, e-Recruitment
Recommender systems emerged as a decision support systems and aims to help stakeholders in finding items
that match their preferences. However, existing solutions do not afford the recruiter to manage the whole
process from different points of view. Thus, the main goal of this paper is to build an accurate and generic data
driven system based on Business intelligence architecture. The strengths of our proposal lie in the fact that it
allows decision makers to (1) consider multiple and heterogeneous data sources, access and manage data in
order to generate strategic reports and recommendations at all times (2) combine many similarity’s measure in
the recommendation process (3) apply prescriptive analysis and machine learning algorithms to offer adapted
and efficient recommendations.
E-recruitment platforms, as decision support system,
has increasingly been used in the industry particu-
larly with the epidemic widespread and accomplished
clear advantages for both recruiters and job-seekers
by reducing the recruitment time and advertisement
cost. Data has been largely widespread in the job
market.Consequently, these platforms suffer from an
inappropriateness of traditional information retrieval
and exploitation techniques (Al-Otaibi and Ykhlef,
2012). Recommendations for e-recruitment systems
generally differ from recommendations generated in
other contexts (e.g. movies, e-commerce), given that
the job seekers’level of competencies and knowledge
rather then their interests is key for suggesting the
most appropriate position. In fact, the problem of
matching jobs and candidates can be enhanced from
two distinct perspectives: (a) find relevant candidates
to a job opening; and (b) select the suitable jobs to a
specific candidate.
Thus, the challenge of making recommendations
and to develop an accurate decision support system
is closely tied to which and how data are extracted,
transformed and conveyed. Over the past few years,
Jobs’ recommendation became an important task for
the modern recruitment process in order to improve
manager experience.
On the other hand, existing recommender solu-
tions do not afford the recruiter to manage the whole
process. Recruiters need to capitalize on clear KPIs
inside of Talent Management through guided and dy-
namic insights. Consequently, Explanation for rec-
ommendation systems through the deployment of a
KPIs dashboard with the purpose of helping the re-
cruitment’s team make more precise decisions is un-
In order to support this research area, we describe,
in this article, a Content based Recommender which
relies on a Business Intelligence architecture, devel-
oped in a computer science company, part of a larger
group, whose objective was to improve the manage-
ment of e-recruitment process searching for more ad-
equate means to attract qualified coworkers. Our pro-
posal is based on prescriptive analytics that assists de-
cision makers in identifying data-driven strategic de-
This article is organized as follows. We ana-
lyze the e-recruiting business requirements and re-
Ayadi, S., Bensassi, M. and Ben Ghezala, H.
“eRReBIS” Business Intelligence based Intelligent Recommender System for e-Recruitment Process.
DOI: 10.5220/0011530200003318
In Proceedings of the 18th International Conference on Web Information Systems and Technologies (WEBIST 2022), pages 373-380
ISBN: 978-989-758-613-2; ISSN: 2184-3252
2022 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
view related recommender systems proposals for e-
recruitment management in section 2. eRReBis Rec-
ommender System overview and architecture are pre-
sented in Section 3. Usage scenario and the inter-
pretation of results are discussed in Section 4. The
evaluation of efficiency of the proposal system are de-
scribed in section 5. Finally, conclusions are drawn
and future works are suggested in Section 5.
In recent years, data are overwhelming, growing in
size and in connections, creating a potential chal-
lenge of information retrieval in a complex environ-
ment. Filtering, extracting, and prioritizing relevant
information is a pressed requirement for data ‘s users
and analytics, especially, decision makers in the busi-
ness management field. In this context, recommender
systems have been developed to search into a large
amount of data and select relevant information ac-
cording to user’s preferences, interest, or observed be-
havior about item (Karimi et al., 2018)
In order to assist the recruiter in looking for
new talents and decision-maker in defining strate-
gic guidelines for the e-recruitment process, recom-
mendation system, as decision support system (Stohr
and Viswanathan, 1998), afford a flexible support
to different stakeholders. In literature, many rec-
ommender system have been proposed. Casagrande
(Casagrande et al., 2017) proposes person-job recom-
mendation system using the profile information from
both candidates and jobs descriptions in order to find
a good match between talents and jobs. Ramanath
(Ramanath et al., 2018) explored machine learning
of candidate potentials to determine a ranked list of
the most relevant candidates from the LinkedIn thou-
sands of candidates. Recently, Norbert (Jiechieu and
Tsopze, 2021) followed the ideas of recommender
systems and proposed a comprehensive job recom-
mender system based on competency prediction and
resume classification. He used classification model
based on neural networks in order to predict high level
skills imprecisely written.
Other works are dedicated to semantic modeling.
For example, Paquette (Paquette, 2014) proposes a se-
mantic modeling of skills which allows to model dif-
ferent level of granularity of skills. Zhao et al. (Zhao
et al., 2015) proposed a combination of Named En-
tity Recognition and Named Entity Normalization to
identify skills from texts, considering them as named
entities. Cheng et al. collected the job-related infor-
mation from various social media sources and con-
structed an inter-company job-hopping network to
demonstrate the flow of talents. (Wang et al., 2013)
predicted the job transition of employees by exploit-
ing their career path data. Xu et al. (Xu et al., 2016)
proposed a talent circle detection model based on a
job transition network which can help the organiza-
tions to find the right talents and deliver career sug-
gestions for job seekers to locate suitable jobs.
However, these systems are not enough sophisti-
cated and presented some limits related to the man-
agement of e-recruitment process. In fact, enterprises
require better management of the data and systems
involved in decision-making. To do so, we conclude
that decision-makers need to:
Deploy an agile IT architecture that can integrate
an increasing number of data sources required for
decision-making, as well as external and big data
Increase benefit from collective intelligence in de-
cision making process by exploring collaboration
and participation in data analysis and planning
processes to improve performance and to eas-
ily move from strategic to operational point of
view (and reverse) to a more democratic style of
Define, pervasively use and visualize KPIs across
the organization through dashboard to achieve a
common foundation for decisions, align measures
of success and focus data governance on impor-
tant data.
From this point of view, presented works still in-
sufficient and do not cover the whole business require-
ment to manage the e-recruitment process. Compa-
nies had to embrace flexible business structures and
to change decision-making cultures from decisions
based on intuition and experience to data-driven
decision-making in order to identify new business
opportunities, predict future trends and behavior.
On the other side, many online recruiting plat-
forms suffer from an inappropriateness of Boolean
search methods for matching applicants with job re-
quirements. Consequently, a large number of candi-
dates missed the opportunity of recruiting (Lang et al.,
2011). Actual practices and theoretical thoughts show
that this search type is insufficient for achieving a
good fit between candidate aptitudes and job require-
ments (F
arber et al., 2003).
Further, given the business requirements, re-
cruiters need to capitalize on clear KPIs inside of
Talent Management through guided and dynamic in-
WEBIST 2022 - 18th International Conference on Web Information Systems and Technologies
sights. Consequently, the deployment of a KPIs dash-
board with the purpose of helping the recruitment’s
team make more precise decisions is underlined.
3.1 Motivations
Our main goal is to provide a theoretical model along
with a proof-of-concept implementation that:
could be adopted and appropriately adapted to en-
hance the e-recruitment management from back to
front office.
assists decision makers in identifying data-driven
strategic decision and help them to avoid the lim-
itation of standard data analytics.
Thus, the design of our approach is dedicated to the
need to exploit capabilities of BI paradigms towards
the enhancement of data driven decision making in
the e-recruitment management process. The blending
of BI tools, machine learning algorithms and recom-
mender system is considered promising for the im-
provement of data driven decision making in the e-
recruitment field and for the development of solutions
able to provide effective and adapted recommenda-
tions by combining different techniques of similarity
3.2 eRReBIS: Conceptual Architecture
In the proposed approach (Fig 1), heterogeneous data
are collected transformed according the following
Step 1: Data Gathering: from multiple data sources
that are structured by Company’s developers. We con-
Byblos data source that contains Company’s em-
ployee information.
Application Tracking System (ATS) Database
from which passive candidate profiles looking for
a new opportunity are created.
LinkedIn from which relevant Company’s job data
is extracted. To do so, we have opted for the web
scraping (screen scraping) method, based on au-
tomatic data retrieval implemented in python.
Step 2: Data Integration: is ensured by the ETL
(Extract, Transform, Load) data integration process,
which takes care of collecting all the necessary infor-
mation from the various sources.
Data Pre-processing and Cleaning: Heterogeneous
data requires a homogenization and integration pro-
cess. The first stage consists of eliminating the ir-
relevant data that represents a noise. In the second
stage, we removed duplicated data, punctuation and
numbers. The tokenization stage is an essential phase
in order to remove stop words and irrelevant words.
In this step,we also integrate an algorithm to measure
the similarity between the descriptions of the offers
and the profiles of the candidates. Transformed and
adapted data to the feeding phase is stored in a spe-
cialized database “data Mart” which is integrated in
the ”Data Warehouse”.
Step 3: Data Analysis: is the graphical represen-
tation of the performance indicators and recommen-
dations in a way that is understandable and helps the
decision maker to make strategic and relevant deci-
sions. For this end, we analyse data with the Tabular
Model which is based on a new in-memory engine for
tables. The in-memory engine loads all table data into
memory and answers questions directly from mem-
ory. This is very fast in terms of response time. Un-
like the multidimensional model, the tabular model
processes the data warehouse in a single processing
block without dividing it into cubes.
Step 4: Data Visualization: whose objective is to
make access to the decision much more immediate
through graphic representations and indicators map-
Step 5: Prediction and Recommendations: Our
”eRReBIS” system provides predictive analytics and
guided recommendations to decision makers in order
to provide additional insights. Given the fact that our
training data is labeled and has output variables cor-
responding to the the input variables, supervised ma-
chine learning algorithm was the most suitable to our
business requirement. In addition, our main goal is
to support decision-maker tp predict the evolution of
strategic KPI: the number of incoming employees, the
number of resignations and hires,...Then, we need a
regression algorithm to build a mathematical model
from the training data. Then for this study, to choose
the most suitable regression model to the extracted
data, a comparative and evaluation study is elaborated
between these two supervised and regression algo-
Random Forest Regression: is a supervised
learning algorithm that combines the prediction
of several automatic learning algorithms to make
a more accurate prediction than a single model.
“eRReBIS” Business Intelligence based Intelligent Recommender System for e-Recruitment Process
Figure 1: Conceptual Architecture.
During the learning process the algorithm builds
decision trees that work in parallel.
Linear Regression: is a supervised prediction al-
gorithm, its objective is to predict from the de-
pendent variables of the independent variables by
finding a function of linear prediction y=f(x) in
order to predict values which are not observed.
There are two types of simple and multiple lin-
ear regression, according to the number of depen-
dent variables. In our case we have two dependent
variables (month, year) so it is the multiple linear
Finally, Data is represented in an e-recruitment dash-
board that aims to make access to the decision much
more immediate through graphic representations and
indicators mapping presented in the following sec-
3.3 e-Recruitment Dashboarding based
on KPI Cartography
To build such a dashboard, the first step is to under-
stand with greater detail the real pains of Talent Man-
agement and the main opportunities that could be cre-
ated for the short and long term by answering these
questions: What is currently not possible to analyze?
Which analyses and information would be necessary
for a decision to be made comfortably? For each
WEBIST 2022 - 18th International Conference on Web Information Systems and Technologies
analysis, which metrics would help take the actions
chosen above? The recruitment process is divided
into two phases: pre and post hire. In order to re-
spond to manage the whole process chronology, we
organise the e-recruitment KPI into:
Pre-hire Indicators (calculated before the re-
cruitment): The first pre-hire indicators are spe-
cific to the application tracking system (SmartRe-
cruters) as the recruitment process begins online.
Then, pre-hire indicators are automatically calcu-
lated as soon as the recruiter receives the applica-
Post-hire Indicators (calculated after the recruit-
ment): The second post-hire indicators are calcu-
lated from the company’s ”Byblos” database af-
ter hiring by using the dimensions ”Personnel”,
”Contract” and their attributes.
In order to assist different stakeholders in decision
making process at different levels (operational, ana-
lytical and strategic), we propose, a cartography of
indicators organised as follow:
Operational Indicators: which show data re-
lated to daily e-recruitment operations. The main
purpose of an “operational dashboard” is to pro-
vide a comprehensive snapshot of e-recruitment
performance. For example, we cite : Number
of applications that represents the total number
of applications received by the recruitment chan-
nels, Hiring rate, Number of inactive or active
Analytical Indicators: such as Expected num-
ber of hires and expected number of resignations,
which use historical company data to identify
trends that can influence future decision-making.
The ideal audience for viewing analytical dash-
boards are database analysts, as they typically re-
quire a level of understanding that a typical busi-
ness user may not possess.
Strategic Indicators: which offers a decision-
maker to track performance in relation to strategic
key performance indicators to better align actions
with strategy such as cooptation indicator which
measures ”participative recruitment” (number of
applications from internal source). In fact, Coop-
tation is a recruitment method that enhances col-
laboration and employee participation in the re-
cruitment process by recommending talent for a
specific position.
3.4 e-Recruitment Recommender
Figure 2 shows the use case breakdown of the eRRe-
BIS System. The roles involved in the process include
both human actors and machine system components
(shown, respectively, as man icons and gray rectan-
gles in Figure 2 ). The recruiter has the possibility to
configure the dashboard, to select the axis of analysis
and to explore operational and tactical KPI through
descriptive analysis.The process of descriptive anal-
ysis is split into job seeker reporting (visualising of-
fers list) and recruiter reporting (visualizing candidate
profile). Recruiter reporting is the process in which
similarity computed and all aspects related to job po-
sition and each candidate profile are visualized . The
recruiter deals with lists of offers, possibly organiz-
ing them hierarchically in sub groups through recom-
mendation engines. Then, he could visualize the most
appropriate profile to a job position and conversely.
The role of the decision-maker is to deal with the
strategic dashboard through predictive analysis when
necessary to define the recruitment requirements and
eRReBIS System is also based on the “Recom-
mendation engine” that is an autonomous module in
charge of executing the recommendation algorithm
and computing the similarity measure.It is based on
predictive analysis which is a process during execu-
tion and it always follows descriptive analysis.
Recommendation engine ensures prescriptive
analysis which is the process by which the job seeker,
recruiter and decision-maker finally makes decisions
about the most suitable position, the most suitable
profile or about recruitment strategies. Based on the
output of predictive analysis and the computation of
the similarity’s measure, recommendations are pro-
posed to the job seeker, recruiter and decision-maker
by eRReBIS solution. To do so, we propose a content
based recommendation system and algorithm which:
Creates vectors from jobs’ databases and profiles.
Combines semantic and syntactic similarity mea-
sures to reconcile documents represented in vec-
tor form.
A comparative study has been conducted, to show
the efficiency of our measure compared with classi-
cal measures, is presented in section 5.0.2.
eRReBIS system was developed by combining
several tools and languages. In fact, we visualize
integrated data using the Power BI tool. Recom-
mender Engine, coded with python language, was in-
jected into the integration process deployed with Tal-
end Data Integration (see figure 2).
“eRReBIS” Business Intelligence based Intelligent Recommender System for e-Recruitment Process
Figure 2: eRReBis Functionnalities.
Our prototype provides several views supporting fil-
tering in the result set and integrates:
KPI cartography for strategic, analytical and oper-
ational decisions based on straightforward charts
to interpret.
Prescriptive analysis based on predictive machine
learning algorithm and adapted recommendations
Building upon these features, we divide our dash-
board into 3 views to support the recommendation
Strategic View: (figures 3 and 4) Main KPIs of
Talent Management, not with the objective of a
micro analysis but a general view of how the met-
rics are doing, being able to identify any potential
issues. KPIs such as Forecast, hired, candidate to
hire, and others. This can all be filtered by job or
Analytical View: also known as tactical view,
shows a more specific view for the leader of the
Figure 3: Analysis of applications.
Figure 4: Turnover analysis.
area (recruiter). Has some analyzes which are
more broken up by date for example.
A snapshot of the dashboard illustrates generic re-
sults on the offers and profiles of the candidates,
such as the total number of offers, number of of-
WEBIST 2022 - 18th International Conference on Web Information Systems and Technologies
fers by location, as well as the total number of
profiles of the candidates and the profiles by se-
Operational View: This view is designed to be
used by recruiters and offer a very micro view of
how each job is doing, the pipeline, forecast etc.
This is all filtered by the recruiter.
Multi-varied Recommendations on Different
Activities and Steps (figures 5, and 6): The re-
cruiter has the possibility to choose an offer from
the list extracted from the linked In. He can visu-
alize a table that contains the profiles with the sim-
ilarity measure. The most suitable profile with the
highest similarity measure to his offer is shown in
figure 5. The selected profile could be more ad-
equate to another position proposed by the com-
pany. eRReBIS prototype offers the possibility to
the recruiter to evaluate the adequation between
this profile and all offered positions. He could se-
lect, as shown in the figure, the same profile re-
sulting from the first search and visualize the sim-
ilarity measure with all the offers of the company.
In this way, he can visualize the most adequate of-
fer for this profile. Comparing scores, the last one
is more appropriate.
Figure 5: Analysis of the offer recommendations.
Figure 6: Analysis of Profile Recommendations.
5.0.1 Prediction Algorithm Performance
For the evaluation process of the prediction algorithm,
we used two measures : The term R score 1 and The
Root mean square error (RMSE) 2
The result of the comparative study is shown in the
figure 7 and 8. The algorithm “Random Forest regres-
sion algorithm” that has the lowest value of RMSE
and the highest value of R score was adopted.
Figure 7: R2 Results.
Figure 8: RMSE Results.
5.0.2 Evaluation of Similarity Measures
The job seeker data is taken from Babylos and ATS
database. The data set contains 24734 candidates
plus 7270 active collaborators . In total we have
32004profiles, where each profile contains weighted
competencies and 70 job positions. For testing pur-
poses, we divide the user ratings dataset into training
(80%) and testing sets (20%). We perform a 4-fold
cross validation and results are averaged over the 4
cycles of execution. We conducted a set of experi-
ments using four different scenarios. Scenario 1 uses
the Content based Filtering method only. We con-
ducted three sub experiments with different syntac-
tic measures(Pearson, Cosine and Jaccard).Scenario
“eRReBIS” Business Intelligence based Intelligent Recommender System for e-Recruitment Process
2 uses the semantic similarity method only, Scenario
3 uses a hybrid of syntactic and semantic similar-
ities methods with equal contribution (both have a
weight of 0.5).We present the evaluation of the pro-
posed similarity measure compared with the other
measure. Similarity measures are computed as shown
in 1. When we combine cosine and semantic similar-
ity, we obtain optimal results and it is more accurate
than a single measurement. We evaluate recommen-
dation accuracy by using Precision as shown in table
Table 1: Similarity measurement results.
Syntactic Similarity Semantic
Hybrid Simi-
Job ID Pearson Cosine Jaccard Spacy
Cosine and
Spacy NLP
1 1 0.34 0.44 0.41 0.51 0.47
2 2 0.45 0.38 0.21 0.32 0.35
3 3 0.74 0.92 0.67 0.99 0.9
4 4 0.56 0.58 0.45 0.62 0.6
5 5 0.33 0.41 0.12 0.38 0.35
Table 2: Evaluation recommendation.
Syntactic similarity Semantic
Hybrid simi-
Jaccard Pearson Cosine Spacy
Cosine &
Spacy NLP
Total of recommenda-
180 166 208 195 248
Pertinent recommen-
77 68 92 84 139
Precision Value 0.42 0.4 0.44 0.43 0.56
“eRReBIS” is data-driven strategic decision system
and differs from the existing approach in that it re-
lies mainly on data management based BI architec-
ture. We propose Dynamic Multiviews dashboard-
ing (strategic, tactical, operational) based on prescrip-
tive analysis in order to assist the recruiter to iden-
tify trends, to create new business opportunities. Our
prototype is a content-based recommendation engines
that deploy a combination of similarity measures in
order to classify the candidates. We show that by
combining the similarity measures between the jobs
and skills, our model provides better recommenda-
tion for both recruiter and candidate. Additionally,
we also show some case studies which validate our
claims. Our system is scalable and can be gradually
enhanced by other massive data sources from social
networks and Business platforms. Furthermore, the
system architecture can be improved by other intel-
ligent features and components of deep learning al-
lowing the inference of strategic recommendations.
Future research will focus more on deep and predic-
tion analysis with machine learning techniques to im-
prove the performance of recommenders and linked
data profiling in the knowledge representation.
Al-Otaibi, S. T. and Ykhlef, M. (2012). A survey of job rec-
ommender systems. International Journal of Physical
Sciences, 7(29):5127–5142.
Casagrande, A., Gotti, F., and Lapalme, G. (2017). Cerebra,
un syst
eme de recommandation de candidats pour l’e-
recrutement. In AISR2017.
arber, F., Weitzel, T., and Keim, T. (2003). An automated
recommendation approach to selection in personnel
Jiechieu, K. F. F. and Tsopze, N. (2021). Skills prediction
based on multi-label resume classification using CNN
with model predictions explanation. Neural Comput-
ing and Applications, 33(10):5069–5087. Publisher:
Karimi, M., Jannach, D., and Jugovac, M. (2018). News
recommender systems–Survey and roads ahead. Infor-
mation Processing & Management, 54(6):1203–1227.
Publisher: Elsevier.
Lang, S., Laumer, S., Maier, C., and Eckhardt, A. (2011).
Drivers, challenges and consequences of e-recruiting:
a literature review. In Proceedings of the 49th SIGMIS
annual conference on Computer personnel research,
pages 26–35.
Paquette, G. (2014). Technology-based instructional de-
sign: Evolution and major trends. Handbook of re-
search on educational communications and technol-
ogy, pages 661–671. Publisher: Springer.
Ramanath, R., Inan, H., Polatkan, G., Hu, B., Guo, Q.,
Ozcaglar, C., Wu, X., Kenthapadi, K., and Geyik,
S. C. (2018). Towards deep and representation learn-
ing for talent search at LinkedIn. In Proceedings of
the 27th ACM International Conference on Informa-
tion and Knowledge Management, pages 2253–2261.
Stohr, E. A. and Viswanathan, S. (1998). Recommendation
systems: Decision support for the information econ-
Wang, J., Zhang, Y., Posse, C., and Bhasin, A. (2013). Is it
time for a career switch? In Proceedings of the 22nd
international conference on World Wide Web, pages
Xu, H., Yu, Z., Yang, J., Xiong, H., and Zhu, H. (2016).
Talent circle detection in job transition networks. In
Proceedings of the 22nd ACM SIGKDD International
Conference on Knowledge Discovery and Data Min-
ing, pages 655–664.
Zhao, M., Javed, F., Jacob, F., and McNair, M. (2015). Skill:
A system for skill identification and normalization. In
Twenty-Seventh IAAI Conference.
WEBIST 2022 - 18th International Conference on Web Information Systems and Technologies