International Journal of Management, IT & Engineering

Vol. 13 Issue 04, April 2023,

ISSN: 2249-0558 Impact Factor: 7.119
Journal Homepage: http://www.ijmra.us, Email: [email protected]
Double-Blind Peer Reviewed Refereed Open Access International Journal - Included in the International Serial Directories Indexed &
Listed at: Ulrich's Periodicals Directory ©, U.S.A., Open J-Gate as well as in Cabell’s Directories of Publishing Opportunities, U.S.A

An Innovative Hashing Scheme and BiLSTM-based

Dynamic Resume Ranking System
1st Sandeepkumar Racherla Prathyusha Sripathi
Amazon Amazon
Data Scientist Data Scientist
Seattle, Washington, United States Seattle, Washington, United States
[email protected] [email protected]
ORCID: 0000-0002-8155-5643 ORCID: 0000-0001-9260-3906

Abstract—Content-based-filtering dependent job The performance of Deep Learning (DL)-based solutions

recommender systems are failing to cope with the dynamic depends on the dataset’s quality [7]. Recommender systems,
nature of the job market by processing the massive amount of also known as recommendation systems, are a sub-branch of
data the Human Resources (HR) professionals handle for talent information filtering syst. Usually, they are used to
hunting. Deep Learning (DL)-based job recommender systems recommend items to the users [8]. A DNN-based
capable of dynamically establishing a relation between recommendation system trained with a proper dataset
candidate profile and job description is a promising solution to identifies the motives of the users based on their interaction
this challenge. This paper proposes an innovative hashing
with the system and recommends items the user may be
scheme-based word embedding dependent recommendation
interested in [9]. This system consists of two parts - feature
system using a Bidirectional Long-Short Term Memory
(BiLSTM) network to intelligently sort the applicants' resumes
extraction and prediction. The feature extraction part extracts
according to the job description and requirements. The the feature from the user interaction, and the prediction part
proposed system ranks the resumes against job description predicts the probability of the user being interested in some
based on skills required with 89.93% accuracy, 88.11% available items [10]. The system learns from the interaction,
precision, 89.17% recall, 6.79% 𝑹𝟐 , 6.05% RMSE, and 7.31% correct and incorrect prediction, and eventually becomes
MAE. The proposed recommendation system effectively better at recommending the appropriate items [11].
handles dynamic job market requirements and the rapid rate of
new candidate entrance. This noble three-module system A. Research Gap Identification
architecture beats the challenges of automating talent The longer a recommendation system performs on a
acquisition responsibilities of HR managers in today's vibrant
similar task, the better it becomes in that task. Moreover, the
and rapidly changing job market.
more enriched the dataset gets, the better performance it can
Keywords—Resume Ranking, CV Shortlisting, Deep ensure
Learning, BiLSTM, Word Embedding [12]. It works well for e-commerce sites, Social media, and
e-Learning platforms [13]. However, the ever-evolving and
I. INTRODUCTION changing environment and shorter interaction period are
significant challenges to the overall efficiency of such
Talent acquisition is one of the responsibilities of Human systems [14]. That is why despite the massive success of
Resource (HR) officials [1]. The resume ranking is at the heart
DNN-based recommendation systems [15], it needs to
of this process [2]. It starts with preparing the job description
improve in assisting HR managers in automating the
and circular publication. Then the interested candidates apply
with their resumes. The HR hiring managers need to review candidate shortlisting process. A recommendation system
these resumes to rank the eligible and potential candidates [3]. still works well if the job description remains the same every
It sounds simple in this context. However, ever-changing job time. However, the job market is more vibrant than ever, and
descriptions, rapid rate of change of new job seekers, a large job descriptions change rapidly [16].
number of resumes for a single circular, and the improper ratio Furthermore, a job circular stays online for a short period.
of relevant and irrelevant resumes make the talent acquisition At the same time, the number of new job seekers changes
process very competitive for HR professionals [4]. The rapidly [17]. That is why personalized recommendation
content-based-filtering dependent job recommender systems systems must work better in this vibrant environment [18].
used to be an effective solution to assist in this process [5]. On the other hand, the customized PageRank algorithms
However, this system is failing to cope up with the vibrant and work well for old users with enough interaction with the
rapidly changing nature of the job market [6]. A noble system. There are well-developed and personalized
recommendation system has been developed and presented in PageRank algorithms [19]. However, there is a research gap
this paper to meet the market demand and HR professionals’ that focuses on a system that works well for rapidly changing
criteria for resume ranking. new and existing users. Data sparsity is another challenge in

9 International journal of Management, IT and Engineering

http://www.ijmra.us, Email: [email protected]
 ISSN: 2249-0558 Impact Factor: 7.119
this domain. When cosine similarities or relational methods 2) Bag of Words (BoW): As the name suggests, in this
are used for filtering and ranking, the different data instances embedding technique, a text is represented as a bag of words.
exhibit a heterogeneous nature [20]. As a result, the data The grammatical rules, word sequence, and multiplicity are
becomes sparse distorting identifiable or measurable patterns. ignored in this method. The equation 2 defines the structure
Incorporating advanced behavioral similarities with the [23].
existing DNN-based recommendation systems is another 𝑁
research gap. 𝐵𝑜𝑊 = ∑ 𝑊𝑜𝑟𝑑𝑖 : 𝑖 (2)
B. Technology Selection 𝑖=1
It is famous for document classification based on the
The deep learning-based recommendation systems are frequencies of certain words. However, the BoW is not an
adequate to abridge the research gaps identified in the appropriate solution for resume ranking because of not
previous section [21]. There are numerous DNN models consider the word sequence.
available. The Feed-Forward Neural Network (FFNN)
architecture [22] and text as a bag of words [23] are potential 3) Continuous Bag of Words (CBoW): The CBoW
solutions to text classification. The Convolutional Neural method is not based on the word frequency only. It is a
Networks (CNN), specially designed for image classification,
contextual approach. The distribution representation of the
are also helpful in text classification. However, it is specific
context of a specific corpus is vectorized in this model [35].
to computer vision, and NLP-based solutions [24]. Most of
the job descriptions come in pure text form. As a result, the
4) Autoencoder: Autoencoder is a DNN used to
CNN for this task can be ignored [25]. The Recurrent Neural
Network (RNN) architecture uses the texts as a sequence of encode unlabelled data. Once encoded, the data can be
words. It maintains the sequence [26]. That is why the regenerated from the code. It is more famous for
semantics and parts of speech relations are preserved [27]. As dimensionality reduction. However, also used for word
a result, the text classification exhibits human intelligence embedding. The autoencoder consists of two components -
when RNN is used. the decoded message x and the encoded message 𝑥 ′ . These
Job circular and application process involves both real- are considered in the Euclidean space where 𝑥 = ℝ𝑚 and
time and offline processing [28]. The RNN requires 𝑥 = ℝ𝑛 where m and n are the Euclidean coordinates. Both
substantial computing resources if the input sequence encoding and decoding are done using the multilayer
becomes lengthier. Moreover, it becomes complicated when perceptron layer defined in equation 3 [36].
long-term dependencies are considered. It is a significant
drawback of RNNs in real-time applications [29]. The 𝐸ϕ (𝑥) = σ(𝑊𝑥 + 𝑏) (3)
solution to the limitations of RNN is Long Short-Term
Memory (LSTM) neural networks. The LSTM networks In this paper, the σ is the sigmoid activation function, W is
promise to process long and short sequences in real-time [30]. the weight matrix, and b is the bias.
However, there is still a gap in identifying the semantic
structure of the job description to relate it to the candidate 5) TF-IDF: The TF-IDF is the abbreviation for Term
profile. The solution to abridge this gap is one of the novel Frequency-Inverse Document Frequency. In the BoW
contributions of this paper. method, frequency plays a significant role in embedding. The
C. Word Embedding Analysis TF-IDF is an extension of the BoW method. In this
embedding, the words are organized according to their
The DL-based text classifiers require word vectors [31]. The
significance. The significance is controlled by the weighted
process of creating a wordvector is called word embedding.
It is a process of vectoring the word tokens to mathematically sum assigned to them during embedding. The equation 4
assign distance among different words so that different simi- defines the procedure [37].
larity measures can be applied [32]. There are different word
embedding methods. These methods have been discussed and 𝑇𝐹 − 𝐼𝐷𝐹 = 𝑇𝐹(𝑤, 𝑑) ∗ 𝐼𝐷𝐹(𝑤, 𝐷) (4)
analyzed below [33]:
1) Word2Vec: The Word2Vec algorithm is a neural Here, w means the frequency of the word, d is the
network dependent algorithm. It uses an existing corpus of document where it showed up, and D is the entire dataset.
text to discover the associations among words. The network Once the term frequency is calculated, the inverse of it is
convolved with it. The inverse is calculated using the IDF (w,
learns to distinguish the words by assigning unique numbers
D) function. The word embedding analysis shows that the
representing vectors. The vectors consist of amplitude and
existing effective methods construct the word vector directly
direction that preserve the semantic meaning and syntactic
using the tokens. It works well as a general solution and is not
structure. The cosine similarities among the word vectors are a practical solution for resume ranking. In this paper, instead
used to retrieve the semantic similarities and classify the of wordvector, hash-vectors have been used, which is another
words. The similarity is measured using equation 1 [34]. noble contribution of this experiment. The following list
highlights the contribution of this research:
𝑉𝑒𝑐𝑖 × 𝑉𝑒𝑐𝑗
𝑆(𝑖, 𝑗) = (1)
|𝑉𝑒𝑐𝑖 ||𝑉𝑒𝑐𝑗 | 1. Semantic structure-based resume recommendation
system development to handle the dynamic and
rapidly changing environment

10 International journal of Management, IT and Engineering

http://www.ijmra.us, Email: [email protected]
 ISSN: 2249-0558 Impact Factor: 7.119
document(D). The system tries to pair up the candidates’
2. Hash-vector dependent noble word embedding skills with the relevant skills of the job description and
technique for Deep Learning-based classifiers. expresses it in probability calculated by equation 6.
3. BiLSTM-based processing module design and
implementation exp(𝑠𝑖𝑔 × 𝑅(𝑄, 𝑑))
4. Achieving a remarkable performance of 89.93% 𝑃(𝑑|𝑄) = (6)
∑ 𝑑 ∈ 𝐷 exp(𝑠𝑖𝑔 × 𝑅(𝑄, 𝑑 ′ ))
accuracy, 88.11% precision, 89.17% recall, 6.79%
𝑅2 , 6.05% RMSE, and 7.31% MAE.
Here is equation 6, the sig represents the sigmoid function,
which is the activation function of the layers used in the
The rest of the paper has been organized into four different
proposed network. The skill of the candidate is d, the query
sections. The second section discusses the methodology. We
performed on the job description (D) is Q, and the P (d|Q) is
presented the experimental results and evaluation in the third
the probability to rank the candidate based on the skills. The
section. The paper has been concluded in the fourth section.
proposed system uses the log-loss function defined by
II. METHODOLOGY equation 7 to measure the loss during the learning process.
The introduction highlights the research gap. This research
𝐿(𝑖, 𝑗) = −𝑙𝑜𝑔 × α × 𝑃(𝑑 + 𝑄) (7)
has been conducted to abridge the gap and develop an
innovative and effective solution for resume ranking. The
Here in equation 7, the L(i, j) is the layer L of the multilayer
methodology starts with the hashing of the datasets. Then the
perceptron model located at index (i, j) and the α is the
hash vectors are sent to the BiLSTM network to train it. The
learning rate.
probability predicted by the network is used to rank the

resumes. The overview of the proposed methodology has

Figure 1: The overview of the proposed methodology

been illustrated in figure 1. Table 1: THE TERM FREQUENCY IDENTIFICATION BY THE

HM
A. System Structure
The proposed methodology processes the document and the TF Softw Engineeri wit Excelle Salar an Benefi
are ng h nt y d ts
query by a specially designed neural network that handles d1 1 1 0 0 0 0 0
them separately and later merges the signals together to make d2 1 1 1 2 1 1 1
the final decision. The proposed system is defined as S(q; Φ)
and the working principle is governed by equation 5. B. Module View of the System
The proposed system has been designed as a combination of
𝑆: 𝑞 ∈ 𝐷 → 𝑦 ∈ 𝑅𝐷 (5) three modules. They are the Hashing Module (HM), the
Processing Module (PM), and the Ranking Module (RM).
According to 5, the system queries(q) on the entire
document(D) and the output(y) establishes a relation(R) with 1) The Hashing Module (HM): The word embedding
analysis presented in the introduction of this paper shows that
converting the text into vectors using the existing state-of-art

11 International journal of Management, IT and Engineering

http://www.ijmra.us, Email: [email protected]
 ISSN: 2249-0558 Impact Factor: 7.119
2) methods is not a feasible solution to the resume Table 2: THE NETWORK ARCHITECTURE DETAILS
ranking challenge because of the dynamic nature of the job
description and short-term interaction of the new candidates Layer Type Output Shape Parameters
entering the job market at a rapid rate [38]. In this paper, an Embedding (None, 50, 16) 12000
innovative hashing method has been employed in the Hashing Bidirectional (None, 256) 194460
Module to process the input text sequence for the PM. The
Dropout (None, 256) 0
HM handles large datasets smoothly because of Word N-
gram scaling. Dense (None, 1) 564

Case Study of HM Application: In the case study, the d1 and The network architecture details have been listed in table II.
d2 are two English sentences. These sentences are The embedding layer, which receives the signal from the HM
experimental. These are not any real job descriptions. They module, processes the input signal with 12,000 learning
have been carefully chosen to explain the HM module’s parameters. The bidirectional layer contains 194,460
operations properly. parameters. And the dense layer consists of 564 parameters.
Together there are 207,024 learning parameters in the
• d1: Software Engineering Salary network. The proposed network uses a Log-Loss function to
• d2: Software Engineering with excellent salary and measure the loss during the learning process defined by
excellent benefits. equation 8 [41].

The HM identifies the term frequency first. It has been listed

in table I. Now the hashes are generated using Word N-gram
as follows:

• d1: software #engineering

• d2: software #engineering #with #excellent #salary
engineering #with #excellent #salary #and with
#excellent #salary #and #excellent excellent #salary
#and #excellent #benefits

The value of N in the N-gram dynamically changes within the

range of 1 to 5. If the lowest value is not applicable, then the
immediate next value is assigned to N using the algorithm 1.

Figure 2: The BiLSTM Network Architecture

𝑁
1
𝐿𝑜𝑠𝑠 = − ∑ 𝐻𝑀(𝑦𝑖 ) × 𝑙𝑜𝑔 (𝑝(𝐻𝑀(𝑦𝑖 )))
𝑁
𝑖=1
+ (1 − 𝐻𝑀(𝑦𝑖 )) × 𝑙𝑜𝑔(1 − 𝑝(𝑦𝑖 )) (8)

Here, HM (y) is the label of the training dataset assigned by

The word hashes are dynamically generated using the N the hashing module, and N is the total number of datasets. In
selected by the algorithm 1. this paper, we experimented with three different optimization
algorithms to train the BiLSTM network. First, we
3) The Processing Module (PM): The Processing experimented with the Adaptive Gradient Descent (Adagrad)
Module (PM) uses a Bidirectional Long-Short Term Memory optimizer. This algorithm updates the weight by equation 9
(BiLSTM) network [39]. The analysis presented in the [42].
Technology Selection section of the introduction justifies that δ𝐿
𝑊𝑡 = 𝑊𝑡−1 − η′𝑡 (9)
the LSTM network is well-suited for the resume ranking we δ𝑤(𝑡 − 1)
are developing in this paper. However, it has been observed Here in the equation 9, the 𝑊𝑡 is the updated weight, and
that the BiLSTM networks perform better than LSTM for a 𝑊𝑡−1 is the updated weight. The η′𝑡 is defined by equation 10.
η
large volume of data [40]. That is why the BiLSTM network η′𝑡 = (10)
illustrated in figure 2 has been used in research. √α𝑡 + ϵ
The second optimizer we experimented with is the Root
Mean Square Propagation (RMS Prop) optimizer.

12 International journal of Management, IT and Engineering

http://www.ijmra.us, Email: [email protected]
 ISSN: 2249-0558 Impact Factor: 7.119
𝑁

It calculates the average of squared gradients for each weight. 𝑅𝑠 = ∑ 𝑠[𝑖] (14)
After that, it divides the graduate by the square root of the 𝑖=1
mean square. First, the output from the network v(w, t) is Here is equation 14, the s is the skillset specified by the HR
calculated using the equation 11 [43]. hiring managers. The skill set of a particular candidate is a
subset of the Rs defined by the equation 15.
2
𝑣(𝑤, 𝑡) = γ𝑣(𝑤, 𝑡 − 1) + (1 − γ)(δ𝑄𝑖 (𝑤)) (11)
𝑠 ∈ 𝑅𝑠 (15)
After that, the v(w, t) is used to update weight using the
equation 12. The comparison between the s and the 𝑅𝑠 is made through the
η probability score obtained from the processing unit. The
𝑤𝑡 = 𝑤𝑡−1 − δ𝑄𝑖 (𝑤) (12) higher the probability, the higher the resume rank for a
√𝑣(𝑤, 𝑡) particular job circular.
After getting the experimental results from the Adagrad and
RMS-Prop algorithms, we experimented with the network
using the Adaptive Moment Estimation (ADAM) algorithm.
The adaptive momentum used to update weight is calculated
using the equation 13 [44].
δ𝐿
𝑚𝑡 = β1 𝑚𝑡−1 + (1 − β1 ) [ ]𝑣 (13)
δ𝑤𝑡 𝑡

Figure 3: The performance analysis of different word embedding approaches

Based on the experimental results, which have been presented

in the Experimental Results and Evaluation section, the
ADAM has been used in this research for better performance.

Table 3: THE PERFORMANCE EVALUATION ON WORD

EMBEDDING ALGORITHMS

4) The Ranking Module (RM): The ranking module

uses the prediction from the processing module to rank the
candidates' resumes. It involves the candidates' skillset and
III. EXPERIMENTAL RESULTS AND EVALUATION
Precision Recall
A. Evaluation Metrics
Model k=5 k=1 k=1 k=2 k=5 k=1 k=1 k=1
0 5 0 0 5 0 There are a variety of job descriptions in the
Word2Ve 28.1 33.4 38.1 43.2 26.7 31.3 36.3 41.1 experimenting datasets. Some descriptions are almost
c 7 2 7 4 4 5 identical, some are similar, and some are unique. Because of
BoW 14.1 21.7 28.0 35.0 18.4 25.7 32.4 39.2
2 4 8 1 8 1
the nonuniform nature of the dataset, an overall performance
CBOW 23.4 30.9 37.4 44.7 25.7 32.4 39.7 46.6 evaluated using the state-of-the-art regression evaluation
1 6 1 1 6 2 1 3 metrics R Squared (𝑅2 ), Root Mean Squared Error (RMSE),
Autoenco 25.6 28.1 31.6 34.1 26.7 29.1 32.4 35.5 and Mean Absolute Error (MAE) is not an appropriate
der 2 4 8 9 4 7
approach [45]. That is why we have taken a different
TF-IDF 32.1 34.1 36.4 38.0 34.6 36.0 36.7 40.7
4 1 4 1 4 8 9 approach to evaluating the performance of the proposed
Proposed 34.1 42.1 50.3 58.6 37.3 45.1 53.4 61.8 resume ranker.
3 6 4 7 4 6 7 8 We have already discussed the dynamic nature of the job
the required skillset mentioned in the job description. The market in the introduction. We created a similar vibrant
skillset necessary is known from the job description. The environment in the experimental setup by randomly selecting
candidates' skillset is unknown. The skillset needed in the job the top 15% of the data from the test dataset. Then we
description is expressed as equation 14. observed the top N -recommendations where N = 5, 10, 15,
20. We used precision and recall of the network on the test

13 International journal of Management, IT and Engineering

http://www.ijmra.us, Email: [email protected]
 ISSN: 2249-0558 Impact Factor: 7.119
dataset for different values on N to evaluate the effectiveness Figure 4 shows the improved precision and recall at different
of the network. values of N. No significant variations between precision and

Figure 5: (a)𝑅2 on 7000 instances, (b) RMSE on 500 instances, and (c) MAE on 500 instances

recall are observed. Both of these are equally improved after

B. Evaluation of Word Embedding
changing the range of N from the Word N-grams.
The hash-based word embedding is one of the noble
contributions of this paper. The first experiment has been C. The Network Performance Evaluation
conducted to determine this embedding scheme’s The network performance has been evaluated using R
effectiveness. The precision and recall for N = 5, 10, 15, 20 Squared (𝑅2 ), Root Mean Squared Error (RMSE), and Mean
have been listed in table III. The experimental results Absolute Error (MAE). 8000 instances from the test dataset
demonstrate that the proposed hash-based word embedding have been used for this evaluation. To plot the 𝑅2 , RMSE,
performs better than other popular word embedding schemes and MAE, relevance scores from 1 to 5 have been prepared
for the experimenting scenario. The comparison is more vivid based on the 2D spatial distribution of the prediction made by
in figure 3. Although the performance is better than other the network expressed in percentage. For ease of
embedding algorithms, it is still insufficient to rank the visualization, the 7000 instances for 𝑅2 , 500 instances for
resumes properly. It has been observed that changing the RMSE, and 500 instances for the MAE have been illustrated
range of N to 3 to 5 in the Word N-gram significantly in figure 5. The values have been calculated from the same
improves precision and accuracy. The improved results have number of instances mentioned earlier. The visualization
been listed in table IV. shows that the hypothesis fits properly for all three data
Table 4: THE IMPROVEMENT AFTER CHANGING THE RANGE distributions. That means the network is well-trained for
correct prediction.
Embedding N Precision Recall
Proposed 5 57.48 62.17 In this experiment, we used k-fold cross validation at k = 5.
Proposed 10 65.41 76.86 The values of the precision, recall, accuracy, 𝑅2 , RMSE, and
MAE have been listed in table V at different values of k. The
Proposed 15 78.1 82.35 average performance is listed in the table as well. The
Proposed 20 88.96 92.74 accuracy, precision, and recall at each fold are almost similar.
The 𝑅2 , RMSE, and MAE are also stable at every fold. That
Figure 4 shows the improved precision and recall at different means the network is not overfitting and good at
values of N . No significant variations between precision and generalization.
recall are observed. Both of these are equally improved after
Table 5: THE AVERAGE PERFORMANCE WITH 5-FOLD
changing the range of N from the Word N-grams. CROSS VALIDATION

K Accurac Precisio Recal R RMS MA

y n l Square E E
1 88.17 87.94 88.45 6.92 5.47 7.48
2 89.67 88.75 89.76 7.02 6.02 7.91
3 90.78 88.05 89.7 6.47 6.75 6.98
4 89.99 87.62 89.9 6.84 5.97 7.04
5 91.05 88.19 88.03 6.74 6.04 7.12
Averag 89.932 88.11 89.17 6.798 6.05 7.31
e

Figure 6 demonstrates the insignificant variations at the

different folds. The R2, RMSE, and MAE are significantly
lower than the accuracy, precision, and recall. That is why
Figure 4: The improved precision and recall there is a sudden change in the slope on 6.

14 International journal of Management, IT and Engineering

http://www.ijmra.us, Email: [email protected]
 ISSN: 2249-0558 Impact Factor: 7.119
career and team development efforts,” in Intelligent Computing, pp.
1149–1166, Springer, 2022.
The variations among the R2, RMSE, and MAE on the figure
[5] T. V. Yadalam, V. M. Gowda, V. S. Kumar, D. Girish, and M.
are ignoble. The experimental results and evaluation Namratha, “Career recommendation systems using content based
demonstrate the excellent performance of the proposed filtering,” in 2020 5th International Conference on Communication and
system in ranking the resumes against the job description Electronics Systems (ICCES), pp. 660–665, IEEE, 2020.
depending on the skill matching. The proposed system is [6] Y. Zheng and D. X. Wang, “A survey of recommender systems with
robust, dynamic, and correctly rank resumes with less than multi-objective optimization,” Neurocomputing, vol. 474, pp. 141–
153, 2022.
7% 𝑅2 error.
[7] Faruqui, N., Yousuf, M. A., Chakraborty, P., & Hossain, M. S. (2020).
Innovative automation algorithm in micro-multinational data-entry
industry. In Cyber Security and Computer Science: Second EAI
International Conference, ICONCS 2020, Dhaka, Bangladesh,
February 15-16, 2020, Proceedings 2 (pp. 680-692). Springer
International Publishing.
[8] A. Gunawardana, G. Shani, and S. Yogev, “Evaluating recommender
systems,” in Recommender systems handbook, pp. 547–601, Springer,
2022.
[9] S. Ahmadian, M. Ahmadian, and M. Jalili, “A deep learning based
trust-and tag-aware recommender system,” Neurocomputing, vol. 488,
pp. 557–571, 2022.
[10] Q. Shambour, “A deep learning based algorithm for multi-criteria
recommender systems,” Knowledge-Based Systems, vol. 211, p.
106545, 2021.
Figure 4: The performance visualization [11] A. M. Elkahky, Y. Song, and X. He, “A multi-view deep learning
approach for cross domain user modeling in recommendation
systems,” in Proceedings of the 24th international conference on world
IV. CONCLUSION wide web, pp. 278–288, 2015.
[12] L. Almonte, E. Guerra, I. Cantador, and J. De Lara, “Recommender
The innovative hashing scheme and BiLTSM-based dynamic systems in model-driven engineering,” Software and Systems
resume ranking system is an extension of the content-based- Modeling, vol. 21, no. 1, pp. 249–280, 2022.
filtering dependent job recommender system. The key to [13] H. Ko, S. Lee, Y. Park, and A. Choi, “A survey of recommendation
technological sustainability is to evolve with the change. The systems: recommendation models, techniques, and application fields,”
job market change rate from both requirements and Electronics, vol. 11, no. 1, p. 141, 2022.
operational perspectives is the highest that every human [14] A. Iana, M. Alam, and H. Paulheim, “A survey on knowledge-aware
news recommender systems,” Semantic Web Journal, 2022.
civilization has experienced. The system presented in this
[15] H. Wang, D. Amagata, T. Makeawa, T. Hara, N. Hao, K. Yonekawa,
paper has been designed to help HR professionals automate and M. Kurokawa, “A dnn-based cross-domain recommender system
their recruitment process by automatically ranking the for alleviating cold-start problem in e-commerce,” IEEE Open Journal
candidates' resumes. It solves the challenges imposed by the of the Industrial Electronics Society, vol. 1, pp. 194–206, 2020.
rate of rapid change of job requirements and the entrance of [16] M. Almgerbi, A. De Mauro, A. Kahlawi, and V. Poggioni, “A
systematic review of data analytics job requirements and online-
new candidates on Deep Learning based automation. This courses,” Journal of Computer Information Systems, vol. 62, no. 2, pp.
research has examined and presented the analysis of the 422–434, 2022.
current context and possible technology to develop a [17] C. Clegg and C. Spencer, “A circular and dynamic model of the process
dynamic, robust, automatic, and accurate resume ranking of job design,” Journal of Occupational and Organizational
system. Based on the analysis, an innovative hashing-based Psychology, vol. 80, no. 2, pp. 321–339, 2007.
word embedding scheme has been developed in a three- [18] S. Khatwani and M. Chandak, “Building personalized and non
personalized recommendation systems,” in 2016 International
module system. This BiLSTM-dependent system ranks the Conference on Automatic Control and Dynamic Optimization
resumes with 89.93% accuracy, 88.11% precision, 89.17% Techniques (ICACDOT), pp. 623–628, IEEE, 2016.
recall, 6.79% 𝑅2 , 6.05% RMSE, and 7.31% MAE. The [19] T. Gr ̈utzmacher, T. Cojean, G. Flegar, H. Anzt, and E. S. Quintana-
proposed system has been cross-validated and examined in Ort ı́ , “Acceleration of pagerank with customized precision based on
mantissa segmentation,” ACM Transactions on Parallel Computing
different ways, and the experimental results, along with (TOPC), vol. 7, no. 1, pp. 1–19, 2020.
thorough evaluation, demonstrate that it is a potential solution [20] N. Sano, N. Machino, K. Yada, and T. Suzuki, “Recommendation
to rank resumes in today's rapidly changing job market. system for grocery store considering data sparsity,” Procedia Computer
Science, vol. 60, pp. 1406–1413, 2015.
REFERENCES [21] A. H. Khan, J. Siddqui, and S. S. Sohail, “A survey of recommender
systems based on semi-supervised learning,” in International
[1] H. H. Larsen and C. Brewster, “Line management responsibility for Conference on Innovative Computing and Communications, pp. 319–
hrm: what is happening in europe?,” Employee relations, 2003. 327, Springer, 2022.
[2] A. Mohamed, W. Bagawathinathan, U. Iqbal, S. Shamrath, and A. [22] G. Bebis and M. Georgiopoulos, “Feed-forward neural networks,”
Jayakody, “Smart talents recruiter-resume ranking and IEEE Potentials, vol. 13, no. 4, pp. 27–31, 1994. [23] C.-F. Tsai, “Bag-
recommendation system,” in 2018 IEEE International Conference on of-words representation in image annotation: A review,” International
Information and Automation for Sustainability (ICIAfS), pp. 1–5, Scholarly Research Notices, vol. 2012, 2012.
IEEE, 2018. [23] R. Ying, R. He, K. Chen, P. Eksombatchai, W. L. Hamilton, and J.
[3] V. Yakubovich and D. Lup, “Stages of the recruitment process and the Leskovec, “Graph convolutional neural networks for web-scale
referrer’s performance effect,” Organization science, vol. 17, no. 6, pp. recommender systems,” in Proceedings of the 24th ACM SIGKDD
710–723, 2006. international conference on knowledge discovery & data mining, pp.
974–983, 2018.
[4] A. Vatousios and A. Happonen, “Transforming hr and improving talent
profiling with qualitative analysis digitalization on candidates for

15 International journal of Management, IT and Engineering

http://www.ijmra.us, Email: [email protected]
 ISSN: 2249-0558 Impact Factor: 7.119
[37] A. Aizawa, “An information-theoretic perspective of tf–idf measures,”
[24] I. Verboncu and L. Zeininger, “The manager and the managerial tools: Information Processing & Management, vol. 39, no. 1, pp. 45–65,
Job description,” Revista de Management Comparat International, vol. 2003.
16, no. 5, p. 603, 2015. [38] K. Erk, “Vector space models of word meaning and phrase meaning:
[25] L. R. Medsker and L. Jain, “Recurrent neural networks,” Design and A survey,” Language and Linguistics Compass, vol. 6, no. 10, pp. 635–
Applications, vol. 5, pp. 64–67, 2001. 653, 2012.
[26] Y. Tan, X. Wang, and T. Jia, “From syntactic structure to semantic [39] Achar, S.; Faruqui, N.; Whaiduzzaman, M.; Awajan, A.; Alazab, M.
relationship: Hypernym extraction from definitions by recurrent neural Cyber-Physical System Security Based on Human Activity
Recognition through IoT Cloud Computing. Electronics 2023, 12,
[27] networks using the part of speech information,” in International
1892. https://doi.org/10.3390/electronics12081892.
Semantic Web Conference, pp. 529–546, Springer, 2020.
[40] Paula, L. P. O., Faruqui, N., Mahmud, I., Whaiduzzaman, M.,
[28] J. Abendroth, A. Heiss, T. Jacobsen, S. R ̈ottger, and J. Kowalski, “Job
Hawkinson, E. C., & Trivedi, S. (2023). A Novel Front Door Security
mobility and job performance: beliefs about social and occupational
(FDS) Algorithm Using GoogleNet-BiLSTM Hybridization. IEEE
(dis) advantages as mediators,” International Journal of Occupational
Access, 11, 19122-19134..
Safety and Ergonomics, vol. 28, no. 1, pp. 199–212, 2022.
[41] Trivedi, S., Patel, N., & Faruqui, N. (2022, October). NDNN based U-
[29] L. Rasmy, M. Nigo, B. S. Kannadath, Z. Xie, B. Mao, K. Patel, Y. Net: An Innovative 3D Brain Tumor Segmentation Method. In 2022
Zhou, W. Zhang, A. Ross, H. Xu, et al., “Recurrent neural network IEEE 13th Annual Ubiquitous Computing, Electronics & Mobile
models (covrnn) for predicting outcomes of patients with covid-19 on Communication Conference (UEMCON) (pp. 0538-0546). IEEE.
admission to hospital: model development and validation using
electronic health record data,” The Lancet Digital Health, 2022. [42] Patel, N., Trivedi, S., & Faruqui, N. (2023, February). An Innovative
Deep Neural Network for Stress Classification in Workplace. In 2023
[30] H. Sak, A. Senior, and F. Beaufays, “Long short-term memory based International Conference on Smart Computing and Application
recurrent neural network architectures for large vocabulary speech (ICSCA) (pp. 1-5). IEEE.
recognition,” arXiv preprint arXiv:1402.1128, 2014.
[43] S. K. Turitsyn, T. Schafer, and V. K. Mezentsev, “Generalized root-
[31] S. Minaee, N. Kalchbrenner, E. Cambria, N. Nikzad, M. Chenaghlu, mean-square momentum method to describe chirped return-to-zero
and J. Gao, “Deep learning–based text classification: a comprehensive signal propagation in dispersion-managed fiber links,” IEEE Photonics
review,” ACM Computing Surveys (CSUR), vol. 54, no. 3, pp. 1–40, Technology Letters, vol. 11, no. 2, pp. 203–205, 1999.
2021.
[44] Trivedi, S., Patel, N., & Faruqui, N. (2023, January). Bacterial Strain
[32] F. Heimerl and M. Gleicher, “Interactive analysis of word vector Classification using Convolutional Neural Network for Automatic
embeddings,” in Computer Graphics Forum, vol. 37, pp. 253–265, Bacterial Disease Diagnosis. In 2023 13th International Conference on
Wiley Online Library, 2018. Cloud Computing, Data Science & Engineering (Confluence) (pp. 325-
[33] P. L. Rodriguez and A. Spirling, “Word embeddings: What works, 332). IEEE.
what doesn’t, and how to tell the difference for applied research,” The
[45] Patel, N., Trivedi, S., & Faruqui, N. (2023, February). A Novel
Journal of Politics, vol. 84, no. 1, pp. 101–115, 2022.
Sedentary Workforce Scheduling Optimization Algorithm using 2 nd
[34] K. W. Church, “Word2vec,” Natural Language Engineering, vol. 23, Order Polynomial Kernel. In 2023 International Conference on Smart
no. 1, pp. 155–162, 2017. Computing and Application (ICSCA) (pp. 1-7). IEEE.
[35] L. Shi, A. Somnali, and P. Guan, “Using continuous bag of words,”
2022.
[36] J. Su, S. Wu, B. Zhang, C. Wu, Y. Qin, and D. Xiong, “A neural
generative autoencoder for bilingual word embeddings,” Information
Sciences, vol. 424, pp. 287–300, 2018.

16 International journal of Management, IT and Engineering

http://www.ijmra.us, Email: [email protected]