Hindawi

Scientific Programming
Volume 2021, Article ID 8387277, 10 pages
https://doi.org/10.1155/2021/8387277

Research Article
Research on Human Resource Recommendation Algorithm
Based on Machine Learning

1,2
Hong Zhu
1
Xi’an Jiaotong University, Xi’an, Shaanxi 710049, China
2
Shaanxi XueQian Normal University, Xi’an, Shaanxi 710100, China

Correspondence should be addressed to Hong Zhu; [email protected]

Received 21 June 2021; Revised 2 August 2021; Accepted 13 August 2021; Published 25 August 2021

Academic Editor: Muhammad Usman

Copyright © 2021 Hong Zhu. This is an open access article distributed under the Creative Commons Attribution License, which
permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
The economic environment has changed dramatically around the world in recent years, generating favorable conditions for the
growth of small- and medium-sized firms. The socioeconomic development and international integration of China are greatly
influenced by the growth in both quality and quantity, the scale of operations, and the internal force of small- and medium-sized
businesses. Moreover, in comparison with other developed countries around the world, Chinese small- and medium-sized
enterprises continue to face many limitations in terms of size and contribution levels and have not yet fully realized their potentials
due to difficulties and poor quality; human resources in this field are still lacking. This study defines the current state of human
resources in small and medium firms, the factors that impede development, and the steps that can be taken to overcome these
obstacles in order to assist human resource development in this sector during the current period. This study uses machine learning
(ML) techniques to manage and analyze human resource data in modern enterprises. The ML techniques realize the functions of
the human resource system and reduce the business volume in human resource in order to improve the efficiency and man-
agement of the human resource work. In this paper, we designed and implemented the wage forecasting model in human
resources that uses a gradient descent algorithm, its types, and backpropagation (BP) neural network to improve the accuracy of
the forecasting model. We performed multiple experiments by using a various number of neurons in the hidden layers, different
number of iterations, and several types of gradient descent algorithms. The BP neural network model was performed brilliantly by
attaining the training accuracy of 89.98% and validation accuracy of 84.05%. The experimental results show the significance and
importance of the proposed work.

1. Introduction the database technology and other requirements missing in

the first generation. The functions of nonfinancial infor-
Human resource management system (HRMS) is an im- mation, salary data analysis, and report production were
portant tool of enterprise management that plays an im- kept in consideration, which greatly improved the practi-
portant role in the development of enterprises. From its birth cability of HRMS in the enterprises. However, there were
to now, HRMS has experienced three stages [1]. The first still some major problems in the second generation of
generation of HRMS appeared in the 1960s. The main ob- HRMS; for example, the management process was not
jective and functionality of HRMS at that time was to au- standardized and the information processing techniques
tomatically calculate the salaries of employees via a were not comprehensive. The third generation of HRMS
computer system, excluding salary data analysis and non- appeared in the late 1990s. Because of the fierce competition
financial information. The main reason for such circum- among enterprises at that time, especially in the introduction
stances at that time was the limited resources and technical and cultivation of talent, HRM has risen to the top of the
requirements. The second generation of HRMS started in the company’s priority list. Due to the popularity of computers
early 1980s. The main drawbacks of the first generation were at that time, different technologies, particularly database
overcome in the second generation with the development of technology and Internet, had made great progress in the
2 Scientific Programming

HRMS. These technologies consolidate human resources for put forward using data mining techniques to extract and
unified management and create integrated data sources, analyze big data. Machine learning (ML) is one of the main
making data analysis and sharing easier. tools to realize data mining, and it has various applications
With the advent of the information age and the devel- in different fields, including the field of HRM and businesses.
opment of enterprises, a vast amount of data is produced The important roles of ML in human resource management
which needs to be managed and controlled in a well-or- include recruitment, talent management, preventing brain
ganized manner. To control, manage, and analyze such a drain, and improving human resource precise management
large amount of data is indeed a tough task for human beings and are analyzed in detail by different researchers [11, 12]. In
to implement at all, so the existence of information man- a study conducted by Chen et al. [13], they presented the
agement system (MIS) came into being. Management in- capability of data mining techniques in the development of
formation system (MIS) is management software having the the decision-making process of HRMS. The significance of
functionality of data collection, database creation, and data data mining in turnover management is analyzed and
management (including data storage and extraction) and implemented by Meng [14]. They presented a knowledge
can analyze a large amount of data [2–4]. The HRMS is an discovery approach to assist firms in predicting the future
information management system which combines both performance of their employees and then assigning relevant
HRM and information technology (IT). It is not only an personnel to the appropriate project positions.
information processing tool but also a resource management In this paper, a fuzzy clustering method based on an
standard. Its main purpose is to standardize the business adaptive genetic algorithm is proposed that provides effective
process of the human resource department through the data analysis for the performance appraisal of employees in
system, concentrate human resource information, and im- human resource management of modern enterprises. The
prove the transparency of HRM. HRMS plays an important potential of AI in HRM is explored in much detail in different
role in enterprises and has the powerful functions of opti- studies, among which six scenarios are turnover prediction of
mizing business processes, improving work efficiency, and the artificial neural network (ANN), candidate search based
improving management mode. Its quality is directly related on a knowledge search engine, personnel scheduling of ge-
to the performance of the enterprise, which affects the netic algorithm, HR emotion analysis, resume data collection
survival and development of the enterprise [5, 6]. Therefore, and extraction of information, and interactive voice response
the HRMS is considered an important tool for enterprises. employee self-service. Some of the basic contributions of this
However, in recent years, the data collected by enterprises paper are given as follows:
are growing rapidly and continuously. Such a large amount
of data has exceeded the processing capacity of HRMS, (1) This study designs and implements the model of
which leads to its failure to conduct data management and human position matching using ML algorithms,
data analysis normally. This situation is described as “rich which can provide a guarantee for effective human
data, but the information is poor.” Furthermore, the ma- resource allocation.
jority of the enterprise-level HRMS has not been fully uti- (2) A new method is proposed to evaluate the candidates
lized, with only data collection and storage, no effective data in an online recruitment system and to solve the
management and analysis, and no use of a large amount of problem of candidate ranking using ML algorithms.
data, wasting not only the enterprise’s limited resources but Further, a new algorithm based on support vector
also the opportunity to use big data to develop rapidly. Due regression is designed for the prediction of the
to the fact that the traditional HRMS cannot fulfill the needs human resource demand of enterprises which has a
of organizations’ while having big data, in order to make the good practical reference value.
system work normally, it must artificially screen and process (3) To solve the problem of staff turnover, an extreme
such a vast amount of data, thereby reducing the volume of gradient boosting (XGBoost) algorithm is used to
big data. However, this kind of processing method is back to predict the employee turnover rate. It is proved with
the old era of manual data processing, which seriously affects the help of experimental results that it is more ac-
the improvement and efficiency of the work. Therefore, curate than the other supervised classification al-
people started the use of data mining techniques to reduce gorithms in predicting employees’ turnover.
the data to be processed [7, 8]. The purpose of data mining is
(4) Based on the work and life situation, this study
to analyze and find out the important data patterns in the
analyzes the factors affecting the retirement of em-
massive amount of raw data and finally convert the original
ployees and uses ML algorithms to create a retire-
data into useful information and knowledge.
ment prediction model, which can provide solutions
The 21st century is the era of big data that brought a
for the outflow of human resources. Further, this
revolution in every aspect of human life such as healthcare,
study realizes a human resource ranking model,
industries, government sectors, social media, and business
which can predict the ranking and classification of
organization and changed the policies of human resource
resumes with high accuracy, and effectively sim-
management [9, 10]. The traditional HRMS can realize the
plifies the work of human resources.
functionality of data entry, query, and statistics but cannot
analyze the relationship between data samples effectively and The rest of the paper is organized as follows. Section 2
cannot predict the future development of enterprises represents the related work, while Section 3 illustrates the
according to the existing data. To solve this problem, people applications of the backpropagation neural network (BPNN)
Scientific Programming 3

in wage forecast. Section 4 demonstrates the experimental management systems that effectively steer human compe-
simulations and results analysis, and finally, Section 5 tencies toward corporate goals since then [23]. With the rise
concludes the research work. of the Internet as a professional platform and technological
advancements, human resource information is evolved into
2. Related Work a central hub for human resource management. Electronic
HR is the more generic term for the use of computer
The term “human resources” was initially coined in the early technology in HR [24]. E-HRM appears to have become the
1970s, when the importance of labor relations became more prevalent word for referring to virtual HR or e-enabled
widely recognized, and concepts like motivation, organi- HRM in the last two decades, in which IT is used to im-
zational behavior, and selection assessments began to plement comprehensive HR functions within a business
emerge. In “The Human Resources Glossary,” Huong [15] [25, 26]. In general, it is regarded to be useful since it is
describes human resource management as “the people and designed to improve HR efficiency, save costs, minimize
staff who administer an organization,” as opposed to an administrative responsibilities, simplify HR planning, and
organization’s financial and material resources. ML is a empower HR practitioners to become strategic or business
subject that studies how to improve the performance of the partners in firms. According to Marler and Parry [27],
system itself using calculation and experience. In general, it e-HRM may not be as cost-effective as it claims to be.
transforms the unordered data into useful information Organizations have been transitioning from traditional face-
through calculation, and this information generally refers to to-face HR procedures to current electronic ones; e-HRM is
the model obtained through training and learning. With the becoming an important theoretical and practical phenom-
use of computer systems for human resource tasks, HRM has enon [28, 29]. Many individuals expected e-HRM to revo-
experienced significant changes since the 1980s. Initially, lutionize the way HRM is done in organizations, moving it
human resource professionals used these systems primarily away from being purely administrative and toward being
as record-keeping systems for simple operations like more strategic. According to Thite et al. [30], the electronic
accepting and storing job applications on behalf of their transition faces two major hurdles in the perspectives of
employers [16]. However, as information technology (IT) business and technology leaders. One is HR’s ability to assist
advanced, HR professionals began to employ sophisticated business leaders in adopting a digital mindset, as well as a
computer systems to do the majority of their everyday tasks digital way of managing, organizing, and leading change.
and make strategic decisions. As a result, HR procedures The second is HR’s ability to transform the whole experience
have been greatly simplified [17]. IT is increasingly and of employee by transforming HR processes, systems, and the
widely used in HRM, and complex computer systems are HR organization via novel digital platforms, applications,
required for both internal and external operations such as and way to deliver HR services.
recruitment, performance assessment, remuneration, out- Information technology (IT) is on many people’s
sourcing, compensation, onboarding, development, strate- agendas, especially professionals and academics, as a crucial
gic analytics, and business-to-employees (B2E). The commercial tool for many firms. What is more intriguing is
advancement in computer engineering has resulted in the that there’s a global effort underway to improve IT’s qualities
emergence of advanced technologies like artificial intelli- using AI, often known as cognitive computing. According to
gence (AI). AI is also bringing up significant changes. HRM, Pan [31], since 2010, several IT behemoths have bought
a tech-dependent phenomenon, has recently seen such a approximately 140 AI start-ups. These businessmen are
shift. These significant shifts in HRM that we are witnessing attempting to combine artificial intelligence (AI) into a
are well worth investigating. The industrial revolution in the variety of computer systems in order to produce seamless
18th century caused a significant shift in industries and experiences during the installation of these technologies. AI
economics [18]. These significant shifts in HRM that we are strives to make machines capable of performing tasks in the
witnessing are well worth investigating [19]. Manufacturers same way that humans do. However, given the fact that
encountered the scientific management paradigm in the people have not yet completely deciphered and grasped the
early 20th century, which promoted standard-performance human intelligence that AI simulates, Shah and Warwick
production as the optimum approach to achieve maximum [32] believe that computers functioning like human beings
profitability [20]. Employers at the time were more con- should be viewed in terms of imitation. The Japanese
cerned with employee production rather than with employee Cabinet has announced a new project called Society 5.0 with
satisfaction, owing to the application of scientific manage- the goal of bringing about societal improvement. This new
ment techniques and the lack of governmental regulation approach appears to be aimed at achieving a high level of
and law governing personnel policies and practices. The human-machine interaction [33]. Its goal is to connect
widespread study of human behavior, as well as dramatic people, objects, and computers in cyberspace, using sensors
changes in the structure and intensity of worldwide com- to detect commands and AI to execute them [34]. Unlike the
petition between large organizations, had a significant im- idea of Business 4.0, which we are approaching, Society 5.0 is
pact on personnel management methods near the end of the not limited to the industrial industry but rather penetrates
20th century, and as a result, HRM emerged [21]. social life as AI, the Internet of Things (IoT), augmented
HRM has become the standard technique for managing reality, and robotics come together [35].
people inside a business since the early 1980s [22]. It has In this paper, a new HR recommendation algorithm
been highlighted as an important method for developing based on the hidden semantic model and the deep forest is
4 Scientific Programming

proposed in document, which can better recommend the neuron then becomes the input for another neuron and this
interesting posts for the users. Various studies made an process continues until it reached the last layer. Therefore,
attempt to understand that whether machine learning (ML) neurons can calculate the total input signal by multiplying
can be effectively used to perform potential assessments itself with the corresponding weights, then compare it with
using preventive methods, thus helping or even replacing the threshold value of neurons, and process it via the ac-
human resource managers. The ML methods are used in the tivation function. At last, the output of the neuron is ob-
earlier studies to provide human resource decision support. tained, as shown in Figure 1.
In this paper, a model of employee turnover based on a A neural network is a topological structure consists of
gradient descent algorithm and BP neural network is various neurons connected via a particular hierarchical
established in document, which can predict the turnover of structure. Neural network models can be classified into two
employees in enterprises and provide enterprises with the types based on the hierarchical structure of neurons: in-
opportunity to solve any problems and improve the re- terconnection network model and hierarchical network
tention rate. To solve the problem of human resource re- model. The neurons in the hierarchical network model are
placement in technical posts, this study uses supervised ML hierarchical and are generally divided into several layers
techniques to predict the appropriate human resources to fill according to their functions, and each layer is connected,
the vacancy of the technical posts. All of these indicate that such as the input layer, the middle layer, and the output
the influence of ML in the field of human resource man- layer. The interconnection network model is that in which
agement is expanding and has a great potential to improve any two neurons can be connected and has certain ran-
the efficiency and performance of HRMS. domness. The hierarchical network has become the most
widely used model structure because of its good structure
3. Applications of ML and Backpropagation and easy analysis.
Neural Network (BPNN) in Wage Forecast As mentioned earlier, the learning process of the neural
network is to adjust the parameters’ values of neurons, and
According to R. Michalski, research in the field of ML can be the backpropagation (BP) algorithm is used to achieve this
divided into “learning from examples,” “learning in prob- purpose in order to reach the target. BP algorithm is based
lem-solving and planning,” “learning through observation on gradient descent strategy, which takes the negative
and discovery,” and “learning from instructions.” According gradient direction of the target as the search direction to
to E. A. Feigenbaum, in the manual of AI, ML can be divided adjust the parameters and finally achieves convergence by
into “mechanical learning,” “teaching-learning,” “analogy continuously iterating and updating the parameters. It is one
learning,” and “inductive learning.” At present, inductive of the most successful neural network algorithms and has
learning is the most widely used learning technique of ML. been used widely in various applications. In general, the
There are three main streams of inductive learning: con- BPNN model refers to the multilayer feed-forward neural
nectionist learning based on neural networks, semiotic network trained by the BP algorithm, as shown in Figure 2.
learning, and statistician learning. Connectionism learning
based on neural networks is learning of the ML model that
simulates the functionality of the neurons of the human 3.1. Model Selection. Salary prediction is a kind of regression
brain, and its work in terms of ML is represented by the problem. The regression prediction models commonly used
perceptron, although connectionism learning started early in ML include linear regression, polynomial regression,
and encountered great obstacles. The main reason was that ridge regression, lasso regression, elastic network regression,
the neural network at that time could only deal with a simple and neural networks. Linear regression is a regression model
linear problem and even could not solve the small problem composed of linear variables. Therefore, the linear rela-
of XOR, which led to the development and learning of the tionship between the independent variables and dependent
neural networks. Semiotic learning was originated from variable must be satisfied; that is, linear regression is only
mathematical logic, and its core idea is to predict the results applicable to solve linearly separable problems. In linear
through the deduction and inverse deduction of symbols, regression problems, there exists a linear relationship be-
which mainly represents decision tree and logic-based tween variables; as a result, the calculation complexity is low
learning. Statistical learning is used to apply statistical and the training speed is faster. The disadvantage of linear
learning knowledge to ML techniques, and its representative regression is that it is very sensitive to outliers. If there is any
technology is SVM and “kernel method.” outlier in the data, the accuracy of the model will be affected.
Neural network is a parallel interconnected network Polynomial regression is the extension of linear regression,
composed of adaptive simple units, and its organization can which refers to the regression model with independent
simulate the interaction of biological neural system to real- variables, having an index greater than 1 in the regression
world objects. The “simple unit” in the definition of a neural equation. When the independent index is equal to 1, the
network refers to the neuron model, and the neuron model polynomial regression equation degenerates and becomes a
generally refers to the “MP neuron model or single-layer linear regression equation. As in most of the cases, inde-
perceptron model” proposed by McCulloch and others in pendent variables and dependent variables are not linear, so
1943 [14]. In the model, firstly, the neurons get input signals the polynomial regression is an automatic choice for it,
from the environment and then multiply them with the which can solve nonlinear and complex problems with ease.
corresponding weights put on the edges. The output of a The disadvantage of the polynomial regression model is that
Scientific Programming 5

X1
W1

W2 n
X2 f(WiXi) Y
i=1
W3

X3

Figure 1: Single-layer perceptron model.

Backward Phase
Y

1 Z
V Z1
X1 E1
1 T1
1

2 Z2
X2 2 E2 T2
2

Target Value

Zn
Xm n En T1
m
h

Input Layer Hidden Layer Output Layer

Forward Phase
Figure 2: General architecture of backpropagation neural network (BPNN) model.

the best index is difficult to determine, the high index is easy shortcomings too. The most important and commonly faced
to cause overfitting, and the low index is prone to problem in neural networks is the topological structure,
underfitting. parameters initialization, and optimization techniques of
Ridge regression is an optimization technique used when neural network. Further, it has a great impact on the training
the data features of linear regression or polynomial re- results of the model; if the selection of superparameters is
gression have multiple collinearity (i.e., there is an ap- not good, it will significantly affect the performance of the
proximately linear correlation between independent model.
variables). It reduces the influence of multiple collinearities
by adding a square bias factor to the regression estimation 3.2. Salary Forecast Model Based on BP Neural Network.
function (loss function). Lasso regression is similar to ridge The BP neural network offers nonlinear mapping, gener-
regression, and it is also an optimization technique used alization, and fault tolerance capabilities. Moreover, the
when the data features of linear regression or polynomial multilayer (three-layer or above) BP neural network can
regression have multiple collinearities and also add bias approximate any precision nonlinear function theoretically,
factors to regression estimation function. Only the square which is a powerful ML algorithm. To establish a salary
offset is added in ridge regression, and the absolute value forecast model based on BP neural network, it is necessary to
offset is added in the lasso regression. Elastic net regression solve the key problems such as topology, activation function
is a mixture of ridge regression and lasso regression. It adds selection, and parameter initialization.
both square bias and absolute value offset in regression The design of the network topology is relatively free and
estimation function, to achieve the common effect of ridge has no fixed form. The hierarchical structure is generally
regression and lasso regression. BP neural network is the designed according to the nature of the problem. In most
most widely used type of‘ neural network. Theoretically, a cases, the network parameters are also selected according to
three-layer BP neural network can approach any continuous the nature of the problem and earlier experience. Salary
function with arbitrary precision, has a strong learning prediction is a mapping of multiple input and a single
ability and simple network structure, and takes low time in output. According to the format of data in demand analysis,
calculation. At the same time, neural networks have some it can be determined that the number of input layer neurons
6 Scientific Programming

in the network is 14 and the number of neurons in the output oscillation phenomenon, but this oscillation also makes it
layer is 1. There is no definite calculation method for the possible to jump out of local optimal position and reach
number of neurons in the hidden layers. Some methods can global optimal position. The minibatch gradient descent
be used to estimate its approximate range, and then the (MBGD) is a compromise method to solve the problems
model can be trained repeatedly using the trial and error existing in BGD and SGD. It divides the whole sample
method. Finally, the number of neurons in a hidden layer dataset into small sample datasets and takes the cumulative
can be confirmed according to the performance of the error gradient of each small sample dataset as the estimation
model. There are many empirical formulas used to calculate value of the real gradient of the whole sample. When the
the number of neurons in the hidden layer. The most sample size is large enough, the local error gradient value is
commonly used formulas to calculate the number of neu- approximately equal to the total sample error gradient value.
rons in the hidden layer are given as follows: The small-batch gradient descent method can have both bad
√����� and SGD advantages.
n h � n i · no , (1)
√�����
4. Experimental Simulations and
nh � ni · no + k, (2) Result Analysis
where ni is the number of input layer nodes, no is the number This section of the paper represents the experimental results
of output layer nodes, k [1, 10]. According to the empirical carried out via different experiments and their analysis. All
formula (2), the range of the number of neurons in the the experiments were performed on a laptop computer
hidden layer is estimated to be [5, 14], then space is slightly system having the following specifications: Intel Core-i5, 7th
expanded to [3, 16] during the training, and the BP neural generation, processor of 2.4 GHz, 16 GB of RAM, and the
network is used to train repeatedly. operating system installed on the system being Microsoft
To sum up, the number of neurons in the input layer, Windows 10). Anaconda Jupyter Notebook is used as an IDE
hidden layer, and output layer is 14, 15, and 1, the activation for carrying out the simulations. Python has been used as a
function of neurons is the sigmoid function, the error language for the implementation and for generating the
function (objective function) is a quadratic mean square simulation results. The list of python packages used in this
function, Gaussian random variable with a mean value of 0 study was Keras, TensorFlow, seaborn, Numpy, Matplotlib,
and standard deviation of nin is selected as initialization and Sklearn, and so on.
weight, and Gaussian random variable with the mean value
of 0 and standard deviation of 1 is selected as initialization 4.1. Results of Salary Forecast Model Based on BP Neural
threshold. Network. This subsection demonstrates the simulation re-
sults carried out via the salary forecast model based on BP
cj � f′ 􏼐netj 􏼑 􏽘 δvjk � lj 􏼐1 − lj 􏼑 􏽘 δvjk , (3)
k k
neural network. All the simulation results were attained
using a different number of neurons and iterations as shown
where yj is the target value of the K nodes and zk is the in Tables 1 and 2. Table 1 shows the training and validation
output value of the K nodes. After the salary forecast model accuracies of the salary forecast model based on the BP
is established, the training can be started. The training neural network, using the data size of 1000 samples. The
flowchart of the model is represented in Figure 3. dataset samples were divided into training and testing
In the previous introduction, BP neural network adopts samples, the size of the training dataset was 700, the size of
the gradient descent method in the reverse propagation the testing dataset was 300, the number of neurons in each
update. According to the amount of data used in each hidden layer was trained 30 times, and finally, the average
update, the gradient descent method can be divided into value of the 30 times results was calculated. The execution
three forms: batch gradient descent method, random gra- process reached its termination when the number of iter-
dient descent method, and small-batch gradient descent ations reaches 250.
method. Each update is called batch gradient descent which From Table 1, it is quite obvious that the training and
uses all samples to update. Because batch gradient descent validation accuracies increase as the number of neurons in
needs to use all training samples at every iteration, when the the hidden layer is increased. Figure 4 shows the training and
number of training datasets is large, the time of one iteration validation accuracies of the salary forecast model based on
update will be longer, the cumulative error of all samples will the BP neural network.
decrease down to a certain extent, and eventually, the Figure 4 illustrates that the training and validation ac-
training process will be slow. As a result, the batch gradient curacies increase with the increase in the number of itera-
descent training process may fall into local optimal but tions. The highest training accuracy of 89.98 was attained at
cannot achieve global optimization. The method of updating the 250th iteration. The highest validation accuracy of 84.05
only a single sample is called stochastic gradient descent was also conquered at the 250th iteration.
(SGD). Because SGD only uses a single sample for each Table 2 shows the training and validation loss of the
iteration update and causes parameter updates very fre- salary forecast model based on the BP neural network. The
quently, the update between different sample data may offset training and validation loss of the BP neural network de-
each other, which makes each update not advance toward crease as the number of neurons in the hidden layer is
the overall optimization direction, and it is prone to increased.
Scientific Programming 7

start

Establishing model and initialization

parameters

Calculate the output of hidden layer and

output layer

Adjusting the parameters of the output

layer cell

Adjusting parameters of hidden layer

cells NO

Learn all the samples

YES

The number of
times to send
generation is Stop condition
reached or the satisfied
error is less than
the threshold
YES

END

Figure 3: Training flowchart of the wage prediction model.

Table 1: Training and testing accuracies of BP neural network We also performed multiple experiments using different
using a different number of neurons in the hidden layer. gradient descent techniques, that is, batch gradient descent
(BGD), stochastic gradient descent (SGD), and minibatch
Number of neurons Training accuracy Validation accuracy gradient descent (MBGD). To show the training effect of
3 86.66 80.04 BGD, SGD, and MBGD in an effective way, the loss function
4 87.04 80.20 curve of these techniques during training is shown in
5 87.30 80.44 Figure 5. Among them, the horizontal coordinate in Figure 5
6 87.44 80.52
is the iteration update times, the ordinate is the loss function
7 87.46 80.72
8 87.80 80.96
(cost), the training sample size is 200, and the small-batch
9 87.94 81.22 sample size is set to 20. So, BGD performed 200 cycles (read
10 88.00 81.44 200 rounds of dataset), MBGD performed 20 cycles, and
11 88.52 81.73 SGD performed a single cycle.
12 88.64 81.93 From Figure 6, it can be seen that when the iteration
13 89.00 82.17 update time is the same, the convergence results of the three
14 89.50 83.08 gradient descent methods are relatively close, but the loss
15 89.80 83.86 change curve is very different: SGD curve has a large
16 89.98 84.05 fluctuation, BGD curve is relatively smooth, and MBGD is
8 Scientific Programming

Table 2: Training and validation loss of BP neural network using a Comparision

different number of neurons in the hidden layer.
Number of neurons Training loss Validation loss 0.027
3 0.009975 0.010725
0.026
4 0.009958 0.010724
5 0.009950 0.010723
0.025
6 0.009954 0.010720

Cost
7 0.009947 0.010716 0.024
8 0.009944 0.010710
9 0.009940 0.010690 0.023
10 0.009938 0.010660
11 0.009932 0.010630 0.022
12 0.009928 0.010595
13 0.009926 0.010580 0 25 50 75 100 125 150 175 200
14 0.009920 0.010553 Iteration update times
15 0.009916 0.010510
16 0.009912 0.010502 SGD
BGD
MBGD
Figure 6: Comparison of SGD, BGD, and MBGD iterative update
Training and Validation accuracy
100 process.

90
80 between SGD and BGD and is relatively stable. The training
70
speed of SGD, BGD, and MBGD is different, and the training
period of SGD, BGD, and MBGD is different, which is
60 different from that of SGD, BGD, and MBGD. If the sample
50 space is large, the training speed of the investigated three
40 techniques will be more obvious, especially SGD and BGD.
In short, the training speed of SGD is fast and iteration
30
update is frequent, but there is a fluctuation that easily
20 produces oscillation. BGD is stable and iteration update is
10 less, but the training speed is slow. MBGD is a compromise
scheme, which has a fast training speed and is relatively
0
50 100 150 200 250 stable.
Iteration # The three-layer BP neural network model used for the
salary prediction consists of a different number of neurons in
Training accuracy
Validation accuracy different layers. The number of neurons in the input layer,
hidden layer, and output layer is 14, 15, and 1, the activation
Figure 4: Training and validation accuracies of BP neural network. function is the sigmoid function, the error function is
quadratic mean square error (QMSE), and the network
parameters are updated by using small-batch gradient de-
scent method (SBGD) and mixed optimization method
Nadam. The method used in this experiment divides the
0.090000 dataset into two parts, that is, training and testing. The total
sample space selected for the experiment was 1000, of which
0.070000 900 samples were selected for training while the rest of the
100 samples were selected for testing and validation purpose.
Model loss

0.050000 Figure 7 shows the adaptation process and comparison

of training effect of various optimization algorithms. In
0.030000 Figure 7, the ordinate is the loss function (cost) and the
abscissa is the period (iteration). As can be seen from
0.010000 Figure 7, Adam and Nadam have better performance than
other optimization algorithms in terms of convergence
0 speed, whereas Nadam is slightly better than Adam in terms
50 100 150 200 250
of overall performance.
Iteration #
To better reflect the performance of Nadam based salary
Training loss prediction model, the comparison results between Nadam
Validation loss
and other ML regression algorithms are given here. The size
Figure 5: Training and validation loss of BP neural network. of the training sample space is 900, and the size of the test
Scientific Programming 9

Comparision results in disparities in many facets of business operations

0.040 and trade, such as business thinking, observations, confi-
0.035
dence, client psychology, behavioral culture, and salary
0.030
prediction. As a result, for sustainable and adaptive devel-
opment in the new context, each small- and medium-sized
0.025
Cost

firm must effectively leverage supportive solutions from the

0.020
government, social media platforms, and small- and me-
0.015 dium-sized enterprises themselves. This study mainly fo-
0.010 cuses on the application and implementation of the BP
0.005 neural network-based salary forecasting model. Multiple
0 20 40 60 80 100 120 140 experiments have been performed by using a different
Epoch number of neurons in the hidden layers and a different
SGD RMSProp
number of iterations in order to check the stability and
Momentum Adam effectiveness of the proposed model. This study also uses
NAG Nadam the gradient descent algorithm and its types by per-
Adagrad forming multiple experiments for each of its types.
Figure 7: Comparison of convergence rates of various optimiza- Among the implemented types of gradient descent al-
tion methods. gorithm, Nadam showed good performance as compared
to its other types. BP neural network-based salary fore-
casting model performed excellently by attaining the
25000 training accuracy of 89.98% and validation accuracy of
84.05%. The proposed BP neural network-based salary
20000 forecasting model is anticipated to be of great help in
human resource management systems.
Salary

15000

10000
Data Availability
The data used to support the findings of this study are in-
5000
cluded within the article.

0 20 40 60 80 100
No of Samples
Conflicts of Interest
target The author declares that there are no conflicts of interest
output
regarding the publication of this study.
Figure 8: Fitting effect of Nadam based salary forecasting model.
References
sample space is 100. Each algorithm is tested 10 times, and the [1] Y. Lu, Research on the Application of Data Mining Technology
best test score is obtained. It can be observed that the results of in Human Resource Management Systems, Beijing University
the Nadam optimized salary prediction model in the test set is of Posts and telecommunications, Beijing, China, 2020.
relatively higher as compared to the other utilized algorithms. [2] L. Y. Wang and C. I. Peng, “Application of decision tree
Finally, the fitting effect of the Nadam salary prediction model algorithm in human resource recommendation technology,”
is calculated and is shown in Figure 8. Modern electronic technology, vol. 44, no. 3, pp. 105–110, 2021.
It can be seen from Figure 8 that the prediction fitting effect [3] T. Liu, M. Liu, H. Lihong, W. Luo, and W. Zheng, “Explo-
of the Nadam based salary prediction model is comparatively ration of human resource evaluation model based on ERP,”
Energy Technology, vol. 18, no. 9, pp. 90–93, 2020.
higher as compared to the rest of the techniques. Although the
[4] Q. Xia and F. Huang, “Data-driven human resource man-
normalization and antinormalization of salary data in model agement: review and prospect,” China personnel science,
training may lead to some errors in calculation, this error is still vol. 12, no. 6, pp. 72–87, 2020.
within the controllable range and has no great influence on the [5] J. Tang, Programmer Recommendation System Based on
final results of the model. Further, the salary prediction model GitHub, Yangzhou University, Yangzhou, China, 2020.
optimized by Nadam has a fast convergence speed and good [6] Y. Li, S. Li, Y. song, F. Zhang, and X. Zhou, “Human Resource
prediction effect and has good accuracy as well, so it has certain Recommendation Based on K-means clustering algorithm of
practical and reference value. spark platform,” Journal of Jinan University (Natural Science
& Medicine Edition), vol. 34, no. 5, pp. 430–435, 2020.
[7] L. Qi, Design and Implementation of a Personalized Recom-
5. Conclusion mendation System for enterprise Training Resources, Harbin
Institute of Technology, Harbin, China, 2018.
Globalization and worldwide economic integration have had [8] G. Liang, Human Resource Recommendation Algorithm Based
a considerable impact on China’s economic development, on Lifting Tree and Neural Network, South China University of
including the business community, in past few years. This Technology, Guangzhou, China, 2018.
10 Scientific Programming

[9] C. Lin, Human Resource Recommendation Algorithm Based on [29] S. Panos and V. Bellou, “Maximizing E-HRM outcomes: a
Hybrid Genetic Ensemble Learning, South China University of moderated mediation path,” Management Decision, vol. 54,
Technology, Guangzhou, China, 2018. 2016.
[10] Z. Gu, Human Resource Recommendation Algorithm Based on [30] M. Thite, e-HRM: Digital Approaches, Directions & Applica-
the Hidden Semantic Model and Deep forest, South China tions, Routledge, Oxford, England, 2018.
University of Technology, Guangzhou, China, 2018. [31] Y. Pan, “Heading toward artificial intelligence 2.0,” Engi-
[11] S.. Research, “On job recommendation of AI companies based neering, vol. 2, no. 4, pp. 409–413, 2016.
on data mining,” Value Engineering, vol. 36, no. 34, pp. 42–44, [32] H. Shah and K. Warwick, “Testing Turing’s five minutes,
2017. parallel-paired imitation game,” Kybernetes, vol. 39, 2010.
[12] Z. Li, Human Resource Hybrid Recommendation Algorithm [33] Y. Shiroishi, K. Uchiyama, and N. Suzuki, “Society 5.0: for
Based on Spark, South China University of Technology, human security and well-being,” Computer, vol. 51, no. 7,
Guangzhou, China, 2017. pp. 91–95, 2018.
[13] W. Chen, Research on Human Resource Recommendation [34] S. I. Team, “Realization OF society 5.0 BY utilizing precision
Algorithm Based on Deep Learning, South China University of agriculture into smart agriculture IN naro, Japan,” in Inter-
Technology, Guangzhou, China, 2017. national Workshop on ICTs for Precision Agriculture, p. 58,
[14] F. Meng, Application Research of Human Resource Recom- Wuhan, China, 2019.
mendation Engine Based on Hybrid Recommendation Algo- [35] P. Skobelev and S. Y. Borovik, “On the way from industry 4.0
rithm, Beijing University of Technology, Beijing, China, 2014. to industry 5.0: from digital manufacturing to digital society,”
[15] P. N. Q. Huong, “The quality OF human resources IN small Industry, vol. 2pp. 307–311, Suceava, Romania, 2017.
and medium enterprises: current situation, reason, recom-
mendations ON solutions,” Journal of Natural Remedies,
vol. 21, pp. 73–81, 2021.
[16] D. L. Stone and J. H. Dulebohn, Emerging Issues in Theory and
Research on Electronic Human Resource Management
(eHRM), Elsevier, Amsterdam, Netherlands, 2013.
[17] J. Heikkila, “A delphi study on e-HRM: future Directions,” in
Proceedings of the 3rd European Academic Workshop on
Electronic Human Resource Management, pp. 229–249,
Bamberg, German, May 2010.
[18] M. Berg, The Age of Manufactures, 1700-1820: Industry, In-
novation and Work in Britain, Routledge, Oxford, England,
2005.
[19] C. Trebilcock, The Industrialisation of the Continental Powers
1780-1914, Routledge, Oxford, England, 2014.
[20] F. W. Taylor, Scientific Management, Routledge, Oxford,
England, 2004.
[21] O. Yabanci, “From human resource management to intelli-
gent human resource management: a conceptual perspective,”
Human-Intelligent Systems Integration, vol. 1, no. 2-4,
pp. 101–109, 2019.
[22] A. Iqbal, “The strategic human resource management ap-
proaches and organisational performance: the mediating role
of creative climate,” Journal of Advances in Management
Research, vol. 8, 2019.
[23] R. Mathis and J. Jackson, Human Resource Management
Mason, Thomson South-Western, Mason, OH, USA, 2008.
[24] M. Armstrong and S. Taylor, Armstrong’s Handbook of Hu-
man Resource Management Practice, Kogan Page, London,
England, 2009.
[25] H. Francis and A. Keegan, “The changing face of HRM: in
search of balance,” Human Resource Management Journal,
vol. 16, no. 3, pp. 231–249, 2006.
[26] S. Strohmeier, “Research in e-HRM: review and implications,”
Human Resource Management Review, vol. 17, no. 1,
pp. 19–37, 2007.
[27] J. H. Marler and E. Parry, “Human resource management,
strategic involvement and e-HRM technology,” International
Journal of Human Resource Management, vol. 27, no. 19,
pp. 2233–2253, 2016.
[28] L. Ma and M. Ye, “The role of electronic human resource
management in contemporary human resource manage-
ment,” Open Journal of Social Sciences, vol. 03, no. 04,
pp. 71–78, 2015.