Risk prediction of enterprise human resource management based on deep learning

1.1Related work

Human resources are the carrier of enterprise knowledge and skills, and also the core resource of the enterprise. Currently, many researchers have agreed that human resources are the cornerstone of enterprise development, and HRM also occupies an important position in enterprise management. Therefore, many scholars have put forward their insights and opinions on HRM in enterprises. Berhil et al. reviewed and analyzed the research results related to the application of artificial intelligence in HRM in enterprises, so as to discuss the application path, application effect and development prospect of artificial intelligence technology in HRM. Their conclusions provide digitalization, intelligence and informatization for HRM Theoretical guidance [6]. Carnevale et al. explored the impact of job adjustment and company welfare on HRM during the new crown, which played an enlightening role for HR stability during the epidemic. It also had implications for the development of enterprises during the new crown epidemic [7]. Yong et al. compiled, counted and analyzed the relevant literature from 2007 to 2019. They discussed the current situation, problems and development prospects of green HRM, and pointed out the direction for the research in the field of green HR [8]. Pham et al. conducted a comprehensive analysis and review of the current problems, development status and challenges encountered in green HRM. They also proposed corresponding strategies to solve the problems, and finally put forward their own views on the future development of green HR. Cross et al. explored the role of HRM in independent work and made a challenging hypothesis, which was verified based on empirical analysis [10]. The importance of HRM work was verified by analyzing and discussing the practical significance of HRM work in a study [11]. Putra et al. provided a comprehensive compilation and elaboration of the literature related to HRM, thus providing a theoretical reference for decision making on business management behavior and providing a skill enhancement and communication skills for HRM work [12]. Ansari et al. studied and explored the potential correlation and influence mechanism between green HRM efforts and the environmental behavior of corporate employees, thus providing new ideas that contribute to the development of green HRM efforts [13].

Deep learning technology is rapidly developing and is having a profound impact on automation, intelligence, and digital transformation in various fields, which is why they are attracting the attention of researchers in various fields. BPNN is an important part of deep learning technology, and its simple model structure, strong classification performance and low algorithm complexity make it an important application in various fields. Han et al. used Genetic Algorithm (GA) to optimize the structure of BPNN and achieve intelligent modeling design of products with GA-BPNN model to improve the efficiency of product design [14]. Li et al. proposed to optimize the current motion management system by using BPNN to improve the management efficiency [15]. Chen et al. constructed a research performance management evaluation model based on the current situation of scientific research and related research. They used BPNN to achieve intelligent evaluation of scientific research performance, showcase existing problems in scientific research, and improve the efficiency of scientific research [16]. Li et al. proposed an improved BPNN model and constructed a smart city construction information model based on this model, providing a new path for urban smart management work and promoting the development of smart cities [17]. Wu et al. used GA to improve the optimization performance of Simulated annealing (SA) algorithm, and then obtained the best parameters of BPNN model based on GA-SA to improve the BPNN model. Finally, a prediction model was constructed based on GA-SA-BPNN to achieve coal and gas prominence prediction [18]. Tang et al. used the BPNN model to achieve intelligent segmentation of retinal vessels in color fundus images and validated the performance of the model. The results showed that the segmentation effect of the model was better than several existing methods [19]. Zhang et al. used BPNN to construct models and achieve intelligent classification and prediction of stock price patterns. Thus providing data support and reference opinions for investors’ decision-making, and contributing to the development of the stock market [20]. Jiang proposed a BPNN-based intelligent estimation method for construction project cost. The method can achieve efficient and accurate estimation of construction project cost based on the existing data, thus providing data reference for the preliminary capital investment of construction projects [21].

In the above content, it can be seen that the current HRM has gradually integrated with artificial intelligence technology to achieve intelligent and digital management. However, at present, there are still few studies on HRM risk intelligent prediction. To make up for this gap, based on the development of deep learning technology in the fields of sports management and investment decision-making, the study combines AI technology with HRM risk intelligent prediction, and proposes a HRM risk intelligent prediction model based on BPNN. The current research results confirm that the traditional BPNN model has many defects. To this end, an ISOA is proposed to improve BPNN. Based on the above content, in order to achieve intelligent and high-precision prediction of HRM risks, and provide data reference for enterprise human resource management. This has positive significance for the stable operation and development of enterprises.

2.1HRM risk evaluation index system construction

HRM refers to the effective use and allocation of human resources, so as to efficiently achieve corporate goals and employees’ self-development. However, in HRM, if the interaction between various factors is not properly handled, it will lead to the occurrence of HRM risks and affect the development of enterprises [2, 3]. Therefore, it is very necessary to achieve HRM risk prediction. The basis of HRM risk prediction is to build a reasonable HRM risk evaluation index system [7, 8]. When constructing a risk evaluation index system for HRM, the selection of indicators should follow the principles of purposefulness, scientificity, feasibility, and hierarchy to ensure the scientific and rational construction of the index system. Based on existing research literature [1, 6, 9], HRM risks are divided into five dimensions. B1: recruitment risk; B2: Training risks; B3: Performance management risks; B4: Compensation management risks; B5: Employee management risk. Based on these five dimensions, indicators are selected and a HRM risk evaluation index system is constructed, as shown in Table 1.

In Table 1, a review of the relevant literature yields a large number of indicators. For example, Berhil S and others mentioned that the lack or surplus of human resources, unreasonable structure, and imperfect development mechanism may lead to the difficulty of realizing the development strategy of enterprises. Unreasonable human resource incentive and constraint systems and imperfect management of personnel in key positions may lead to brain drain, low operational efficiency or leakage of key technologies, trade secrets and state secrets [6]. After sorting out the management risk factors in the literature and seeking expert advice, the preliminarily constructed HRM risk evaluation index system includes a large number of indicators. It has a total of 24 secondary indicators, and the relationship between indicator systems is relatively complex. In actual HRM, the degree of impact of different risk factors is also different. Spending a lot of time on the research of risk factors with less impact will not greatly improve the accuracy of HRM risk prediction. In addition, to realize the intelligent prediction of HRM risks, it is necessary to input the indicator data into the deep learning model. When the number of indicators is too large and the number of dimensions is too high, it will lead to an increase in the amount of calculation and model complexity of the deep learning model. It will affect the accuracy and efficiency of the model. Therefore, it is necessary to streamline the indicator system. The study combines analytic hierarchy method and expert scoring method to determine the weights of each index in Table 1, and selects the top 13 indicators with weight to construct an index system. The aim is to realize the simplification of the HRM risk evaluation index system, as shown in Table 2.

After streamlining, the number of indicators is reduced to 13, which is a significant reduction compared to the original 24. Based on the above, the construction of HRM risk evaluation index system is completed.

2.2PCA-based public factor extraction

This study constructed a risk evaluation index system for HRM and calculated the corresponding weights of each index. The data corresponding to each indicator is input to the BPNN model as an input sample, and sufficient training fitting is performed to achieve intelligent evaluation and prediction of HRM risk. However, after simplification, the number of indicators in the HRM risk assessment indicator system is still 13. This means that BPNN requires the construction of 13 input nodes, resulting in a complex network structure of the model that affects its risk prediction performance. Therefore, it also needs to further reduce the dimensionality of the indicators in the HRM risk evaluation index system. The study uses principal component analysis (PCA) to realize the dimensionality reduction of indicator data. Firstly, factor analysis was conducted on the HRM risk assessment indicator system shown in Table 2, and the most correlated factor was extracted as the common factor, as shown in Table 3.

In Table 3, a total of 4 factors have a combined value greater than 1, so these 4 factors are taken as public factors. The cumulative total variance contribution of the four common factors exceeds 85%. This indicates that the four common factors extracted have good variable explanatory power and can more intuitively and comprehensively reflect the HRM risk situation of the enterprise. By using descriptive statistical methods to construct a factor component matrix for extracting factors, the correlation between each common factor and each indicator is analyzed, and the corresponding indicators of the four common factors are obtained, as shown in Table 4.

In Table 4, four public factors are obtained corresponding to the indicators C₈: Unreasonable salary design, C₁: Unreasonable post setting, C₁₁: Risk of staff turnover, and C₆: Unfair performance appraisal. The data and weights corresponding to these four the data and weights corresponding to these four indicators are input to the BPNN as the input nodes of the BPNN. Thus, the BPNN model can be trained, and the output of the BPNN can be used as the prediction result of HRM risk.

2.3ISOA-based BPNN optimization

In the BPNN model, the initial settings of the network weights and thresholds directly affect the training effect of the BPNN, which in turn affects the accuracy of the model. The traditional method of setting BPNN parameters is to select the best performing set of parameters after several manual experiments. However, this method is inefficient and cannot guarantee that the parameters are the best values. Therefore, the study proposes to use SOA to find the best BPNN parameters, obtain the optimal parameter values, improve the performance of the BPNN model, and realize the optimization of the BPNN model. In SOA, seagull individuals have two behaviors in different stages, i.e., migratory behavior and spiral attack behavior. In the migration phase, the position of an individual gull is represented by Equation (1) after the individual has flown for a period of time according to certain rules.

In Equation (1), t indicates the current number of iterations of SOA; C_s (t + 1) is the new location where the individual will not hit other gulls after t + 1 iterations; P_s (t) is the location of the individual gull at the t iteration; α indicates the migratory behavior of the individual, which can be represented by Equation (2).

In Equation (2), f_c is a decreasing parameter that takes values in the range [0,2] and is used to control the frequency of migratory behavior; T means the maximum number of iterations of SOA. If the gull individual does not collide with other individuals at this time, the individual flies toward the current position where the optimal individual is located. Thus obtaining the new position M_s (t), expressed as Equation (3).

In Equation (3), P_bs (t) indicates the current optimal position; B is the random number introduced to regulate the global optimality seeking performance of the algorithm. When the direction is determined, the individual moves in that direction, as shown in Equation (4).

In Equation (4), D_s (t) indicates the distance between the optimal and the current position of the individual gull. The spiral attack behavior of the gull is actually the foraging process of the gull, and the behavior of the individual gull in the x plane during this phase is represented by Equation (5).

In Equation (5), r is the radius of the spiral; θ is a random number taking values in the range [0,2π]. The behavior of the individual in the y plane is represented by Equation (6).

The behavior of the individual in the z plane is represented by Equation (7).

The attack position of individual gulls is achieved by Equation (8).

In Equation (8), Ps′(t) is the position of the individual seagull after updating. SOA has the drawbacks of low accuracy and easy prematureness, resulting in the results obtained from its optimization search may not be the best parameter values for BPNN. Therefore, the study proposes strategies to optimize it. First, the bracketing strategy of Whale optimization algorithm (WOA) is introduced to optimize the SOA. The bracketing strategy of WOA is shown in Equations (9) and (10).

In Equations (9) and (10), D denotes the distance between the individual’s location and the optimal position at t iterations. X^* denotes the optimal position. C and A are two coefficient vectors. WOA has a strong global search capability, which can effectively compensate for the premature defect of SOA. In addition, the Levy flight mechanism optimizes the position update strategy of individual gulls when they perform prey search, as shown in Equations (11) and (12).

In Equations (11) and (12), D is the distance of an individual from a random individual in the population; d denotes the individual dimension; levy is a random vector following the distribution of levy. P_rand (t) is the position of a random individual in the population. Combining the above elements, the IPSO algorithm is constructed as shown in Fig. 1.

Fig. 1

Basic process of IPSO algorithm.

ISOA-BPNN is used to build a model to achieve intelligent prediction of HRM risks, improve the efficiency of HRM in enterprises, and ensure stable and sustainable development of enterprises. The ISOA-BPNN HRM risk intelligent prediction model is shown in Fig. 2.

Fig. 2

ISOA-BPNN HRM risk intelligent prediction model.