Abstract
The efficiency and effect of enterprise economic management activities depend on whether the performance evaluation can be implemented and the evaluation results can effectively feedback and improve the enterprise knowledge management process. Allocate enterprise resources to create more value. The evaluation of enterprise economic management activities depends heavily on the weights of different evaluation indicators of enterprises. In order to solve this problem, this paper proposes an evaluation method that integrates entropy weight analytic hierarchy process and deep neural network algorithm. Through objective entropy weight and imitation, the deep neural network fusion of subjective judgment of experts combines the subjective and objective factors of the evaluation index, then uses the AHP method as the evaluation index of performance evaluation according to the index weight, and finally realizes the accurate evaluation of enterprise economic management.
1. Introduction
At present, China is in the critical period of the era of economic restructuring, development power shift, innovation, and R&D upgrading; in this context, high-tech enterprises represented by high-level talent-intensive, innovation-driven development have become an important force to carry economic development. In recent years, the high-tech industry with its realistic function of transforming development, promoting innovation and pulling the economy, has been widely concerned by the political, industrial, academic, and research communities and generally welcomed by the scientific and technological entrepreneurial groups [1]. High-tech industries, represented by electronic information technology, material and chemical technology, new medicine and biotechnology, aerospace technology, etc., have been emerging, the sales revenue and output value of high-tech enterprises have surged, and the number of enterprises has also shown a rapid growth trend. The high-tech industry has become an important driving force in the development of the national economy, from the initial industrial output value of 152.8 billion yuan to the present 20.5 trillion yuan, which has increased 134 times. 2021 high-tech enterprises accounted for 26.8% of China’s GDP, with a growth rate of 12.4%. This shows that high-tech enterprises have achieved significant development, which is important for promoting the process of new industrialization in China, building an innovative country and gaining an advantageous focus in the fierce international competition [2].
Enterprise economic management is the economic regulation and control of enterprises in specific operations, so as to reduce costs, improve efficiency, and achieve the purpose of enterprise development. Enterprise economic management is a comprehensive concept. It includes not only specific production and operation activities, but also the configuration optimization of finance and human resources. The main purpose of enterprise development economic management is to maximize benefits and obtain higher benefits at extremely low costs. In today’s market, in order to ensure a certain degree of innovation in the production of enterprises, many enterprises are improving quality and upgrading products to improve their core competitiveness. However, in order to achieve the above measures, good economic management is the basic premise, which must be achieved through new technologies. Enterprise innovation and transformation calmly face the challenges of the times.
Management master Peter Drucker once said “If you cannot measure it, you cannot manage it,” so performance evaluation is one of the core elements of internal business management, and in order to further improve the level of business performance evaluation our government has issued a number of relevant policies and established industry standards [3], for example, a series of policy specifications, such as “establishing 6 economic indicators to assess industrial enterprises,” “enterprise economic efficiency evaluation index system” (for trial implementation), and “adjusted 7 indicators to assess the economic efficiency of industrial enterprises”. At this stage, the financial performance indicators highlighting economic efficiency are the main focus. With the rapid development of the economy and the deepening of the research on enterprise business performance, it was gradually found that the influence of nonfinancial elements on enterprise business performance should not be underestimated. The early 20th century extended the evaluation of enterprise business performance to nonfinancial fields and carried out comprehensive evaluation of enterprise business performance [4]. Accordingly, the Ministry of Finance issued a series of policy specifications, such as the Operating Rules for Enterprise Business Performance Evaluation, the Interim Measures for Performance Evaluation of Financial State-Owned and State-Controlled Enterprises, and the Interim Measures for Management of Business Performance Evaluation of Central Enterprises [5]. It can be seen from the existing policy norms of enterprise business performance evaluation that most of them are based on the whole industry or popular traditional industries and do not involve high-tech enterprises. The above policy norms have some practical guidance but lack relevance for guiding the innovation-driven development and high-level talent-intensive high-tech enterprises. Therefore, it is imperative to create a set of business performance evaluation systems and evaluation models that fit the business development law of high-tech enterprises, being scientific, reasonable, and effective.
2. State of the Art
Enterprise economic management is a multicriterion decision-making problem, and scholars have put forward different evaluation methods for different application fields. These evaluation methods generally consist of two parts: weight determination method and evaluation method. At present, the methods of weight calculation are mainly divided into view empowerment method and objective empowerment method. Common subjective empowerment method includes Delphi method and hierarchical analysis method. Objective empowerment method mainly includes principal component analysis method and entropy weight method. In these common weight determination methods, the entropy weight method can handle more sample data, avoid the subjective interference of the evaluation index system, and retain more original information, and the results are more objective. Therefore, this paper uses the entropy weight hierarchy analysis method and the deep neural network algorithm combination to determine the evaluation index weight. Scholars have put forward different economic management methods for enterprises. Kan used the event research method when exploring the impact of the investment announcement on the enterprise market value and found that the industrial investment announcement can significantly improve the enterprise market value [6]. Sayjadah et al. used the DEA method to complete the evaluation content of Internet companies [7]. Zhou et al., respectively, constructed the data envelope model and the neural network model for the empirical analysis and evaluation of the innovative enterprise [8]. Zhao et al. completed the [9] research of enterprise green business performance evaluation based on BP neural network model. Among these common evaluation methods, the TOPSIS comprehensive evaluation method is not limited by the sample size and indicators and can rank the target units, with simple calculation and accurate results, which is suitable for the comprehensive evaluation in various situations. Therefore, this paper uses the TOPSIS method to evaluate the enterprise economic management. However, the traditional evaluation method only considers the difference of the evaluation indicators, without considering other information. Guo et al. [10] used the method of grey association to deeper study and analyze the relevant information of the index system and pointed out that the difference and variation information covered by the indicators is equally important. Therefore, this paper proposes a TOPSIS evaluation method considering both the difference and the changing trend of indices. Moreover, static time points cannot reflect the overall development state [11] of multiple multiattribute objects to be evaluated over the time period. The role of time factor cannot be ignored, and different time points and different near and far time point results have different effects on investment value and investment decision. In order to analyze the change of enterprise investment value over time and explore the development trend of related industries, the research perspective of dynamic timing is particularly necessary. Kairouz et al. [12] dynamically evaluate the hybrid information based on the double excitation model. Yang et al. [13] combined performance evaluation and prediction to construct DEA-TOPSIS time series as a 3-stage dynamic evaluation model and fitted a simple exponential smoothing model and ARIMA (Autoregressive Integrated Moving Average Model) to the evaluation results of TOPSIS method to explore the ranking of prediction performance evaluation. Nock et al. [14] took the temporal arithmetic average operator as the proxy variable of time factors and used the method of weighting to construct the dynamic evaluation of time variables.
3. Methodology
3.1. AHP Method
AHP method is a multiobjective decision analysis method combining qualitative and quantitative, which divides various influencing factors of a complex problem into interrelated orderly levels, quantitatively describes the importance of two elements of a level, and calculates the weights of the influencing factors through the judgment matrix to determine the relative importance of the factors in relation to the total objective [15]. The AHP method has the advantage of simplicity and flexibility and is more accurate in determining the relative weights of indicators than other methods. Nowadays, the AHP method has become a multicriteria method commonly used by decision-makers, is widely used in economic management, energy system analysis, forecasting, and project evaluation, and is usually used in combination with other methods such as fuzzy logic, neural network, and data mining for better results. In this paper, we mainly use the AHP method to filter the indicators for knowledge management evaluation [16].
3.2. Entropy Method
Through research, some scholars believe that the AHP hierarchical analysis method still has certain problems, such as being limited by experts’ experience and knowledge level, and at the same time it often has a strong subjective color when determining the weights, which in turn brings great uncertainty to the current evaluation results [17]. Therefore, how to improve the traditional AHP hierarchical analysis method is the focus of current thinking and research [7]. To solve this problem, many scholars have proposed the objective weighting method. That is, in the determination of the weights, the weights are determined from a more objective perspective to improve the accuracy of the weight calculation. And, among the objective assignment methods, the entropy weight method is a commonly used objective assignment method. This method is a comprehensive evaluation method used for multiple indicators and multiple objects based on the large amount of information provided by each quality assurance observation to determine the weight of the indicator in the overall set of factors. And, in this method, the evaluation results are mainly based on objective information, which can avoid human interference to the maximum extent [8].
In the process of evaluation, the entropy value is determined by the degree of variation of the values of the evaluation indicators [9]. Thus, in the process of evaluation, the weights are generally determined based on the degree of variation of each indicator combined with the entropy value. If the entropy value of the index is larger, then the degree of variation is less obvious and the weight of the index is larger; if the entropy value is smaller, then the degree of variation of the index value is more obvious and the weight is smaller. It can be seen that the essence of the entropy method is to determine the weight according to the meaning of its own value and is not affected by the evaluation data; relatively speaking, the evaluation results are more objective, and this is seen as a dynamic method of assigning weights. Therefore, this paper chooses the entropy method, an objective assignment method, to determine the weights. Its specific steps are as follows.(1)The original matrix is dimensionless so that Xij belongs to the interval 0 to 1.(2)Calculate the weight Pij of the ith object indicator under the jth indicator.(3)Calculate the entropy Hiof the ith evaluation indicator:where k = 1/ln n. The purpose of adding a constant k to the equation is to ensure that each weight Pij of the ith indicator can be equal and meet Hi = 1. This indicator cannot provide any information and does not play any role in the comprehensive evaluation.
Define the entropy weight of the ith indicator.
3.3. Neural Network Theory
The topology of the BP neural network model consists of three layers: the input layer, the hidden layer, and the output layer, each of which is composed of a large number of simple neurons operating in parallel. The neurons in each layer of the network are fully interconnected, while there is no connection between neurons in the same layer. The topology of the BP neural network is shown in Figure 1. The main role of the implicit layer is to extract features and pass the extracted features to the output layer below, which is actually the process of “self-organization” of the weights between the input and implicit layers.
The learning of BP algorithm can be divided into two processes: forward propagation of input information and backward propagation of error information. In forward propagation, the input signal is passed backward from the input layer through the implicit layer to the output layer, and finally an output signal is generated in the output layer. The weights of the network are kept constant during the forward propagation of the signal, and the upper layer neurons only have an effect on the lower layer neurons. If the output does not meet the predesired requirements, it enters the backpropagation phase. In this process, the error is propagated forward by the output layer in the reverse direction, while the weights of the network are adjusted by feedback according to the error, and finally the error is distributed within the desired output range. The learning of the whole network ends when the error meets the requirements. These two propagation processes usually do not reach the goal at once and need to go through several iterative cycles.
The advantages of BP neural network, the most widely used ANN, are as follows.(1)Nonlinear mapping capability: after training the BP network, a large number of input-output mapping relations are stored in the network, so there is no need to know the function expression of the mapping relation before using it, and the nonlinear mapping from input space to output space can be realized as long as there are sufficient samples to train the network. Particularly, in the real engineering field, it has unparalleled advantage to use this method before the internal laws are mastered.(2)Generalization ability: after the nonlinear mapping relationship is stored in the weight threshold of the trained neural network, if fresh sample data is input to the network, the network can correctly map the relationship between input and output according to its own stored knowledge. The generalization ability of BP network makes it superior to other nonlinear system models in solving practical problems. General principle: in order to improve the generalization ability of the network, the simplest network that can reflect the implicit laws inside the system should be selected.(3)Fault tolerance: since the training process of the network is to present the input-output relationship from a large number of samples, individual noisy samples cannot have an impact on the weights and thresholds of the network.
4. Result Analysis and Discussion
4.1. Combined Empowerment Evaluation Model
Subjective weighting method (AHP) has more advantages than objective weighting method (entropy weight method) in determining the weight according to the intention of decision-makers, but its objectivity is relatively poor and its subjectivity is relatively strong; objective weighting method has objective advantages, but it cannot reflect the importance of decision-makers to different indicators, and there will be a certain weight and the opposite degree to the actual indicators.
In view of the advantages and disadvantages of subjective and objective weighting methods, we also strive to control the subjective randomness within a certain range to realize the neutrality in subjective and objective weighting. The indicators are weighted and fair, which realizes the internal unity of subjectivity and objectivity, and the evaluation results are true, scientific, and credible.
Combining the objective assignment method mentioned above, this paper proposes using the combination of subjective and objective assignment methods to determine the weights in order to improve the accuracy of weight calculation. The reason for this idea is that the traditional AHP method uses subjective judgment, and the judgment matrix and the final weights are also very subjective. The weights calculated by the entropy weighting method are objective, but they are not good for systematic analysis. Thus, we first use AHP to calculate the subjective weights, then use entropy weight method to calculate the objective weights, and finally use the combined weight formula to get the comprehensive weight results. The main block diagram of the algorithm is shown in Figure 2.
In the traditional combination weight calculation, the weights are calculated by using the average, and this calculation and weighting method also leads to errors in the calculation. In this regard, this paper uses the following method to calculate the weights.where Wj is the entropy weight value calculated by entropy weight method and is the weight calculated by AHP.
On the basis of the above-mentioned evaluation index framework of the enterprise human resource information management system, the indexes after normalization are firstly scored, then the weights of each index are calculated by hierarchical analysis method and entropy weight method, respectively, and finally the final index weights are calculated by using the subjective and objective combination weight calculation formula. On the basis of the weight determination, the fuzzy comprehensive evaluation model is used to evaluate the HR management system. The specific evaluation steps are as follows.
First, the index system of enterprise human resource management evaluation is screened according to the index screening principle.
Second, subjective weighting is assigned to HRM evaluation indexes by hierarchical analysis method.
Third, use the entropy value method to objectively assign the HRM evaluation indexes.
Fourthly, the weights are calculated by using the combination weighting calculation method.
Fifth, according to the above-mentioned Z = w-r comprehensive evaluation following formula, we get
4.2. Application of Combinatorial Assignment Method
In the analytic hierarchy process, although the pairwise comparison data can be converted by objective absolute data, it is generally given subjectively by domain experts, so AHP is generally a subjective weighting method. The entropy weight method is an objective weighting method, which uses the existing objective data to obtain the weight of each evaluation index (the lower the entropy of the evaluation index, the greater the weight).
In order to verify the feasibility of the above method, with reference to the evaluation index system developed by Heray, this paper uses Heray’s evaluation index system as the object of research validation, the relevant scientific research results, and company development of the Institute of Spatial Information from 2012 to 2018 as sample data to verify the above constructed model. Meanwhile, in order to compare the advantages of the algorithm constructed in this paper, it is compared with the neural network evaluation algorithm. The specific index evaluation system indicators are shown in Table 1:
By consulting experts in relevant fields from the Institute of Spatial Information, based on the relevant scientific research results, etc., from 2012 to 2018, and the development of the company as sample data, the importance scales according to the above-mentioned index system were constructed and the two-two judgment matrices from the first level index layer to the second level index layer were set.
We compare the four methods of obtaining weights to obtain the result graph in Figure 3.
In Figure 3, the horizontal coordinates indicate the corresponding 20 secondary indicators, and the vertical coordinates indicate the weight values of each indicator. We connect the weight values of the 20 secondary indicators in turn with a dash, and we can see that the AHP method and the entropy weight method are two extremes in the treatment of this problem. For the AHP method, its advantage is that it verifies whether the indicators conflict with each other and provides a consistency test. But, for the problem of this paper, the variability between the weights obtained by the AHP method is too large, which means that there is a strong reliance on a priori knowledge, and then after giving the experts’ scores the idea of decomposition, comparison, discrimination, and summation is followed to obtain the weights. The entropy weighting method, on the contrary to the AHP method, is an objective assignment method which is based on the amount of information provided by each observation to determine the weight of the indicator in the overall factor. We can see that the curve of the obtained entropy weights tends to be flat, unlike the great variability of the weights obtained by the AHP method. These are the shortcomings of single weights, which can be overcome by the combined weighting method. Figure 3 shows that both the method in the literature and the method in this paper are based on the combined assignment method and the weights that they obtain are between those obtained by the AHP method and the entropy weight method. In contrast, the one calculated by the reference is only the average of the two weights. The weights obtained by the combined assignment calculation method used in this paper are more accurate and more inclined to the real indicator weights. This shows that the combined assignment method proposed in this paper is feasible.
4.3. BP Neural Network Evaluation
There are many influencing factors in economic management evaluation and the mapping relationships between them are complex, so evaluating performance is a rather difficult task. How to draw on the experience and intuitive thinking patterns of evaluation experts while reducing the subjective factors of individuals in the evaluation process and making the evaluation method normative is the crux of accurate evaluation.
Using neural network for enterprise’s knowledge management performance evaluation actually simulates the way of human brain to process information and realize the parallel processing and nonlinear transformation of information. First, the input and output of the network are determined, which represent the characteristic information of the object to be evaluated and the final evaluation target of the network; then, a sufficient number of known samples are selected to train the network, and the input and output are one-to-one mapping relationships, at which the weights and thresholds obtained through adaptive learning can correctly represent the internal structure of the network; the trained network can simulate the knowledge management performance evaluation because it has the memory capacity, adaptive capacity, and recognition capacity similar to the human brain. The trained network can simulate the relationship between knowledge management evaluation indexes and performance because it has the memory ability, adaptive ability, and recognition ability similar to the human brain, and then it can be tested by testing samples until satisfactory results are obtained. At this time, the network becomes an effective prediction tool, and the evaluation results of an unknown sample can be obtained by inputting that sample. This network not only simulates expert evaluation, but also excludes personal subjective factors in evaluation.
The evaluation indexes of knowledge management are input from the input layer and passed to the final output layer through the intermediate implicit layer, and the obtained output results are compared with the desired output data, and if the accuracy requirements are not met, then the error backpropagation stage is entered. In this process, the weight threshold of the network is modified according to the output and other data, so that the network error gradually decreases and finally meets the requirements. With this error backpropagation correction, the correct rate of the neural network for input vector mapping gradually increases.
In this paper, 20 enterprises are selected for knowledge management performance evaluation research, and the results of expert evaluation are shown in Table 2, and the input data through normalization process are shown in Table 3.
The BP neural network algorithm is a search algorithm based on the gradient method, which fully reflects the parallel processing characteristics of neural networks in its implementation and is divided into two stages: forward transmission of input and backward propagation of error. The input data is processed by the neural units in the implicit layer and the output layer to obtain an evaluation result, which is also the output result of the network. When the difference between it and the desired result is outside the error range, the error signal starts to backpropagate, and the weight threshold of each neuron is continuously adjusted, so that the actual output is as close to the desired value as possible after repeated iterative process, which means that the average value of the squared error of the network reaches the desired minimum value, and the network is constructed. The learning process of the network is completed.
The BP neural network algorithm is implemented as follows.(1)The weights and thresholds of the network are assigned to small random numbers between [−1, 1].(2)Input a learning sample Xk and calculate the output value of each node in the hidden layer. Calculate the output value of the output layer node.(3)Calculate the connection weight corrections between the output layer nodes and the implicit layer nodes:(4)Calculate the connection weight corrections between the implicit layer nodes and the input layer nodes:(5)Correct the weights and thresholds of the network with error correction quantities:(6)Find the value of the error function E and determine whether E is less than the desired error accuracy. If the error requirement is satisfied, the algorithm ends; otherwise, the network returns to step (2) to continue training until the requirement is satisfied where ok is the desired output. Here, the weights are corrected in a batch approach, where the total error is calculated after all samples are input and the weights are adjusted according to the error. The batch approach allows E to move along the error reduction direction and converges faster when dealing with problems with more samples.
4.4. BP Neural Network Model
The evaluation of enterprise performance mainly adopts linear and nonlinear evaluation models. The nonlinear evaluation model can better simulate economic phenomena, and the evaluation results are objective and accurate, which has more practical reference value. From the perspective of assessing input-output efficiency, this paper selects the basic financial indicators to form the evaluation system. On this basis, this paper establishes the performance evaluation model of high-tech enterprises based on error backpropagation artificial neural network (BP neural network).
If the number of nodes in the hidden layer is too large, the network will be too complex and the learning time will be too long; if the number of nodes in the hidden layer is too small, the network will be less fault tolerant and the ability of identifying untrained samples will be lower. The trial-and-error method is a common method to determine the optimal number of nodes in the hidden layer. First, the network can be trained with a small number of nodes, and if the error does not meet the desired requirements, the number of nodes can be increased gradually based on the same sample set, and the number of nodes can be repeated until the network error is minimized. In this paper, the results are shown in Tables 4-5 after several iterations of testing based on the trial-and-error method. By comparing the training times and errors in the table, it is found that when the number of cells in the hidden layer is 11, the convergence speed and stability of the network are optimal, so the number of nodes in the hidden layer is set to 11.
The error performance index of network convergence is set as MSE, and the learning accuracy of the network is 0.0001.
Maximum training step: the maximum training step is set to 10000, and the training process of the training set can be stopped when the training step of the network reaches 10000.
Determine the momentum factor : the momentum factor can avoid the network training into local minima; according to experience, usually the value of the momentum factor is 0.95.
In the selection of transfer function, according to the general principle of BP neural network design, the S-type tangent function tansig is chosen as the transfer function of neurons in the hidden layer, and the S-type logarithmic function logsig is chosen as the transfer function of neurons in the output layer; as mentioned above, its output value is contained between [0, 1], which just meets the requirements of the network for the output.
Learning samples: if the network structure and training algorithm are determined, then the mapping and generalization ability of the network will eventually be determined by the training samples of the network. Generally, the samples are divided into training samples and testing samples, the training samples are to get a suitable internal representation, and the testing samples are to test the trained network to see if it can meet the desired requirements, generally 80% and 20% of the total number of samples each. In this paper, the knowledge management performance evaluation indexes of 20 enterprises are selected as input data, of which 16 groups are used as training sets to train the neural network and 4 groups are used as testing sets in order to improve the generalization ability of the network.
By training the network with learning samples, the adjustment of the network weights is achieved so that the output error of the network is minimized. Learning rate is too large or too small to have this adverse effect on the network; in general, the learning rate of the network takes a range between [0.001, 0.9], and the learning rate of this paper is chosen as lr = 0.05.
4.5. Simulation Analysis of BP Network Model
The network is trained according to the standard BP network algorithm. The horizontal coordinate represents the number of iterations and the vertical coordinate represents the training error of the network. From the figure, it can be seen that the training step size has reached the predetermined maximum step size of 10000, but the error has not converged to the expected minimum error, indicating that the network converges very slowly using the standard BP algorithm and the network falls into a local minimum, which leads to the failure of the network training. The correlation coefficient is an index indicating the linear correlation of the two variables; when R is closer to 1, this represents the closer relationship between the two variables. From the figure, it can be seen that the correlation coefficient of the standard BP algorithm, R = 0.52797, has a poor correlation between the predicted and expected values.
The above phenomenon occurs because there are certain flaws in the standard BP algorithm itself. One of the reasons for the long iteration steps and low convergence accuracy is that the learning rate is difficult to choose, too large rate will trigger oscillations, and too small rate will lead to slow convergence.
4.6. Simulation Analysis of the Evaluation Model of the Momentum-Adaptive Algorithm
The network is trained using the momentum-adaptive learning rate algorithm, and the results of 7 training sessions are shown in Table 5.
From the results of the seven experiments in the above table, mean absolute and relative errors can be seen, but the step size is large. In a comprehensive comparison, the overall results of each error of the 1st training are relatively the best. Therefore, the BP neural network of the 1st training was used for this enterprise knowledge management performance evaluation. The training error performance curve of the network is shown in Figure 4, and it can be seen that the network converges after 438 times of learning. Four sets of test data are input to test the trained network, and the test results of the network are shown in Figure 5. The solid line indicates the desired output result, and the dashed line indicates the result of the network training output; and, compared with Figure 5, it can be seen that the test error is small and basically meets the requirements. Figure 6 shows the regression analysis of the desired results of the improved momentum-adaptive algorithm and the actual training results. Compared with Figure 5, it can be seen that the correlation coefficient of the improved momentum-adaptive BP algorithm is R = 0.98747, the correlation between the predicted and desired values is large, and the regression performance of the network is better, thus indicating that the network has better generalization performance.
5. Conclusion
Enterprises can only manage the economy, and economic management performance evaluation can realize reasonable allocation of enterprise resources by evaluating the current situation of knowledge management and finding out the deficiencies therein, so as to maximize the profits of enterprises. The primary problem of implementing economic management performance evaluation is the construction of index system, and a scientific and reasonable index system can produce scientific and reasonable evaluation results.
The evaluation of enterprise performance mainly adopts linear and nonlinear evaluation models. The nonlinear evaluation model can better simulate economic phenomena, and the evaluation results are objective and accurate, which has more practical reference value. From the perspective of assessing input-output efficiency, this paper selects the basic financial indicators to form the evaluation system. On this basis, this paper establishes the performance evaluation model of high-tech enterprises based on error backpropagation artificial neural network (BP neural network).
The artificial neural network method used in this paper is a nonlinear mapping method. Unlike other evaluation methods such as analytic hierarchy process (AHP), fuzzy comprehensive evaluation method, and grey clustering method, which have obvious subjective judgment, it only needs to input the processed data into the network and produce the results through calculation. It does not need to determine the weight artificially, which does reduce the human factors in the evaluation process and improve the reliability of the evaluation. It makes the evaluation results more effective and objective, but it also has some disadvantages as follows:(1)BP neural network model requires certain learning samples. The number and quality of learning samples affect the learning performance of neural network model to a great extent. However, it is not easy to select appropriate learning samples.(2)The selection of the number of layers and hidden neurons of the network affects the learning ability and learning efficiency of the whole network to a great extent. However, at present, there is no certain guiding principle on this problem. When determining the number of layers and hidden neurons, there are often human factors, which will inevitably reduce the performance of BP network.
In this paper, a two-level evaluation index system is designed by reading relevant literature and, according to enterprise practice, eight important indicators are selected as evaluation indicators by applying hierarchical analysis. Then, the BP neural network method is used to evaluate the knowledge management performance of enterprises. Firstly, the input and output vectors and the transfer function are determined according to the research problem of this paper and the filtered evaluation indexes. After the model parameters are selected, we first use the standard BP algorithm to train the network, but the training effect is not very good, and sometimes even the training fails, which is caused by the inherent defects in the standard BP neural network algorithm, i.e., the existence of local minima, slow convergence speed, and poor generalization ability. To address these problems, the momentum-adaptive learning rate algorithm, LM algorithm, and genetic neural network algorithm are proposed to train the network. The momentum-adaptive learning rate algorithm integrates the advantages of both adaptive learning rate and additional momentum term algorithm to improve the convergence performance of the network, suppress falling into local minima, and increase the convergence speed more; the genetic neural network algorithm uses genetic methods to search for the optimal weight threshold. The convergence accuracy and convergence speed are also improved. The example verification shows that the momentum-adaptive algorithm has the best performance and the BP network model can be used to effectively evaluate the economic management of the enterprise, so as to guide the implementation of economic management activities.
Data Availability
The labeled data set used to support the findings of this study is available from the corresponding author upon request.
Conflicts of Interest
The authors declare that there are no conflicts of interest.
Acknowledgments
This work was supported by the Hebei Vocational University.