Producing and teaching fusion talent cultivation effect evaluation method based on index weight verification and discrete data convolution calculation

Abstract

The invention relates to the technical field of production and teaching fusion talent cultivation effect evaluation, and aims to solve the problem that intelligent evaluation cannot be realized due to lack of specific evaluation standards and intelligent evaluation methods in existing production and teaching fusion talent cultivation effect evaluation. The invention provides a production and education fusion talent cultivation effect evaluation method based on index weight verification and discrete data convolution calculation, and the method comprises the steps: constructing a subjective and objective multi-stage evaluation index system and robustness verification, and carrying out the cross-domain feature enhancement of an initial one-dimensional discrete sequence evaluation sample; feature expressions of discrete sequence samples are enriched, a feature domain suitable for evaluation model input is screened out, a multi-dimensional multi-target evaluation self-adaptive model is constructed, a local power index search evaluation method is utilized, iterative search and evaluation are conducted continuously, an equilibrium solution on the two targets of model classification precision and parameter quantity is achieved, and the evaluation accuracy is improved. Optimal model parameters are determined globally through an improved optimization algorithm, the generalization performance of the model is improved, and finally the efficient production and education fusion talent cultivation effect evaluation model is constructed.

Description

Technical Field

[0001] The present invention belongs to the technical field of talent cultivation effectiveness evaluation of industry-education integration, and specifically is a talent cultivation effectiveness evaluation method of industry-education integration based on indicator weight verification and discrete data convolution calculation. Background Art

[0002] Under the background of rapid economic and technological development, the cultivation of talents through industry-education integration has become a key force in promoting national scientific and technological progress and industrial upgrading. Industry-education integration cultivates applied talents in line with the development of modern industries through in-depth cooperation between industry-related organizations and education and scientific research-related organizations in collaborative education and collaborative innovation. This model not only improves the quality of talent cultivation between industry-related organizations and education and scientific research-related organizations, but also promotes the deep integration and development of both parties. However, although current talent cultivation research has widely focused on the optimization of training mechanisms, it is still insufficient in the evaluation system of results, which limits the continuous improvement of quality and the precise realization of goals. Therefore, building a set of intelligent, comprehensive and scientific evaluation system methods plays a vital role in ensuring the effectiveness of training, improving the quality of education and promoting the sustainable development of the country.

[0003] Traditional evaluation methods are often subject to subjectivity and limitations in the evaluation of the effectiveness of talent training in the integration of industry and education, while machine learning methods are being applied to performance evaluation to improve objectivity and accuracy. For example, Li et al. used sparse principal component analysis to determine the main aspects of takeover performance based on multiple original objective takeover performance indicators to obtain the driver's original subjective self-assessment of takeover performance. Radovanovic et al. incorporated the social and economic impact of bankruptcy into machine learning modeling, introduced two evaluation indicators, "financial costs associated with bankruptcy" and "social impact represented by unemployment", and compared them with traditional methods such as multivariate discriminant analysis and logistic regression based on logistic regression, verifying the potential of machine learning models in improving prediction accuracy. Baser et al. proposed a clustering-based fuzzy classification method for credit risk assessment, aiming to improve the ability to predict potential non-repayment risks through machine learning methods and fuzzy theory, and use the membership values ​​obtained by the fuzzy K-means clustering algorithm to more accurately grasp the risk structure, so as to issue credit to practitioners and decision makers. Wu et al. proposed a risk assessment method based on dual Bayesian networks, established a risk assessment index system for arch settlement with groundwater status, construction environment complexity, tunnel arch bottom and arch soil as the main risk indicators affecting arch deformation, and then used the proposed model to evaluate the arch settlement risk of Tuanshan Tunnel in East Lake, Wuhan. Through the analysis of the above research, it is found that there is a lack of evaluation of the effectiveness of the integration of production and education for talents, and the evaluation system still has a single evaluation subject and evaluation angle, and lacks robustness verification of weights. In addition, traditional machine learning methods have difficulties in mining nonlinear features and correlations, which makes it difficult to effectively guarantee the evaluation effect.

[0004] Deep learning model design is complex and relies on professional knowledge and experience, with high computational costs and limited generalization ability. Neural network architecture search (NAS) has attracted much attention as an adaptive modeling method to solve these problems. Gao et al. designed a differentiable architecture search with partial channel connections to achieve automated architecture search instead of manual design. Sun et al. used genetic algorithms to automatically evolve convolutional neural networks and optimized the neural network structure by accelerating fitness evaluation in evolutionary deep learning. Lu et al. proposed a NAS multi-objective evolutionary algorithm based on NSGA-II, which approximated the entire Pareto frontier through genetic operations that gradually reorganized and modified architectural components. Zhong et al. proposed BlockQNN to accelerate the NAS process using q-learning. Li et al. proposed the DARTS algorithm to optimize the blocks in the proxy network to optimize the target network. From the above studies, it can be seen that although NAS shows advantages in image classification tasks, it is rarely used in low-value density discrete sequence decision-making; secondly, when optimizing the network architecture, the NAS method fails to fully evaluate the complexity of the network when searching for cell structures, resulting in unnecessary complexity redundancy in cell structure design, which in turn makes the final model complexity high, affecting its generalization ability and practicality. Summary of the invention

[0005] In view of the fact that the existing evaluation of the effectiveness of talent training in the integration of industry and education cannot be realized intelligently due to the lack of specific evaluation standards and intelligent evaluation methods, the purpose of the present invention is to provide an evaluation method for the effectiveness of talent training in the integration of industry and education based on indicator weight verification and discrete data convolution calculation, which is mainly used to scientifically and comprehensively evaluate and optimize the talent training effect of the integration of industry and education projects.

[0006] To achieve the above object, the present invention adopts the following technical solution:

[0007] An evaluation method for the effectiveness of talent training in the integration of industry and education based on indicator weight verification and discrete data convolution calculation is proposed. Firstly, an evaluation data set is constructed by combining multi-level evaluation indicator combination weights with the TOPSIS method, and the combined weights are robustly verified by the MIV algorithm and the R-Dropout model to improve data reliability. Then, multi-dimensional space cross-domain mapping is used to enhance the feature representation of discrete sequence evaluation samples, and a spatial domain metric screening mechanism is designed to select feature domains with better performance to provide effective data input for the model. Then, the local power index search and evaluation method is used to achieve a balanced solution for the two goals of network classification accuracy and parameter quantity through continuous iterative search and evaluation. Finally, the optimal model parameters are determined by the optimization algorithm to improve the model performance, thereby constructing an efficient evaluation model for the effectiveness of talent training in the integration of industry and education.

[0008] The above method comprises the following steps:

[0009] Step 1: Robustness verification of the weights of evaluation indicators: Comprehensively use qualitative and quantitative research methods to summarize evaluation indicators that can fully reflect the multi-dimensional characteristics of cooperation performance, combine subjective and objective factors for comprehensive weighting, successfully build an evaluation system for the effectiveness of talent training in industry-education integration, design a robustness verification mechanism to verify the determined weights, and solve the weights of each evaluation indicator by quantifying the average impact of input changes on output results, and then implement weight reliability verification;

[0010] Step 2: Enhance the cross-domain features of discrete sequence evaluation samples and build a metric screening mechanism: Map the one-dimensional discrete sequence evaluation samples to the multi-dimensional space domain for cross-domain feature enhancement, build a metric screening mechanism by comprehensively considering the key evaluation indicators of multi-domain data, optimize feature extraction and classification performance, and screen out input data suitable for the evaluation task;

[0011] Step 3: Automatically construct a multi-dimensional and multi-objective evaluation model: Use the local power index evaluation strategy to maximize network classification accuracy and minimize the number of parameters as two goals, build an evaluation model for the effectiveness of industry-education integration talent training, and then use the improved optimization algorithm to adaptively optimize the model to improve model performance and achieve the final optimized evaluation model for the effectiveness of industry-education integration talent training.

[0012] In order to verify the effectiveness of the method of the present invention, the improved training effectiveness evaluation model OCEEM-NAS based on neural network architecture search is compared with the manually designed baseline network. The OCEEM-NAS network performs well in the classification task of talent training effectiveness evaluation, with a classification accuracy of 98.74%, and can effectively distinguish evaluation samples of different categories. Compared with other models on the EITI data set, the model constructed by the design method shows obvious advantages in search time, complexity and performance, proving the superiority of the construction method. The present invention is applicable to various prediction, recognition, classification, diagnosis, detection and other tasks based on discrete sequences, providing new impetus for the development of automation, education and other fields, promoting the deep integration of industry and education, achieving a win-win situation between education and industry, helping intelligent manufacturing, leading innovative development, and having practical engineering significance and theoretical research value.

[0013] The invention has reasonable design and good practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A flow chart showing the present invention.

[0015] Figure 2 It represents the flow chart of the evaluation system of talent training effectiveness through integration of industry and education.

[0016] Figure 3 Represents the weight robustness verification flow chart.

[0017] Figure 4 Represents the R-dropout module diagram.

[0018] Figure 5 Represents a robust verification radar chart.

[0019] Figure 6 Represents the cross-domain feature enhancement graph of discrete sequence evaluation samples.

[0020] Figure 7 Schematic diagram of the automatic architecture of the multi-dimensional and multi-objective evaluation model.

[0021] Figure 8 Represents a cell structure diagram.

[0022] Fig. 9 Schematic diagram of the spiral search strategy.

[0023] Fig.10 T-SNE diagram showing the evaluation results of different models (MobileNetV3, ESPNetV2, ConvNext, RepVGG).

[0024] Fig.11 T-SNE diagram showing the evaluation results of different models (MnasNet, WDD-NAS, CEEM-NAS, OCEEM-NAS). DETAILED DESCRIPTION

[0025] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0026] The method for evaluating the effectiveness of talent training for industry-education integration based on indicator weight verification and discrete data convolution calculation described in the present invention constructs an evaluation data set by combining multi-level evaluation indicator weights with the TOPSIS method, and robustly verifies the combined weights through the MIV algorithm and the R-Dropout model to improve data reliability. Multidimensional space cross-domain mapping is used to enhance the feature representation of discrete sequence evaluation samples, and a spatial domain metric screening mechanism is designed to select feature domains with better performance to provide effective data input for the model. The local power index search and evaluation method is used to achieve a balanced solution for the two goals of network classification accuracy and parameter quantity through continuous iterative search and evaluation. Finally, the optimal model parameters are determined through the optimization algorithm to improve the model performance, thereby constructing an efficient network for evaluating the effectiveness of talent training for industry-education integration. The specific steps are as follows.

[0027] Step 1. Verify the robustness of the evaluation index weights: Comprehensively use qualitative and quantitative research methods to summarize the evaluation indicators that can comprehensively reflect the multi-dimensional characteristics of cooperation performance, combine subjective and objective factors for comprehensive weighting, successfully build an evaluation system for the effectiveness of talent training in industry-education integration, design a robustness verification mechanism to verify the determined weights, and scientifically solve the weights of each evaluation index by quantifying the average impact of input changes on output results, and then implement weight reliability verification.

[0028] 1.1. Taking talents as the evaluation object, based on the opinions of experts in related fields and existing relevant literature, combined with subjective and objective factors, an evaluation index system for the effectiveness of talent training in industry-education integration was established, which includes 2 first-level indicators, 12 second-level indicators, and 51 third-level indicators. According to the constructed third-level indicators, data on the number of master's and doctoral students enrolled and the number of students on campus in each province of China from 2009 to 2019, the number of students enrolled in each discipline, and human capital accumulation were collected. Among them, the employment data comes from the series of books "Annual Report on the Development of China's Degrees and Postgraduate Education" from 2009 to 2019; the data related to human capital accumulation comes from the project data of the China Human Capital and Labor Economics Research Center of the Central University of Finance and Economics; other data are calculated based on the data in the "China Statistical Yearbook" and "China Education Statistical Yearbook" officially released by the Ministry of Education and the National Bureau of Statistics from 2010 to 2020.

[0029] 1.2. Based on the above indicator system, the subjective and objective weights of each indicator are calculated and weighted comprehensively. The specific weight calculation is as follows: The improved fuzzy hierarchical analysis method is a subjective weighting method that determines the relative importance of each indicator based on expert opinions. This method uses a five-level scale system from 0.1 to 0.9 to construct a judgment matrix. Then, through a series of calculation processes, including constructing a reciprocal judgment matrix, calculating a ranking vector, and necessary iterative processing, the subjective weight vector W of each indicator is finally obtained. 1 =(w 11 ,w 21 Λ,w n1 ) T The coefficient of variation method is an objective weighting method. The weight is determined by calculating the coefficient of variation of each indicator. First, the original data is dimensionless, that is, the data is standardized to the range of [0,] to eliminate the dimension effect. Then, based on the processed data, the standard deviation σ and mean μ of each indicator are calculated to obtain the coefficient of variation. The final objective weight of each indicator is W 2 =(w 12 ,w 22 Λ,w n2 ) T, which is determined by the ratio of the coefficient of variation of the indicator to the sum of the coefficients of variation of all indicators. In the framework of game theory, the optimal combination weight is solved by integrating subjective and objective weights. The objective function is first set to minimize the weight vector difference, and then it is converted into a linear differential equation using matrix differentiation to solve it. Then the solution is normalized to obtain the final combination weight W * =μ 1 * w 1 T +μ 2 * w 2 T , ensuring the balance and optimization of subjective and objective weights in decision-making.

[0030] 1.3. According to the comprehensive weight obtained by TOPSIS method combined with game theory weighting, quantitative evaluation of the training effect of industry-education integration is carried out. The process includes data normalization, construction of weighted matrix, calculation of the distance between each indicator and positive and negative ideal solutions, and determination of sample closeness K. i , according to K i The values ​​are in the interval (0,0.3], (0.3,0.4], (0.4,0.5], (0.5,1], and the samples are classified into D (unqualified), C (qualified), B (good), and A (excellent). Finally, 30 A-class data, 151 B-class data, 104 C-class data, and 15 D-class data were summarized, totaling 300 data, and the Evaluation of Industry-Education Integration Training Effectiveness (EITI) data set was successfully constructed.

[0031] 1.4. Nonlinear robustness verification mechanism increases and decreases each indicator of the input feature set to obtain the input feature set before and after the change, and then uses the model to predict these feature sets, calculates the difference in the prediction results and averages them, and compares the calculated weights with the original combined weighting method weights to analyze and verify the robustness of the weights. The detailed process of robustness verification is as follows: Figure 3 As shown. The nonlinear robustness verification mechanism combines the mean influence value (MIV) algorithm and the R-Dropout model. The verification process is as follows:

[0032] Take the indicator sample set X as input and combine the weights W * As labels, train the R-Dropout model, and then increase and decrease each indicator of the input feature set X by n% and n% at the same time to obtain two input feature sets X add and X sub , and then use the trained model to input feature set X add and X sub Make predictions and get the prediction result set Y add and Y sub, then calculate the difference in the prediction results before and after the change of each indicator, and find the average value to get the MIV of the indicator. Repeat this process to obtain the MIV of all indicators, and ensure that the sum of the weights is 1 through normalization. Finally, these calculated weights ω i Compared with the original combined weighting method weight W * Compare to analyze and verify W * The robustness check weight formula is as follows:

[0033]

[0034] Among them, R-Dropout is proposed as an improved method, which is more suitable for the robustness verification of the industry-education integration evaluation system. The R-Dropout module is shown in the figure. Figure 4 As shown in the figure. R-Drop performs i Perform two forward propagations to obtain the two distributions P predicted by the model 1 (y i |x i ) and P 2 (y i |x i ), and regularize the model prediction D by minimizing the bidirectional Kullback-Leibler (KL) divergence KL (P 1 ||P 2 ), this method not only enhances the robustness of the model to the Dropout operation, but also further improves the generalization ability of the model by ensuring the consistency of the output under different Dropout configurations. Therefore, the training objective of the model consists of two parts, one is the KL divergence between the two outputs as shown in formula (2), and the other is the model's own loss function cross entropy as shown in formula (3). The final loss function is the sum of the two formulas as shown in formula (4).

[0035]

[0036]

[0037]

[0038] In order to more intuitively show the changes in weight distribution under different n values ​​(n = 10 and n = 20), a radar chart is used for presentation, as shown below: Figure 5 (a) and Figure 5By comparison, it can be found that the weights calculated by the MIV algorithm are slightly different from the weights obtained by the combined weighting method (i.e., the combination of the coefficient of variation method and the improved fuzzy analytic hierarchy process). This finding fully proves that the combined weighting method proposed in this invention has high robustness in determining the weights of the evaluation indicators of the integration of production and education.

[0039] Step 2: Perform cross-domain feature enhancement on discrete sequence evaluation samples and construct a metric screening mechanism: Map the one-dimensional discrete sequence evaluation samples to the multi-dimensional space domain for cross-domain feature enhancement, build a metric screening mechanism by comprehensively considering the key evaluation indicators of multi-domain data, optimize feature extraction and classification performance, and screen out input data suitable for the evaluation task.

[0040] The samples collected for the evaluation of the effectiveness of industry-education integration training at this stage are one-dimensional discrete sequence samples with low value density. In traditional discrete sample data processing, feature extraction usually relies on manually designed statistical mathematical indicators. However, these traditional methods often cannot capture the complexity and nonlinear characteristics of the data well, which limits the effect of feature extraction. Now with the development of deep learning technology, convolutional neural networks (CNNs) automatically learn the hierarchical features of data through convolutional layers, can process high-dimensional data and extract features that are useful for classification and regression tasks, and compared with traditional methods, CNN does not require manually designed feature extractors, and can learn complex feature representations from large amounts of data. Therefore, the present invention performs multi-spatial domain conversion on discrete evaluation samples, and designs a spatial domain metric screening mechanism to select input spatial domains suitable for evaluation tasks. The specific steps are as follows:

[0041] 2.1. Cross-domain feature enhancement: Perform multi-space domain conversion on discrete evaluation samples and map one-dimensional discrete sequence samples to high-dimensional space domain;

[0042] Convolutional neural networks (CNNs) automatically learn hierarchical features of data through convolutional layers, can process high-dimensional data and extract features useful for classification and regression tasks, and compared with traditional methods, CNNs can learn complex feature representations from large amounts of data without manually designing feature extractors. Therefore, discrete evaluation samples are transformed into multi-space domains, and one-dimensional discrete sequence samples are mapped into high-dimensional space domains. Through this transformation, discrete samples are converted into image forms that can be processed by CNNs, so that the network can recognize and extract more complex feature patterns.

[0043] At the current stage, the collected evaluation samples belong to the type of one-dimensional discrete sequence. The feature expression of these samples is relatively simple and the value density is low, making it difficult for the model to fully explore their intrinsic characteristics. In order to enrich the feature representation of these evaluation samples, the one-dimensional discrete sequence samples are mapped to high-dimensional spatial domains such as Gram angle difference field (GADF), Gram angle sum field (GASF), Markov transfer field (MTF), recursive graph (RP), and relative position matrix (RPM). Through this transformation, the discrete samples are converted into image forms that can be processed by CNN, so that the network can recognize and extract more complex feature patterns. In order to effectively avoid the overfitting phenomenon that may occur during the model training process, the high-dimensional spatial images obtained by multi-domain mapping are subjected to various data enhancement processes such as rotation, mirroring, and noise addition. The four categories of data A, B, C, and D in each spatial domain data set are expanded to 860, 1380, 1080, and 680, a total of 4000, and divided into training sets and test sets at a ratio of 4:1. In the MTF space domain, the specific distribution of the four types of samples A, B, C, and D is as follows: Figure 6 shown.

[0044] At the same time, in order to verify the effectiveness of high-dimensional spatial domain conversion in discrete evaluation samples, the performance of mainstream lightweight networks such as ShuffleNetV1, ShuffleNetV2, GhostNet, ResNet-D and MnasNet on classification tasks is compared. The detailed comparison results are shown in Table 1. When the data of the same network is processed by mapping techniques such as MTF, GASF, GADF, RP and RPM, the average accuracy of these processed data is improved compared with the original discrete sequence. This result strongly proves the effectiveness of the spatial domain conversion method.

[0045] Table 1 Analysis of the impact of multi-domain mapping technology on the classification performance of discrete evaluation samples

[0046]

[0047] 2.2. Construct a spatial domain metric screening mechanism: Comprehensively consider the key evaluation indicators of multi-domain data, such as average accuracy and mean square error, to optimize the feature extraction and classification performance, and then screen the spatial domain.

[0048] After processing the discrete sequence evaluation samples and mapping them to the high-dimensional space domain, in order to ensure the validity of the model input, a spatial domain screening measurement mechanism is constructed. This mechanism is designed to adapt to the generalization requirements of different discrete evaluation samples, and optimizes the feature extraction and classification performance by comprehensively considering the key evaluation indicators of multi-domain data, such as average accuracy and mean square error. Specifically, the high-dimensional spatial domain data obtained by different mapping methods are input into multiple convolutional neural network models, and the average accuracy (Equation (5)) and mean square error (Equation (6)) of each mapping method under different models are calculated. The gain of high-dimensional mapped data in feature extraction and classification performance relative to the original data can be quantitatively evaluated, and the contribution is calculated accordingly. The calculation formula is shown in Equation (7). It is used to measure the specific contribution of different domain features to the classification performance. It can be observed from Equation (7) that as the average accuracy increases or the mean square error decreases, the corresponding feature contribution will also increase, thereby screening the spatial domain.

[0049]

[0050]

[0051]

[0052] In the formula, i represents the neural network model number, n represents the total number of models, represents the average accuracy of the mapping method under different models, s 2 It represents the mean square error of the mapping method under different models, and con represents the contribution of multi-spatial domain features of discrete evaluation samples.

[0053] The calculation results are shown in Table 2, which shows the specific results of the contribution of each feature domain. In order to make full use of the advantages of multi-domain mapping technology, the MTF method with the highest contribution was finally selected as the input data for network training, thereby ensuring that the experimental data processing and network training process can maximize the performance gain brought by multi-domain mapping technology.

[0054] Table 2 Contribution results of each feature domain

[0055]

[0056] Step 3: Automatically construct a multi-dimensional and multi-objective evaluation model: Use the local power index evaluation strategy to maximize network classification accuracy and minimize the number of parameters as two search goals, build an evaluation model for the effectiveness of industry-education integration talent training, and then use the improved optimization algorithm to adaptively optimize the model to improve model performance and achieve the final optimized industry-education integration talent training effectiveness evaluation model.

[0057] After cross-domain feature enhancement and screening of discrete sequence evaluation samples, the MTF with the highest contribution was selected as the input data for network training. In order to effectively extract and utilize key abstract information such as intersections, colors, shapes, etc. to improve feature representation performance, adapt to diverse cross-domain features and realize the generalization ability of the model, the present invention uses NAS technology to give the model the ability of self-learning. However, in the field of NAS, the search method based on gradient descent fails to fully evaluate the complexity of the network when searching for cell structures, which leads to the high complexity of the generated cell structure, and then makes the complexity of the final model high, affecting its generalization ability and practicality.

[0058] To solve the above problems, a method for automatic architecture search of evaluation models from a multi-dimensional perspective is proposed, aiming to improve network performance and reduce network complexity at the same time. Figure 7 As shown in the figure, it is mainly divided into two stages: local search and global search. In the local search stage, a multi-objective power index (PI) evaluation strategy is proposed, which can accurately evaluate and select low-complexity and high-performance cell structures, thereby constructing a cultivation effectiveness evaluation model (CEEM-NAS) based on neural network architecture search. In the global search stage, the parallel spiral golden sine slime mold optimization algorithm (PSG-SMA) is used. With its efficient parallel processing capability and global search capability, the algorithm finds the optimal model parameters and significantly improves the model performance. The specific steps are as follows:

[0059] 3.1. Search space of evaluation model based on Cell:

[0060] In NAS, a directed acyclic graph (DAG) containing N nodes is defined as the basic cell structure. Nodes represent feature representations, such as feature maps in convolutional networks. Edges represent operations from one node to another, such as convolution and pooling. In order to build an efficient feature extraction network, two types of key blocks are designed for this purpose: normal cell keeps the input spatial resolution unchanged, and reduction cell returns a feature map with half the spatial resolution through an operation with a stride of 2.

[0061] In order to construct these two types of key blocks, a series of candidate operations are collected from the CNN research field, such as standard convolution, depthwise separable convolution, dilated convolution, etc. The specific candidate operations are as follows:

[0062] 1) No operation;

[0063] 2) 3*3 maximum pooling operation (max_pool_3*3);

[0064] 3) 3*3 average pooling operation (avg_pool_3*3);

[0065] 4) Residual connection (skip-connect);

[0066] 5) 3*3 depth separable convolution (sep_conv_3*3);

[0067] 6) 5*5 depth separable convolution (sep_conv_5*5);

[0068] 7) 3*3 dilated convolution (dil_conv_3*3);

[0069] 8)5*5 dilated convolution (dil_conv_5*5).

[0070] 3.2. Power Index-based Evaluation Strategy: The proposed power index aims to quantify the value of each search primitive in the search space to identify search primitives that have a positive impact on model accuracy and complexity.

[0071] In the local search phase, the goal is to design two basic units, normal cell and reduction cell, with high performance and low complexity from the search space O in step 3.1. Traditional NAS is a black-box optimization process based on discrete space. The present invention uses DARTS to weaken the search space into a continuous spatial structure, so that performance optimization can be performed through the gradient descent algorithm. In addition, it is proposed to use the power index (PI) to evaluate the contribution of each component in the search space, and use these evaluation results to guide the search process.

[0072] As mentioned above, the proposed PI aims to quantify the value of each search primitive in the search space to identify search primitives that have a positive impact on model accuracy and complexity. In this method, each search primitive contained in the search space is (i,j) It is regarded as a participant in the cooperative game. After the network is initialized, these search primitives work together to determine the complexity C and accuracy A of the network. The complexity and accuracy together constitute the so-called "revenue". Based on this, according to the definitions of "participants", "collaboration" and "revenue", the search primitive o (i,j) PI value It can be expressed as formula (8).

[0073]

[0074] In the formula, N represents the set of all participants, represents all subsets S that do not contain participant i, 2 |N|-1 represents the probability of subset S being randomly selected, represents the marginal contribution of participant i to the alliance S. ΔA and ΔC can be calculated by formulas (9) and (10) respectively:

[0075]

[0076]

[0077] For each participant, all possible subsets need to be considered. The number of subsets that do not contain i is 2 |N|-1 -1 (minus the empty set). Since 2 |N-1| As |N| increases, it grows exponentially, and the computational complexity is high, especially when the number of participants is large. Therefore, the set of all possible subsets is approximated by randomly sampling subsets. For each randomly selected subset, the marginal contribution of the participant is calculated, and the power index is approximated by the average of a large number of random samples. The calculation formula is shown in formula (11):

[0078]

[0079] In the formula, S k is the kth randomly selected subset, and m is the number of random samplings. Therefore, in determining the value of the search primitive P i After that, the following formula (13) is used to update the architecture parameter α of the i-th training cycle i :

[0080]

[0081]

[0082] in, It represents the momentum of accelerating the coverage average, preventing large fluctuations during the search process, while allowing the gradual integration of new information, thereby smoothing the search process and improving stability. η represents the step size.

[0083] By calculating and analyzing the formula for the power index, the present invention successfully searches out the normal cell and reduction cell structures with superior performance. These cell structures have high model performance and generalization ability while maintaining low complexity. By repeatedly stacking these cells, CEEM-NAS is constructed. This model not only has powerful feature extraction and representation capabilities, but can also adaptively adjust the network structure according to specific task requirements, thereby achieving an accurate evaluation of the effectiveness of industry-education integration training. The specific structures of normal cell and reduction cell are as follows: Figure 8 (a) and Figure 8 (b) as shown.

[0084] To verify the effectiveness of this strategy, Table 3 compares the mean ± standard deviation of the accuracy of several strategies run 3 times using the NAS-Bench-201 benchmark search space and the CIFAR-10, CIFAR-100, and ImageNet16-120 datasets. Optimal represents the best performing architecture in the search space. For NAS-Bench-201, our method achieves excellent performance with 94.37%, 73.51%, and 46.24% test accuracy on CIFAR-10, CIFAR-100, and ImageNet16-120, respectively. It is worth noting that the global optimal structure is obtained on CIFAR-10 and CIFAR-100, which shows that the proposed method can identify and obtain the best architecture from a large search space. On the ImageNet-16-120 dataset, a near-optimal solution is obtained, which outperforms the listed algorithms, once again verifying the effectiveness of our method.

[0085] Table 3 Comparison results with NAS methods on NAS-Bench-201

[0086]

[0087] 3.3. High-dimensional hyperparameter optimization strategy for the evaluation model: This strategy is based on the Slime Mould Algorithm (SMA), which achieves the optimization goal by simulating the foraging process of slime molds. When slime molds sense food sources, they will show oscillatory contraction. The quality of the food source directly affects the state of the venous network, and the presence of multiple food sources will also cause slime molds to form venous networks of different thicknesses. Slime molds have three main behaviors: approaching food, wrapping food, and obtaining food.

[0088] The present invention designs an improved PSG-SMA algorithm, which is used to optimize the hyperparameters of the CEEM-NAS model in the neural network architecture search. Specifically, the algorithm first uses the population parallel mechanism to initialize the population (Population1 and Population2), and then repeats the following steps within the maximum number of iterations: for each time step, the fitness of all slime molds is calculated, and the optimal fitness and corresponding parameters are updated. Then, the position update of the slime mold is calculated according to the formula of the slime mold optimization algorithm, and each search solution is updated. During the update process, the algorithm randomly selects to use the spiral search strategy or the original strategy according to probability, and combines the golden sine search strategy to enhance the global search capability. Finally, the algorithm returns the optimal fitness and corresponding parameters, thereby completing the adaptive optimization of the hyperparameters of the CEEM-NAS model.

[0089] Among them, the traditional SMA has the characteristics of few parameters and strong optimization ability. However, when dealing with complex global optimization search problems, it faces the problems of slow convergence speed and easy to fall into local optimum. Therefore, the present invention will optimize and improve the SMA algorithm from the following aspects.

[0090] (1) Population parallel mechanism based on fitness replacement. The present invention adopts a parallel mechanism to improve the search speed of the SMA algorithm for the optimization component of the evaluation model. Specifically, the initial population is divided into two sub-populations for parallel update, and the low-quality solution is replaced by the high-quality solution to accelerate the search speed.

[0091] (2) The original strategy is that the position update formula of the slime mold individual is shown in (14):

[0092]

[0093] Where rand takes a random value in the interval [0, 1]; UB and LB are the upper and lower boundaries of the search range respectively; and are the positions of two random individuals; vb and vc are oscillation parameters; W represents the weight factor of the slime mold individual, which is also the mass of the slime mold; t represents the current number of iterations; parameter p is a conditional parameter used to control the way the slime mold position is updated; z is the probability that the separated slime mold individual searches for other food sources, that is, the probability of mutation, which is generally set to 0.03.

[0094] In view of the shortcomings of the SMA algorithm in local optimization, the present invention proposes a spiral search strategy. This strategy expands the search range by exploring in depth in the local search area in a spiral path, thereby avoiding falling into the local optimum. The schematic diagram of the spiral search strategy is shown in Fig. 9 As shown in: The spiral search strategy and the original strategy are randomly selected according to the probability of updating the population position. Therefore, the modified position update formula is shown in formula (15):

[0095]

[0096] Where l is a random number in [-1,1], which is used to generate a spiral search path. This random number affects the direction and step size during the search process, allowing the search to explore in different directions and distances.

[0097] (3) Golden sine search strategy. The golden sine search strategy, which combines the golden ratio with the sine wave dynamic adjustment mechanism, makes up for the shortcomings of the population parallel mechanism in global optimization. As shown in formula (16), this strategy guides the algorithm to efficiently explore in the global search space, effectively enhancing the ability to escape from the local optimum, thereby significantly improving the global optimization ability of the SMA algorithm.

[0098]

[0099] In the formula, is the optimal position of the tth iteration of the i-th individual, r 1 is a random number in the range [0,2], r 2 is a random number in the range [0,2], x 1 and x 2 is the golden ratio coefficient.

[0100] 3.4. Self-optimization of the evaluation model based on the improved optimization algorithm: The PSG-SMA algorithm is used to adaptively optimize the 10 key hyperparameters of CEEM-NAS. Globally, the model is lightweight compressed by taking the number of normal cells that appear repeatedly in the model as the decision variable, removing the redundant structural modules and reducing the model parameters; locally, the hyperparameters (learning rate, batch size, activation function, number of iterations and optimizer) within the model are self-optimized to further improve the model's adaptability, thereby effectively reducing research and time costs.

[0101] In the local search stage, a multi-objective power index evaluation strategy is used to evaluate and select low-complexity and high-performance cell structures, thereby constructing a talent training effectiveness evaluation model based on neural network architecture search. In the global search stage, an improved parallel spiral golden sine slime mold optimization algorithm is used to find the optimal model parameters, thereby significantly improving the model performance and constructing the final evaluation model. When testing the effect of the trained model, the fitness function value f(i) of individual i in the population is calculated by formula (17), and then the population is dynamically updated according to f(i) and the optimal individual is selected. Finally, the optimized adaptive classification model OCEEM-NAS is obtained, which is the talent training effectiveness evaluation model for industry-education integration.

[0102]

[0103] Where T represents the number of iterations, t(i) and p(i) represent the number of samples with correct category prediction and the number of samples with incorrect category prediction, respectively.

[0104] In summary, the method for evaluating the effectiveness of talent training in the integration of industry and education based on indicator weight verification and discrete data convolution calculation described in the present invention constructs a subjective and objective multi-level evaluation indicator system and robustness verification, performs cross-domain feature enhancement on the initial one-dimensional discrete sequence evaluation samples, enriches the feature expression of discrete sequence samples and screens out feature domains suitable for evaluation model input, constructs a multidimensional and multi-objective evaluation adaptive model, uses a local power index search and evaluation method, continuously iterates search and evaluation, and achieves a balanced solution for the two objectives of model classification accuracy and parameter quantity. The optimal model parameters are determined globally through an improved optimization algorithm to improve the generalization performance of the model, and finally constructs an efficient model for evaluating the effectiveness of talent training in the integration of industry and education.

[0105] To verify the performance of PSG-SMA, the present invention conducts simulation experiments on five comparison algorithms, namely, dung beetle optimizer (DBO), Great Wall Construction Algorithm (GWCA), bald eagle search algorithm (BES), Marine Predators Algorithm (MPA), and Slime Mould Algorithm (SMA), on 12 standard test functions. The 12 standard test functions include 7 unimodal test functions and 5 multimodal test functions.

[0106] The algorithm parameters are uniformly set as follows: population size is 30, dimension is 30, and maximum number of iterations is 1000. To avoid the randomness of the optimization results, each algorithm is independently run 30 times on 12 test functions, and the mean (Mean) and standard deviation (Std) of the 30 results are selected as the final evaluation indicators of the algorithm. The optimization results of each algorithm are shown in Tables 4 and 5, where the best results are marked in bold.

[0107] Table 4 Experimental results of unimodal test function

[0108]

[0109] Table 5 Experimental results of bimodal test function

[0110]

[0111] As can be seen from Table 4, PSG-SMA can find the optimal solution when solving the unimodal test functions F1, F2, F3, F4, F6, and F7. Only when solving F5, BES obtains the optimal solution and DBO obtains the minimum standard deviation. As can be seen from Table 5, PSG-SMA can get the ideal objective function value by running 30 times independently on the five bimodal test functions, and the average value and standard deviation of the solution are better than the other five comparison algorithms. It is proved that on the bimodal test function, the designed PSG-SMA has higher optimization ability than other comparison algorithms and the best stability. Therefore, the PSG-SMA algorithm is applied to the CEEM-NAS adaptive optimization problem.

[0112] Based on the test results of the optimization algorithm in the optimal segment of the 12 peak test functions, the high-performance PSG-SMA algorithm was selected to adaptively optimize CEEM-NAS, and the optimized model was called OCEEM-NAS. The 10 dimensions of the individual were set to correspond to the 10 selected parameters, the initial population size was set to 10, and the maximum number of iterations was set to 5. The parameter range to be optimized and the optimal parameters are shown in Table 6, where V1, V2, and V3 represent the number of normal cells between reduction cells, and n and m represent the increment of the channel when the feature is downsampled.

[0113] Table 6 Parameters to be optimized and search range

[0114]

[0115] In order to verify the effectiveness of the method for constructing the evaluation model of the invention, four widely used lightweight models MobileNetV3, ESPNetV2, ConvNext and RepVGG and two NAS models MnasNet and WDD-NAS are used for comparison with the model proposed in the present invention. Table 7 lists the number of parameters, reasoning time, accuracy and search time of each model in detail. The results show that CEEM-NAS is higher than all other models in the table in terms of classification accuracy, with a maximum improvement of 21.5%. In terms of parameter quantity, CEEM-NAS has a more obvious advantage over the traditional ConvNext model. Compared with the MnasNet and WDD-NAS models designed with the NAS method, CEEM-NAS reduces the number of parameters by up to 52.5% and shortens the reasoning time by up to 30%. In addition, in terms of search time, CEEM-NAS also achieves a reduction of up to 81.2%. These results fully demonstrate the effectiveness and efficiency of the method proposed in the present invention. At the same time, compared with the original CEEM-NAS, OCEEM-NAS after adding PSG-SMA has improved performance in all aspects. This further verifies the feasibility and effectiveness of the proposed PSG-SMA algorithm in adaptive optimization of the evaluation model.

[0116] Table 7 Parameters of different models

[0117]

[0118] Based on the different lightweight models in Table 7, the output results of the optimized industry-education integration talent training effectiveness evaluation model and other models are mapped onto a two-dimensional plane by using the t-SNE method, thereby realizing the visualization of the evaluation results. Fig.10 , Fig.11As shown in the figure, the obvious inter-class distance shows that the optimized industry-education integration talent training effectiveness evaluation model performs well in classification tasks and can effectively distinguish evaluation samples of different categories. That is, the multi-dimensional and multi-objective evaluation model automatic architecture method proposed in the present invention can realize intelligent evaluation of the industry-education integration talent training effectiveness with a higher classification accuracy.

[0119] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A method for evaluating the effectiveness of talent training based on indicator weight verification and discrete data convolution calculation, characterized by: Firstly, the evaluation data set is constructed by combining multi-level evaluation index combination weighting with TOPSIS method, and the combined weights are robustly verified through MIV algorithm and R-Dropout model to improve data reliability; then, the feature representation of discrete sequence evaluation samples is enhanced by multi-dimensional space cross-domain mapping, and the spatial domain metric screening mechanism is designed to select the feature domains with better performance to provide effective data input for the model; then, the local power index search evaluation method is used to achieve a balanced solution for the two goals of network classification accuracy and parameter quantity through continuous iterative search and evaluation. Finally, the optimal model parameters are determined by optimization algorithm to improve model performance, thereby constructing an efficient evaluation model for the effectiveness of talent training in industry-education integration.

2. According to claim 1, a method for evaluating the effectiveness of talent training based on indicator weight verification and discrete data convolution calculation for integrating industry and education, characterized in that: The steps include: Step 1: Robustness verification of the weights of evaluation indicators: Comprehensively use qualitative and quantitative research methods to summarize evaluation indicators that can fully reflect the multi-dimensional characteristics of cooperation performance, combine subjective and objective factors for comprehensive weighting, successfully build an evaluation system for the effectiveness of talent training in industry-education integration, design a robustness verification mechanism to verify the determined weights, and solve the weights of each evaluation indicator by quantifying the average impact of input changes on output results, and then implement weight reliability verification; Step 2: Enhance the cross-domain features of discrete sequence evaluation samples and build a metric screening mechanism: Map the one-dimensional discrete sequence evaluation samples to the multi-dimensional space domain for cross-domain feature enhancement, build a metric screening mechanism by comprehensively considering the key evaluation indicators of multi-domain data, optimize feature extraction and classification performance, and screen out input data suitable for the evaluation task; Step 3: Automatically construct a multi-dimensional and multi-objective evaluation model: Use the local power index evaluation strategy to maximize network classification accuracy and minimize the number of parameters as two goals, build an evaluation model for the effectiveness of industry-education integration talent training, and then use the improved optimization algorithm to adaptively optimize the model to improve model performance and achieve the final optimized evaluation model for the effectiveness of industry-education integration talent training.

3. The method for evaluating the effectiveness of talent cultivation based on indicator weight verification and discrete data convolution calculation in industry-education integration according to claim 2 is characterized by: Step 1 is as follows: 1.

1. Taking talents as the evaluation object, based on the opinions of experts in related fields and existing relevant literature, combined with subjective and objective factors, an evaluation index system for the effectiveness of talent training in industry-education integration has been established, which includes 2 first-level indicators, 12 second-level indicators, and 51 third-level indicators; 1.

2. Based on the above index system, the subjective and objective weights of each index are calculated and weighted comprehensively. The specific weight calculation is as follows: The subjective weighting adopts the improved fuzzy hierarchical analysis method, and the relative importance of each index is determined according to the expert opinions. The five-level scale system of 0.1 to 0.9 is used to construct the judgment matrix. Then, by constructing a reciprocal judgment matrix, calculating the ranking vector and iterative processing, the subjective weight vector W1 of each index is finally obtained. 11 ,w 21 Λ,w n1 ) T ; Objective weighting adopts the coefficient of variation method, and the weight is determined by calculating the coefficient of variation of each indicator. First, the original data is dimensionless, that is, the data is standardized to the range of [0,1]. Then, based on the processed data, the standard deviation σ and mean μ of each indicator are calculated to obtain the coefficient of variation. The final objective weight of each indicator is W2 = (w 12 ,w 22 Λ,w n2 ) T , which is determined by the ratio of the coefficient of variation of the indicator to the sum of the coefficients of variation of all indicators; finally, the optimal combination weight is solved by integrating the subjective and objective weights. The objective function is first set to minimize the weight vector difference, and then it is converted into a linear differential equation using matrix differentiation to solve it. Then the solution is normalized to obtain the final combination weight W * =μ1 * w1 T +μ2 * w2 T , ensuring the balance and optimization of subjective and objective weights in decision-making; 1.

3. According to the comprehensive weight obtained by combining the TOPSIS method with game theory, the effectiveness of industry-education integration training is quantitatively evaluated; specifically, data normalization, construction of weighted matrix, calculation of the distance between each indicator and the positive and negative ideal solutions, and determination of the sample closeness K i , according to K i The value range is (0,0.3], (0.3,0.4], (0.4,0.5], (0.5,1], and the samples are classified into D, C, B, and A. Finally, a data set for evaluating the effectiveness of industry-education integration training is successfully constructed; 1.

4. The nonlinear robustness verification mechanism increases and decreases each indicator of the input feature set to obtain the input feature set before and after the change, and then uses the model to predict these feature sets, calculates the difference in the prediction results and averages them, and compares the calculated weights with the original combined weighting method weights to analyze and verify the robustness of the weights.

4. The method for evaluating the effectiveness of talent cultivation based on indicator weight verification and discrete data convolution calculation for integration of industry and education according to claim 3 is characterized by: The specific verification process of the nonlinear robustness verification mechanism described in step 1.4 is as follows: Take the indicator sample set X as input and combine the weights W * As labels, train the R-Dropout model, and then increase and decrease each indicator of the input feature set X by n% and n% at the same time to obtain two input feature sets X add and X sub , and then use the trained model to input feature set X add and X sub Make predictions and get the prediction result set Y add and Y sub , then calculate the difference in the prediction results before and after the change of each indicator, and find the average value to get the MIV of the indicator. Repeat this process to obtain the MIV of all indicators, and ensure that the sum of the weights is 1 through normalization. Finally, these calculated weights ω i Compared with the original combined weighting method weight W * Compare to analyze and verify W * Robustness; Robustness check weight formula (1): Among them, R-Dropout is performed by i Perform two forward propagations to obtain the two distributions P1(y i |x i ) and P2(y i |x i ), and regularize the model prediction D by minimizing the bidirectional KL divergence KL (P1||P2); The training objective of the model consists of two parts, one is the KL divergence between two outputs as shown in formula (2), and the other is the model's own loss function cross entropy as shown in formula (3). The final loss function is the sum of the two formulas as shown in formula (4); 5. The method for evaluating the effectiveness of talent cultivation based on indicator weight verification and discrete data convolution calculation in accordance with claim 2 is characterized by: Step 2 is as follows: 2.

1. Cross-domain feature enhancement: Perform multi-space domain conversion on discrete evaluation samples and map one-dimensional discrete sequence samples to high-dimensional space domain; Map one-dimensional discrete sequence samples to Gram's angle difference field, Gram's angle sum field, Markov transition field, recurrence graph, relative position matrix high-dimensional space domain, and convert discrete samples into image form that can be processed by CNN; 2.

2. Constructing a spatial domain metric screening mechanism: After processing discrete sequence evaluation samples and mapping them to a high-dimensional spatial domain, a spatial domain screening metric mechanism is constructed; The high-dimensional spatial domain data obtained by different mapping methods are input into multiple convolutional neural network models, and the average accuracy (Equation (5) and mean square error (Equation (6)) of each mapping method under different models are calculated to quantitatively evaluate the gain of high-dimensional mapping data in feature extraction and classification performance relative to the original data, and the contribution is calculated accordingly. The calculation formula (7) is used to measure the specific contribution of different domain features to classification performance; In the formula, i represents the neural network model number, n represents the total number of models, represents the average accuracy of the mapping method under different models, s 2 It represents the mean square error of the mapping method under different models, and con represents the contribution of multi-spatial domain features of discrete evaluation samples.

6. The method for evaluating the effectiveness of talent cultivation based on indicator weight verification and discrete data convolution calculation for integrating industry and education according to claim 2 is characterized by: Step 3 is as follows: 3.

1. Search space of evaluation model based on Cell: In NAS, a directed acyclic graph containing N nodes is defined as the basic cell structure. Nodes represent feature representations, and edges represent operations from one node to another. Two types of key blocks are designed: normal cell and reduction cell. The normal cell keeps the input spatial resolution unchanged, and the reduction cell returns a feature map with half the spatial resolution through an operation with a stride of 2. To construct the above two types of key blocks, the candidate operations are as follows: 1) No operation; 2) 3*3 maximum pooling operation; 3) 3*3 average pooling operation; 4) Residual connection; 5) 3*3 depth-wise separable convolution; 6) 5*5 depth separable convolution; 7) 3*3 dilated convolution; 8) 5*5 dilated convolution; 3.

2. Evaluation strategy based on power index: In the local search phase, the goal is to design two basic units, normal cell and reduction cell, with high performance and low complexity from the search space O in step 3.1; Each search primitive o contained in the search space (i,j) is regarded as a participant in the cooperative game. After the network is initialized, these search primitives work together to determine the complexity C and accuracy A of the network. The complexity and accuracy together constitute the "revenue". Based on this, according to the definitions of "participants", "collaboration" and "revenue", the search primitive o (i,j) PI value It is expressed as formula (8); Where N represents the set of all participants, represents all subsets S that do not contain participant i, 2 N-1 represents the probability of subset S being randomly selected, represents the marginal contribution of participant i to the alliance S. ΔA and ΔC can be calculated by formulas (9) and (10) respectively: For each player consider all possible subsets; the number of subsets that do not contain i is 2 N-1 -1, approximate the set of all possible subsets by randomly sampling subsets, calculate the marginal contribution of the participant for each randomly selected subset, and approximate the power index by the average of a large number of random samples; calculation formula (11): In the formula, S k is the kth randomly selected subset, and m is the number of random sampling; In determining the value P of the search primitive i After that, the following formula (13) is used to update the architecture parameter α of the i-th training cycle i : in, represents the momentum of the acceleration covering the average value, and η represents the step size; By calculating and analyzing the formula of the power index, we searched for normal cell and reduction cell structures with superior performance, and constructed the CEEM-NAS model by repeatedly stacking these cells, thus achieving an accurate evaluation of the training effect of industry-education integration. 3.

3. High-dimensional hyperparameter optimization strategy for evaluation models: The PSG-SMA algorithm is designed to optimize the hyperparameters of the CEEM-NAS model in the neural network architecture search. Specifically, the population is initialized using the population parallel mechanism, and then the following steps are repeated within the maximum number of iterations: for each time step, the fitness of all slime molds is calculated, and the optimal fitness and corresponding parameters are updated; then, the position update of the slime mold is calculated according to the formula of the slime mold optimization algorithm, and each search solution is updated; during the update process, the algorithm randomly selects to use the spiral search strategy or the original strategy according to probability, and combines the golden sine search strategy to enhance the global search capability; finally, the optimal fitness and corresponding parameters are returned, thereby completing the adaptive optimization of the hyperparameters of the CEEM-NAS model; 3.

4. Self-optimization of the evaluation model based on the improved optimization algorithm: The PSG-SMA algorithm is used to adaptively optimize the hyperparameters of the CEEM-NAS model; In the global aspect, the model is lightweight compressed by taking the number of normal cells that appear repeatedly in the model as the decision variable, removing the redundant structural modules and reducing the model parameters. In the local aspect, the hyperparameters inside the model are optimized automatically. When testing the effect of the trained model, the fitness function value f(i) of individual i in the population is calculated by formula (17), and then the population is dynamically updated according to f(i) and the optimal individual is selected. Finally, the optimized adaptive classification model OCEEM-NAS is obtained, which is the evaluation model of talent training effectiveness of industry-education integration. Where T represents the number of iterations, t(i) and p(i) represent the number of samples with correct category prediction and the number of samples with incorrect category prediction, respectively.

7. The method for evaluating the effectiveness of talent cultivation based on indicator weight verification and discrete data convolution calculation for integration of industry and education according to claim 6 is characterized by: In step 3.3, the original strategy is that the position update formula of the slime mold individual is as follows (14): Where rand takes a random value in the interval [0, 1]; UB and LB are the upper and lower boundaries of the search range respectively; and are the positions of two random individuals; vb and vc are oscillation parameters; W represents the weight factor of the slime mold individual, which is also the mass of the slime mold; t represents the current number of iterations; The parameter p is a conditional parameter used to control the update mode of the slime mold position; z is the probability of the separated slime mold individual searching for other food sources, that is, the probability of mutation; The spiral search strategy is that the modified position update formula is shown in formula (15): Where l is a random number in [-1,1], which is used to generate the spiral search path; The golden sine search strategy is a golden sine search strategy that combines the golden ratio with the sine wave dynamic adjustment mechanism, as shown in formula (16); In the formula, is the optimal position of the tth iteration of the i-th individual, r1 is a random number in the range [0,2], r2 is a random number in the range [0,2], x1 and x2 are the golden section coefficients.

Images