Research on Accurate Training Mechanism of Digital Talents for Higher Vocational Innovation and Entrepreneurship Education under Industry-Teaching Integration Driven by Big Data

The academic world hasn’t yet defined the concept of talent portrait, which is derived from the current hot technology in the Internet industry, namely user portrait. Talent portrait is similar to the existing user portrait technology, the core of which is based on real people, based on real data, according to the demand side of the goals, behaviors and views of the differences in the target person model. Talent profiling is a visual representation of talent modeling, which is a serious business issue. Each refined label phrase is based on the analysis and mining of massive data and is closely related to specific businesses. The label is the integration of a single multi-source heterogeneous talent information together, thus forming the unique information of each talent, that is, the unique “portrait” of the talent. The purpose of talent portrait technology is to collect, extract, analyze, interpret and present talent characteristics, the essence of which is to analyze talent prototypes based on their basic natural attributes, learning and working experiences, relevant papers and writings, patents and monographs, as well as behavioral characteristics such as awards and honors. At the same time, the talent model is categorized according to types, and characteristics and common features are extracted from each category, which are described or supplemented by names, texts, and demographic elements and scenarios, and a talent portrait is thus drawn.

The talent portrait model construction described in this paper mainly includes the following 3 stages: talent basic data collection, behavioral modeling and multidimensional talent portrait construction. The 3 stages of talent portrait construction are shown in Figure 1.

Figure 1.

Building talent portrait 3 stage

Segmentation is the basic technique for performing a variety of text mining [23]. The feature values of utterances need to be extracted for vector computation in big data processing. Before that, in this paper, we will use a disambiguation tool to break up long utterances into short phrases.

In this paper, it is very critical to calculate the label weight according to the user’s behavior when portraying the talent portrait, and the algorithm used is the TF-IDF label weighting algorithm.TF-IDF is a commonly used weighting technique for information retrieval and data mining. It is often used to calculate the importance of a phrase for a document dataset or one of the documents in a corpus. In this algorithm, the importance of a lexical item increases as the number of times it appears in an article increases, and decreases as the number of times it appears in the entire document set increases.

As the platform stores tags brought by each activity (from log data), change (from business data) and other behaviors of talents. With the accumulation of time, the labels accumulated by each talent on the platform will be hundreds of millions, so it is necessary to classify and divide the labels to find the category to which each label belongs. Currently, the clustering algorithm is an unsupervised machine learning method for data categorization based on the basic idea that things and people are grouped together. Without knowing the relevant information about the target class, clustering algorithms are able to divide all labeled data into clusters with some metric, making the data items in the same cluster as similar as possible.

At present, there are a variety of clustering analysis algorithms. Users can choose the appropriate clustering algorithm according to the type of business, data structure, classification purposes, and application platform. 1)

Hierarchical clustering algorithm: Hierarchical clustering algorithm is based on some kind of linkage rules will be the data objects will be a collection of data objects for hierarchical structure division or aggregation. The algorithm is divided into two modes: “top-down” split method and “bottom-up” aggregation method. The split method is similar to the idea of divide- and-conquer, in which all objects with clustering are first used as a cluster, and then the least similar classes are split to form another cluster. This is repeated until all the individual objects form their own cluster.

In this paper, based on the characteristics of talent data of talent portrait, the text information of talent portrait system is represented as V (f₁, w₁; f₂.w₂;…; f_n, w_n) by vector space model, where f_i is used to represent each feature word in the talent information data after the word splitting operation, in which each feature word is a dimension of the talent information data. w_i represents the weight of each feature word, and the size of its value can indicate the importance of the feature word in the talent information data.

Combined with the characteristics of talent information data in the talent profiling system, this paper selects the feature selection method of TF-IDF. This method is not only able to assign different weights to different feature items based on the frequency and distribution range of the feature items, but also can extract more representative feature words to help distinguish different talent information data to a greater extent.

The TF-IDF feature selection method is based on calculating the frequency of occurrence of feature words in the data as well as the inverse document frequency of the feature words in all the data, while these two data are multiplied to get the weight value of the corresponding feature words: 1wki=tfki×idfi w_ki is the weight value of the feature word t_i in the data V_i, tf_ki represents the frequency of the feature word in different data texts, and idf_i represents the inverse document frequency of the feature word in all data texts. The formula for calculating the inverse document frequency is as follows: 2idfi=logNni

Where N denotes the number of texts of all the data, and n_i denotes the number of data texts in which the feature words appear, and because of the need to avoid the possible interference of idf_i ‘s value of 0 on the calculation when each data text contains the same feature item, it is necessary to include a constant in the calculation, then the formula is shown below: 3idfi=log(Nni+a)

Usually limit the constant a ∈ (0,1), but because the length of the data text will also have an impact on the weight value of the feature items, so the TF-IDF formula to make the processing of the weight value of the feature items will be limited to [0,1] within the modified formula for: 4wki=tfki×log(Nni+a)∑i=1ntfki×log(Nni+a)

For the data text of talent information, it is necessary to increase the importance of feature words in the data text of talent’s professional field, i.e., the weight value assigned to them should be increased accordingly. At the same time, it is also necessary to limit the weight value of the feature words, the weight value is less than the threshold value μ of the feature words filtered to remove so as to achieve the purpose of the feature space dimensionality reduction, which not only can retain the feature words that are more effective in the extraction of the theme, but also can reduce the complexity of the operation.

In this paper, the method of constructing talent portrait labeling system based on theme model is selected, which firstly needs to train the theme dataset through the LDA [24] theme model, and extract the theme dimensions of various types of talent professions, so as to obtain the labeling system framework of the information of the talent professional field, which helps to solve the problem of the lack of dimension of the labeling system brought about by the usual method of constructing the user’s portrait. In addition, since the construction method of talent portrait is mainly derived from the construction method of user portrait, and the traditional user portrait construction method mostly establishes the labeling system of the platform based on the actual business requirements of the platform used by the user, there exists the problem of poor expansion of the characterization of multiple dimensions of the user. To solve this problem, this paper combines the construction method of statistical analysis, quantitative indexing analysis of talent information data to help establish a more comprehensive, scientific and accurate talent portrait labeling system, so that the labeling system does not have to be confined to the actual business, and enhances the expandability of the labeling system, so that the talent portrait is more detailed and accurate.

Fuzzy C-mean (FCM) [25–26] is a clustering algorithm based on the objective function, the FCM algorithm is improved from the ordinary C-mean algorithm, FCM is usually used to analyze the data for clustering. The hard division of the ordinary C-mean algorithm will strictly classify the original data into a certain class cluster, using 0 and 1 to indicate whether the data belongs to the class cluster, instead of using the value between 0 and 1 to indicate the affiliation of the data belonging to the class cluster, so the results obtained from the ordinary C-mean analysis tend to be in error with the real situation. However, the FCM algorithm obtains a fuzzy delineation result that is more in line with the relationship between data objects and classes in reality.

Assuming that the set of data samples to be analyzed is X, the data samples need to be divided into C classifications, C_j ∈ X, 1 ≤ j ≤ C, c_j is the center of the C_j, the i data sample x_i belongs to C_j affiliation is u_ij, the goal of the FCM clustering algorithm is to divide all the data samples into C classifications, so that the degree of each sample’s affiliation to the classification in which it is located as high as possible. FCM algorithm’s objective function is shown in equation (7). The objective function of the FCM algorithm is shown in equation (7), the product of the degree of affiliation of each data sample and the Euclidean distance of the sample to the center of the class cluster, and then summed to obtain the objective function, in which m is the parameter controlling the degree of flexibility of the algorithm, i.e., the degree of the sample’s gentleness, which is generally set to 2. That is: 7J=∑i=1c∑j=1nuijm‖ xj−ci ‖2 8∑i=1cuij=1,j=1,2,⋯,n

The constraints of affiliation of FCM algorithm are shown in Eq. (8), through the constraints, it can be guaranteed that the affiliation of samples belonging to different classes and the sum of the affiliation is 1. Therefore, for the initial affiliation matrix, it is necessary to satisfy the requirements through normalization. An example of the analysis of the affiliation u_ij is shown in Fig. 2, and an example of the analysis of the clustering centroid c_i is shown in Fig. 3. The iterative process of the algorithm is to obtain the attribution probability u_ij and the center of mass c_j, and the termination of the iteration is based on the need to set a termination value e, so that the following equation is valid, i.e., the maximum value of the difference between the u_ij difference between the first k and k + 1 rounds of the iteration should be less than e, whose geometrical meaning is to make the minimum of the J_m or to be located in the position of the saddle shape. 9max{ | uijk+1−uijk | }<e,1≤i≤n,1≤j≤C

Figure 2.

Analysis of membership degrees u_ij

Figure 3.

Example of cluster center point c_i analysis

The specific steps of the traditional FCM algorithm are: 1)

Determine the number of categories for clustering n, the flexibility parameter m of the algorithm, and the termination condition of the algorithm.

One of the most difficult tasks in cluster analysis is to determine the appropriate number of clusters. The fuzzy partition coefficient can be calculated using the number of classifications in fuzzy clustering, which is one of the most challenging tasks in cluster analysis. In fuzzy clustering, the fuzzy partition coefficient can be calculated by using the number of classifications AA and the attribution probability matrix, comparing the size of the fuzzy partition coefficient for different numbers of classifications, and then determining whether the number of clusters is appropriate or not. The formula for calculating the fuzzy partition coefficient without normalization is as follows: and the attribution probability matrix, comparing the size of the fuzzy partition coefficient for different numbers of classifications, and then determining whether the number of clusters is appropriate or not. To calculate the fuzzy partition coefficient without normalization, the formula is as follows: 12F(U)=1N∑i=1N∑j=1cuij2

F(U) can be normalized by the following formula, after normalization, the F_c value range is 0 to 1. When F_c is 0, it represents that the data samples are completely fuzzy, and the classification effect is poor according to the clustering center under the current number of clusters: when F_c is 1, it represents that the classification effect is optimal according to the clustering center under the current number of clusters. F_c(U) has a unique maximum value within a certain interval of determined number of clusters, and the number of clusters corresponding to the maximum value is the optimal number of clusters. Then there are: 13Fc(U)=F(U)−(1N)1−(1N)

The construction process of equipment group portrait is to determine the optimal number of clusters for clustering based on the source data of equipment group portrait and F_c(U) indicators, and then to divide the clusters, and finally to get the classification result and the center of clustering is the result of equipment group portrait. The flow chart of FCM-based equipment group portrait construction is shown in Figure 4, which mainly includes the following seven steps. 1)

Set a range of values for the number of clusters according to the sample size of the source data of the device population portrait.

Figure 4.

The device group portrait builds the flowchart

The FCM clustering results of higher vocational innovation and entrepreneurship students are shown in Table 1. This paper investigates and analyzes the precise cultivation effect of higher vocational innovation and entrepreneurship digital talents in 5 aspects of “Theoretical Knowledge (TK), Practical Ability (PA), Professional Skills (PS), Brainstorming (B), and Competitiveness (C)” under the background of the integration of big data industry and education. The results of FCM clustering for cultivating innovative and entrepreneurial students are shown in Table 1. The results show that students of innovation and entrepreneurship education, based on talent profiling, are divided into four clusters under the teaching mode of industry-teaching integration. Cluster 2 (114 students, 39.41%) and Cluster 3 (104 students, 35.86%) accounted for a relatively large proportion of the number of people, and Cluster 1 (27 students, 9.31%) accounted for the least proportion of the number of people. This is followed by cluster 4 with 45 persons or 15.52% of the total number of persons.

In order to verify the validity of the clustering results, one-way ANOVA was performed according to the category grouping. The results of one-way ANOVA of the clustering results are shown in Table 2, and it can be found that the clustering results on each dimension show significance, and according to the size of the F-value, we can get the importance of each dimension to the clustering results in an approximate way, which are as follows: the practical ability (F=321.3819) is in the first place, and has the most significant influence on the clustering results; the brainstorming skills (F=93.405) are in the second place, and also have a significant impact on the clustering results which is also significant; competitiveness (F=74.3753) is in the third place and is relatively less important to the clustered results; theoretical knowledge (F=52.3691) is in the fourth place and its effect on the clustered results is relatively less; and in the last place is professional skills (F=36.4204) which is the least important to the clustered results among the dimensions. An in-depth analysis of the students in each of the clusters reveals that Cluster 1 performs slightly less well in each of the dimensions of theoretical knowledge, practical skills, professional skills, brainstorming skills, and competitiveness, Cluster 2 performs more consistently in all dimensions, Cluster 3 does not perform as well as the other dimensions in terms of practical skills and professional skills, and Cluster 4 underperforms in terms of brainstorming skills and competitiveness.

Thus, taxon 1 is a more specific group that represents students who perform poorly in theoretical knowledge, practical skills, professional skills, brainstorming skills, and competitiveness. Cluster 2 can represent a group of students who emphasize both theory and practice, who value brainstorming and who have an eye for development. Cluster 3 may represent a group of students with inadequate mastery of practical and brainstorming skills. Cluster 4 can represent the group of students with insufficient mastery of brainstorming skills and lack of competitiveness. Based on the characteristics of the above four clusters and the understanding of the real situation, cluster 1 is specifically categorized as students who need to be strengthened, cluster 2 is categorized as students with balanced development, cluster 3 is categorized as students with lack of skills, and cluster 4 is categorized as students with rich imagination.

The results of entropy value of information technology students in each dimension are shown in Fig. 5, which can be analyzed that the entropy value of “to be strengthened” students in the theoretical knowledge dimension ranges from 0 to 0.0089, the entropy value of the practical ability dimension ranges from 0 to 0.0113, the entropy value of the professional skills dimension ranges from 0. 0008 to 0.0090, the entropy value on the brain power dimension ranges from 0. 0012 to 0. 0119, and the entropy value on the competitiveness dimension ranges from 0. 0012 to 0.0119.

Figure 5.

The result of the entropy value for each dimensional student

The results of entropy value of Balanced Development (BD) students in each dimension are shown in Figure 6. The entropy value results of balanced development (BD) students in the theoretical knowledge dimension ranged from 0.0072 to 0.0118, in the practical skills dimension ranged from 0.0073 to 0.0115, in the professional skills dimension ranged from 0.0067 to 0.0124, in the brainstorming skills dimension ranged from 0.0066 to 0.0123, and in the entropy results on the competitiveness dimension ranged from 0.0108 to 0.0177.

Figure 6.

The entropy of a balanced development student in each dimension

The results of entropy value of skill-deficient (U) students in each dimension are shown in Figure 7. The entropy value results of skill-deficient students in the theoretical knowledge dimension ranged from 0.0057 to 0.0123, in the practical skills dimension ranged from 0 to 0.0101, in the professional skills dimension ranged from 0.0026 to 0.0127, in the brainstorming skills dimension ranged from 0.0059 to 0.0132, and in the competitiveness dimension The entropy results on the competitiveness dimension ranged from 0.0083 to 0.0123.

Figure 7.

The result of the problem of the problem of the skill of the lack of the skills

The results of the entropy value of the imaginative (IS) students in each dimension are shown in Figure 8. The entropy value results of Imaginative Students (IS) in the theoretical knowledge dimension ranged from 0.0086 to 0.0139, the entropy value results in the practical skills dimension ranged from 0.0022 to 0.0169, the entropy value results in the professional skills dimension ranged from 0.0065 to 0.0131, the entropy value results in the brainstorming skills dimension ranged from 0 to 0.0118, and the entropy value results in the competitiveness dimension the entropy results ranged from 0 to 0.0123.

Figure 8.

The entropy of the imaginative student in each dimension

The results of the final scores of various groups of senior students are shown in Figure 9. The results show that the entropy values of various groups of student portraits are small, which indicates that the classification results are not prone to change and are relatively stable. Even the individual dimensions appeared to have an entropy value of 0, indicating that the weight distribution of each dimension is very clear and the contribution of these dimensions to the goal is consistent. The weights of each evaluation dimension are obtained through calculation: the theoretical knowledge dimension is 0.0970, the practical ability dimension is 0.5796, the professional skills dimension is 0.097, the brainstorming skills dimension is 0.1212, and the competitiveness dimension is 0.1216. Balanced developmental students had the highest final score range (0.769) while students to be strengthened had the lowest final score range (0.344). The final score mean of the skill-deficient students (0.449) was slightly higher than that of the numerically conservative students (0.407). The results are consistent with the weights assigned to the dimensions, which suggests a more accurate classification of the student clusters based on their characteristics, and reasonable clustering results.

Figure 9.

The results of the final scores of the students in higher vocational colleges

This study used SPSS22.0 software to compare whether there is a significant difference in the cultivation of innovation and entrepreneurial ability between spring 2022 and 2021 data using an independent samples t-test. A virtual simulation experimental course on innovation and entrepreneurship offered by higher vocational colleges and universities was awarded as a first-class undergraduate course of the national online and offline blending in 2022, and more than 100 colleges and universities came to exchange and learn from the virtual simulation experimental teaching experience. In this study, an online questionnaire survey was conducted in senior vocational college A as an example. From September 2021 to June 2022, 4985 questionnaires were received. After data cleaning and excluding duplicates of answers, there were 4,670 valid questionnaires, including 1,348 in the spring of 2021, 1,659 in the fall of 2021, and 1,663 in the spring of 2022. Their mean scores for entrepreneurship and innovation were 4.297 and 4.3419, respectively, and their mean standard deviations were 0.681 and 0.6433, respectively.

The results of the t-tests for innovativeness and entrepreneurship for the 2021 and 2022 survey data are shown in Table 3. Both innovation ability and entrepreneurship ability passed the significance test in 2021 and 2022, indicating that they are significantly different. The above comparative analysis shows that there is a significant difference between the virtual simulation experiment for the cultivation of students’ innovation and entrepreneurship ability during the epidemic and before the epidemic, but the teaching effect in spring 2022 is very close to the effect in 2021 and within the acceptable range.

Next, this study estimates the results of the model test for spring 2022 and 2021 using multi-group comparisons to specifically analyze the differences in the relationships between variables at different times. The results of the regression coefficient test for each path of the model for the school year 2021 are shown in Table 4. It can be seen that the relationship between the effects of each variable in 2021 passed the significance test. In terms of the impact on knowledge acquisition, the experimental design variable has the greatest impact, followed by blended teaching and teamwork, with little difference in the impact of the latter two. Mixed instruction had a greater direct effect on innovation than on entrepreneurship; teamwork had a greater direct effect on entrepreneurship than on innovation; and overall, teamwork had a greater direct effect on innovation and entrepreneurship than mixed instruction. Knowledge acquisition had a greater effect on entrepreneurial competence than on innovative competence.

The results of the test of regression coefficients for each path of the spring 2022 model are shown in Table 5. It can be seen that the effects of blended instruction on knowledge acquisition did not pass the test. The effect of blended instruction on entrepreneurship passed the test at the 0.01 level, but did not pass the test at the 0.001 level. Experimental design had the greatest impact on knowledge acquisition, followed by teamwork. The direct effect of blended instruction on innovation competence was greater than the direct effect on entrepreneurship competence, and the direct effect of teamwork on entrepreneurship competence was greater than the direct effect on innovation competence. Overall, teamwork had a greater direct effect on innovation and entrepreneurship than blended instruction, and knowledge acquisition had a greater impact on entrepreneurship than on innovation. The comparison between spring 2022 and 2021 shows that for the impact of knowledge acquisition, experimental design was the most influential in both cases, and mixed instruction was less influential in spring 2022 than in 2021. In terms of the impact of knowledge acquisition on innovation and entrepreneurship, 2021 had a lesser impact than spring 2022 on innovation competence, and 2021 had a higher impact than spring 2022 on entrepreneurship competence. All aspects of the impact of blended instruction on innovation and entrepreneurship competencies were better in 2021 than in spring 2022, and all aspects of the impact of teamwork on innovation and entrepreneurship competencies were better in spring 2022 than in 2021.