Mobile QR Code QR CODE

2024

Acceptance Ratio

21%


  1. (Jiangsu university of Technology, Changzhou 213001, China)
  2. (Changzhou Vocational Institute of Engineering, Changzhou 213001, China Jianbin_Zhen@outlook.com)



Association rules, RBF neural network, Employment forecast, Correlation analysis

1. Introduction

With the vigorous development of the Internet and computers, as well as the improvement of people's awareness of information technology promoting teaching, the update of analytical technology in the field of education is continuously updating and advancing at a rapidly changing speed [1]. The employment situation of students is a key issue of concern in the field of education, which has attracted numerous researchers to construct various prediction models from different perspectives. These models also take into account the differences and characteristics of the educational environment. Although various prediction models have their own characteristics and excellent results, there are limitations in adaptability and generalization, and the accuracy of the models also needs to be improved. Existing models often overlook the correlation between the various projects involved in prediction, which to some extent restricts the accuracy and reliability of the model. There are certain shortcomings in achieving the accuracy and effectiveness of model prediction [2,3,4]. Therefore, it is an urgent need to construct an efficient learning and employment prediction model that has broad adaptability and can enable educational researchers to predict from multiple perspectives based on the characteristics of their own educational environment.

Based on the above development background, educational researchers and students hope to use learning analysis technology to better enhance the sense of learning experience and teaching experience [5]. By utilizing learning analysis technology and educational data mining technology, while promoting the continuous optimization and improvement of the learning environment, better improvements can also be made to the problem of low teacher-student ratio in traditional education systems. At the same time, appropriate models can be established to achieve predictive analysis of student employment rates. At present, artificial neural networks have made some widespread attempts in educational prediction, and the prediction results obtained by Artificial neural network (ANN) can provide decision-maker suggestions for educators to improve students' academic performance [6,7]. A scholar has designed a scholarship allocation model that allows the school admissions office to maximize the motivation for students' academic performance [8]. Their research proposes a design model for a multi-layer feedback structured student output neural network. Research has shown that the application of artificial neural networks in the field of educational prediction has achieved satisfactory results, and the ANN method has better accuracy compared to traditional linear regression prediction methods. Scholars have established some dynamic models to analyse and predict students' satisfaction with courses. The methods used in the models include Artificial neural network algorithm and statistical analysis techniques. Researchers use artificial neural networks to predict the online course selection behaviour of high school students. The research focuses on exploring potential factors that affect students' satisfaction with online course selection, establishing a model for students' course selection behaviour, and using Artificial neural network algorithms to fit the training data set to train the model [9]. After the course selection period ends, the obtained model is used to predict the final registration number for each course [10,11]. In addition, some scholars have proposed a redesigned RBF neural network that can fit the model due to its hybrid learning method, which uses a genetic algorithm to calculate the centre matrix and Markov distance to construct the maximum decision coefficient R2. This statistic becomes an evaluation function in genetic algorithms, improving the accuracy of RBF [12].

Based on the analysis of existing learning models, it can be seen that RBF neural networks usually independently intervene in predicting content, lacking optimization analysis for specific problems. This article proposes an employment prediction model based on association rules and RBF neural network algorithm, and based on the application characteristics of this model, predicts and analyses students' academic performance. On the basis of knowing the correlation between predicted data and results, this model changes the selection method of data centres to probabilistic random selection, which can significantly improve the learning speed of the model. Through the analysis of this model in employment and performance prediction, the differences in learning speed and accuracy of optimized neural network algorithms were further studied using baseline experiments. Through comparative analysis, it can be found that the employment prediction model constructed in this article can effectively improve the prediction accuracy of the model, providing more theoretical guidance for the prediction and analysis of employment data.

2. Association Rules and Optimization of RBF Neural Network Theory

2.1 Data Mining and Association Rule Analysis

2.1.1 Data Mining

There is a significant difference between traditional data analysis methods and data mining. The most fundamental manifestation of the difference between the two is that data mining is the process of mining and analysing relevant data information without clear assumptions, in order to discover relevant internal correlations. In addition, the data obtained during the data mining process typically has three very important characteristics, namely unknowingness, effectiveness, and practicality. The meaning of unknown characteristics refers to the fact that the information mined is unexpected, which means that through data mining, relevant information or knowledge content that cannot be predicted before can be discovered. The mining information presented during this process may deviate from conventional judgment logic to a certain extent, or even exceed imagination. However, the greater the unpredictability of information, the more valuable and high-quality data information may be realized or excavated [13,14,15].

Based on the above analysis, it can be seen that the process of data mining is an important process in determining whether the information we obtain is effective. Therefore, it is necessary to control the quality of data mining. Usually, the quality of data mining is influenced by two main factors: firstly, the feasibility and effectiveness of relevant data mining techniques in the information mining process; The second is the quality and quantity of basic data, which requires special attention to the quality and quantity of data used for mining. Unreasonable selection or conversion of basic data used for data mining can lead to poor quality of data mining. On the contrary, if the selection of basic data is reasonable and reasonable data conversion forms can be adopted, it can provide strong guarantees for high-quality data mining results [16,17]. As shown in Fig. 1, a basic flowchart of data mining is provided.

Fig. 1. Basic process diagram of data mining.

../../Resources/ieie/IEIESPC.2025.14.1.33/image1.png

2.1.2 Association Rule Analysis

The current hot field of association rule analysis is concentrated in online education, and in terms of application issues, it mainly focuses on learning evaluation and teaching evaluation. By exploring the influencing factors related to academic performance, a comprehensive evaluation of learning and teaching can be formed from multiple aspects. The first stage of the model proposed in this article is the association rule mining stage, which aims to achieve better prediction by mining the factors associated with the prediction results [18].

Association rule mining can usually be divided into two steps. Firstly, the excavation of high-frequency projects needs to be combined with data analysis to determine the relevant high-frequency project groups; Secondly, based on the discovered high-frequency project group, analyse the relevant data information, and then discover association rules. A high-frequency project group refers to a project group that appears at a higher frequency or reaches a certain level compared to all recorded transactions. Such a project group is called a high-frequency project group and needs to be discovered first. With a containing $\alpha$ and $\beta$ for example, the 2-item-set of two projects can be obtained by formula (1), which includes $\{\alpha$, $\beta\}$ The support level of the project team, if the support level is greater than or equal to the set minimum support level, then $\{ \alpha$, $\beta \}$ The project team is called the high-frequency project team. A $k$-project group that meets the minimum support level is called a high-frequency $k$-project group. And according to the usual algorithm, it will continue to excavate and generate large $k+l$ or frequency $k+1$ from the project group of large $k$ or frequent $k$, looping forward until a longer high-frequency project group cannot be found.

(1)
$ Support(\alpha \Rightarrow \beta )=\frac{Freq(\alpha \cap \beta )}{N} . $

Among them, $Freq(\cdot)$ represents the number and size of transactions contained in the project, and N represents the number of all transactions. In this process, the generation of association rules occurs in the second step of association rule analysis. Association rules are generated from high-frequency project groups and utilize the frequency k project groups generated in the previous stages. If a rule meets the minimum confidence threshold requirement, then the rule is called an association rule [19].

The calculation formula for confidence is also based on the analysis of the $ \{ \alpha$,$\beta \}$, project group. The calculation formula for the confidence of the association rules generated by it is shown in formula (2):

(2)
$ Confidence(\alpha \Rightarrow \beta )=\frac{Freq(\alpha \cap \beta )}{Freq(\alpha )} . $

Another important parameter for measuring association rules is the degree of improvement, which is used to determine whether the rules have practical value. If it is greater than 1, the rule is valid, and if it is less than 1, it is invalid. The calculation process is given by formula (3):

(3)
$ Lift(\alpha \Rightarrow \beta )=\frac{Support(\alpha \Rightarrow \beta )}{Support(\alpha )\times Support(\beta )} . $

2.2 RBF Neural Network Theory

The basic structure of RBF neural networks is similar to that of forward neural networks, with a three-layer forward neural network as the main structure. Firstly, the first layer is the input layer, which is composed of signal source nodes. The second layer is the hidden layer, in which the problem to be described determines the size of node data, meaning that the number of nodes depends on the problem described [20,21]. The third layer structure is the output layer, whose function is to make corresponding responses based on the input node. In this process, the transformation from input space to hidden layer space is based on nonlinear transformation, while the transformation from hidden layer to output layer is a linear transformation process, which is a topology structure of feedforward networks. The transformation process of the hidden layer unit adopts a radial basis function, which is a topological structure and exhibits a non-negative nonlinear function RBF neural network. In addition, the local distribution of this function exhibits a symmetric attenuation characteristic towards the centre point radially [22].

There are many methods for analysing the mapping or fitting relationship between student employment rate and multivariate factors. Considering that regression and other methods mostly assume a certain correlation between employment rate and variable factors, this article considers the Artificial neural network method that can approach nonlinear functions with arbitrary accuracy to construct a model [23,24]. Due to different excitation functions, different ANN network structures are determined. Given the characteristics of multiple variable factors and large data volume in this article, there is a high requirement for convergence speed during operation. Therefore, a radial basis function neural network (RBF) with the characteristics of optimal approximation, global optimal performance, and fast learning convergence speed is selected in this article. As shown in Fig. 2, a schematic diagram of the structure of the RBF neural network is provided.

Fig. 2. Structure diagram of RBF neural network.

../../Resources/ieie/IEIESPC.2025.14.1.33/image2.png

3. Construction of Employment Model Based on Association Rules and Optimized RBF Neural Network

3.1 Hybrid Prediction and Optimization Algorithm

3.1.1 Basic Structure of Hybrid Prediction Model

To address the issue of correlation between multiple influencing factors and prediction results in employment prediction, this study proposes a method for selecting influencing factors in prediction models based on association rules [25]. On the basis of traditional expert analysis and field observation methods, the selection of influencing factors in the prediction model is further determined by exploring strong association rules between influencing factors and learning results, thereby discovering a scientific and reliable set of influencing factors. Due to the higher accuracy of this influencing factor selection mode and its ability to greatly improve the efficiency of model design, reduce the probability of unsatisfactory model prediction results, and reduce repeated prediction experiments to prove that the waste of resources in the prediction project set can better prevent the risk of both favourable and unfavourable factors in the remaining projects.

In the model proposed in this article, unlike the previous model, after determining the potential set of related items based on expert opinions or actual research, these items are not directly used for prediction. Instead, association rules are applied between these items and the predicted items to obtain association rules, support, confidence, and other parameters [26], in order to ensure the validity and rationality of the data. Then, select projects with high correlation as prediction inputs, and use the obtained support and confidence as parameters for optimizing the prediction algorithm to analyze the prediction results. The prediction model designed based on this pattern ensures the quality of the set of projects involved in the prediction. Therefore, compared to previous models, repeated experiments can be avoided to verify whether the projects involved in the prediction can serve as effective inputs for the prediction results, ensuring the quality of the model and eliminating the tedious verification process. The framework diagram of the prediction model in this article is shown in Fig. 3.

Fig. 3. Framework diagram of prediction model.

../../Resources/ieie/IEIESPC.2025.14.1.33/image3.png

3.1.2 Optimization algorithm for selecting RBF neural data centres

The selection of data centres is usually done randomly from the samples, or by using multiple clustering centres from the training sample set, such as the k-means algorithm to cluster and collect more data centres in densely populated areas [27,28]. This article proposes a data centre selection method based on association rules: in the first stage of association rule mining, factors with strong correlation with predicted results have been identified, and specific values of support S and confidence C have been obtained. The significance represented by support and confidence is the degree of correlation between these input values and the predicted results [29]. The higher the value of a certain support and confidence level, the closer the value of the predicted result is to the value of the input sample. Therefore, if the data centre of the neuron is selected near these strongly correlated input samples, there is a greater chance of the neuron being closer to the predicted result, and samples that are also strongly correlated with the predicted result have a greater chance of being activated, this can effectively improve learning speed and prediction accuracy [30,31]. From the following two figures, it can be seen more vividly and intuitively that the selection of data centers has an impact on learning. Fig. 4 shows the image of a radial basis function with two-dimensional input. The right figure shows that the RBF network activates sample points only by nearby inputs. The center point of the circle in Fig. 5 is the data center of the neuron, and only the input within the circle can be activated, and the points closer to the data center are more easily activated. Therefore, it can be recognized that an excellent strategy for selecting a data center should be based on the relationship with the output sample.

Fig. 4. 2D input radial basis function.

../../Resources/ieie/IEIESPC.2025.14.1.33/image4.png

Fig. 5. RBF network activation range.

../../Resources/ieie/IEIESPC.2025.14.1.33/image5.png

Based on this, support and confidence are used as the basis for selecting data centres, and a probabilistic random selection method is used to select data centres, as shown in formula (4):

(4)
$ P_{C} (X)=Support(X)\times Confidence(X) . $

If the optimized data centre selection method is used to select the centre, assuming that f optimal data centres are input into the sample according to the probability distribution of P${_{C}}$, that is, the probability of the optimal data centre distribution in the input sample of project is P${_{c}}$(X${_{n}}$), then the probability of idealized selection of the optimal data centre is shown in formula (5):

(5)
$ P=\sum _{i=1}^{n}P_{C} (X_{n} )\times \frac{f}{m} . $

Among them, $P_{c} ( X_{i} )$ represents the probability of the first sample being selected as the data centre among $n$ inputs, $m$ represents the number of data. In addition, for analysing different probabilities, the formula (6) can be obtained:

(6)
$ P_{c} (X_{1} )+P_{c} (X_{2} )+...+P_{c} (X_{n} )=1 . $

3.2 Analysis and Construction of Employment Prediction Model

The structure of Radial Basis Function (RBF) neural network is based on multivariate interpolation of radial basis functions. By applying this method to neural network design, the RBF neural network was constructed. The prominent feature of RBF neural networks is mainly reflected in the output function of hidden layer neurons being defined as a basis function with radial symmetry [32,33]. Based on the analysis of the basic structure, it can be seen that the input layer to hidden layer units of the neural network have a nonlinear mapping relationship, while the process from hidden layer to output layer has a linear weighted summation relationship. Through the structural analysis of the RBF neural network, its mathematical expression is shown in formula (7):

(7)
$ \stackrel{\rightharpoonup}{f}\left[\stackrel{\rightharpoonup}{x}(n)\right]=\sum _{i=1}^{m}w_{i} y =\stackrel{\rightharpoonup}{W}^{T} \stackrel{\rightharpoonup}{Y} . $

For the output of RBF networks, formula (8) can be used to describe:

(8)
$ \hat{y}=\sum _{i=0}^{m}w_{i} \exp \left[-\frac{\left\| X-C_{i} \right\| ^{2} }{2\delta _{i}^{2} } \right] . $

In the formula, ${m}$ is the number of hidden layer nodes; $w_{i}$ represents the weight coefficient from the $i$-th hidden layer node to the output node; $C_{i}$ and $\delta_{i}$ are the centres and widths of the Gaussian function of the second hidden layer node.

Using unsupervised self-organizing learning method to select the centre $C_{i}$ and $\delta_{i}$. Based on the k-means clustering algorithm, $C_{i}$ and $\delta _{i}$ can be determined using a fixed method. When the centre is determined by training data, the width of the RBF can be determined by formula (9):

(9)
$ \delta _{i} =\frac{d_{m} }{\sqrt{2M} }. $

Among them, $d_{m}$ is the maximum distance of all classes; M is the number of RBF centres. The weight $w_{i}$ from the hidden layer to the output layer of the network is determined using the gradient descent method. Consider the sum of squares error function, as shown in formulas (10) and (11):

(10)
$ E=\sum _{p=1}^{s}E_{p} =\frac{1}{2} \sum _{p=1}^{s}\left[\hat{y}(X_{p} )-y(X_{p} )\right]^{2},\\ $
(11)
$ \Delta w_{ik} =-\eta \frac{\partial E}{\partial w_{ik} } =-\sum _{p=1}^{s}\frac{\partial E_{p} }{\partial w_{ik} } \\ \hskip 1.7pc =-\eta \sum _{p=1}^{s}\frac{\partial E_{p} }{\partial \hat{y}(X_{p} )} \times \frac{\partial \hat{y}(X_{p} )}{\partial E_{p} } . $

Among them, $\eta$ is the learning rate, with a value range of (0$\mathrm{<}$$\eta$$\mathrm{<}$1) According to the above process, the final adjustment amount for each step of the weight value can be obtained, as shown in formula (12):

(12)
$ \Delta w_{ik} =-\eta \sum _{p=1}^{s}\left[\hat{y}(X_{p} )-y(X_{p} )\right] \exp \frac{-\left\| X-C_{i} \right\| ^{2} }{2\delta _{i}^{2} } . $

According to the gradient descent method, at a certain time t, the adjustment of network parameters can be carried out using the following formula. As shown in formula (13):

(13)
$ \left\{\begin{aligned} w_{ij}^{t+1} =w_{ij}^{t} -\eta _{w} \frac{\partial E_{t} }{\partial w_{ij}^{t} },\\ c_{j}^{t+1} =c_{j}^{t} -\eta _{c} \frac{\partial E_{t} }{\partial c_{j}^{t} },\\ \delta _{j}^{t+1} =\delta _{j}^{t} -\eta _{c} \frac{\partial E_{t} }{\partial \delta _{ij}^{t} } .\end{aligned}\right. $

In this process, if ${N}$ vectors $R_{i}$, (${i} =1$, $2$, ..., $N_{h}$) are regarded as ${N}$ points in the $R_{s}$ space, this similarity can be described by the distance between points. When the distance between two points is close, the difference between the two vectors is small. The merged neurons ${i}$ and ${j}$ satisfy the formula (14):

(14)
$ H_{i,j} =\min \left\{-\eta \sum _{p=1}^{s}\left\| R_{i} (X_{p} )-R_{i} (X_{p} )\right\| ^{2} \right\} . $

In the above equation, the $R_{i} (X)$ in the $R_{s}$ space can be described using formula (15):

(15)
$ R_{i} (X)=\exp \left[-\frac{\left\| X-C_{i} \right\| ^{2} }{2\delta _{i}^{2} } \right] . $

4. Analysis of the Experimental Results of the Employment Prediction Model

4.1 Predictive Analysis of Academic Performance

Before analyzing the employment forecast data, the model was first validated and analyzed by predicting academic performance. In the process of predicting and analyzing academic performance through the constructed model, the data for model training and performance prediction was based on relevant data from 3000 junior high school students, and the grades of students in different grades were classified and organized to eliminate the influence of invalid data. In addition, this article combines relevant data and randomly selects data from 100 students to analyze the predictive performance of the model (to avoid the influence of gender factors, 50 males and 50 females were selected in the sample). Combined with their relative error and absolute comparison, the basic performance of the model is further analyzed.

Based on the above data preparation, in order to verify the effectiveness and accuracy of the employment model based on association rules and optimized RBF neural network, the actual grades of 100 students were determined according to randomly selected rules. The initial RBF neural network model and the optimized RBF neural network employment prediction model were used for prediction and analysis, and the relative and absolute errors of the two models were compared based on the prediction data of the two models, the specific experimental comparison results are shown in Fig. 6. From Fig. 6, it can be seen that the relative and absolute errors of the predicted values of learning scores given by the optimized RBF neural network prediction model are relatively small compared to the initial RBF neural network prediction model. At the same time, the error distribution is relatively stable and the volatility is relatively small. For example, the fluctuation range of the relative error of the optimized RBF neural network prediction model is between 0.1 and 0.3, while the fluctuation range of the relative error of the initial RBF neural network prediction model is between 0.15 and 0.42. In addition, for absolute error, the fluctuation range of the former is basically between 3-18, while the fluctuation range of the latter is relatively large, basically between 3-23. Therefore, it can be seen that using the optimized RBF neural network prediction model for predicting student grades can achieve higher prediction accuracy and accuracy, and the error is small.

Fig. 6. Comparison of error between initial RBF prediction value and optimized RBF employment prediction model prediction value.

../../Resources/ieie/IEIESPC.2025.14.1.33/image6.png

Under the same other conditions, the performance of the employment prediction model based on association criteria and optimized RBF neural network constructed in this paper was compared and analysed with other models. At the same time, the algorithm model of this model was compared with other similar neural network models, as shown in Tables 1 and 2.

From the Tables, it can be seen that in the dataset, the optimized RBF neural network prediction model exhibits certain advantages over the other four neural networks of the same type in five performance indicators, with the highest accuracy of around 87%, while the accuracy and recall are both around 85%. The average decision error is also higher than the other four models, reaching over 80%. At the same time, by comparing with other indicators, the LM neural network algorithm and XGBoost algorithm have better predictive performance than the initial RBF neural network and BP neural network. However, overall, it can be concluded that the optimized RBF neural network prediction model performs better in other aspects. Therefore, it can be concluded that among the same type of neural networks, the optimized RBF neural network is more suitable for current academic performance prediction work.

Compared with traditional prediction algorithms, the optimized RBF neural network has significantly higher performance indicators in various aspects compared to traditional performance prediction methods. This further verifies that this method has high accuracy and universality in both the same type of prediction methods and traditional prediction methods.

Table 1. Performance comparison of different models.

Model type

Evaluating indicator

Precision

Recall

Accuracy

MAE

F1-score

Back Propagation model

0.76

0.68

0.65

0.65

0.62

Initial RF model

0.78

0.72

0.71

0.70

0.63

Expected Goals model

0.81

0.80

0.78

0.76

0.67

Levenberg-Marquardt model

0.83

0.81

0.80

0.76

0.68

Model in this article

0.88

0.84

0.86

0.85

0.80

Table 2. Performance comparison of different algorithms.

Algorithm type

Evaluating indicator

Precision

Recall

Accuracy

MAE

F1-score

Decision Tree algorithm

0.72

0.67

0.74

0.63

0.70

Support Vector Machines algorithm

0.74

0.71

0.76

0.67

0.73

Naïve Bayes algorithm

0.77

0.00

0.76

0.73

0.77

Logistic Regression algorithm

0.81

0.83

0.83

0.74

0.82

Model in this article

0.88

0.84

0.86

0.85

0.80

4.2 Employment Forecast Analysis

4.2.1 Mining Association Rules

Take the organized transaction dataset as input and analyse it using the Apriori algorithm written in RStudio. The purpose of mining association rules in this study is to use information from talent cultivation plans and actual training record data to assist in predicting employment situations. Therefore, the leader was selected as the project in the talent cultivation plan, followed by the project in the employment situation. Specific association rules were determined and summarized and analysed based on the degree of correlation. The association rules mined are only applicable to the talent cultivation plan and graduate employment situation of a certain school, college, or major. Due to the differences in the curriculum structure and talent cultivation structure of the school, as well as the different nature of each major, the results of this association rule cannot be used as the association rule results of other schools or majors. Therefore, corresponding association rules should be mined based on the specific environment and data characteristics of the application.

After determining the association rules, enter the second stage, which is the prediction stage. At this stage, an optimized RBF neural network model is used for prediction. The parameters that need to be determined in this stage include the input of the model, the probability of each sample being selected as the neural data centre, and the output of the model. Firstly, set the dataset for network training: In this example, due to the sparsity of the data and the limited dataset, the interpolation method is used to interpolate the experimental data to increase the amount of training data, and then improve the performance of the network through multiple training sessions.

4.2.2 Model Prediction Data Analysis

When predicting employment data under different influencing factors, combined with the relevant relevance of association rule mining, a correlation analysis was conducted on the relationship between academic performance, internship duration, book borrowing, and professional competition scores in the basic data and employment rate. As shown in Fig. 7, the changes between different influencing factors and employment rate were presented through the processing of normalized data. From Fig. 7, it can be seen that among the four influencing factors analyzed above, there is a nearly linear fit between book borrowing volume and employment rate. When the normalized value of the borrowing volume factor is 0.6, the normalized value of the employment rate is about 0.7. When the borrowing volume increases to 0.9, the normalized value of the employment rate increases to 0.92, which means that the normalized value of the employment rate increases by 31.4%. In addition, for the analysis of the other three influencing factors, it can be seen that all three show a change relationship that first increases and then tends to flatten out, and the normalized value of the employment rate is basically stable between 0.9 and 0.95. Therefore, by analysing the selected influencing factors, the employment model constructed in this article can be further combined to analyses relevant factors.

Fig. 7. Comparison of effects of different factors on changes in employment rate.

../../Resources/ieie/IEIESPC.2025.14.1.33/image7.png

Fig. 8. Visual correlation analysis between different projects.

../../Resources/ieie/IEIESPC.2025.14.1.33/image8.png

In addition, in order to further analyse the correlation between different parameter projects in the process of predicting employment rate using a prediction model based on association rules and optimized RBF neural network, as shown in Fig. 8, the correlation between different parameter projects in the model prediction process is presented. The cross grid of the horizontal and vertical coordinates corresponding to two projects represents their correlation. The darker the colour, the larger the shaded area within the circle, indicating stronger correlation. From the graph, it can be seen that the relationships between these projects exhibit a positive correlation. Based on the analysis, it can be seen that there is a strong correlation between academic performance and competition performance, while the weaker correlation is between internship duration and academic performance.

5. Conclusions

The development of the Internet and computer technology has promoted the application and promotion of different big data models in employment prediction, providing more diversified analysis methods for employment rate analysis. Based on association rules and optimized RBF neural network algorithm, this paper constructs an employment prediction model based on association rules and optimized RBF neural network, and tests the effectiveness of the model by predicting students' academic performance. At the same time, the basic performance of this model was compared with other models, and further research was conducted on the changes in different influencing parameters of this model in employment prediction. The correlation between various projects was analysed, and visual analysis was conducted. The main conclusions obtained are as follows:

Firstly, by combining association rules and optimizing the RBF neural network, the constructed employment prediction model can achieve more advantageous prediction performance for academic performance. After optimization, the RBF neural network's prediction model exhibits certain advantages over the other four neural networks of the same type in five performance indicators, with the highest accuracy of around 87%, while the accuracy and recall rates are both around 85%. The average decision error is also higher than the indicators of the other four models, reaching over 80%.

Secondly, the prediction results of this model for employment show that among the influencing factors, when the normalized value of the borrowing volume factor increases from 0.6 to around 0.7, the normalized value of the employment rate increases from 0.7 to 0.92, an overall increase of 31.4%. The impact of academic performance, competition performance, and internship duration on the prediction of employment rate shows a trend of first increasing and then flattening out, with the normalized values of employment rate basically stable between 0.9 and 0.95.

Thirdly, the constructed RBF neural network optimization algorithm based on association rules optimizes the initialization selection method for the data centre of hidden layer neurons. The structure and algorithm of RBF determine that parameters play an important role in the learning of RBF neural networks. In the hidden layer, the data centres and standardization constants of each basis function are also known as extension constants. In the output layer, the third important parameter is the weight of the output node. The optimization strategy in this article is aimed at the data centre, and there are certain shortcomings in optimizing the other two important parameters. In future research, optimization of neural network algorithms from other perspectives will be considered to further improve the performance of prediction models.

Funding

Research on the Linkage Mechanism of College Enrollment, Employment, and Talent Cultivation based on the concept of Outcome Based Education , the special subject of ideological and political education of philosophy and social science research among Jiangsu provincial colleges and universities in 2023 (Project No.2023SJSZ0682); Research on double cycle and triple action mechanism of College Enrollment, Training and Employment based on OBE concept, research project on employment and entrepreneurship of college graduates in Jiangsu Province in 2021 (Project No.JCKT-A-20210101).

REFERENCES

1 
M. L. Manuel, E. S. Antonio, A. I. Juan, et al., ``Network untrusion detection based on extended RBF neural network with offline reinforcement learning,'' IEEE Access, vol. 9, pp. 153153-153170, 2021.DOI
2 
H. Wu, Y.-P. Zhao, and H.-J. Tan, ``Novel radial basis function network based on dynamic time warping and Kalman filter for real-time monitoring of supersonic inlet flow patterns,'' Journal of Aerospace Engineering, vol. 34, no. 4, 2021.DOI
3 
X. Zheng, Y. Zhang, H. Zhang, et al., ``An RBF neural network–based dynamic virtual network embedding algorithm,'' Concurrency and Computation: Practice and Experience, vol. 31, no. 23, 2019.DOI
4 
Y. P. Goh, C. S. Tan, P. W. Cheah, et al., ``Reducing the complexity of an adaptive radial basis function network with a histogram algorithm,'' Neural Computing and Applications, vol. 28, no. 1s, pp. 365-378, 2017.DOI
5 
X. Zheng, Y. Zhang, H. Zhang, et al,. ``An RBF neural network–based dynamic virtual network embedding algorithm,'' Concurrency and Computation: Practice and Experience, vol. 31, no. 23, 2019.DOI
6 
] A. Guillé, H. Pomares, I. Rojas, et al., ``Studying possibility in a clustering algorithm for RBFNN design for function approximation,'' Neural Computing and Applications, vol. 17, no. 1, pp. 75-89, 2008.DOI
7 
J. H. McGrew, J. K. Johannesen, M. E. Griss, D. L. Born, and C. H. Katuin, ``Performance-based funding of supported employment for persons with severe mental illness: Vocational rehabilitation and employment staff perspectives,'' The Hournal of Behavioral Health Services Research, vol. 34, no. 1, pp. 1-16, 2007.DOI
8 
L. Beatrice, R. Stuart, R. Antonio, et al., ``The association between hope and employment among individuals with multiple sclerosis: A hierarchical logistic regression model,'' Work (Reading, Mass.), vol. 74, no. 2, pp. 531-538, 2022.DOI
9 
F. Doménech-Betoret, A. Gómez-Artiga, and L. Abellán-Roselló, ``The educational situation quality model: A new tool to explain and improve academic achievement and course satisfaction,'' Frontiers in Psychology, 101692, 2019.DOI
10 
G. Revell, D. O'Brien, H. J. McGrew , et al., ``Performance-based funding of supported employment: A multi-sitecontrolled trial,'' Journal of Vocational Rehabilitation, vol. 23, no. 2, pp. 81-99, 2005.DOI
11 
A. M. Shokouhi, M. Sadeghi, and R. Asadi, ``Analysis, modeling and optimization of regional development in Sistan Balochestan combining labor market, shift-share and genetic algorithm models,'' Management Science Letters, vol. 3, no. 1, pp. 281-290, 2013.DOI
12 
Y. Wang and D. Yang, ``The design of psychological education intervention system in universities based on deep learning,'' Computational Intelligence and Neuroscience, 20229208172-9208172, 2022.DOI
13 
S. M. Dol and P. M. Jawandhiya, ``Classification technique and its combination with clustering and association rule mining in educational data mining - A survey,'' Engineering Applications of Artificial Intelligence, 122, 2023DOI
14 
M. Amini, J. Rezaeenour, and E. Hadavandi, ``A neural network ensemble classifier for effective intrusion detection using fuzzy clustering and radial basis function networks,'' International Journal on Artificial Intelligence Tools, vol. 25, no. 2, 2016.DOI
15 
M. Siddique and P. S. Mishra, ``Prophecy of stock price of Bharat immunological biological corporation Ltd using hybrid ANN and PSO model,'' Journal of Economics, Management and Trade, pp. 27-35, 2021.DOI
16 
W. Li, ``An association mining method based on LSTM algorithm for massive online news hotspots,'' Computer Informatization and Mechanical System, vol. 6, no. 6, pp. 115-119, 2023.DOI
17 
K. D. Ramakrishna, I. D. Verma, I. Goyal, et al., ``Artificial intelligence: Future employment projections,'' Journal of Critical Reviews, vol. 7, no. 5, 2020.URL
18 
F. Petry and R. Yager, ``Data mining using association rules for intuitionistic fuzzy data,'' Information, vol. 14, no. 7, 2023DOI
19 
R. A. G. Locateli and E. A. S. D. Antunes, ``Short-term relation between air pollutants and hospitalizations for respiratory diseases: Analysis by temporal association rules,'' Environmental Monitoring and Assessment, vol. 195, no. 7, 2023.DOI
20 
F. Iztok, F. Iztok, F. Dusan, et al., ``A comprehensive review of visualization methods for association rule mining: Taxonomy, challenges, open problems and future ideas,'' Expert Systems with Applications, 233, 2023.DOI
21 
W. Yu, F. Zhao, Z. Ren, D. Jin, X. Yang, and X. Zhang, ``Mining attention distribution paradigm: Discover gaze patterns and their association rules behind the visual image,'' Computer Methods and Programs in Biomedicine, 230, 2023.DOI
22 
A. Abdirahman and P. Marek, ``Quantum algorithm for mining frequent patterns for association rule mining,'' Journal of Quantum Information Science, vol. 13, no. 1, 2023.DOI
23 
A. A. A. Dahab, M. R. Haggag, A. A. S. Fotouh, ``Enhancing customer relationship management using fuzzy association rules and the evolutionary genetic algorithm,'' International Journal of Advanced Computer Science and Applications (IJACSA), vol. 14, no. 4, 2023.DOI
24 
K. Biresh, R. Sharmistha, S. Anurag, et al., ``E-commerce website usability analysis using the association rule mining and machine learning algorithm,'' Mathematics, vol. 11, no. 1, 2022.DOI
25 
Z. Jintan, ``Discussion on redundant processing algorithm of association rules based on hypergraph in data mining,'' Journal of Robotics, 2022.DOI
26 
M. A. Hogo, ``The design of academic programs using rough set association rule mining,'' Applied Computational Intelligence and Soft Computing, 2022.DOI
27 
A. C. Faiyaz, P. Shahrear, A. R. Shamim, et al., ``Comparison of different radial basis function networks for the electrical impedance tomography (EIT) inverse problem,'' Algorithms, vol. 16, no. 10, 2023.DOI
28 
J. He, Y. Dong, J. Wnag, and Y. li, ``Intelligent optimization models based on hard-ridge penalty and RBF for forecasting global solar radiation,'' Energy Conversion and Management, vol. 95, 2015.DOI
29 
U. Ravale, N. Marathe, and P. Padiya, ``Feature selection based hybrid anomaly intrusion detection system using K means and RBF kernel function,'' Procedia Computer Science, vol. 45, 2015.DOI
30 
Z. Huang and F. Liu, ``Prediction model of college student’s entrepreneurship ability based on artificial intelligence and fuzzy logic model,'' Journal of Intelligent Fuzzy Systems, vol. 40, no. 2, 2021.DOI
31 
K. Pang, S. Li, Y. Lu, N. Kang, L. Zou, and M. Lu, ``Association rule mining with fuzzy linguistic information based on attribute partial ordered structure,'' Soft Computing, vol. 27, no. 23, pp. 135-148, 2023.DOI
32 
G. R. S. Husain, ``Apriori algorithm based approach for improving QoS and SLA guarantee in IaaS clouds using pattern-based service-oriented architecture,'' SN Computer Science, vol. 4, no. 5, 2023.DOI
33 
P. Yadav, ``Association rule mining in big datasets using improved cuckoo search algorithm,'' Cybernetics and Systems, vol. 54, no. 6, pp. 787-808, 2023.DOI