1. Introduction

Development and study of novel teaching technology for vocational education have emerged as relatively recent research areas due to the rapid rise of information data. The grades of students can serve as a starting point for sustainable and creative growth in vocational education because they reflect the quality and performance of teachers' daily instruction and students' learning ^[1-3]. At present, evaluation of academic performance by scholars at home and abroad is usually highly subjective and lacks rationality. The ant colony algorithm is a bionic evolutionary and combinatorial optimization algorithm inspired by ant colony efforts to find food. The advantage of this algorithm is that it is simple and easy to implement, and it has been widely used in neural networks, combinatorial optimization, and other fields. Despite the benefits of positive feedback, self-organization, widespread collective computing, etc., the basic ant colony algorithm has a poor convergence rate and is prone to stagnation. At the same time, the cluster analysis algorithm is a relatively efficient data mining technique. Each cluster analysis algorithm has unique properties and applications, although this approach falls short when dealing with vast amounts of data.

Based on research and analysis of the ant colony algorithm, improvements have been made through the clustering algorithm to avoid the randomness of repeated access to the same data objects and the similarity calculation process. The aim is to contribute to comprehensive improvement in the overall quality of students' education ^[4,5]. Based on the benefits of ant colony algorithms, combining them with clustering analysis in data mining creates a clustering analysis algorithm. A program based on swarm intelligence analyzes the characteristics of the algorithm in line with the clustering results. At the same time, introducing student grades for clustering analysis based on a slave group algorithm allows analyzing clustering results and proposing a dynamic method to achieve student grade evaluations. The input of this model is the original student grade, and the output is the student grade level. The main contributions of this research are as follows. Analyzing and improving the performance of clustering analysis algorithms based on the ant colony algorithm reduces the random movement of ants to achieve the target. Achieving a fast search for similar data thereby improves the convergence speed of the algorithm. To prevent a local optimal solution from the clustering analysis method based on the slave group algorithm in cluster analysis, the final number of clusters can reach optimal accuracy. Grades for students are grouped, and extensive examination of the clustering is done along with dynamic analysis of student grade evaluations. This is the problem that this manuscript seeks to resolve and validate.

2. Literature Review

To analyze a multi-objective mixed-load school bus routing model, Mokhtari and Ghezavati proposed an objective ant colony optimization algorithm combining a mixed integer linear programming scheme with multiple objectives. An accuracy and diversity measurement analysis of small-scale and large-scale test problems showed that the proposed hybrid multi-objective solution has extremely high practicability ^[5]. Goel and Maini analyzed and studied the optimal problem of vehicle routing through an ant colony system and firefly optimization algorithm. Between them, the ant colony algorithm uses a pheromone shaking process to prevent pheromones from staying in the development area. The standard test set results showed that the algorithm finds the optimal solution faster than the current metaheuristic algorithm ^[6]. Gong et al. analyzed the problem of a digital road network, and proposed use of an ant colony optimization algorithm to simulate the process of ants transporting food. The test results showed that the algorithm can achieve relatively accurate map-matching results in a short time ^[7]. Aiming at the current situation of intrusion detection, Rais and Mehmood proposed a new feature selection algorithm, namely, three-level update feature selection in a dynamic ant colony system. The test results showed that the proposed feature selection algorithm can improve the accuracy of facial recognition for intrusion detection ^[8]. The problems of falling into a local optima and of slow convergence were resolved by Ge et al. after they examined users’ quality of service requirements and the load balance of a cloud platform. They then designed a multi-dimensional quality of service cloud computing task scheduling algorithm using an improved ant colony algorithm, and they dynamically changed the strength of the pheromone using expectation heuristic functions and pheromone update methods. Experimental test results showed that the algorithm outperformed other algorithms in terms of cloud platform load and user satisfaction ^[9].

Ahmad et al.’s performance evaluation of Sri Lankan university scholars is often completed by students, but students’ feedback does not represent comprehensive abilities of scholars. The systems they employed include an institutional perspective, a student perspective, an administrative perspective, and a learning and growth perspective. Their method conducts qualitative research through content analysis to provide reference data for future institutional development and student career development ^[10]. A course recommendation system by Yz et al. enables students to select courses from a wide range of resources based on their individual needs. They offer a hybrid recommendation model that incorporates tensor decomposition, graph neural networks, user interaction activities, and network structural properties and features ^[11]. Shen proposed a learning evaluation system for offline and online teaching modes. The evaluation results showed the feasibility of this teaching model ^[12]. To improve teaching quality, Wei and Fang put forward a teaching quality evaluation system for college teachers. The test results showed that the system has high practicability ^[13]. Dong et al. designed four modeling tasks, and confirmed through examples that each type has certain possibilities and limitations, which can provide decision-making suggestions for future classrooms ^[14]. Ansari et al. adopted an approach based on socio-environmental synthesis of how to adapt to a teaching-oriented teaching practice, plus a research method, and problem-solving in the Internet age ^[15].

Ant colony optimization and combinations with other algorithms have been used in many linked sectors, including banking and engineering, and was combined with the research status of many academics domestically and internationally. However, there are still many unresolved research issues with ant colony algorithms, and at this point, ant colony algorithms and cluster analysis algorithms are hot study topics. At the same time, achievements by students in sustainable and innovative development of vocational education are still important indicators for evaluating students’ academic performance and teaching quality. This study offers a method for assessing the performance of students in vocational education using an ant colony algorithm as a basis, with the goal of offering technological support for the implementation of creative teaching in sustainable education.

3. Evaluation of Vocational Education Student Achievement

3.1 Ant Colony Algorithm Improvement

This research integrates the ant colony algorithm and the classification rule algorithm into an adaptive ant colony classification rule algorithm. Fig. 1 is a simplified diagram of the ant colony’s method for finding the best path. Fig. 1(a) shows that the ant colony finds the best path from the starting point to the goal. Fig. 1(b) depicts placement of an obstacle between the starting point and the target. A short while after the obstacle is placed, the number of ants on either side of the obstruction is about equal. Fig. 1(c) shows that as the time after placing the obstacle increases, ants will take the path with the higher pheromone concentration, and the number of times they pass is proportional to the pheromone concentration. Fig. 1(d) shows that most ant colonies can find food in the fastest way after a certain period of time. A rule classification algorithm is a technology that uses rules to classify data and build a rule-based classifier. The Ant-Miner algorithm is a traditional classification rule algorithm. It uses information entropy theory to establish a heuristic function with high classification accuracy and simple rules ^[4]. But the disadvantage of this algorithm is that the entropy calculation is complicated, the running speed is slow, and local stagnation is likely to occur.

Based on the original Ant-Miner algorithm, the optimization heuristic factor and the pheromone update method are added to the adaptive mechanism to construct an optimization algorithm ^[16,17]. For the decision classification problem in the coverage algorithm, there are $k$ pieces of data, and the attribute variables in the data set are $A_{i}$, $i=1,2,\ldots a$, $V_{ij}$, and $b_{i}$, representing $A_{i}$, the first $j$ attribute value, and the number of attributes, respectively. Attribute item $term_{ij}$ refers to $V_{ij}$ and $A_{i}$. For the best rule selected from the coverage algorithm in the process of rule construction, each time the selected rule, $term_{ij}$, is applied to the current rule, the probability calculation formula of the selected attribute item is formula (1):

(1)

$P_{ij}(t)=\tau _{ij}(t)^{\alpha }\cdot \eta _{ij}(t)^{\beta }/\sum _{i}^{a}\sum _{j}^{b_{i}}\tau _{ij}(t)^{\alpha }\cdot \eta _{ij}(t)^{\beta }$,

where $\eta _{ij}(t)$and $\tau _{ij}(t)$, respectively, refer to the pheromone concentration and the heuristic factor of the time $t$ path $ij$, while $\alpha $ and $\beta $ denote the relative importance of the pheromone concentration and the heuristic factor, respectively, and both are not less than 0. The higher the value of $\alpha $, the more ants choose the path affected by the concentration of pheromone, but if the value is too large, it will cause the phenomenon of local stagnation ^[16-18]. A larger $\beta $ indicates that the ants’ chosen path is greatly affected by the current attribute items of the coverage sample. When establishing a new classification rule, it is necessary to iteratively update the data information attributes, where the updated information attributes are obtained in accordance to formula (2):

(2)

$\tau _{ij}(t+1)=(1-\rho )\tau _{ij}(t)+\tau _{ij}(t)\cdot Q$.

$Q$refers to the validity of the rule; $\rho $ represents the information exertion factor, and its value is within the interval [0, 1), where the pheromone and the validity of the rule have a positive-effect relationship. Attributes that do not contain this rule are updated by formula (3):

(3)

$\tau _{ij}(t+1)=\tau _{ij}(t)/\sum _{i}^{a}\sum _{j}^{b_{i}}\tau _{ij}(t)$.

Due to the computational complexity of the original information entropy update approach using attribute items, formula (4) is used in this research as a density-based heuristic factor:

(4)

$\eta _{ij}(t)=\left| T_{ij}(t)\right| /\left| T(t)\right| $,

in which $\left| T_{ij}(t)\right| $ and $\left| T(t)\right| $, respectively, refer to the number of attribute items and the total amount of data contained in the dataset. The pruning strategy prevents overfitting, and the rules for judging the validity before and after pruning are shown in formula (5):

(5)

$Q=\frac{TP}{TP+FN}\cdot \frac{TN}{TN+FP}$.

$TP$ refers to the number of data samples that conform to the rule before and after the change. $FP$ refers to the number of samples that conform to the rule before the change but that do not conform after the change. The $FN$ and $TN$ sum is the opposite of the $FP$ and $TP$ sum. After pruning, $Q$ increases and the research can be pruned. The rules after pruning are completed are valid. Two optimization techniques (deterministic selection and random selection) are utilized to dynamically change the volatility of information in order to address the issues with the ant colony method. The least-traveled route raises the likelihood of random selection. Previous studies have pointed out that $\rho $ in the following requirements need to be met. Set a small $\rho $value in the initial stage of the path search, and increase the value of $\rho $ appropriately in a later stage to avoid local stagnation ^[19,20]. The adaptive adjustment mechanism adopted in the study is shown in formula (6):

(6)

$\rho (t)=\frac{3}{2}\int _{0}^{t}f(\tau )d\tau $,

in which $f(\tau )$ is $\mu $ (a normal distribution function of 0), the $\rho $ maximum value is 0.75, and the standard deviation is 10. Compared with the standard deviation of 4, the $\rho $value decreases more smoothly with time, increasing the convergence speed, and helping to find the best path.

Fig. 1. Schematic of an ant colony seeking the optimal path to the target.

3.2 Ant Colony Clustering Improvement

Ants foraging (the basic model), ant self-aggregation behavior, the formation of ant piles, and chemical identification of ants based on ant nest classification are the four categories into which the principles of clustering methods based on ant colony algorithms can be divided. Each principle has corresponding application scenarios, and can finally achieve clustering of similar data. The basic process of the cluster analysis method combined with the ant colony algorithm is shown in Fig. 2. First, each individual ant in the colony is initialized, the goal of the ant colony is $N$, and the maximum number of iterations for the entire algorithm is $M$. Set other parameter values in the ant colony clustering technique, as well as the side length of the local area, that were obtained from similarity of the data object identified. The coordinates are $\left(x,y\right)$, and all data objects that exhibit clustering are projected into the relevant interval. Third, no data object appears for each ant in the initial setting, and a data object within range is randomly selected. Fourth, calculate the average similarity of data objects. When the ants do not appear to be loaded, the probability that the ants obtain the object is determined. When $P_{p}$ exceeds a certain random probability, and the data object is not picked up by other ants, it can be marked as loaded, and can be set to move randomly in the two-dimensional plane. Otherwise, the ants will refuse to pick up the object, other objects are randomly selected at the same time, and the above process is repeated. But when the ants have loaded the object, the probability of dropping the object, $P_{d}$, needs to be calculated. The ants will not pick up the object, and will mark its state as unloaded when $P_{d}$ surpasses a specific random object. The algorithm will then randomly choose a new object. Additionally, the ants will load the object and relocate it to a new position at random if the object that dumps the data does not fulfill the requirements. Fifth, when an object is isolated, it needs to be marked as an isolated point; otherwise, the cluster sequence needs to be assigned to this object, and the same sequence number is recursively given to other objects in the neighborhood.

For calculation of the average similarity, if a certain time, $t$, is set, then the data object found by the ants at location $r$ is $o_{i}$, and then, the calculation of the average similarity with neighboring objects, $o_{j}$, is formula (7):

(7)

$f\left(o_{j}\right)=\max \left\{0,\frac{1}{S^{2}}\sum _{{O_{j\in Neig{h_{sxs}}\left(r\right)}}}\left[1-\frac{d\left(o_{i},o_{j}\right)}{\varepsilon \left(1-\left(v-1\right)\right)/v\max }\right]\right\}$.

In formula (7), $\varepsilon $ refers to the similarity parameter; the moving speed of the ants is $v$, the maximum speed is $v\max $, $r$ is the square local area, $o_{i}$, next to location $o_{j}$, and $s$ is the distance of the neighborhood object in the attribute space kernel $d\left(o_{i},o_{j}\right)$. Cosine distance and Euclidean distance make up the distance calculation. Formula (8) is the computation for Euclidean distance:

(8)

$d\left(o_{i},o_{j}\right)=\sqrt{\sum _{k=1}^{m}\left(o_{ik}-o_{jk}\right)^{2}}$.

In formula (8), the number of attributes for the object and the neighboring object is $m$. The calculation of cosine distance is formula (9):

(9)

$d\left(o_{i},o_{j}\right)=1-sim\left(o_{i},o_{j}\right)$.

In formula (9), the value of the function is formula (10):

(10)

$sim\left(o_{i},o_{j}\right)=\frac{\sum _{k=1}^{m}\left(o_{ik}\cdot o_{jk}\right)}{\sqrt{\sum _{k=1}^{m}\left(o_{ik}\right)^{2}\cdot \sum _{k=1}^{m}\left(o_{jk}\right)^{2}}}$.

For the probability conversion function, it is essentially $f\left(o_{i}\right)$, which converts the average similarity of data objects into a drop function. The probability conversion function's foundation is a symmetric function in accordance with the function conversion principle, which states that the smaller the average similarity between a data object and its neighborhood, the less likely it is that the object will be in the neighborhood, and the higher the probability of picking it up (a sigmoid process). The calculation of probability $P_{p}$ for randomly moving ants picking up an object when they are not loaded with an object is (11):

(11)

$P_{p}=1-Sigmoid\left(f\left(o_{i}\right)\right)$.

Similarly, for $P_{d}$, the calculation is formula (12):

(12)

$P_{d}=Sigmoid\left(f\left(o_{i}\right)\right)$.

In formula (11) and formula (12), the $Sigmoid\left(f\left(o_{i}\right)\right)$ expression is formula (13).

(13)

$Sigmoid\left(f\left(o_{i}\right)\right)=\frac{1-e^{-cx}}{1+e^{-cx}}$.

The $Sigmoid$ function is a natural exponential function. If $c$ (the parameter value) is larger, the curve saturation will be faster, and the convergence speed of the algorithm will be faster. However, the ant colony clustering algorithm has the following problems. First, it cannot ensure that the ants can extract or load all the data to avoid repeated access to the same data object in the similarity calculation. Second, in extracting data and finding neighborhood objects, the ant similarity calculation is random. Low accuracy in clustering occurs when the ants are distracted by randomness in only a few iterations, and they fail to consistently locate the vicinity of the cluster, even when adding more iterations to the clustering process does not improve the results. The algorithm's performance cannot then produce outcomes that are satisfactory. In response to these two problems, the acting skills are improved and optimized with strategy 1 and strategy 2. Strategy 1 is as follows. The introduction of a recording mechanism ensures that all data are found, and at the same time, can prevent the same data from being accessed multiple times, affecting the clustering of data. Strategy 2 is establishing a ``global memory bank of historical positions'' for the data dropped by the ant colony to complete a comparison of all ant data. The primary functions of this structure are to keep track of the location where the ants placed their data, to establish the memory bank's contents, and to continuously update the database's data in each iteration.

The expression of the pheromone concentration on the time $t$ path is formula (14):

(14)

$\tau _{ij}(t)=\left\{\begin{array}{l} 1d_{ij}\leq R\\ 0d_{ij}>R\enspace \enspace \end{array}\right.$.

In (14), $i$ is the weighted Euclidean distance of the first $d_{ij}$ object, the first object is $j$, and the maximum radius of the cluster after clustering is $R\enspace \enspace $. The dissimilarity of data objects in the same class is shown in formula (15):

(15)

$d_{ij}=d\left(x_{{_{i}}},x_{j}\right)=\sqrt{\sum _{i=1}^{m}P_{i}\left(x_{i1}-x_{j1}\right)^{2}}$.

Based on this, the optimization diagram of the proposed ant colony clustering algorithm is shown in Fig. 3. The main steps are as follows. Calculate the historical coordinate location in the calendar memory after the ants have loaded the data, as well as its similarity and drop probability.

Fig. 2. Basic processes of the clustering analysis method based on the ant colony algorithm.

Fig. 3. Basic processes of the clustering analysis method based on the ant colony algorithm.

4. Evaluation of Vocational Education Students’ Performance

In the experimental environment, the operating system was Windows 10 Professional running on an Intel i7-4710MQ CPU at 2.5GHZ with 8GB of memory, and using the PyTorch 1.10 software environment. This study used MATLAB 6.5 to construct the ant colony clustering algorithm; the test set to training set ratio was 1:3. Fig. 4 displays the outcomes of the experiment that first establishes the ideal number of ant colonies for the ant colony optimization method and the ant colony clustering technique. The three evaluation indicators are mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE). The three error values of the two ant colony optimization techniques are all at minimum when the number of ant colonies is nine. The minimum MAE, RMSE, and MAPE for the ant colony optimization algorithm are 0.32, 0.56, and 0.65, respectively, and the minimum MAE, RMSE, and MAPE of the ant colony clustering algorithm are 0.41, 0.51, and 0.59, respectively.

Figs. 5(a) and (b) illustrate the loss values of the ant colony optimization algorithm under different numbers of training runs and batch sizes. From the graph, we can see that as the amount of training increased, the loss value of the ant colony optimization algorithm gradually decreased and eventually converged to a fixed value. The loss values of the models tended to settle at a training frequency of 200, and the ant colony algorithm, the ant colony clustering method, and the ant colony clustering improvement algorithm's associated convergence values were 1.136, 1.002, and 0.498, respectively. Compared with the other two algorithms, the loss value of the improved ant colony clustering algorithm decreased by 47.53% and 56.35%. For training batch sizes, the loss values of the improved ant colony clustering algorithm gradually decreased with an increase in the batch size. At 1000 pieces at a time, the corresponding convergence values for the three error values of the three algorithms were 2.121, 2.032, and 1.523. Compared with the other two algorithms, the loss value of the improved ant colony clustering algorithm decreased by 48.26% and 61.23%, respectively. Therefore, after optimizing the ant colony algorithm with the optimization and clustering strategies, the performance of the combined algorithm was better. The loss values of the ant colony optimization technique based on testing frequency and batch sizes are shown in Figs. 5(c) and (d). The test results show that the improved ant colony clustering algorithm exhibited the most significant decrease in the loss curve under different testing frequencies and batch sizes, with a maximum difference in amplitude of 32.4% and 41.7% compared to other algorithms. And when the test batch size reached 1000, the three algorithms gradually converged, with the improved ant colony clustering algorithm converging to 1.620. The improved ant colony clustering algorithm had lower loss values than the other two algorithms, and the testing and training results were consistent.

Student performance evaluation is an important aspect of teaching management, and an important source of information for evaluating teaching quality. Scientific evaluation of students' academic performance and dynamic adjustments and changes based on their actual situations can help students have a more intuitive and clearer understanding of their own learning situation. The use of the ant colony clustering algorithm for student grade classification and evaluation can effectively mine the data behind the hierarchy, thereby mastering students' learning abilities, and can provide relevant suggestions. Because there are significant differences in students' learning abilities, learning styles, and learning habits behind different grades, the evaluation results can help us obtain dynamic evaluation and classification results from a lateral perspective. The better the classification effect of the algorithm on students' grades, the better it indicates that it can effectively distinguish students' learning status, and thus demonstrate good practicality.

Table 2 lists the running times of the ant colony clustering algorithm under different training frequencies. Different ant colony optimization techniques gradually take longer to perform as the amount of training data grows. The improved ant colony clustering algorithm, however, ran the fastest overall when compared to the ant colony algorithm and ant colony clustering algorithm. When the training frequency was 200, the running times of the ant colony algorithm, the ant colony clustering algorithm, and the improved ant colony clustering algorithm were 4.6s, 4.1s, and 3.5s, respectively.

Figs. 6(a) and (b) show the classification results of students’ grades on the training set and test set by the ant colony clustering algorithm, respectively, in which 0, 1, and 2 respectively refer to low, medium, and excellent grades. The classification results of the training set and the test set show that the number of misjudgments made by the ant colony optimization algorithm for good grades was 14 and 4, respectively, and the number of misjudgments for middle grade students was 8 and 44, respectively. The number of false positives was 6 and 4, respectively. The classification results of the training set and the test set showed that the number of misjudgments made by the ant colony clustering optimization algorithm for outstanding students was 4, and the number of misjudgments for students with corresponding grades was 1. The number of misjudgments of grade students was 2 and 2, respectively. As a result, the ant colony clustering optimization technique classified student grades remarkably accurately for both the training set and the test set.

Fig. 7 shows the evaluation results of the ant colony clustering optimization algorithm for student performance.

We selected a standard color type as the clustering center, and manually classified colors into six categories based on the colors reproduced. By using the closest value in each category as the clustering center and calculating the distance between all samples and the corresponding clustering center, cluster classification can be achieved. Figs. 7(a) and (b) both show six clusters. From Fig. 7(a), we can see that for the clustering algorithm, 200 pieces of data can be divided into six clusters with 24, 38, 42, 31, 40, and 25 members. The 200 bits of data that were clustered using the ant colony clustering optimization algorithm were consistently clustered, with more concentrated clusters showing superior performance. This suggests there is a sizable difference in the comprehensive grades of the students.

The precision/recall (PR) curve is made with precision as the vertical axis and recall as the horizontal axis. It is widely used in classification, retrieval, and other fields, and the classifier can obtain different PR values by adjusting the threshold. When the classification threshold changes from large to small, precision decreases, and recall increases. The closer the PR curve to coordinates (1,1), the better the performance. Precision represents the proportion of positive samples to the sample size. The smaller the precision, the higher the probability that the model will predict false samples, which makes classifier performance prone to failure in multiple experiments. If classification is committed to improving precision, in order to ensure the feasibility and accuracy of the prediction results, the threshold will be a higher value. For binary classifiers, after completing training, each sample in the model receives a class probability value, which is sorted in descending order. The first 10% probability value after sorting is used as the threshold, and the classification threshold can sort the prediction results for the samples. The classification threshold is used to adjust the size of the prediction probability. The larger the classification threshold and the closer the PR curve is to the right, the better the classification predictions of the algorithm under high probabilities. Therefore, the classification thresholds set in the study were 0.3, 0.5, 0.6, and 0.8. Fig. 8 shows the PR curves at classification thresholds of 0.3 and 0.5. Overall, the student performance evaluation model combined with the ant colony clustering optimization algorithm performed better, followed by CF, FOA, and BP. KNN had the worst performance. When the classification threshold was 0.3 and 0.5, the proposed method predicted correct values of 0.72 and 0.78, respectively.

Figs. 9(a) and (b) show the classification performance at two thresholds of 0.6 and 0.8. For different classifications, when the threshold was 0.6, the performance of the five student achievement evaluation models was the best (0.49 and 0.44). Therefore, the combination of the ant colony clustering optimization algorithm proposed for student performance evaluation has high practicality and application value.

Fig. 4. Numbers of ant colonies in the ant colony optimization and ant colony clustering algorithms.

Fig. 5. Loss values of the clustering optimization algorithms.

Fig. 6. Ant colony clustering algorithm performance in student classification.

Fig. 7. Results from ant colony clustering optimization of students’ performance.

Fig. 8. PR curves at thresholds of 0.3 and 0.5.

Fig. 9. PR curves at thresholds of 0.6 and 0.8.

5. Conclusion

This study combined the ant colony algorithm and a cluster analysis algorithm in order to analyze its application in sustainable innovation for vocational education. The improved ant colony clustering optimization algorithm's values were 0.41, 0.51, and 059, whereas the lowest MAE, RMSE, and MAPE of the ant colony optimization algorithm were 0.32, 0.56, and 0.65, respectively. When the training batch was run 200 times, the convergence values of the ant colony algorithm, the ant colony clustering algorithm, and the ant colony clustering optimization algorithm were all lowest at 1.136, 1.002, and 0.498, respectively. When the training batch size was 1000, the convergence values of the three algorithms were lowest at 2.121, 2.032, and 1.523, respectively. The 200 pieces of data could be divided into six clusters of 24, 38, 42, 31, 40, and 25. Both the ant colony clustering analysis algorithm and the traditional clustering algorithm had consistent results. The five student performance evaluation models performed the best, with correct prediction values of 0.85, 0.67, 0.55, 0.49, and 0.44 for the FOA, BP, KNN, and CF ant colony clustering optimization techniques, respectively. This research provides technical support for sustainable and innovative development of vocational education, and can provide educators with more professional and targeted teaching strategies for students. The proposed scheme has not undergone extensive theoretical verification due to time constraints and experimental conditions, necessitating further optimization.