Construction of Artificial Intelligence-driven Education Management Informatization System and Its Practice in College Management

Collecting and analyzing students’ course learning information is the basis for educational administrators to better understand their professional course learning needs. By describing the “association rules”, “Apriori algorithm”, “Apriori algorithm with interest degree” and other theories of information management, the information system is applied to the relevance of courses. The theory of information management aids educational administrators in comprehending the origins of information.They must grasp the rules of operation, then deeply understand the conclusions derived from a principled perspective, and scientifically and objectively enhance the relevance of courses based on these conclusions.

One of the theories of information management is the rule of association. Association rules pertain to uncovering dependencies and correlations among a vast quantity of ostensibly connected data, serving to reveal latent relationships embedded within objects. Furthermore, association rules suggest that when numerous events are interconnected, the occurrence of specific events may predict the manifestation of others.

In association rule analysis, the most commonly used are the two metrics of support and confidence, through which we can determine which are the strong association rules and which are the weak association rules, and the extraction of association rules also depends on the thresholds of these two metrics, and the design of the thresholds has a great impact on the results of mining. In the next section, the two thresholds are further introduced.

The support degree reflects the proportion of the number of transactions in which X and Y appear simultaneously to the number of all transactions in the transaction database, i.e., the frequency of the number of transactions in which X and Y appear simultaneously in the set of transactions, which is denoted as Support(X ⇒ Y), and is calculated as shown in equation (1): (1)Support(X⇒Y)=P(XY)=number(XY)number(D)

The confidence level is usually used together with the support level, reflecting the frequency of transactions in which X and Y occur at the same time when transaction X occurs, denoted as Confidence(X ⇒ Y), and defined by the formula in equation (2). (2)Confidence(X⇒Y)=P(Y|X)=Support(X⇒Y)Support(X)

The second information management theory is the Apriori algorithm.The Apriori algorithm model can be broadly categorized into three parts: firstly, constructing a dataset to extract all frequent itemsets that meet the mining criteria; secondly, deriving association rules from these itemsets that adhere to the minimum support and confidence thresholds; thirdly, analyzing the derived association rules.de some guidance for the development of this field.The flow chart of the Apriori algorithm is shown in Figure 1. Thirdly, the mined association rules are parsed to provide some guidance and help the development of this field.The flowchart of Aprioni algorithm is shown in Fig. 1.

Figure 1.

Flowchart of the Apriori algorithm

In this algorithm model, the efficiency of the algorithm execution is significantly impacted by the extraction of all frequent itemsets. Traditional methods require comprehensive data scanning each time, which is time-consuming. However, more efficient algorithms, such as MPFI and FP-Growth, have been proposed. MPFI’s efficiency is greatly improved by using a binary matrix and prefix tree to compress and store information about frequent itemsets in just one scan of the transaction database. Similarly, FP-Growth reduces the number of scans to two and uses a specialized data structure, the FPTree, to enhance the process of frequent itemset extraction.

The traditional approach to enhancing algorithms involves applying the a priori theorem. This theorem states that if a subset of a term set is infrequent, the entire set is also infrequent. Conversely, an infrequent alarm term set implies that all its subsets are infrequent. According to the a priori theorem, when extracting frequent itemsets, the time for generating frequent itemsets is greatly reduced, which greatly improves the execution efficiency of the Apriori algorithm.

The third of the information management theories is the Apriori algorithm with the addition of the degree of interest. The traditional Apriori algorithm uses only support and confidence thresholds to extract association rules, which can lead to uninteresting or misleading results. To minimize the creation of non-interesting rules, it is essential to incorporate an interest degree measure into the judgment process. Today, several widely adopted interest degree models include the Apriori algorithm, which has been effectively applied in various domains such as optimizing association rule mining, enhancing user interest models for college student employment recommendations, and evaluating customer interest in CRM systems. 1)

Probabilistic interest degree model as shown in equation (3): (3)Interest(X⇒Y)=Support(X∪Y)Support(X)Support(Y)−1

From equation (3), it is evident that the extent to which the presence of X affects the presence of Y is represented in the model. The extraction of rules is mainly based on the interest level of the value of the judgment, as per formula (3), which enables the determination of the interest level values for positive and negative judgments. When the value of the interest level is positive, it indicates that the rule preceding the item can promote the occurrence of the rule following the item; when the value of the interest level is negative, it suggests that the rule preceding the item can inhibit the occurrence of the rule following the item. If the value of the interest level is zero, it signifies that the rule preceding the item and the rule following the item are independent of each other, with no relationship of promotion or inhibition. 2)

The informative interest degree model as shown in equation (4): (4)Interest(X⇒Y) =P(X)[P(Y|X)log(P(Y|X)P(Y)) +(1−P(B|A))log(1−P(Y|X)1−P(Y))]

The model primarily enhances the Apriori association rule algorithm by utilizing the amount of information. The value of the interest degree directly influences the quantity of information in the mining rules: a higher value results in more information, while a lower value leads to less information, making the rule mining less significant. 3)

As shown in equation (5) the difference interest degree model: (5)Interest(X⇒Y)=P(Y|X)−P(Y)max{P(Y|X),P(Y)}

Among them, max{P(Y|X),P(Y)} is a criterion and a constraint on the value of the interest degree, which ensures that the value range of the interest degree can only be between -1 and 1. This model avoids the confidence problem caused by too much support in the posterior term of the rule in the association rule, and it can effectively reduce the generation of some of the weak association rules, so as to improve the efficiency of the mining of the association rules, so in this paper, we choose to use the interest degree model based on difference So this paper chooses to use the difference-based interest level model to carry out the association rule extraction between the course evaluation indicators.

To facilitate readers’ understanding, this paper briefly explains the methods used in extracting association rules among course evaluation indicators using the difference-based interest degree model.

For course evaluation, an association rule can be described as follows: let the data set I = {i₁, i₂, ..., i_m} be the set of all attribute data in the course evaluation attributes, and m be the number of attributes in the course evaluation data. An association rule is an implication of the form X ⇒ Y, X ⊂ I, Y ⊂ I, and X ∩ Y = ∅. Here, X is called the rule antecedent and Y is the consequent.

The extraction of association rules is mainly measured by support and confidence. Support refers to the ratio of the number of transactions containing itemsets X and Y to the total number of transactions in transaction dataset D, denoted as Support(X ⇒ Y), and confidence refers to the ratio of the number of transactions containing both itemsets X and Y to the number of transactions containing X in D, denoted as Confidence(X ⇒ Y), as shown in equations (1) and (2).

To identify a rule as a strong association rule depends on whether the threshold conditions of minimum support and minimum confidence are satisfied. Minimum support is used to measure the minimum frequency of occurrence of the itemset in the statistics, denoted as min_sup(0 < min_sup < 1), and minimum confidence is used to measure the minimum reliability of the association rule, denoted as min_conf(0 < min_conf < 1).

The traditional Apriori algorithm, which relies solely on support and confidence thresholds to extract association rules, may not always produce intriguing results and may even be misleading. This is because support defects often result in the loss of valuable rules due to their limited support. Meanwhile, a confidence defect may lead to meaningless results by disregarding the significance of the subsequent item in a transaction. Furthermore, as illustrated in Equation (5), the incorporation of the differential interest degreee for the interestingness judgment of rules, and then filtering the frequent item sets, thus can suppress the generation of misleading association rules and make the extracted association rules more reliable.

Combined with the content mentioned in the previous section, this paper uses the Apriori algorithm, which introduces the differential interest degree, to extract the course association rules. To assess the efficacy and dependability of the selected methodology, this study conducts experiments to contrast the traditional Apriori algorithm with its enhanced versions, specifically the Apriori-ni algorithm and a vector-based improved Apriori algorithm, in the context of mining association rules. before specifically analyzing the course relevance of related majors.

Stacked histograms are used to display the number of association rules obtained by mining two algorithms under different levels of interest, given the conditions of a minimum support of min-support=0.05 and a minimum confidence of min-confidence=0.15. The traditional Apriori algorithm does not require the setting of an interest variable because it does not inherently have a measure of interest.

The comparative analysis of rule numbers derived from mining two algorithms under varying interest degree thresholds is depicted in Fig. 2, demonstrating the effectiveness of interest degree thresholds in refining the association rule mining process.

Figure 2.

Illustrates the changing trend in the number of rules for two algorithms, indicating a shift in their performance over time.

As illustrated in Fig. 2, the improved Apriori algorithm yields a progressively diminishing number of rules as the interest degree increases, indicating a more focused and efficient mining process. By analyzing the trend of the number of association rules, it has been found that when the interest degree increases to a certain value, some rules will gradually be lost. Through the above comparative analysis, it can be seen that improving the Apriori algorithm can reduce invalid rules and improve the reliability and validity of association rules obtained by mining.

In students’ curriculum, the correlation between courses is often very large, Correlation analysis generates numerous valuable association rules, and the results obtained from mining these rules increasingly serve as significant references for teaching managers to optimize curriculum arrangements. After verifying the reliability and validity of the association rules obtained by mining with the improved Apriori algorithm, the following section adopts the improved Apriori algorithm to carry out a detailed analysis of the results obtained by mining the discretized course performance data, and applies the valuable conclusions obtained to the work of teaching administrators to give suggestions for them to improve the teaching plan or formulate the corresponding system.

Following extensive experimentation, we set the minimum support threshold to 0.05, the minimum confidence threshold to 0.15, and the minimum interest threshold to 0.2. This led to the discovery of 157 association rules through data mining. After filtering out the less significant rules, 71 strong association rules were identified, exemplifying the practical application of association rule mining. Some of these rules are illustrated in Fig. 3.

Figure 3.

Part of the course grade association rules

where numbers 1-9 represent some of the association rules selected for display.1 is General Studies Course Grade A → Accounting; 2 is General Studies Course Grade B → Professional Core Grade B; 3 is Professional Core Grade C → Information Management; 4 is Professional Foundation Grade A → Accounting; 5 is General Studies Course Grade C → Practical Grade C; 6 is Professional Foundation Grade C → Information Management; 7 is General Studies Course Grade B, Professional Core Grade C → Practical Grade C; 8 is General Studies Course Grade B → Soft Work; 9 is Professional Core Grade C → Soft Work.

Studies have shown that general education courses have a significant impact on the grades of professional and practical courses among undergraduate students, as evidenced by the high frequency of occurrence in correlation rules mined from data. If the grades of this kind of courses fall behind, it may affect the learning mentality of the students and reduce the students’ interest in learning, which may lead to the unsatisfactory grades of the following some courses.

The general education courses primarily encompass higher mathematics, probability theory, linear algebra, university physics, and university English, therefore, the school should enhance the emphasis on teaching these courses.

Fig. 3 demonstrates that students pursuing accounting majors outperform those in other majors in both general and professional courses. Consequently, the school should uphold the management standards of accounting majors while enhancing the oversight of other majors and elevating the quality of instruction.

Finally, it has been observed that in the correlation rule, the professional core grade of C frequently appears, suggesting that the professional core courses are relatively challenging and students’ performance is not satisfactory. Therefore, the school should improve the teaching quality of these core courses.

To further analyze the correlation rules between specific courses in detail, the course grades of accounting majors are utilized as the database for the subsequent correlation rule analysis.

For this reason, we continue to use the improved Apriori algorithm to analyze the association rules in the same specialty, set the minimum support min-support = 0.05, the minimum confidence min-confidence = 0.15, the minimum interest min-interest = 0.2, and we get 92 association rules by mining, and 40 association rules are obtained by removing the useless rules. In the realm of data mining, association rules are pivotal for uncovering interesting relationships among items within large databases. These rules are quantitatively described by parameters such as support, confidence, and lift, which help in assessing the significance and reliability of the discovered patterns. Some of these association rules are visually represented in Figure 4, providing a clearer picture of the relationships between different items.

Figure 4.

Association rules of Computer science professional grades

where numbers 1-11 represent some of the association rules selected for display.1 is Programming Fundamentals A → Java Programming A; 2 is Advanced Mathematics E → Database Principles and Applications E; 3 is Algorithm Analysis and Design B → Software Engineering B; 4 is Programming Fundamentals D → Database Principles and Applications D; 5 is Probability Theory and Mathematical Statistics A → Java Language Programming A; 6 is Database Principles and Applications B, Programming Fundamentals B→Linear Algebra B; 7 is Programming Fundamentals E→Operating Systems E; 8 is Programming Fundamentals C, Software Engineering B→Data Structures B; 9 is Data Structures C→Java Programming C; 10 is Java Programming B→Computer Networks B; 11 is Java Programming C→Data Structures C.

Based on the generated association rules, the rules pertaining to programming courses and mathematics courses are the most prevalent, suggesting a significant impact of both these courses within the overall curriculum system. The fact that these two courses are mandatory for computer-related majors and their content is often challenging means that students must possess strong theoretical learning abilities. Students mastering these two types of courses often indicate a solid theoretical foundation, thus school teaching managers should focus on enhancing the teaching quality of these courses, especially given teacher shortages, to indirectly and efficiently elevate students’ overall learning level.

Accounting majors find that the correlation rule highlights a strong link between Programming Fundamentals and Java Programming with other core courses, suggesting that students struggling in programming may perceive these subjects as challenging and adopt a negative stance towards future similar courses. Furthermore, courses such as higher mathematics, probability theory and mathematical statistics, as well as linear algebra, also demonstrate significant correlations.Courses, especially those with many association rules, often involve computer-related professional directions that cannot be separated from data processing and analysis. Therefore, a solid mathematical foundation is crucial for learning subsequent professional courses. Lecturers must focus on motivating students and ensuring they have a strong mathematical foundation to better comprehend the content of these courses. Teachers should study how to stimulate students’ interest in studying these courses seriously and build a good foundation in mathematics so that they can learn other specialized courses better.

A comprehensive analysis of the association rules in the courses of different majors reveals that the number of association rules involving programming courses is the highest, which also implies that programming courses, as a mandatory course for all students in this college, have the greatest impact on students’ subsequent grades. Programming courses frequently serve as a foundational prerequisite for all related major courses in the teaching plan, yet many students fail to appreciate the significant impact they have.ourses on the subsequent courses, and the related lecturers should let the students realize that the quality of learning in programming courses directly affects the learning of the subsequent courses. Students should recognize the significance of such courses and approach them with a positive and diligent attitude, regularly practicing programming exercises and engaging in discussions with classmates and teachers, thereby establishing a solid foundation for future related studies. Educational administrators should cTo ensure the effectiveness of the curriculum, it is essential to continuously evaluate and refine the course design. This involves actively adjusting the teaching content and optimizing the teaching format to reduce the difficulty of students in learning these courses. In addition, students who perform poorly in these courses should actively seek help from their teachers to adjust their learning attitudes in time to minimize the impact on other related courses.

As a crucial component of the AI-powered education management information system, the effectiveness and precision of performance management and analysis systems in higher education are closely linked to students’ learning outcomes and development. Grasping the core algorithms and fundamental improvement ideas can empower educational administrators to enhance their data analysis capabilities and provide more insightful suggestions. The following section briefly introduces the idea of improving the C4.5 decision tree algorithm.

In view of the requirements of the outcome analysis system, the C4.5 algorithm needs to be improved accordingly. The algorithm calculates the information gain ratio of each attribute and selects the test attribute as set S. Create nodes and label them, then create branches for each attribute value to complete the sample division. The idea of improving the C4.5 decision tree algorithm is as follows: 1)

Use the information gain ratio to select attributes to avoid using information gain to select attributes, which in turn improves the speed of operation.

Let E = F₁*F₂...*F_n be a n -dimensional exhaustive vector space, F_j denote an exhaustive set of discrete symbols, and the elements e = <w_I, w₂..., w_n> in E, are called examples, where w_j ∈ F, 1, 2, 3, ...n. Suppose Y_E and N_E are the two example sets of E. Suppose that the positive example set Y_E and the negative example set N_E in the vector space E are of sizes y and n, respectively, and that: The decision tree has the same probability of positive and negative instances in the classification probability E of any sample set on the vector space E. E_i denotes the sample data for which attribute A takes on a value A_i.

The amount of information required for a decision tree to be able to judge a sample set according to the correct category is derived from Eq. Let the number of positive and negative examples in E_i be y_i and n_i, respectively, and the amount of information required for E_i is calculated by Eq: (6)I(yi, ni)=yiyi + nilog2yiyi + ni−niyi + nilog2yiyi + ni $$I({y_i},\>{n_i}) = {{{y_i}} \over {{y_i}\> + \>{n_i}}}{\log _2}{{{y_i}} \over {{y_i}\> + \>{n_i}}} - {{{n_i}} \over {{y_i}\> + \>{n_i}}}{\log _2}{{{y_i}} \over {{y_i}\> + \>{n_i}}}$$

Therefore the information entropy of having attribute A as the root node is Eq. (7): (7)E(A)=∑i=1yyi+niy+nI(yi+ni)

Simplification leads to equation (8): (8)E(A)=1(y+n)ln(2)∑i=1y(−yilnyi+niy+n)−nilnniy+n

In the training set 1(y+n)ln(2) is a constant value and the function E(A) is assumed to satisfy equation (9) (9)E(A)=∑i=1y(−yilnyiyi+ni)−nilnniyi+ni

Based on Taylor’s formula, McLaughlin’s formula for the calculation of information entropy simplified by the principle of equivalent infinitesimal, when x is very small, can be approximated that the simplification of the formula can be obtained formula (10) and formula (11): (10)lnn1y1+n1=ln(1−y1y1+n1)≈n1y1+n1 (11)lnniyi+ni=ln(1−yiyi+ni)≈niyi+ni

Substituting both of the above formulas yields the information entropy formula (12): (12)E(A)=∑i=1y(−yilnyiyi+ni)−nilnniyi+ni≈∑i=1y2yiniyi+ni

The splitting information is Eq. (13): (13)SplitI(A)=−niyi+nilog2yiyi+ni−yiyi+nilog2yiyi+ni≈∑i=1y2ln2yini(yi+ni)2

The information gain rate is equation (14): (14)Gain(A)−ratio(A)=Gain(A)spliti(A)=I(A)−E(A)spliti(A)

The formula of information entropy is simplified with the split formula and the simplified information splitting formula is ∑i=1y2ln2yini(yi+ni)2 .

Then we calculate the amount of split information according to the new formula, and the attribute with the largest information gain rate obtained through the calculation is used as the root node. Following the improvement of the calculation method, numerous logarithmic operations have been transformed into quadratic mixed operations, facilitating faster computation. 2)

Adoption of processing continuous numerical attributes. As previously discussed, the C4.5 algorithm is capable of handling both discrete and continuous attributes. When selecting branching attributes on a node, C4.5 and ID3 algorithms are used for discrete attribute description. Their objective is to analyze and calculate the number of attribute values. For continuous attribute description of Ac, we assume the total number of samples in its dataset. C4.5 algorithms will perform the following processing:

All the sample data, with reference to the descriptive attribute values of the continuous type, are arranged from smallest to largest, and the resulting sequence of values is {A1c, A2c, ……Atolalc}. In the fetch sequence generates total-1 segmentation points. The value of the ith (0<i<total) split point is set as Vi=(Aic+A(i+1)c)/2, and the node dataset is divided into two subsets with reference to this method. Finally the more appropriate split point is selected among the total-1 split points.C4.S calculates the information gain ratio by dividing the dataset with different split points and selects the split point with the largest value and then divides the dataset.

In the above formula, N represents the number of instances, f = EN is the error rate (E is the number of classification errors among N instances), q is the true error rate, c is the confidence level, and z is the standard deviation corresponding to the confidence level c, whose value can be obtained by checking the table of normal distribution according to the set value of c.

By using the above formula, we can calculate the upper limit of a confidence interval for the true error rate q, which we use here to make an estimate for the error rate e of the node: (16)e =f+z22N+ZfN−f2N+z24N21+z2N $$e\> = {{f + {{{z^2}} \over {2N}} + Z\sqrt {{f \over N} - {{{f^2}} \over N} + {{{z^2}} \over {4{N^2}}}} } \over {1 + {{{z^2}} \over N}}}$$

As shown in Equation (16), the size of e before and after pruning is determined to decide whether pruning is required or not. 4)

Handling of missing values In some cases, there is a possibility of missing attribute values in the data available for use, if < x, c(x) > is a trained instance in sample S, but the value A(x) of attribute A is unknown. How should this be handled? The way to deal with missing attribute values is to assign the most common value of the attribute in the training instance corresponding to node n. If given a Boolean attribute A, such that node n contains 4 instances of known A = 1 and 6 instances of A = 0, the probability of A(x) = 1 is then 0.4, and the probability of A(x) = 0 is then 0.6. It can then be seen that the branch assigned to A = 1 is then 40% of the instances of x, and the other branch is assigned 60%. If a second attribute with a missing value exists and has to be tested.

Considering the enhancements proposed for the C4.5 decision tree algorithm in section 3.2.1, it remains evident that despite these improvements, missing values may still occur, necessitating the testing of these missing data points. When compared to the single PSO algorithm and the XGBoost algorithm, the C4.5 decision tree algorithm offers a more precise means of testing and verifying missing relevant data, thereby enhancing the accuracy of student performance warnings.a basis for student management.

And in order to verify the effectiveness of C4.5 decision tree algorithm, this paper compares the performance of C4.5 decision tree algorithm with PSO particle swarm algorithm and XGBoost algorithm respectively through experiments. The experimental results, generated using the C4.5 algorithm, are detailed in Table 1.

According to Table 1, the C4.5 decision tree algorithm demonstrates higher accuracy compared to both the PSO algorithm and the XGBoost algorithm when evaluated as single algorithms. This is consistent with the principles of C4.5, which is known for its simplicity and strong interpretability, as well as its ability to handle missing attributes and irrelevant features effectively. However, it is important to note that while C4.5 may outperform XGBoost in some scenarios, XGBoost is designed to address the limitations of traditional decision trees by incorporating regularization to prevent overfitting and by supporting parallel computation for improved efficiency. This is mainly because the academic achievement early warning model based on the C4.5 decision tree algorithm can convert missing values in the dataset into sparse matrices. In terms of feature selection, the academic performance early warning model based on the C4.5 decision tree algorithm utilizes the individual memory and rapid convergence characteristics of particle swarm optimization. This algorithm is capable of screening the optimal subset during the feature selection stage, and it locates the global optimum of the particle swarm via its internal memory and group communication, thereby enhancing the speed and precision of feature selection. Compared to a single algorithm, the academic achievement early warning model using the C4.5 decision tree algorithm has a higher accuracy.

For feature selection, this model combines the features of XGBoost algorithm classification regression and PSO algorithm with fast convergence speed to effectively screen the optimal subset. The experimental results show that compared with the single PSO algorithm and the single XGBoost algorithm, the method improves the accuracy by 8.5% and 5.4% on the mat dataset, 7.2% and 5.7% on the por dataset, and 7.4% and 5.5% on the dataset of students’ grades of the School of Computer Science of our university, thus indicating that, compared with the existing studies, the The model better improves the accuracy of student performance warning and provides data basis for student management. The ROC curves of various algorithms are displayed in Fig. 5.

Figure 5.

ROC curve test results

According to the content of the ROC curve depicted in Figure 5, it is evident that the ROC area for the academic performance warning model utilizing the PSO algorithm is 0.7, while the ROC area for the model based on the XGBoost algorithm is 0.84, and the ROC area for the model using the C4.5 decision tree algorithm stands at 0.89. These results further confirm that the C4.5 decision tree-based algorithm exhibits superior performance in the academicc alert model.