Abstract
The personalized recommendation method of higher education resources currently cannot carry out multidimensional association analysis of learners, situations, and resources and cannot extract accurate resources for learners, resulting in a large error. This study constructs a personalized recommendation method for higher education resources based on multidimensional association rules. This algorithm clarifies the multidimensional association rules, extracts the key data from massive educational resources, and groups the same kind of data by using a frequent itemset algorithm for mining association rules, namely, the Apriori algorithm. Combined with traditional data mining technology, this study constructs a new personalized recommendation model for education resources based on multidimensional association rules, which achieves the accurate extraction of higher education resources and ensures the matching degree between learners and resources. The experimental results show that the personalized recommendation model of educational resources in this study effectively makes up for the disadvantages of the traditional data mining algorithms, with a small root mean square error and short data mining time, within 20 ms.
1. Introduction
With the rapid development of the application of “Internet Plus Education,” abundant online learning resources have emerged. Although online learners in higher education can easily access a large number of learning resources, the diversity and relevance of online learning resources are less taken into account, and it is increasingly difficult for learners to obtain appropriate resources for learning [1, 2]. At the same time, it is very difficult for learners to understand their own knowledge level and learning status, so it is very easy to get lost in the process of learning. The ideal online learning system can recommend learning resources according to the characteristics of learners and use an intelligent algorithm to implement learning resource recommendations. Since the characteristics of learners are changeable and difficult to quantify, and the online learning resources are massive and complex, the recommendation of online learning resources is required to be higher, which leads to the slow speed and low matching of online learning resource recommendation methods [3].
Reference [4] studies and constructs a learner model suitable for the personalized recommendation of online learning resources according to the education and teaching theory and the relevant data of learners in the online education platform. Taking the collaborative filtering recommendation method as the starting point, the collaborative filtering method is improved by integrating the static and dynamic characteristics of the learner model, and a collaborative filtering recommendation method of online learning resources integrated into the learner model is established. Reference [5] proposes that in the mixed learning scenario, collaborative recommendation technology is used to match learners with high-quality audio-visual resources to dynamically customize personalized learning content for learners. Through quantitative analysis of the large sample data movielen 100 k, it is found that users are generally familiar with only a small number of resources and need the recommendation mechanism to help expand the cognitive field, and this method can improve the recommendation effect. Through teaching practice in student associations, it is found that students with different backgrounds have different preferences for audio-visual resources, and this method has a positive impact on audio-visual learning. Reference [6] proposes a personalized learning resource recommendation method based on the dynamic collaborative filtering algorithm. The Pearson correlation coefficient is used to calculate the data similarity between learning users or project resources in the network, and the unscored value is obtained. In order to solve the problems of sparse data and poor scalability in the collaborative filtering algorithm, dynamic -nearest-neighbor and Slope One algorithm are used to optimize it, and the sparsity of learning resource data in the network is analyzed according to the result of neighbor selection. The bidirectional self-equalization of stage evolution is used to improve the personalized recommendation of resource push, and the fuzzy adaptive binary particle swarm optimization algorithm based on the evolution state judgment is used to solve the problem of the optimal sequence recommendation, so as to realize the personalized learning resource recommendation.
However, the above methods cannot obtain the multidimensional association rules of data, and there are large errors in the extraction of key information of educational resources, resulting in a long time-consuming personalized recommendation method of resources. Therefore, this paper designs a personalized recommendation method of higher education resources based on multidimensional association rules. The structure of this paper is as follows: (1)The problem of personalized learning resource recommendation is described, and a higher education resource model is built by combining the characteristics of multiple learners and learning resources(2)Calculate the multidimensional support and the data confidence of higher education resources, mine the frequent itemsets of association rules by using the Apriori algorithm and correction algorithm, and obtain multidimensional association rules(3)The key data in educational resources are extracted, and the interval value attribute dataset clustering algorithm is used to cluster the same kinds of data together(4)Due to the excessive amount of data in the higher education resource set of interval value attribute and the poor data mining effect of traditional methods, the traditional data mining methods are improved to realize the optimal mining of educational resources(5)Construct the interest matrix of college students, use the matrix decomposition to design the personalized recommendation model of higher education resources, and realize the personalized recommendation of higher education resources(6)The experimental results show that the root mean square error of educational resource clustering of the proposed method is lower, and the time consumption of data mining is less than 50 ms. The time consumption of the proposed method in the process of resource information qualification recommendation is always kept below 20 ms(7)Summarize and analyze the content of the full text, and get relevant conclusions
2. Personalized Recommendation of Higher Education Resources Based on Multidimensional Association Rules
2.1. Problem Description and Model Construction
The essence of personalized learning resource recommendation is to match the characteristics of learning resources with the characteristics of learners and then obtain the optimal solution. The current model of recommending learning resources is aimed at recommending multiple learners. The main factors to be considered are the following: first, whether the difficulty of learning resources matches the ability level of learners; second, whether the learning time of learning resources is consistent with the expected learning time of learners [7]; third, whether the learning concepts corresponding to learning resources meet the learners’ expected learning objectives; fourth, whether learners’ preference for the content type of learning resources matches the content type of learning resources; and fifth, whether the type of learning resource media meet the learners' preference for the type of learning resource media. This factor is also in line with learners’ online learning preferences [8].
The differences of learning resource sequences among learners are reduced. After adding the personalized characteristics for a single learner, the recommended learning resource sequence is more in line with the personalized needs; that is, the group recommendation stage (GRS) evolves to the personalized recommendation stage (PRS), which can be described as
This phase is composed of learners features (LF) and resources features (RF), incorporating personalized features (PF) with content and media as the core, forming a balanced state with features based on group recommendations, refining the association of learner and learning resources and two-way effects to make the recommendation of personalized learning resources more accurate and in line with needs, as described in [9]: where .
2.2. Multidimensional Support Calculation of Association Rules
Association rules can effectively reflect the degree of association between objects in a large number of data and can mine the important information contained in the dataset, which is widely used in many fields [10].
In the relevant rules, the number of objects in a data itemset is regarded as the itemset length, and the data itemset whose length is [11] is regarded as the itemset . The whole pattern of frequent itemsets is the number of itemsets it contains, and the most frequent itemsets can be found by computing the multidimensional support of each itemset. Suppose there are items, the number of items in many different sets is increased to ; in many datasets, the minimum support threshold is determined according to the frequency generated in the network, and the effective association rules in the frequent patterns of multidimensional support data are mined [12, 13].
If object is a subset of data items, it is , and each item has an identification number. Each subset obtained in dataset is the itemset contained in . Assuming , the itemset is the itemset of object . If , the itemset is contained in dataset .
Firstly, the support expression of dataset is
In the formula, is the transaction number of dataset , and is the transaction number of itemset contained in dataset .
If is greater than or equal to the specified minimum support, then is called a frequent set, whereas if is less than the specified minimum support, then is called a nonfrequent set [14, 15]. An association rule is a logical expression similar to , in which , , and .
Secondly, if the support in the transaction set , the ratio of the number of and transactions contained in the dataset to the number of transactions, the formula is
In formula (4), refers to the multidimensional support of rule .
2.3. Confidence Calculation of Higher Education Resource Data
Association rules can accurately reflect the existence of a certain association or dependency between various transactions. Usually, association rules are used to describe the hidden relationship between data attributes and variables in the distributed energy database, and the conditions of strong association rules can be obtained [16, 17].
is defined as a set of terms , and is a subset of , , where . Given and , if they are both terms or sets in and satisfy the condition , then the association rule is . has the following meanings:
When the preset threshold is greater than the minimum support and the minimum confidence , the rule meets the conditions of strong association rules [18].
2.4. Explicit Multidimensional Association Rules
In the association rule algorithm, it is divided into the single dimension association rule and multidimensional association rule. As the name implies, a one-dimensional association rule contains only one key predicate, while a multidimensional association rule contains multiple predicates such as location, person, and reason, and more effective rules for predicates can be obtained by pair-wise combination between predicates. A dimensional association rule refers to a key predicate, but the predicate has many different attributes [19].
The Apriori algorithm has obvious advantages in mining frequent itemsets of association rules. This algorithm makes use of the iterative way of searching layer by layer to explore the generation of itemsets. According to the actual needs of the distributed energy system, this paper partially improves the Apriori algorithm and realizes the data mining of the distributed energy system by using the frequent pattern-growth (frequent pattern tree) algorithm without generating candidate options [20, 21]. The implementation is as follows:
Generate (1 itemset) frequent itemsets. The object mined is regarded as a database, which is scanned for the first time, and the frequent itemsets (1 itemsets) of all dimensions are generated to get the count of support (frequency) of each dimension. The frequent itemsets shall be divided according to different dimensions, and the items with the highest support in each group shall be arranged in descending order, and the intragroup support shall also be arranged in descending order. The result set of the arrangement may be described in a table [22]. Table 1 shows a set of multidimensional transaction tables.
If there is a node where the total support count for all child nodes is less than the support count for that node, then the support count for that node and for all parent nodes for that node needs to be calculated [23, 24].
FP-tree construction. The second time you scan the database, you sort the items in the transaction. Delete the infrequent items contained in it and create a separate branch. The FP-tree built in this article is shown in Figure 1.
Pruning treatment. For frequent FP-tree, the modified algorithm is used to mine frequent patterns, and the redundant data is removed by pruning, which makes the final mining results more accurate. Pruning node is to delete node and subtract the original support count from the count of all the parental nodes of the node. After pruning, the remaining node information is the final result of data association of higher education resources. The data mining process based on the multidimensional association rule algorithm is shown in Figure 2.
2.5. Data Clustering of Higher Education Resources
The concepts of different data in the interval-valued higher education resource set are different, so it is necessary to summarize, compare, and analyze different data and give a comprehensive description method. Because of the huge amount of data, it is difficult to find the associated data in the general algorithm. The mining algorithm used in this paper can analyze the relationship between data and data and find out the correlation and dependency between different data [25].
The representative data are extracted from numerous data, and the data are divided into several models to predict the future trend. Classify the data according to the forecast results, cluster the data of the same kind together, and ensure that the data in one set have very high similarity and the data in different sets do not have similarity [26, 27]. In the process of clustering, the biased data are found by the detection box, and these biased data are eliminated to ensure that the mined data are accurate and reliable. At the same time, the time series graph is used to arrange the data visually, which makes mining easier and improves mining efficiency.
Clustering algorithms need to be run many times because they need to be iterated over and over again to find valuable data from a large volume of data [28]. The workflow of the interval-valued attribute data clustering algorithm is shown in Figure 3.
Step 1 (problem definition). Find the information of higher education resources which is useful to people in the mass of higher education resources, so it is necessary to define these higher education resources and determine the types of information to be found before searching.
Step 2 (higher education resource preparation). Higher education resource preparation is directly related to the follow-up mining quality and has an obvious impact on the accuracy. So, data preparation can be divided into three parts: data selection, data preprocessing, and data conversion.
The data selection of higher education resources can identify the objects that need to be mined and search for high-quality data in the database to ensure the collection density. Data preprocessing transforms data in different formats into a unified format, reduces the noise and redundancy in the data, and fills in the missing units. After the pretreatment, the data is different, the conversion method is different, and the dimension of the data is reduced, so the data is simplified.
Step 3 (extract higher education resources). After confirming that the results of the preparation of higher education resources meet the requirements, it is best to extract the knowledge in an intuitive and understandable way.
Step 4 (model prediction). The extracted higher education resources are analyzed and measured to predict the user’s demand.
Step 5 (cluster of higher education resources). At the same time, we choose the association algorithm, classification algorithm, and prediction algorithm to cluster the data and complete the clustering work by looking for potential information and rules.
Higher education resource concentration contains data from many fields, so the computation process of the clustering algorithm faces a lot of challenges. The clustering algorithm must be scalable and high dimensional and can deal with abnormal and complex resource data.
2.6. Mining Method
Due to the huge amount of data in the resource set of higher education and the multidimensional data, parallel algorithms should be used to reduce the mining time and storage space. The parallel algorithm will produce a variety of computing results, so it is necessary to find effective mining results in many computing results; in the study of data attributes, we should analyze both numerical attributes and typed attributes, through preprocessing to obtain valid data [29]. There are many abnormal attributes among interval attribute datasets, some of which are null or erroneous. If the mining algorithm uses erroneous values, the mining results will be inaccurate. Therefore, these erroneous data should be uniformly listed in the same dataset, and the dataset included in the list is called the “outlier dataset.” The remaining data is divided into different intervals by means of discrete, looking for the attributes between data and mapping these data in a unified manner to facilitate the later mining. In this paper, the mining algorithm can deal with both classified attribute data and numerical attribute data, which is simpler and more practical [30].
Due to the large amount of data in the higher education resource set of interval value attribute, the traditional mining algorithm is difficult to achieve the mining effect in a short time. Therefore, this paper improves the traditional mining algorithm, and the mining algorithm flow is shown in Figure 4.
In Figure 4, after all the higher education resource sets are loaded, the logical address of the editor is analyzed; the file name, type, and operation time of the interval attribute dataset are obtained; and the data concept is analyzed. The logical address is edited, the target that can be mined separately is found, and the information is transmitted after the data is mined. If the mining results can be successfully transmitted, the mining work is successful; if the mining results cannot be successfully transmitted, the mining work fails and needs to be reexcavated. For the dataset with no minimum threshold, after the minimum support is determined, the work is directly ended and no mining is needed.
2.7. Personalized Recommendation of Higher Education Resources
Suppose that the interest matrix of students in a given university is
Among them, represents the elements in row and column of the matrix, and and represent the number of college students and the number of recommended objects, respectively [31, 32]. Assume that the weight matrix is where is a positive real number greater than 0 and less than 1. It is necessary to find a low rank matrix to approximate matrix [33].
In order to find that can approach matrix to the greatest extent, the norm loss function of minimizing weighting is given, and the calculation formula is
In the equation, represents the square error in the low rank matrix, and represents the contribution of the data point to the loss function. For positive examples, , and for missing values, when all missing values are negative, . Because the confidence of positive examples is high, the weights of this kind of data are set to 1. Not all mixed data is treated as a negative example [34, 35], but as a negative example with a small weight assigned to the data.
In the personalized recommendation algorithm for higher education resources based on matrix decomposition, firstly, the matrix is initialized by the Gaussian stochastic number with deviation of 0.01 and mean of 0, and then, the square error is updated by formula (8) until the number of iterations is sufficient. The specific steps of the Algorithm 1 are as follows.
| ||||||
To sum up, the time complexity of higher education resource recommendation is combined with the collection of higher education resources to realize personalized recommendations of higher education resources. The expression of the personalized recommendation model is
To sum up, this paper uses the Apriori algorithm and modified algorithm to mine frequent itemsets of association rules, obtain multidimensional association rules, improve the traditional data mining method, use the improved data mining method to optimize the mining of educational resources, and combine the collection of higher education resources with the time complexity of recommendation of higher education resources to build a personalized recommendation model, in order to achieve the ultimate goal of personalized recommendation of higher education resources.
3. Experimental Test and Result Analysis
In order to make the experiment more comprehensive and effective, the learning resource recommendation method integrating the learner model proposed in Reference [4] and the personalized recommendation method based on hybrid collaborative filtering proposed in Reference [5] are used as comparison methods. The experimental results are compared with the experimental test results of the fast push method to highlight the application performance of the proposed method. The simulation environment is the Windows 10 operating system, and the programming language environment is matlab 2020b. The hardware environment is Intel Core Processor i5-4570, the main frequency is 3.20 GHz, and the memory is 4 GB. Take the students of three universities as the research object; recommend personalized higher education resources to these students by different methods; calculate the changes of indicators such as the clustering effect of educational resources, the efficiency of data mining, and the number of time-consuming association rules in the process of educational material recommendation; integrate and count these data; and take the integrated and counted data as the experimental sample data. The final experimental sample data size is 15.32 G. In this paper, two methods, cross-validation and increasing the number of experiments, are used to reduce the experimental error, so as to improve the authenticity and reliability of the experimental results.
3.1. Clustering Effect Test of Educational Resources
The learning resource recommendation method integrating the learner model proposed in Reference [4] and the personalized recommendation method based on hybrid collaborative filtering proposed in Reference [5] are compared with the proposed method, and the RMSE (root mean square error) index is used to evaluate the clustering effect, as shown in Figure 5.
It can be seen from the curve trend in Figure 5 that the RMSE of the clustering results of educational resources in this method is always higher than the learning resource recommendation method integrating the learner model proposed in Reference [4] and the personalized recommendation method based on hybrid collaborative filtering proposed in Reference [5], which shows that the clustering effect of the proposed method is better. This is because this method has more significant advantages than other methods by constantly updating the information of attraction and attribution.
3.2. Comparison of Data Mining Efficiency of Educational Resources
The proposed method, the learning resource recommendation method integrating the learner model proposed in Reference [4], and the personalized recommendation method based on hybrid collaborative filtering proposed in Reference [5] are compared with the proposed method in terms of mining efficiency. The number of experiments is 500. Set the threshold of minimum support to 0.2 and calculate until the frequent 4 itemsets. The three methods carry out data mining in the above environment, and the comparison of the results is shown in Figure 6.
As can be seen from Figure 6, this method shows obvious advantages in data mining efficiency. Compared with the learning resource recommendation method integrating the learner model proposed in Reference [4] and the personalized recommendation method based on hybrid collaborative filtering proposed in Reference [5], data mining efficiency is higher and time consumption is shorter.
3.3. Comparison of Time-Consuming Indicators
Figure 7 is a comparison of the running time generated by the proposed method, the learning resource recommendation method integrating the learner model proposed in Reference [4], and the personalized recommendation method based on hybrid collaborative filtering proposed in Reference [5].
As can be seen from Figure 7, three methods are used for comparative analysis. The time consumption of the proposed method in the process of completing resource information qualification recommendation is always kept below 20 ms. The fluctuation range of time consumption of the learning resource recommendation method integrating the learner model proposed in Reference [4] and the personalized recommendation method based on hybrid collaborative filtering proposed in Reference [5] is 160 ms~100 ms. The experimental results show that the proposed method has the property of rapid convergence, and its final running time is always balanced.
3.4. Number Test of Association Rules under Different Supports
Figure 8 shows the number of association rules under different minimum supports when different methods run.
As can be seen from Figure 8, the proposed method can maintain fewer association rules under different minimum supports, so that the difficulty of resource recommendation is kept at a low level. Experimental results show that the proposed method can automatically and flexibly adjust the minimum support in each dataset according to different situations and converge to the selected set and frequent set.
3.5. Comparison of Key Information Extraction Errors
The error rates of the proposed method, the learning resource recommendation method integrating the learner model proposed in Reference [4], and the personalized recommendation method based on hybrid collaborative filtering proposed in Reference [5] in extracting the key information of educational resources are compared. The comparison results are shown in Table 2.
By analyzing the data in Table 2, it can be seen that the average error rate of key information extraction of educational resources in the learning resource recommendation method integrating the learner model proposed in Reference [4] is 85.6%, and the average error rate of key information extraction of educational resources in the personalized recommendation method based on hybrid collaborative filtering proposed in Reference [5] is 83.5%. It is the lowest of the three methods. Compared with these two methods, the average extraction error rate of the key information of educational resources of the proposed method is 95.5%, and the extraction error rate is higher, indicating that this method can accurately extract the key information of educational resources.
To sum up, the root mean square error of educational resource clustering of the proposed method is lower, and the time consumption of data mining is less than 50 ms. The time consumption of the proposed method in completing the qualification recommendation of resource information is always kept below 20 ms. The average error rate of key information extraction of educational resources is 95.5%, and the personalized recommendation effect of higher education resources is better.
4. Conclusion
Based on the big data environment, this study constructs a personalized recommendation method of higher education resources based on multidimensional association rules, so as to realize the accurate extraction and recommendation of education resources. By calculating the support degree and confidence degree in the association rules of data mining of higher education resources, and using the frequent itemset Apriori algorithm based on mining association rules, the key data are uniformly extracted and divided in the massive higher education resources, and the model is constructed to scientifically predict the change trend. This recommendation model can reasonably classify data types combined with the prediction results, perform cluster analysis on the data to ensure strong similarity of the same group of data, and realize the effective cluster processing of higher education resources. This recommendation model further optimizes the traditional data mining algorithm and introduces learners’ preferences as weights in the process of computing data similarity, so as to provide accurate personalized education resources for learners. Through the experimental research, it can be found that the personalized recommendation model of higher education resources based on multidimensional association rules has high clustering accuracy and short time consumption of education resource recommendation. For multiple minimum supports, this personalized recommendation model can preserve the rule content that meets the minimum confidence and support the adjustment of the minimum support of the dataset.
Data Availability
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.
Conflicts of Interest
The authors declared that they have no conflicts of interest regarding this work.