Multi-dimensional learner portrait label construction and analysis method based on data fusion

Abstract

The invention discloses a multi-dimensional learner portrait label construction and analysis method based on data fusion, which comprises the following steps of: firstly, effectively fusing a multi-source learner data set by adopting a data fusion method based on normal distribution in consideration that most learner score data appears in a normal distribution form; secondly, in order to effectively process the skew data, Mlog normalization is provided, and the digital portrait label generated through the normalization has high discrimination. And finally, fusing the social cognition theory into the education field, and mining behavior characteristics of the learner from two perspectives of trajectory analysis and time sequence analysis so as to construct a scientific and multi-dimensional learner portrait label system for the learner. The method is experimented on two disclosed data sets, and experimental results prove that the method is effective in the aspects of constructing digital tags and generating learner portraits.

Description

Technical Field

[0001] The present invention belongs to a method for building and analyzing digital labels for learners, and in particular, it relates to data fusion technology, label construction based on social cognitive theory, HDBSCAN clustering based on MLog normalization method, and label prediction based on PyTorch deep framework. The purpose of expanding the learner information dimension is achieved by fusion of multi-source data, and then the clustering algorithm is used to divide the group, and the deep learning model is used to predict the group to which the learner belongs based on the division result, and finally the learner characteristics are analyzed in combination with social cognitive theory and personalized guidance is given. Background Art

[0002] In recent years, in order to provide learners with precise and personalized guidance, both domestic and foreign researchers have begun to pay attention to the construction of learner portrait labels and learner feature analysis methods. With the popularization of digital technology, educational data has experienced explosive growth, making it possible to obtain and process multi-source information such as student grades and student behavior. The development of artificial intelligence technology has promoted the application of advanced technologies such as learner portraits, deep learning, and personalized digital education, enabling models to better understand student behavior and trends. Some studies use label grading methods to divide learner labels into primary and secondary levels based on statistical analysis of learner behavior data. Statistical analysis helps to classify learners' behaviors, so that a readable label system can be simply constructed to help teachers understand learners more efficiently.

[0003] However, the existing research still has the following shortcomings:

[0004] (1) Previous studies have been unable to construct comprehensive, multi-dimensional learner labels due to incomplete data content and single information dimensions. This defect leads to incomplete learner portraits and limits educators’ in-depth understanding and analysis of learners.

[0005] (2) Previous studies did not fully address the problem of data skew when conducting statistical analysis on learner behavior. Skewed data will reduce the model's ability to distinguish samples, ultimately affecting the division of learner groups. Statistical results show that learner behavior data such as test scores, study time, and study frequency often present a skewed distribution. Traditional log normalization can mitigate this effect, but it depends largely on the distribution of the original data.

[0006] (3) Current researchers often show strong subjectivity when constructing labeling systems and selecting data content, ignoring the changes in learner behavior characteristics over time. This subjectivity may lead to labels that lack scientific rigor and deviate from educational needs and goals. Learners' learning habits, psychological states, and living environments are usually unstable. Ignoring the changes in learner behavior characteristics will lead to incomplete learner portraits, thereby affecting a comprehensive understanding of learner characteristics.

[0007] To solve the above problems, the present invention proposes a method for constructing and analyzing multi-dimensional digital profile Tags based on data fusion (CAMTD). This method can analyze learner behaviors from multiple dimensions and construct a comprehensive and accurate learner portrait to help educators understand learners' characteristics and personalized needs efficiently. Summary of the invention

[0008] The present invention provides a method for constructing and analyzing multidimensional learner digital portrait labels based on data fusion. The purpose of increasing the learner information dimension is achieved by using a data fusion method based on normal distribution, and a method for constructing learner digital portrait labels based on social cognitive theory is proposed to enhance the scientific nature of label construction. The MLog normalization method combining traditional Log normalization and Min-max normalization methods improves the processing ability of skewed data, and uses the HDBSCAN clustering method to divide learner groups to reduce the influence of subjective selection of model parameters. Finally, a multi-label multi-classification model based on the PyTorch deep learning model is used to predict the group to which a single learner belongs, and personalized guidance is provided for learners in combination with social cognitive theory. The effectiveness of the proposed method in the task of predicting student grades is verified through experiments on public data sets. Compared with other existing digital portrait construction methods, it has been improved in all indicators.

[0009] To achieve the above object, the present invention adopts the following technical solution:

[0010] A method for constructing and analyzing multidimensional learner digital portrait labels based on data fusion, comprising the following steps:

[0011] Step 1: Select raw data from two public datasets. First, select the major 8 with a large number of students in the campus behavior information. This information reflects some of the students' consumption and commuting behaviors on campus. The locations where the behaviors occur include 14 locations such as canteens, dormitories, supermarkets, and libraries. Second, select the course A with a large number of learners in the learning behavior information. This information reflects information such as the frequency of students' study of learning materials.

[0012] Step 2: Divide the learners’ scores. Divide the learners according to the K percentile method based on their professional ranking or course score ranking. The selected K percentile points are [0, 2.5, 16, 50, 84, 97.5, 100]. Each data set is divided into 6 different data subsets, and the score levels increase in sequence. The data subsets obtained by dividing multiple different data sets using the same boundary points are considered as a group.

[0013] Step 3: Learner data fusion based on normal distribution. According to the results generated in step 2, randomly select the same number of learner samples from each group, and concatenate the learners with the same ranking. Given that the two datasets contain 700 and 677 learner samples, respectively, 500 learner samples are extracted from each dataset. The number of samples extracted from each data subset is determined by the data distribution law of normal distribution.

[0014] Step 4: Construct a scientifically based learner label based on social cognitive theory. The theory mainly includes five core elements: observational learning, imitative learning, self-efficacy, self-regulation, and reproduction determinism. Analyze the meaning of the core elements and summarize the behavior categories contained in each core element: 8 behavior categories such as observational behavior, imitative behavior, persistence behavior, and situational behavior. Combined with the fused data set, analyze the specific learner behaviors contained in each behavior category, such as classroom classes, online classes, autonomous learning, and other 10 specific behaviors. Finally, abstract a single or partial specific behavior into a label.

[0015] Step 5, use the MLog method to construct digital labels. The present invention combines the traditional Log normalization method and the Min-Max normalization method and proposes the MLog normalization method, which can effectively process skewed data and normalize the values ​​to [0,1] while maintaining the numerical relationship so as to be more flexible for data analysis or machine learning. The present invention uses a statistical method to initialize the label value of each learner, and then inputs the initialized labels of all learners into the MLog normalization method and uses the calculated results as digital labels.

[0016] Step 6: Use the HDBSCAN algorithm to divide the learner groups. When dividing students into groups, the HDBSCAN algorithm uses a density-based tree pruning method to determine how to cut the hierarchical clustering tree into flat clusters. In addition, the algorithm does not require the specification of parameters such as the number of clusters and the minimum distance between samples. Therefore, there is almost no need to adjust parameters to obtain good clustering results, which helps to determine the number of learner groups more objectively.

[0017] Step 7: Use a multi-label multi-classification model based on the PyTorch deep learning framework to predict digital labels. Use the deep learning module to extract deep features between sample data to better discover the rules between samples, and use multiple MLP classifiers to achieve multi-label multi-classification prediction to enhance the accuracy of each label prediction.

[0018] Step 8: Use the Minkowski distance algorithm to determine the group to which a single learner belongs. Calculate the distance between a single learner and each cluster center point based on the prediction results of the multi-label multi-classification model, and regard the cluster with the closest cluster center point as the learner group to which the single learner belongs.

[0019] Step 9: Compare the experimental results of different methods on learner group segmentation to verify the effectiveness of this method.

[0020] Step 10: Verify the generalization ability of this method in learner profile and digital label prediction tasks by predicting unknown students in the dataset.

[0021] Step 11: Use the PCA method to reduce the dimension of the digital labels and visualize the clustering results to further analyze the reasons.

[0022] Step 12: Use Plotly to visualize the portrait results, combine social cognitive theory to analyze the characteristics of each learner group and provide personalized suggestions.

[0023] The present invention proposes the construction and analysis of multi-dimensional learner portrait labels based on data fusion. This data fusion method can effectively integrate multi-source data sets from different scoring standards to enrich and improve learners' information. Through this multi-dimensional digital label, the present invention can conduct in-depth analysis of learners from multiple perspectives and establish a comprehensive and accurate learner portrait to better understand their characteristics and needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is the overall framework diagram of the technical solution of the present invention;

[0025] Figure 2 It is a data fusion diagram based on normal distribution;

[0026] Figure 3 Clustering diagrams based on different normalization methods; (a) Clustering based on original features; (b) Clustering based on L2 regularization; (c) Clustering based on Log normalization; (d) Clustering based on MLog normalization.

[0027] Figure 4 It is a multi-label multi-classification prediction model;

[0028] Figure 5The results of learner portraits based on label values: (a) balanced development explorers; (b) fully engaged learners; (c) social growth improvers; (d) persistent learners; (e) dedicated and excellent learners; (f) comprehensive comparison of learners. DETAILED DESCRIPTION

[0029] like Figure 1 , Figure 2 As shown, the present invention provides a method for constructing and analyzing multidimensional learner portrait labels based on data fusion, comprising the following steps:

[0030] Step 1: Select raw data from two public datasets. First, select the major 8 with a large number of students in the student campus behavior information. This information reflects some of the students' consumption and commuting behaviors on campus. The locations where the behaviors occur include 14 locations such as canteens, dormitories, supermarkets, and libraries. Second, select the course A with a large number of students in the student learning behavior information. This information reflects the frequency of students' study of learning materials.

[0031] Step 2: Divide the learners’ scores. Divide the learners according to the K percentile method based on their ranking in the major or the course score. The selected K percentile points are [0, 2.5, 16.50, 84, 97.5, 100]. Each data set is divided into 6 different data subsets, and the score levels increase in sequence. The data subsets obtained by dividing multiple different data sets with the same boundary points are regarded as a group. Figure 2 As shown:

[0032]

[0033] count i =n * ×(k i+1 -k i ), 0≤i<num(K), 0≤n * <min(Datas) (2)

[0034] Among them, n j represents the number of samples of the jth learner data, num(Datas) represents the number of learner data, which is 2 in this case, and k i Represents the percentage value of the i-th boundary point, the calculation result Represents the specific location of the ith boundary point of the ith learner data set. In formula (2), represents the number of samples in the final fused learner data set, num(K) represents the number of elements in the cut point set K, min(Datas) represents the minimum number of samples between multi-source learner data, and the calculation result count iRepresents the number of samples drawn from the i-th cut-point data set.

[0035] Step 3: Learner data fusion based on normal distribution. According to the results generated in step 2, randomly select learners with a certain number of data from each group, and splice the data of learners with the same ranking. Given that the two data sets contain 700 and 677 learner samples respectively, 500 learner samples are extracted from each data set. The number of samples extracted from each data subset is determined by the data distribution law of normal distribution. Figure 2 As shown, in order to further merge into complete learner data, the feature matrices of randomly sampled instances in the same group are merged horizontally to increase the information dimension of each learner. Then, these groups are merged vertically to form the final fused data set. The specific calculation process is as follows:

[0036]

[0037]

[0038] data=[data0;data1;…;data num(k)-1 ] (5)

[0039] Among them, GNSUN (n, a, b) means randomly selecting n learner samples from the starting cut point a and the ending cut point b. i It is the result of horizontal splicing of the features of the i-th learner, and data is the fused data set after vertical splicing of all learners.

[0040] Step 4: Construct a scientifically based learner label based on social cognitive theory. This theory mainly includes five core elements: observational learning, imitative learning, self-efficacy, self-regulation, and reproduction determinism. Analyze the meaning of the core elements and summarize the eight behavior categories contained in each core element, such as observation behavior, imitative behavior, persistence behavior, and situational behavior. Combined with the fused data set, analyze the specific learner behaviors contained in each behavior category, such as classroom classes, online classes, autonomous learning, and other 10 specific behaviors. Finally, abstract a single or multiple behaviors into a label.

[0041] Step 5. Use the MLog method to construct digital labels. The traditional Log normalization method and the Min-Max normalization method are combined to propose the MLog normalization method, which can effectively process skewed data and normalize the values ​​to [0,1] while maintaining the relationship between the numerical values ​​so that it can be more flexibly used in data analysis or machine learning. The present invention uses a statistical method to initialize the label value of each learner, and then inputs the initialized labels of all learners into the MLog normalization method and uses the calculated results as digital labels. There are two methods for generating initial labels: one is based on a single learner behavior data, and the other is based on multiple learner behavior data. This study uses the formula MLog to generate the first type of digital labels. The main labels include learning costs, dining expenses, living expenses, learning activities, social interactions, and self-management. The calculation process is:

[0042] Mlog(x)=log(λ+1 / x) / log(λ+1 / x min ),x∈{SC,DC,LC,SF,DF,LF} (6)

[0043] Among them, x is one of the elements in the initial label set, x min is the minimum value among all sample labels, and λ is the correction parameter. The calculation result of Mlog(x) is used as a digital mark. As can be seen from the formula, the value of Mlog(x) decreases as x increases. Therefore, any data set of data size can be normalized to [0,1] while retaining the original numerical size relationship. Subsequently, the second category of labels is calculated. Given that the second category of labels is affected to different degrees by each initial value, we assign different weights to each initial value, mainly including reproduction imitation, knowledge mastery, and self-persistence. The detailed calculation process is:

[0044] Reappearance=w1×Mlog(WB)+w2×Mlog(RB),w1+w2=1 (7)

[0045]

[0046] Persistence=w1×(1-Mlog(ST))+w2×Mlog(CL),w1+w2=1 (9)

[0047] Among them, WB, RB, TS i , ST, and CL are the initial data of various behaviors of learners.

[0048] Step 6. Use the HDBSCAN algorithm to divide the learner groups. The HDBSCAN algorithm is an optimization and improvement of the traditional DBSCAN algorithm. It uses the HDBSCAN clustering algorithm to divide the learner groups. The minimum spanning tree MST constructs a weighted minimum spanning tree based on the kth distance and the actual distance between each team of data points, and constructs a hierarchical structure from low density to high density. It can not only handle the noise in the data, but also discover clusters of different densities. This can help educators discover different learner groups and conduct targeted analysis, so as to better provide personalized education. The specific calculation formula is:

[0049] N ε (p)={q∈D|dist(p, q)≤ε} (10)

[0050] core(p)={p∈D||N ε (P)|≥MinPts} (11)

[0051] Where ε is the distance threshold for defining the domain, D is the dataset, and dist(p, q) is the distance measure between points p and q. If there are at least MinPts points in the ε domain of point p, then p is a core point. If q is in the ε domain of point p at this time, then q is said to be directly density reachable from p. Select a point p that has not been visited. If p is a core point, start a new cluster and add all density reachable points from p to the cluster. Otherwise, the point is a noise point.

[0052] Step 7: Use a multi-label multi-classification model based on the PyTorch deep learning framework to predict digital labels. Considering that in actual application scenarios, new learners may join the group, reprocessing the data and dividing the groups will consume a lot of resources. Therefore, the present invention proposes a multi-label multi-classification prediction model based on the PyTorch deep learning framework. Use S = (s1, s2, ..., s n ) represents the set of all learners in the fusion data, and n represents the total number of learners in the fusion data set. i =(x1, x2, ...x m , z1, z2, …, z j ) represents the feature tensor of a single learner in the fusion data, x i , and z i represents the features extracted from the campus card swiping dataset and the online course dataset respectively, and m and j represent the number of extracted features respectively. Represents the model for learner s i The prediction result of , g represents the number of initialization labels.

[0053] Considering that the learner's label value is a linear value, the learner label value is divided into four ranges when predicting the learner label, namely [0, 0.25], [0.25, 0.5], [0.5, 0.75] and [0.75, 1], and each range is a category. If the predicted value and the label are in the same range, the label prediction is considered correct. The specific calculation formula is:

[0054]

[0055]

[0056] Where ReLU(W×s i +b) represents a fully connected layer, and a nonlinear activation function ReLU is added after the matrix mapping is completed. This function can reduce the problem of gradient disappearance and better learn the nonlinear relationship between features.

[0057] Step 8: Combine the clustering results obtained in step 6 and take the most dense part of the cluster as the center point of the cluster, and take the center point as the representative point of each group. Calculate the average distance from each point in the cluster to other points To measure the density of the point, the point with the smallest density is taken as the center point of the cluster. The calculation process is:

[0058]

[0059] Wherein, p is a variable that determines different distance metrics. In the present invention, p=2, and the Minkowski distance is the Euclidean distance.

[0060] Step 9: Compare the experimental results of different methods for learner group segmentation to verify the effectiveness of this method. The specific process is: In order to prove the effectiveness of the MLog normalization method proposed in this chapter, we compared Mlog with five traditional normalization techniques: L2 regularization, minimum-maximum normalization, maximum-absolute value normalization, Z-score and Log normalization.

[0061] The experimental results are shown in Table 8. It can be observed from the table that compared with the most effective traditional method (logarithmic normalization), the SC value of HDBSCAN reached 0.7682, an increase of 0.36, and other indicators were also improved. In addition, compared with the campus card swiping dataset, the fusion dataset proposed in this chapter has also been greatly improved. The reasons for this may be: minimum-maximum normalization and maximum-absolute value normalization limit the original features to the range of [0, 1], thereby reducing the impact of inconsistent feature scales on data stability. However, these methods are susceptible to outliers. In contrast, Z-score and logarithmic normalization are opposite to the previous two algorithms. Although L2 normalization can reduce these two effects, it only retains the directional information of the feature vector and ignores the size relationship of the feature value, so it is not suitable as a digital label.

[0062] Step 10: By classifying and predicting unknown students, the generalization ability of this method in learner portrait and digital label prediction tasks is verified. Compared with other digital models, the effectiveness of the proposed method is demonstrated. The experimental results are shown in Table 9. The classification model is as follows Figure 4 As shown. In order to further verify the effectiveness of the constructed labels proposed in this study, this chapter compares it with the box plot K-percentile digital label generation method. This invention introduces a multi-label multi-classification model based on the PyTorch deep learning framework. The model consists of multiple MLP (multi-layer perceptron) classifiers. In order to demonstrate the predictive performance of the model, the present invention uses two different methods to generate digital labels, and divides the digital labels into four different categories in proportion (1:1:1:1). According to the feature extraction method detailed in Tables 6 and 7, a multi-label multi-classification model is used to predict each label simultaneously. The experimental results show that the Mlog digitization method has an average improvement of about 2% in accuracy, recall, precision and F1 score. This shows that the Mlog digitization method can effectively distinguish learners' behavioral patterns and learning preferences, and shows excellent performance in quantifying these features.

[0063] Step 11: Use PCA method to reduce the dimension of digital labels and visualize the clustering results to further analyze the feasibility of the method proposed in the present invention. The normalized clustering results are as follows: Figure 3 shown.

[0064] It can be observed from the figure that the sample characteristics of the original data show a skewed distribution pattern. At the same time, the DBSCAN algorithm divides the population into two clusters, but the boundaries between the clusters are not clear. Figure 3 (a) and Figure 3(b), and considering the results of Experiment 1 in Table 1, it can be seen that L2 normalization significantly improves the inter-cluster SC, but has poor processing capabilities for skewed data. This shortcoming may be due to the fact that L2 normalization focuses on retaining the direction of the original feature vector, thereby reducing the noise of a single sample feature, but does not consider the numerical size relationship between multiple sample features.

[0065] Comparing Experiment 1 and Experiment 5 in Table 8, we can find that logarithmic normalization does not significantly improve CH, which indicates that the distribution of sample points in the cluster is not uniform. Figure 3 As shown in (b), the feature data show strong distinguishability on the x-axis. The possible reason is that the effectiveness of this method is limited to certain types of data features.

[0066] Further analysis found that the logarithmic normalization method can only show strong discrimination ability when the eigenvalue is close to 1. Therefore, we introduced the inverse of the eigenvalue into the traditional logarithmic normalization method, as shown in formula (6). Figure 3 (c) and Figure 3 (d) It can be found that Figure 3 In (d), the distribution of sample points in the clusters is denser and more uniform, and there is a clear distinction between clusters on both the x-axis and the y-axis. This observation verifies that the proposed method can effectively distinguish data of different scales.

[0067] Step 12: Use Plotly to visualize the portrait results, combine social cognitive theory to analyze the characteristics of each type of learner and give personalized suggestions. According to the results of Experiment 6 in Table 7, we divide the learner group characteristics into five categories: balanced development explorers, fully engaged learners, social growth progressers, persistent practitioners, and dedicated learners. In order to show these group characteristics more clearly and intuitively, we use radar charts to visualize the digital trajectory labels of each learner group. The visualization results are shown below: Figure 5 shown.

[0068] Balanced development explorers, who have no obvious weaknesses but also lack significant strengths. Teachers should adopt strategies to provide diversified resources, offer a variety of learning materials, and encourage participation in various activities. This approach will not only help these learners explore and determine the most effective learning methods to meet their individual needs, but it is also likely to discover and stimulate their potential abilities or preferences.

[0069] Fully engaged learners who demonstrate excellent self-management skills and are very active in social interactions and learning environments. Educators should provide them with richer opportunities for teamwork and collaborative learning, and can promote the continuous progress and growth of these learners by setting specific and challenging learning goals.

[0070] Social progressors, who usually excel in social interaction but may have difficulty completing learning tasks independently and maintaining learning motivation, should learn through collaboration by integrating them into group learning or projects. In addition, teachers should help these learners set short-term goals and provide them with positive feedback and support. This approach aims to cultivate their willingness to persist in learning.

[0071] Educators should give greater autonomy to perseverants, who demonstrate high levels of diligence, engagement, and academic achievement. This includes helping them set challenging but achievable goals and encouraging them to persist in creative and critical thinking to achieve those goals. In this way, educators can further enhance and maintain their sense of self-efficacy, thereby promoting continued personal growth and motivation to learn.

[0072] Students who are focused and good at learning like to work alone and may seem a little withdrawn. While retaining their ability to self-regulate, educators need to cultivate their ability to interact and cooperate by providing a safe interactive space. Online learning platforms with relatively low social barriers can help these learners strengthen their interaction with others while maintaining their independence.

[0073] Table 1

[0074]

[0075] Table 2

[0076]

[0077] Table 3

[0078]

[0079] Table 4

[0080]

[0081] Table 5

[0082]

[0083] Table 6

[0084]

[0085]

[0086] Table 7

[0087]

[0088] Table 8

[0089]

[0090] Table 9

[0091]

Claims

1. A method for constructing and analyzing multidimensional learner portrait labels based on data fusion, characterized in that: The following steps are involved: Step 1: Select raw data from two public datasets. First, select the majors with a large number of students in the learners’ campus behavior information to reflect the students’ consumption and commuting behaviors on campus. The locations where the behaviors occur include canteens, dormitories, supermarkets, and libraries. Second, select course A with a large number of students in the learners’ learning behavior information to reflect the learning frequency of students’ access to learning materials. Step 2: Divide the learners’ scores. Divide the learners according to the K percentile method based on their professional ranking or course score ranking. The selected K percentile points are [0, 2.5, 16, 50, 84, 97.5, 100]. Each data set is divided into 6 different data subsets, and the score levels increase in sequence. The data subsets obtained by dividing multiple different data sets using the same boundary points are considered as one group. Step 3: Learner data fusion based on normal distribution law; According to the generated results of step 2, randomly select the same number of learner samples from each group, and splice the data of learners with the same ranking; Given that the two data sets contain 700 and 677 learner samples respectively, a total of 500 learner samples are extracted from each data set, and the number of samples extracted from each data subset is determined by the data distribution law of normal distribution; Step 4: Construct a scientifically based learner label based on social cognitive theory; social cognitive theory includes five core elements: observational learning, imitative learning, self-efficacy, self-regulation, and representational determinism; analyze the meaning of the core elements and summarize the behavior categories contained in each core element: observational behavior, imitative behavior, persistence behavior, and situational behavior; analyze the specific learner behaviors contained in each behavior category based on the fused data set, and finally abstract a single or partial specific behavior into a label; Step 5: Use the MLog method to construct digital labels; combine the Log normalization method with the Min-Max normalization method and propose the MLog normalization method; use a statistical method to initialize the label value of each learner, then input the initialized labels of all learners into the MLog normalization method and use the calculated results as digital labels; Step 6: Use the HDBSCAN algorithm to divide the learner groups. When dividing the students into groups, the HDBSCAN algorithm uses a density-based tree pruning method to decide how to cut the hierarchical clustering tree into flat clusters. Step 7: Use a multi-label multi-classification model based on the PyTorch deep learning framework to predict digital labels; use a deep learning module to extract deep features between sample data to discover the rules between samples, and use multiple MLP classifiers to achieve multi-label multi-classification prediction to enhance the accuracy of each label prediction; Step 8: Use the Minkowski distance algorithm to determine the group to which a single learner belongs; calculate the distance between a single learner and the center point of each cluster based on the prediction results of the multi-label multi-classification model, and regard the cluster with the closest cluster center point as the learner group to which the single learner belongs; Step 9: Compare the experimental results of different methods on learner group division to verify the effectiveness of this method; Step 10: Verify the generalization ability of the learner profile and digital label prediction tasks by performing classification prediction on unknown students; Step 11: Use the PCA method to reduce the dimension of the digital labels and visualize the clustering results to further analyze the reasons; Step 12: Use Plotly to visualize the portrait results, combine social cognitive theory to analyze the characteristics of each learner group and provide personalized suggestions.

2. The multi-dimensional learner portrait label construction and analysis based on data fusion as claimed in claim 1, characterized in that: Step 2 is as follows: Use K = [0, 2.5, 16, 50, 84, 97.5, 100] to represent the threshold for dividing the learner groups, n represents the total number of students in the campus card swiping data set, and n j represents the number of samples of the jth learner group, num(Datas) represents the number of learner data sets, k i Represents the percentage value of the i-th cut point, the calculation result represents the specific location of the i-th cut point of the j-th learner group; in formula (2), represents the number of samples in the final fused learner data set, num(K) represents the number of elements in the cut point set K, min(Datas) represents the minimum number of samples between multi-source learner data, and the calculation result count i Represents the number of samples extracted from the i-th cut-point data set; count i =n * ×(k i+1 -k i ),0≤i<num(K),0≤n * <min(Data) (2)。 3. The multi-dimensional learner portrait label construction and analysis based on data fusion as claimed in claim 1, characterized in that: Step 3 is as follows: data fusion based on normal distribution; to fuse into complete learner data, the feature matrices of randomly sampled instances at the same performance level are horizontally spliced ​​to increase the information dimension of each learner; then, all learner samples are vertically spliced ​​to generate the final fused data set; the specific calculation process is as follows: data=[data0;data1;…;data num(k)-1 ] (5) Among them, GNSUN(n,a,b) means randomly selecting n learner samples from the starting cut point a and the ending cut point b; data i It is the result of horizontal splicing of the features of the i-th learner, and data is the fused data set after vertical splicing of all learners.

4. The method for constructing and analyzing multidimensional digital portrait labels based on data fusion according to claim 1, characterized in that: Step 5 is as follows: using the learner's initialization label as input, there are two methods to generate the initial label: one is based on single-category data and the other is based on multi-category data; using formula MLog to generate the first category of digital labels, including learning cost, dining expenses, living expenses, learning activities, social interaction and self-management; the calculation process is: Mlog(x)=log(λ+1 / x) / log(λ+1 / x min ),x∈{SC,DC,LC,SF,DF,LF} (6) Among them, x is one of the data in the initial label set, x min is the minimum value among all sample labels, λ is the correction parameter; the calculation result of Mlog(x) is used as the digital mark; the value of Mlog(x) decreases as x increases, and the data set of any data size is normalized to [0,1], while retaining the original numerical size relationship; calculate the second category label. In view of the different degrees of influence of each initial value on the second category label, different weights are assigned to each initial value, including reproduction imitation, knowledge mastery and self-persistence labels. The specific calculation process is as follows: Reappearance=w1×Mlog(WB)+w2×Mlog(RB),w1+w2=1 (7) Persistence=w1×(1-Mlog(ST))+w2×Mlog(CL),w1+w2=1 (9) Among them, WB, RB, TS i , ST, and CL are the initial data of various behaviors of learners.

5. The method for constructing and analyzing multidimensional digital portrait labels based on data fusion according to claim 1, characterized in that: Step 6 is as follows: HDBSCAN clustering algorithm is used to divide the learner group. The minimum spanning tree (MST) constructs a weighted minimum spanning tree based on the kth distance and the actual distance between each group of data points, and constructs a hierarchical structure from low density to high density. The specific calculation formula is: N ε (p)={q∈D|dist(p,q)≤ε} (10) core(p)={p∈D||N ε (P)|≥MinPts} (11) Where ε is the distance threshold for defining the domain, D is the data set, and dist(p,q) is the distance between points p and q. If there are at least MinPts samples in the ε domain of point p, then p is a core point. If q is in the ε domain of point p at this time, then q is said to be directly density reachable from p. Select a point p that has not been visited. If p is a core point, start a new cluster and add all points that are density reachable from p to the cluster, otherwise the point is a noise point.

6. The method for constructing and analyzing multidimensional digital portrait labels based on data fusion according to claim 1, characterized in that: Step 7 is as follows: propose a multi-label multi-classification prediction model based on the PyTorch deep learning framework; use S = (s1, s2, ..., s n ) represents the set of all learners in the fusion data, n represents the total number of learners in the fusion data set; s i =(x1,x2,…x m ,z1,z2,…,z j ) represents the feature tensor of a single learner in the fusion data, x i and z i They represent the features extracted from the campus card swiping dataset and the online course dataset, respectively, and m and j represent the number of extracted features; Represents the model for learner s i The prediction result of , g represents the number of initialization labels; Considering that the learner's label value is a linear value, the learner label value is divided into four ranges when predicting the learner label, namely [0, 0.25], [0.25, 0.5], [0.5, 0.75] and [0.75, 1], each range is a category; if the predicted value and the label are in the same range, the label prediction is considered correct; the specific calculation formula is: Where ReLU(W×s i +b) represents a fully connected layer, and a nonlinear activation function ReLU is added after the matrix mapping is completed.

7. The method for constructing and analyzing multidimensional digital portrait labels based on data fusion according to claim 1, characterized in that: Step 8 is as follows: Combine the clustering results obtained in step 6 to take the densest part of the cluster as the center point of the cluster, and take the center point as the representative point of each group; calculate the average distance from each point in the cluster to other points To measure the density of the point, the point with the smallest density is taken as the center point of the cluster. The calculation process is: Among them, p is a variable that determines different distance metrics, and Minkowski distance is Euclidean distance.

8. The method for constructing and analyzing multidimensional digital portrait labels based on data fusion according to claim 1, characterized in that: Step 12 is as follows: Based on the social cognitive education theory and combined with the learner group portrait, learners are divided into five categories: balanced development explorers, comprehensive participation learners, social growth progressers, persistent learners and dedicated learners.

Images