Learning style identification method based on fusion label and stacked machine learning model

By integrating labels and stacking machine learning models, combining the Kolb learning style scale and the online learning behavior scale, the problems of high subjectivity, staticity and inaccurate recognition of learning styles in the existing technology are solved, and the online education environment is realized The personalized learning style recognition in the system improves the recognition accuracy and dynamics.

CN113408576BActive Publication Date: 2024-11-08SHANGHAI NORMAL UNIVERSITY
View PDF 14 Cites 0 Cited by

Patent Information

Application Number
CN202110517025.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-05-12
Publication Date
2024-11-08
Estimated Expiration
2041-05-12

AI Technical Summary

Technical Problem

The existing learning style recognition methods have problems such as high subjectivity, staticity, inaccurate recognition and inability to change dynamically, especially in the online education environment, which is difficult to achieve personalized learning.

Method used

Using a method based on fusion labels and stacked machine learning models, the learning style is divided through the intersection of the Kolb learning style scale and the online learning behavior scale, and combined clustering and correlation testing, a stacked machine learning model is constructed for dynamic recognition of learning styles.

Benefits of technology

It realizes more objective and accurate learning style recognition, reduces subjective problems caused by single-component scales, can dynamically identify learning styles, improves recognition accuracy, and avoids the problem of insufficient data interpretability in deep learning.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113408576B_ABST
    Figure CN113408576B_ABST
Patent Text Reader

Abstract

The present invention relates to a learning style identification method based on fusion label and stacked machine learning model, comprising the steps of: 1) using the Kolb learning style scale and the online learning behavior survey scale to respectively calculate the learning style corresponding to the learner, and taking the intersection of the two calculation results to obtain the divided and undivided learning styles; 2) clustering the divided and undivided learning styles, determining the learning styles and labeling them, and supplementing the undivided learning styles; 3) performing a correlation test on the obtained learning style labels and the online learning behavior characteristics collected by the online learning platform; 4) selecting training data and test data, and training them using the stacked machine learning model to obtain a complete stacked model; 5) performing a comprehensive performance evaluation on the trained stacked model. Compared with the prior art, the present invention has the advantages of reducing the difficulty of model training and improving the recognition accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of information recognition technology, and in particular to a learning style recognition method based on fusion labels and stacked machine learning models. Background Art

[0002] Online education eliminates the time and space limitations of traditional education, allowing teachers and students to communicate anytime and anywhere, which makes it possible to realize Confucius's "teaching students in accordance with their aptitude". Scholars such as Dunn, Kolb, Felder, and Keefe have long recognized that students have different styles of learning new knowledge. The differences among students include personality traits, knowledge levels, learning abilities, and learning styles. Among them, learning styles include learning preferences and learning characteristics. Finding a learning style that suits students can guide students' learning, so the automatic recognition of learning styles is essential to promote personalized learning in online education environments.

[0003] The traditional method of identifying learning styles is to ask learners to fill out a learning style scale. Although this type of method is effective, it still has some shortcomings. First, the designers of traditional learning style scales cannot avoid subjective factors when preparing questions; second, learners' lack of clear understanding of themselves when filling out the scale will directly lead to inaccurate learning style identification; third, when the evaluation indicators of multiple learning styles have consistent scores, the scale may not be able to identify the learning style; finally, students' learning styles have the characteristics of dynamic changes, and scale-based identification is a static method. Therefore, under the current situation, using multi-source heterogeneous online education data to implicitly and dynamically identify learners' learning styles can solve a series of problems in traditional learning style identification, such as inability to identify, inaccurate identification, high subjectivity, and static identification, and provide methodological support for the realization of personalized learning in the context of online education.

[0004] At present, scholars at home and abroad have conducted various studies on the identification of learning styles. There are two methods for identifying learning styles: the scale identification method based on custom rules and the automatic identification method based on machine learning technology. When it is inconvenient to collect learners' online learning behavior data, the scale identification method is a traditional method. The automatic identification method can classify learners' learning styles by cleaning and integrating students' online learning behavior data. BrahimHmedna et al. designed a method that can automatically identify learning styles using the learning behavior data generated by students in MOOC. The essence of this model is to cluster the data to obtain learning style labels, and then use classification algorithms to predict learning styles. The problem with this method is that it is not interpretable enough, because the actual meaning of the clustering results needs to be explained by manual methods, and this method directly uses the clustering results obtained by the distribution of the data itself as the learner's learning style label. Chia-Cheng Hsu et al. proposed a FIS model based on neuron fuzzy reasoning to identify the learning style of online learners; its essence is to use custom rules to identify learning styles through a single hidden layer neural network. Since the number of learners in the experimental verification part is only 102, the interpretability of this method is insufficient when the learning style distribution is uneven. The EENN-PSO model proposed by Song Lai et al. achieved a high accuracy in identifying learning styles, but the problem with this study is that only the NEO-FFI scale is used to obtain learning style labels, and the use of a single scale is too subjective. In addition, deep learning has recently been applied to identifying learning styles. Zhang Hao et al. constructed a deep belief neural network for learning style detection, and the highest recognition accuracy (Vis / Vrb) of a single learning style was 0.89. However, deep learning methods require a large amount of data to train the model, which will take a lot of time and cannot guarantee the interpretability of the data. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a learning style identification method based on fusion label and stacked machine learning model.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A learning style identification method based on fusion label and stacked machine learning model, the method includes the following steps:

[0008] S1: Use the Kolb Learning Style Inventory and the Online Learning Behavior Survey Scale to calculate the corresponding learning styles of learners respectively, and take the intersection of the two calculation results to obtain the divided and undivided learning styles.

[0009] S2: Cluster the divided and undivided learning styles, determine the learning styles and label them, and supplement the undivided learning styles.

[0010] S3: Perform a correlation test on the learning style label obtained in step S2 and the online learning behavior characteristics collected by the online learning platform.

[0011] S4: For the online learning behavior features that pass the correlation test, select training data and test data, and use the stacked machine learning model for training to obtain a complete stacked model.

[0012] S5: Conduct a comprehensive performance evaluation on the trained stacking model, optimize the stacking model, and use the optimized stacking model to predict the learning style of actual learners.

[0013] The specific steps of step S1 include:

[0014] 11) Obtain the learner's Kolb Learning Style Scale and calculate the scores of each learning style;

[0015] 12) Obtain the learner's online learning behavior survey scale, calculate each learning style score, take the intersection with each learning style score obtained based on the learner's Kolb learning style scale, and obtain a set of successfully divided learning styles and a set of unsuccessfully divided learning styles.

[0016] The calculation formula of the learning style score is:

[0017]

[0018] Where: LS k is the learning style score, a i,j is the score of the jth option of the ith question, and k represents one of all the learning styles of the learner.

[0019] The specific content of step S2 is:

[0020] After feature dimensionality reduction of the online learning behavior survey scale data, K-Means++ was used to cluster the reduced data to obtain four clusters. Then, a cluster meaning determination method based on expert annotation was constructed to determine the position of the centroid of each cluster. Then, the Euclidean distance from the remaining points to the centroid was calculated, and the centroid selection method was used to select the sample points within the threshold closest to the centroid in the cluster. The learning style of the sample was determined and labeled based on the Kolb learning style scale data and the online learning behavior scale data of the sample point, and the unclassified learning style was supplemented.

[0021] Furthermore, the learning styles of the samples are determined and labeled based on the Kolb Learning Style Scale data and the Online Learning Behavior Scale data of the sample points, and the specific contents of the unclassified learning styles are supplemented as follows:

[0022] If it is determined that more than 50% of the sample points have the same learning style, the intrinsic meaning of the cluster is determined to be the learning style. If the condition of more than 50% is not met, the threshold will be expanded at a rate of 1 until the learning style can be successfully divided; finally, the learning styles that cannot be divided will be supplemented.

[0023] In step S3, the learning style label obtained in step S2 and the online learning behavior characteristics collected by the online learning platform are tested for correlation using the Spearman correlation coefficient.

[0024] In step S4, the specific steps of using the stacked machine learning model for training include:

[0025] 41) Preprocessing the behavioral data generated by learners in the online teaching platform;

[0026] 42) Combining the learning style annotation results obtained in step S2, constructing a stacked machine learning model;

[0027] 43) Perform model training and parameter adjustment on the constructed stacked machine learning model to obtain a complete stacked model.

[0028] The stacked machine learning model is a fusion model based on a two-layer model. The first layer includes four basic classifiers: random forest, gradient boosting decision tree, support vector machine and multilayer perceptron. The input of the first layer is the original training set and test set. The second layer is a logistic regressor. The second layer uses the output of the basic classifier of the first layer as input and is retrained by adding the training set to obtain a complete stacked model.

[0029] Furthermore, before step 43), a sample resampling step using a SMOTE algorithm is also included.

[0030] In step S5, the accuracy, recall, precision, F1 score and area under the curve are used to perform a comprehensive performance evaluation on the trained stacking model to optimize the stacking model.

[0031] The learning style identification method based on fusion label and stacked machine learning model provided by the present invention has at least the following beneficial effects compared with the prior art:

[0032] 1) The present invention proposes a method for dynamically identifying learning styles. The method obtains the learner's learning style label by processing two questionnaires filled out by the learner, and then uses the data generated by the student on the online teaching platform to predict his or her learning style. The amount of data processed is large and rich, and the data structure is a second-order tensor, which can reduce the difficulty of model training and the time required for identification;

[0033] 2) The method of the present invention adopts learning style clustering based on fusion labels. By fusing the annotation results of two scales, the highly subjective problem caused by a single scale can be reduced. By using the Spearman correlation coefficient to compare the learning style labels divided by LSDM, Kolb learning style scale, and online learning behavior scale and the learning behavior characteristics on the online learning platform, a correlation analysis is performed, which is conducive to obtaining more objective and accurate recognition results;

[0034] 3) The method of the present invention constructs a two-layer stacking model based on resampling technology, which can predict learning style by integrating learners' learning behavior data into the online teaching platform, thereby reducing the problem of poor prediction performance caused by category imbalance, and has higher recognition accuracy than traditional machine learning methods.

[0035] 4) The present invention is based on a machine learning model, which involves a large amount of data and has a second-order tensor structure. All features of the experiment are interpretable, that is, there is no need to use the existing deep learning model to conduct experiments, thus avoiding the problem of destroying interpretability by using deep learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A schematic diagram of a flow chart of a learning style identification method based on fusion labels and stacked machine learning models in an embodiment;

[0037] Figure 2 It is a principle block diagram of a learning style identification method based on fusion labels and stacked machine learning models in an embodiment;

[0038] Figure 3 A schematic diagram of the overall framework of the learning style label division method constructed by the present invention in the embodiment;

[0039] Figure 4 A schematic diagram of a cluster meaning determination method constructed by the present invention in an embodiment;

[0040] Figure 5 It is intended to represent the online learning behavior quantity constructed by the present invention in the embodiment;

[0041] Figure 6 Schematic diagram of the overall framework of the learning style prediction method constructed by the present invention in the embodiment. DETAILED DESCRIPTION

[0042] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work should belong to the scope of protection of the present invention.

[0043] Example

[0044] The purpose of the present invention is to target the application field of learning style identification. The proposed method based on fusion labels and stacked machine learning models can implicitly and dynamically identify the learning style of learners in the online learning process. According to Kolb's learning style theory, learners' learning styles are divided into four types: divergent, centralized, accommodative, and assimilative. Since learning styles change with the increase of learners' age, cognitive level, and environment, this variability makes the method of statically obtaining learning styles with scales unreliable. In addition, because it is impossible to guarantee that the collected data samples have balanced learning style labels, in order to ensure the accuracy of recognition, a resampling technique is used to improve the problem of data sample imbalance while using a stacked machine learning model to identify learning styles, so as to obtain a higher recognition rate.

[0045] The present invention relates to a learning style identification method based on fusion label and stacked machine learning model. The basic process of the method is as follows: Figure 1 As shown in the figure, the overall method is mainly divided into two steps: label acquisition of learner learning style and dynamic identification of learner learning style. The specific contents are as follows:

[0046] Step 1: Initially, the learner's learning style is unknown. The learning style label acquisition method (Learning Style Dividing Method) proposed in this invention is used to solve the problem of labeling the learner's learning style. The specific process of the method is as follows: Figure 3 shown. Figure 3In the method, first, multiple learners fill out two scales to obtain the corresponding scale data, and then use objective and subjective methods to divide the learning styles. The subjective method is described as a learning style division method based on given rules. When there are multiple highest learning style scores, both rules cannot effectively divide the learning styles. The learning style under the subjective method is obtained by taking the intersection method; the objective method is described as a learning style division method based on a clustering algorithm. The optimal clustering algorithm is selected according to the adjusted silhouette coefficient (Silhouette score), and then the clustered clusters are determined according to the EAM method to determine the specific meaning of each cluster to obtain the learning style under the objective method. Finally, the learning style division results of the subjective method and the objective method are intersected to obtain the final learning style label. That is, LSDM reduces the problem of excessive subjectivity caused by obtaining labels from a single scale by integrating the rule-based labeling results of the two scales and the labeling results of the clustering algorithm. Learners need to fill out two scales in this method: the Kolb Learning Style Scale and the Online Learning Behavior Survey Scale (such as Figure 4 As shown in the figure), the design idea of ​​the online learning behavior scale and the specific method of integrating learning style labels are then explained respectively.

[0047] 1) Kolb Learning Style Inventory Design Idea

[0048] Since learning style and learning ability are closely related, the present invention has designed a new online learning behavior scale as a supplement to the Kolb learning style scale. The online learning behavior scale contains a total of 27 questions. Behind each preset question corresponds to a different learning ability of the students, and these learning abilities can reflect the students' implicit learning behavior. The scale uses the Likert 5-level scale structure (strongly agree, agree, uncertain, disagree, strongly disagree) as the answer to each preset question. There are four basic abilities hidden behind the Kolb learning style: specific and experiential abilities (solution-solving ability, imagination, execution ability), action and application ability (attention, purposefulness, independence), reflection and observation ability (explanation, willpower, feedback), and abstract and summarizing ability (abstraction, logic, practical ability). According to the rationality principle of scale design, each learning ability is measured by at least 3 questions. The overall principle of "ensuring the quality of the scale while minimizing the number of questions as much as possible" is followed. For specific scales, see Figure 4 .

[0049] 2) Methods that integrate learning style labels

[0050] For a total of 12 questions in the Kolb Learning Style Inventory, learners are required to rank the four options in each question according to their matching degree, and then the score corresponding to each learning style of the learner is calculated using the given rules. k The calculation formula is as follows:

[0051]

[0052] where a i,j Represents the score of the jth option of the ith question, and k represents one of the four learning styles. The score of each learning style is the sum of the scores of the corresponding options of the 12 questions. Then, the learning style with the highest score is confirmed as the learner's learning style. However, when multiple learning styles obtain the same highest score, this method cannot divide the learning styles. For example, the scores of divergent and accommodative are both 33, which are greater than the other two learning styles. The method of calculating the learning style scores obtains the set of learning styles that are successfully divided and the set of learning styles that are not successfully divided. The online learning behavior scale is calculated in the same way, and then the intersection of the results obtained from the two is taken. If the same learning style is successfully divided by both, it is regarded as the learner's learning style at the moment. The remaining unsuccessful learning styles are supplemented by the clustering-based learning style label division method.

[0053] First, too high a number of features will have a negative impact on the clustering effect. Therefore, it is necessary to use linear transformation to reduce the feature dimension of the online learning behavior scale data before clustering, while retaining the original meaning of the data. Preferably, the original scale has 27 dimensions, and the clustering effect is best when it is reduced to 24 dimensions. Therefore, principal component analysis is selected as the feature dimensionality reduction algorithm, and K-Means++ is used to perform clustering experiments on the reduced data to obtain four clusters. Subsequently, a cluster meaning determination method based on expert annotation (Expert annotation method, EAM) is constructed, see Figure 5 The centroid is a value that is continuously adjusted by the K-Means++ algorithm during the iteration process. The clustering algorithm can determine the location of the centroid of each cluster after completing the clustering. Suppose that cluster C is divided into {c1,c2,...,c m}, m is the number of cluster divisions, then the goal of K-Means++ is to minimize the square error E:

[0054]

[0055] Among them, u n is cluster c n The centroid of cluster c n The sampling in , ||*||2 represents the Euclidean distance between two vectors. The formula for the centroid is as follows:

[0056]

[0057] After obtaining the centroid of the cluster, calculate the Euclidean distance from the remaining points to the centroid. Then, set the initial distance threshold to 5, that is, select the 5 sample points closest to the centroid in the cluster. Then, the experts need to determine their respective learning styles based on the Kolb Learning Style Scale data and the Online Learning Behavior Scale data of these 5 sample points. If the experts determine that more than half of the sample points have the same learning style, the intrinsic meaning of the cluster will be determined as this learning style. If the above conditions are not met, the centroid selection method will expand the threshold at a rate of 1 until the learning style can be successfully divided. This method is called EAM, see Figure 5 Finally, the learning styles divided in this part supplement the learning styles that cannot be divided in the rule-based method.

[0058] Step 2: Based on the learning style labels obtained in step 1, use the data collected in the online learning platform to construct reasonable features, and then use the Spearman correlation coefficient to test the correlation between the two. Then select appropriate training and test data to initialize the stacked machine learning model. The specific structure and training process of the model are as follows: Figure 6 shown. Figure 6 In the paper, the behavioral data generated by learners in the online teaching platform are first preprocessed, and then SMOTE is used to solve the sample imbalance problem in combination with the learning style annotation results of the LSDM model. The entire stacked machine learning model (hereinafter referred to as SMLM) consists of two layers, including 4 basic classifiers and 1 regressor. And the training method of the model is demonstrated in detail. Stacking is a model fusion strategy in ensemble learning, which can improve the integrity by fusing multiple single models. The SMLM model is a fusion model based on a two-layer model. The first layer consists of four basic classifiers: random forest (RF), gradient boosted decision tree (GBDT), support vector machine (SVM) and multilayer perceptron (MLP). The input is the original training set and test set. The logistic regressor (LR) of the second layer uses the output of the basic classifier of the first layer as input to join the training set for retraining, thereby obtaining a complete stacked model. The overall process is divided into data preprocessing, model training and parameter adjustment, and model evaluation. Specifically:

[0059] 1) Data preprocessing

[0060] This part includes log processing, data standardization and resampling. First, the log represents the interactive records left by students on the online teaching platform. The log records are processed according to the specified rules (behavior labels are marked) to obtain the online learning behavior characteristics of students. At this time, converting the online learning behavior log data into statistics based on duration and number of times can reduce the difficulty of training the stacking model. Then, due to the different structures of online learning behavior characteristics, each feature is standardized using the following formula:

[0061]

[0062] In the formula, y represents an online learning behavior vector, y * is the standardized value, y mean is a sample y in y a The mean value of y std is a sample y in y a The standard deviation of .

[0063] In addition, the distribution of learning styles obtained by LSDM is used to decide whether to use the resampling method. The SMOTE algorithm is used to synthesize new examples by analyzing the distribution rules of several types of samples, which can reduce the overfitting problem caused by random oversampling of repeated samples. Specifically, for each sample y′ in the minority class, the Euclidean distance is used to calculate the distance from this class to all samples in the minority class sample set to obtain its n nearest neighbors, and then the randomly selected adjacent samples are Construct a new sample together with the original sample, the formula is as follows:

[0064]

[0065] In the formula, y′ represents all online learning behavior record samples, y′ new is a newly generated sample, rand(0,1) is a random number between 0 and 1.

[0066] These synthetic data will be added to the training set of the original dataset to solve the problem of class imbalance. Finally, the dataset is divided into training set and test set according to 70% and 30% respectively.

[0067] 2) Model training and adjustment

[0068] The visualization process of model training is as follows Figure 6 Specifically, first, the preprocessed data is trained in a stacked model using a 5-fold cross-validation strategy, which divides the input training set into five training set subsets: i ,i∈[1,5]. Then, for each basic classifier in the first layer, the following operations are performed: in the 5-fold cross validation, each fold uses sequence i as the validation set, and the remaining four sequences j,j∈[1,5]∧j≠i as the training set. The classifier is trained using this data. At the same time, the trained model is used to predict the test set. At this point, each basic classifier in the first layer will obtain five different prediction results. Then, P is obtained by vertically stacking these five prediction results. n1 , and obtain P by averaging the predicted data of the validation set n1 Finally, P with corresponding labels n1is used as a training set, and P with corresponding labels n2 Used as a test set, the final classification is performed through the meta-classifier Logistic, and the final prediction results are obtained.

[0069] The parameters of the model are adjusted one by one. For the parameters of individual classifiers except the multilayer perceptron in the first layer, a combination of "random search coarse tuning" and "grid search fine tuning" is used to find the optimal parameters under the current data set. The multilayer perceptron optimizes the network parameters through the back propagation algorithm to improve the prediction effect of the model.

[0070] 3) Model Evaluation

[0071] The performance of the trained stacked classification model was evaluated using accuracy, recall, precision, F1 score and area under the curve (AUC).

[0072] To prove the effectiveness of the method of the present invention, this embodiment conducted an actual experiment, which selected the learning data of college students from a university on an online teaching platform, including students from 4 grades and 18 colleges. The data takes into account as many students with different learning backgrounds as possible. Since the learning style of students changes dynamically, the "scale data and learning platform data" collected in this experiment are continuous teaching semesters to ensure the timeliness of the research. The period of student online learning data is from March 1, 2020 to July 1, 2020. During the process of collating the scale data, it was found that some students repeatedly filled in the scale, some students did not fill in the scale completely, and some students could not match on the teaching platform. According to psychological research, people's first impression when filling in the scale is effective, so this experiment only selects the information of students filling in the scale for the first time, and the information of students will be deleted in the latter two cases. The number of valid students collected in the end was 2056, and the experiment designed a total of 4 learning styles: divergent, concentrated, compliant, and assimilated.

[0073] The following is a learning style identification method based on fusion label and stacked machine learning model of the present invention, and the specific steps are as follows:

[0074] 1. Learning style labeling: Use two rules to calculate the learning styles corresponding to the two scales, and take the intersection to get the divided and undivided learning styles. Then cluster the online learning behavior scale to determine the learning style, and supplement the undivided learning style.

[0075] 1.1. Two scales were distributed to students of a certain university for answering. After collecting the data, the following formula was used to calculate the scores of each learning style for the Kolb Learning Style Scale:

[0076]

[0077] where a i,j represents the score of the jth option of the ith question, and k represents one of the four learning styles. The score of each learning style is the sum of the scores of the corresponding options of the 12 questions. This method is used to obtain the set of learning styles that are successfully divided and the set of learning styles that are not successfully divided. For the online learning behavior scale, the following formula is used to calculate the scores of each learning style:

[0078]

[0079] where w i represents the weight of question i set by the researcher, and The highest learning style is determined as the student's learning style. Similarly, the set of learning styles that are successfully divided and the set of learning styles that are not successfully divided are obtained by this method. The intersection of the above two sets is obtained to obtain LS know and LS unknow .

[0080] 1.2. After comparing four dimensionality reduction methods, including low variance filtering, Pearson correlation coefficient, factor analysis, and principal component analysis, in the online learning behavior scale data, it was found that principal component analysis had the best dimensionality reduction effect on this data set. Subsequently, by comparing the clustering effects of K-Means++, Birch, Agglomerative, and K-Means, the effect evaluation indicators used were Silhouette coefficient, Calinski-Harabasz coefficient, Davies-Bouldin index, and cluster balance index. The cluster balance index is an evaluation indicator used to measure the volatility of the number of clusters, because an imbalance in the number of clusters will lead to an imbalance in the learning style categories. The specific formula is as follows:

[0081] CBI=std<count(c1)|...|count(c4)>

[0082] Where CBI stands for cluster balance index, std<*|...|*> stands for the standard deviation of all values ​​in the set, and count(c k ) represents the pairs belonging to the same cluster c k The sample sum of .

[0083] After comprehensive evaluation, it was found that K-Means++ had the best clustering effect, which could better meet the requirements of making the distance between clusters as large as possible and the distance within clusters as small as possible. Then, EAM was used to determine the meaning of the four clusters obtained by KMeans++, in which five sample points around the centroid were selected as thresholds. The learning style depicted by the clustering algorithm was successfully obtained through expert annotation. By taking the form of intersection again, the learning style distribution of 2056 students was finally divided into: 32.93% for accommodative type, 19.94% for decentralized type, 10.51% for assimilative type, and 36.62% for centralized type.

[0084] 2. Dynamic prediction of learning style: First, use the Spearman correlation coefficient to test the correlation between the learning style label obtained in step 1 and the data collected in the online learning platform to construct reasonable features. Then, select appropriate training and test data, train the stacked machine learning model, and build a model with reasonable parameters.

[0085] 2.1. First, process the log records (label the behavior) according to the specified rules to obtain the online learning behavior characteristics of the students. At this time, converting the online learning behavior log data into statistical data based on duration and frequency can reduce the difficulty of training the stacked model. Use the Spearman correlation coefficient to compare the learning style labels divided by LSDM, Kolb learning style scale, and online learning behavior scale with the learning behavior characteristics on the online learning platform. That is, the Spearman correlation analysis is performed on the learning behavior feature x and the label Y obtained by LSDM, Kolb scale Y2, and online learning behavior scale Y3, that is, x and y * between.

[0086] Taking the learning behavior feature x and the label Y obtained by LSDM as an example, the calculation formula of the Spearman correlation coefficient ρ is as follows:

[0087]

[0088] Where, d i It represents the position difference of the two variables obtained by calculating the Pearson correlation coefficient of the features and labels after sorting them respectively, and n represents the number of samples.

[0089] The subsequent correlation results show that the LSDM-based learning style classification method can obtain 4 significantly correlated indicators (this number is higher than the Kolb Learning Style Scale and the Online Learning Behavior Scale), indicating that the LSDM method is more reasonable in obtaining learning styles. Although the correlation coefficient is valuable in identifying important indicators, the relationship between the data is very complex. The correlation coefficient that describes how two variables change together cannot directly determine the student's learning style.

[0090] 2.2. Then we enter the data preprocessing part. Due to the different structures of online learning behavior characteristics, we use the following formula to standardize them:

[0091]

[0092] In addition, according to the imbalanced distribution of learning styles obtained by LSDM, a resampling method is needed. After comparing the effects of five resampling techniques including random undersampling, cluster centroid sampling, random oversampling, BorderLine SMOTE and SMOTE, SMOTE is selected as a resampling method more suitable for the samples of the present invention. Specifically, for each sample y′ in the minority class, the Euclidean distance is used to calculate the distance from the class to all samples in the minority class sample set to obtain its k nearest neighbors, and then the randomly selected adjacent samples are Construct a new sample together with the original sample, the formula is as follows:

[0093]

[0094] These synthetic data will be added to the training set of the original dataset to solve the problem of class imbalance. Finally, the dataset is divided into training set and test set according to 70% and 30% respectively.

[0095] 2.3. Next, a two-layer stacked machine learning fusion model is constructed. The first layer consists of four basic classifiers: random forest, gradient boosted decision tree, support vector machine and multilayer perceptron, and the second layer is a logistic regressor. The input is the original training set and test set. The output of the first layer basic classifier is used as input to join the training set for retraining. Specifically, first, the preprocessed data is trained in the stacked model using a 5-fold cross-validation strategy. The model divides the input training set into five training set subsets: train i ,i∈[1,5]. Then, for each basic classifier in the first layer, the following operations are performed: in the 5-fold cross validation, each fold uses sequence i as the validation set, and the remaining four sequences j,j∈[1,5]∧j≠i as the training set. The classifier is trained using this data. At the same time, the trained model is used to predict the test set. At this point, each basic classifier in the first layer will obtain five different prediction results. Then, P is obtained by vertically stacking these five prediction results. n1 , and obtain P by averaging the predicted data of the validation set n1 Finally, P with corresponding labels n1 is used as a training set, and P with corresponding labels n2 Used as a test set, the final classification is performed through the meta-classifier Logistic, and the final prediction results are obtained.

[0096] 2.4. Subsequently, in order to further evaluate the performance of the proposed model, it is compared with the baseline machine learning methods: K-nearest neighbor classification algorithm (KNN), Gaussian-based naive Bayes (Gaussian NB), Bernoulli-based naive Bayes (Bernoulli NB), entropy-based decision tree, Gini-based decision tree, support vector machine (SVM), random forest (RF), Adaboost, stochastic gradient descent classifier (SGD), bagging method with random forest (Bagging), extreme random tree (ET), gradient boosted decision tree (GBDT), voting method and multilayer perceptron (MLP). At the same time, these methods are trained with SMOTE processed data. The basic classifier used in the voting method is the same as the classifier of the first layer of the stacked machine learning model, and the parameter settings are also consistent. The remaining methods use the corresponding default parameters in sklearn. Compared with the baseline machine learning methods, the proposed stacked machine learning model has better prediction performance.

[0097] 2.5. Then, the parameters of the model are adjusted one by one. For the parameters of individual classifiers except the multilayer perceptron in the first layer, a combination of "random search coarse tuning" and "grid search fine tuning" is used to find the optimal parameters under the current data set. The multilayer perceptron optimizes the network parameters through the back propagation algorithm to improve the prediction effect of the model.

[0098] 2.6. Finally, the trained stacked classification model is comprehensively evaluated by using accuracy, recall, precision, F1 score and area under the curve (AUC). After obtaining the optimal model, the online learning behavior data is input into the model for recognition. The online learning behavior data can be the number of video views, the length of video views, the number of discussions participated in, the number of hand raises, etc. Compared with the traditional machine learning method, under the same circumstances, the model of the present invention can produce more accurate results.

[0099] In summary, the learning style prediction model constructed by the present invention is based on 5 machine learning classification models and classic clustering algorithms, all of which are time-tested machine learning algorithms in different scenarios. The characteristics and advantages of these machine learning algorithms are used to construct an overall framework for learner learning style identification. The Kolb learning style scale used to obtain learning style annotations also exists in previous studies and has been proven to have high reliability and validity. Therefore, the present invention makes full use of existing research results, and aims to address the problems that previous studies have not considered the high subjectivity of a single scale, static recognition cannot meet the dynamic changes in learning styles, and the lack of recognition accuracy. A learner learning style identification method based on fusion labels and stacked machine learning models is proposed. This method utilizes the characteristics of the stacked classification model to reconstruct data and extract important features, remove redundant information in the data, and train the data in an integrated manner. At the same time, in order to improve the problem of decreased accuracy caused by sample imbalance, it is proposed to use SMOTE technology to improve this problem. Finally, the model achieved accurate recognition results in four learning styles, that is, the recognition accuracy of this embodiment in divergent, concentrated, accommodative, and assimilated types can reach 97.6%, 96.3%, 96.1%, and 95.4%, respectively. Therefore, it has practical application prospects.

[0100] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person familiar with the technical field can easily think of various equivalent modifications or substitutions within the technical scope disclosed by the present invention, and these modifications or substitutions should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A learning style identification method based on fusion label and stacked machine learning model, characterized in that: The following steps are involved: 1) Use the Kolb Learning Style Inventory and the Online Learning Behavior Survey Scale to calculate the corresponding learning styles of learners respectively, and take the intersection of the two calculation results to obtain the set of divided learning styles and the set of undivided learning styles; 2) Clustering the online learning behavior survey scale without learning style classification, determine the learning style and label it, and supplement the unclassified learning style; the specific process is as follows: After the online learning behavior survey scale data is subjected to feature dimensionality reduction, the reduced data is clustered using K-Means++ to obtain four clusters, and then a cluster meaning determination method based on expert annotation is constructed to determine the position of the centroid of each cluster, and then the Euclidean distance from the remaining points to the centroid is calculated, and the centroid selection method is used to select the sample point within the threshold closest to the centroid in the cluster, and the learning style of the sample is determined and labeled according to the Kolb learning style scale data and the online learning behavior survey scale data of the sample point, and the unclassified learning style is supplemented; 3) Conducting a correlation test between the learning style label obtained in step 2) and the online learning behavior characteristics collected by the online learning platform; 4) For the online learning behavior features that have passed the correlation test, select training data and test data, and use the stacked machine learning model for training to obtain a complete stacked model; 5) Conduct a comprehensive performance evaluation on the trained stacking model, optimize the stacking model, and use the optimized stacking model to predict the learning style of actual learners.

2. The learning style identification method based on fusion label and stacked machine learning model according to claim 1 is characterized in that: The specific steps of step 1) include: 11) Obtain the Kolb learning style scale of the learner, calculate the scores of each learning style, identify the learning style with the highest score as the learner's learning style, and obtain the set of learning styles that are successfully divided and the set of learning styles that are not successfully divided; 12) Obtain the learner's online learning behavior survey scale, calculate the scores of each learning style, confirm the learning style with the highest score as the learner's learning style, obtain the set of successfully divided learning styles and the set of unsuccessfully divided learning styles, and take the intersection with the results obtained based on the learner's Kolb learning style scale.

3. The learning style identification method based on fusion label and stacked machine learning model according to claim 2 is characterized in that: The calculation formula for the learning style score of the Kolb learning style scale is: Where: LS k is the learning style score, a i,j is the score of the jth option of the ith question, and k represents one of all the learning styles of the learner.

4. The learning style identification method based on fusion label and stacked machine learning model according to claim 1 is characterized in that: Based on the Kolb Learning Style Scale data and the Online Learning Behavior Survey Scale data of the sample points, the learning styles of the samples are determined and labeled, and the specific contents of the unclassified learning styles are supplemented as follows: If it is determined that more than 50% of the sample points have the same learning style, the intrinsic meaning of the cluster is determined to be the learning style. If the condition of more than 50% is not met, the threshold will be expanded at a rate of 1 until the learning style can be successfully divided; finally, the learning styles that cannot be divided will be supplemented.

5. The learning style identification method based on fusion label and stacked machine learning model according to claim 1 is characterized in that: In step 3), the learning style label obtained in step 2) and the online learning behavior characteristics collected by the online learning platform are tested for correlation using the Spearman correlation coefficient.

6. The learning style identification method based on fusion label and stacked machine learning model according to claim 1, characterized in that: In step 4), the specific steps of training using the stacked machine learning model include: 41) Preprocessing the behavioral data generated by learners in the online teaching platform; 42) Combining the learning style annotation results obtained in step 2), a stacked machine learning model is constructed; 43) Perform model training and parameter adjustment on the constructed stacked machine learning model to obtain a complete stacked model.

7. The learning style identification method based on fusion label and stacked machine learning model according to claim 6, characterized in that: The stacked machine learning model is a fusion model based on a two-layer model, the first layer includes four basic classifiers: random forest, gradient boosting decision tree, support vector machine and multi-layer perceptron, and the input of the first layer is the original training set and test set; The second layer is a logistic regressor. The second layer uses the output of the first layer basic classifier as input and is retrained by adding the training set to obtain a complete stacking model.

8. The learning style identification method based on fusion label and stacked machine learning model according to claim 6, characterized in that: Before step 43), a sample resampling step using the SMOTE algorithm is also included.

9. The learning style identification method based on fusion label and stacked machine learning model according to claim 1, characterized in that: In step 5), the accuracy, recall, precision, F1 score and area under the curve are used to conduct a comprehensive performance evaluation of the trained stacking model to optimize the stacking model.

Citation Information

Patent Citations

  • Question and answer method and question and answer device for adaptive learning of learners

    CN105512349A

  • Adaptive learning method based on learning style context perception

    CN107423851A

  • An adaptive recommendation method and system based on learning style

    CN109213863A

  • Music style merging method based on coupled generative adversarial networks

    CN110085203A

  • Knowledge point learning path recommendation method and device

    CN111125640A