Student learning behavior analysis method and system based on big data

Abstract

The invention provides a student learning behavior analysis method and system based on big data, and relates to the technical field of data analysis, and the method comprises the steps: obtaining the learning behavior data of a student, and carrying out the data preprocessing and feature extraction, and obtaining a learning behavior data set after dimension reduction; behavior relations among the students are extracted and analyzed through the potential structures, and behavior pattern classification results of the students are obtained based on clustering of the relations; performing smoothing processing on the behavior data by applying a seasonal mixed extreme student residual algorithm and local weighted regression, and analyzing the smoothed data by using an automatic regression condition heterovariance model to obtain a student academic record volatility index; based on a bidirectional long-short-term memory network and a random forest algorithm, a relationship between student learning behaviors and scores is analyzed, a learning behavior analysis report is generated, and key influence factors of behavior characteristics and score fluctuation are provided. According to the method, the core behavior characteristics of the students can be extracted more accurately, and the behavior relationship between the students is analyzed.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and more particularly, to a method and system for analyzing students' learning behaviors based on big data. Background Art

[0002] With the advancement of educational informatization and the popularization of big data technology, the collection and analysis of students' learning behavior data have gradually become an important means of educational research. Currently, schools and educational institutions have accumulated a large amount of students' learning behavior data through various platforms and systems, including students' attendance rates, homework completion situations, classroom participation degrees, etc. However, existing analysis methods are usually limited to simple statistical and descriptive analysis, and cannot deeply explore the relationship between students' learning behaviors and learning effects. Traditional methods mainly rely on teachers' subjective judgments and experiences, have limited data processing capabilities, and are insufficient in dealing with large-scale and diverse data. When dealing with learning behavior data, existing technologies usually adopt a single clustering or regression model, but these methods cannot effectively identify complex learning behavior patterns, and have large errors in predicting the volatility of students' grades and identifying potential problem students. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and system for analyzing students' learning behaviors based on big data to improve the above problems. To achieve the above purpose, the technical solutions adopted by the present invention are as follows:

[0004] In a first aspect, the present application provides a method for analyzing students' learning behaviors based on big data, including:

[0005] Obtaining students' learning behavior data, and performing data preprocessing and feature extraction on the students' learning behavior data to obtain a reduced-dimensional learning behavior data set;

[0006] Performing latent structure extraction on the reduced-dimensional learning behavior data set, analyzing the behavioral relationships between students through the extracted latent features, and classifying behavioral patterns based on the analysis of the behavioral relationships between students to obtain the classification results of students' behavioral patterns;

[0007] Performing seasonal hybrid extreme studentized residual algorithm and locally weighted regression processing on the classification results of students' behavioral patterns to obtain smoothed learning behavior trend data, and analyzing the smoothed learning behavior trend data using an autoregressive conditional heteroskedasticity model to obtain an index of the volatility of students' learning performance;

[0008] Analyze and process the smoothed learning behavior trend data based on the bidirectional long short-term memory network and the random forest algorithm to obtain the relationship data between students' learning behaviors and grades, and generate a learning behavior analysis report for the relationship data and the volatility index of students' learning grades using the deep self-attention network.

[0009] In a second aspect, the present application also provides a big data-based student learning behavior analysis system, including:

[0010] An acquisition unit, configured to acquire students' learning behavior data, perform data preprocessing and feature extraction on the students' learning behavior data to obtain a reduced-dimensional learning behavior data set;

[0011] A classification unit, configured to extract the latent structure of the reduced-dimensional learning behavior data set, analyze the behavioral relationships between students through the extracted latent features, and classify the behavioral patterns based on the analysis of the behavioral relationships between students to obtain the behavioral pattern classification results of the students;

[0012] An analysis unit, configured to perform seasonal hybrid extreme studentized residual algorithm and locally weighted regression processing on the behavioral pattern classification results of the students to obtain the smoothed learning behavior trend data, and analyze the smoothed learning behavior trend data using the autoregressive conditional heteroskedasticity model to obtain the volatility index of students' learning grades;

[0013] A processing unit, configured to analyze and process the smoothed learning behavior trend data based on the bidirectional long short-term memory network and the random forest algorithm to obtain the relationship data between students' learning behaviors and grades, and generate a learning behavior analysis report for the relationship data and the volatility index of students' learning grades using the deep self-attention network.

[0014] The beneficial effects of the present invention are:

[0015] The present invention deeply explores the complex relationship between students' learning behaviors and learning grades through a variety of advanced algorithms, including fuzzy C-means clustering, variational autoencoders, graph convolutional networks, bidirectional long short-term memory networks, random forest algorithms, etc. Compared with traditional methods, the present invention can more accurately extract the core behavioral characteristics of students, analyze the behavioral relationships between students, identify behavioral patterns, predict the volatility of learning grades, and finally generate a comprehensive learning behavior analysis report. The present invention has significant improvements in data processing capabilities, analysis depth, and prediction accuracy, and can provide scientific decision-making basis for educators and promote the development of personalized teaching.

[0016] Other features and advantages of the present invention will be described in the subsequent specification, and, in part, will be apparent from the specification, or will be understood by implementing the embodiments of the present invention. The objectives and other advantages of the present invention can be achieved and obtained by the structures specifically pointed out in the written specification, claims, and drawings. Brief Description of the Drawings

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other relevant drawings can also be obtained based on these drawings.

[0018] Figure 1 It is a schematic flowchart of the method for analyzing students' learning behaviors based on big data described in the embodiments of the present invention;

[0019] Figure 2 It is a schematic structural diagram of the system for analyzing students' learning behaviors based on big data described in the embodiments of the present invention.

[0020] In the figure: 701, acquisition unit; 702, classification unit; 703, analysis unit; 704, processing unit. Detailed Embodiments

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Usually, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed present invention, but only represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0022] It should be noted that: Similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of the present invention, terms such as "first", "second", etc. are only used for distinguishing descriptions and cannot be understood as indicating or implying relative importance.

[0023] Embodiment 1:

[0024] This embodiment provides a method for analyzing students' learning behaviors based on big data.

[0025] See Figure 1 , the figure shows that this method includes step S1, step S2, step S3 and step S4.

[0026] Step S1: Obtain the learning behavior data of students, and perform data preprocessing and feature extraction on the learning behavior data of the students to obtain a reduced-dimensional learning behavior data set;

[0027] It can be understood that in this step, missing values are filled by using locally weighted regression, outliers are detected by the isolation forest, and feature extraction is combined with PCA and sparse coding. This step effectively improves the quality and processing efficiency of the data. This method of combining traditional and innovative algorithms can retain the essential characteristics of students' learning behaviors to the greatest extent and reduce the interference of noise data on the analysis results. In this step, step S1 includes step S11, step S12, step S13, step S14 and step S15.

[0028] Step S11: Perform operations of deleting missing values, deleting outliers and standardizing the behavior data of the students, calculate the similarity between pairwise data based on the standardized data, and retain the data with similarity greater than a preset threshold to obtain preprocessed behavior data;

[0029] It can be understood that this step uses the Mahalanobis distance to delete outliers, which is superior to the traditional Euclidean distance method. The Mahalanobis distance is a multi-dimensional measurement method that can consider the covariance relationship of the data and is suitable for processing correlated data. In the student behavior data, some data points (such as the extreme behaviors of a certain student in all learning activities) may affect the stability of the overall data. Using the Mahalanobis distance can identify these outliers that are significantly different from most data points and eliminate them. In this way, we can effectively identify and remove extreme or unreasonable learning behavior data (such as extremely high learning duration or too low participation), thereby improving the overall quality of the data set. This step also uses the Z-score standardization method to transform the data into a standard normal distribution (mean of 0, standard deviation of 1) to eliminate the influence of different feature dimensions and scales. This step also eliminates the differences in dimensions and scales between different students by calculating the cosine similarity between pairwise data, and focuses on measuring the similarity of students' behavior patterns.

[0030] Step S12: Perform fuzzy C-means clustering on the preprocessed behavior data to obtain a clustering result;

[0031] It can be understood that the greatest advantage of this step of fuzzy C-means clustering is that it can provide a fuzzy membership degree for each student, rather than simply classifying students rigidly into a certain cluster. This is especially important for educational data analysis because the behavior patterns of students are often variable and it is impossible to accurately represent students' learning behaviors using traditional hard clustering methods. Through fuzzy C-means clustering, we can more precisely capture the transitions between multiple behavior patterns of students and then discover potential behavior groups. For example, some students may be active in class but show low participation in homework submission. Fuzzy clustering can better identify this cross-behavior pattern, while hard clustering methods may not be able to effectively express such complex behavior characteristics.

[0032] In addition, the fuzzy C-means clustering algorithm can flexibly control the granularity and accuracy of clustering by adjusting the fuzziness coefficient and the number of clusters. According to the requirements of different educational scenarios, the parameters can be adjusted to obtain appropriate clustering results. For example, for large-scale student behavior data, a higher fuzziness coefficient can make the behavior performance of students more "diverse", while a lower fuzziness coefficient value can more strongly limit the possibility of students belonging to a single cluster.

[0033] Step S13: Generate fuzzy rules based on the clustering results, use fuzzy logic regression to analyze the student behavior data, identify the learning attitudes and participation degrees of students, and obtain a student behavior data set, where the student behavior data set includes the behavior characteristics and learning characteristics of students;

[0034] It can be understood that generating fuzzy rules is the first step of this step. The key lies in determining the relationships between student behaviors through the clustering results and extracting fuzzy rules from them. Fuzzy rules are generally represented in the IF-THEN form. For example, "If a student has a high class participation degree and good homework completion, then the student has a relatively positive learning attitude". The specific steps for generating fuzzy rules include: According to the results of fuzzy C-means clustering, each cluster represents a type of student behavior pattern. Each cluster contains a group of students with similar learning behavior characteristics and learning performance. For the students in each cluster, analyze the relationships between their behavior characteristics (such as class participation degree, homework submission rate, learning duration, etc.) and their learning performance. Through expert knowledge, several fuzzy rules can be extracted for each cluster.

[0035] After obtaining the fuzzy rules in this step, the student behavior data is fuzzified, and the original numerical data (such as learning duration, homework completion rate, etc.) is converted into fuzzy sets. For example, the learning duration can be divided into three fuzzy sets: "short", "medium", and "long", and the homework completion rate can be divided into three levels: "low", "medium", and "high". Then, membership functions are defined for each fuzzy set to describe the degree to which each student behavior data point belongs to that fuzzy set. Fuzzy logic regression establishes a relationship between the fuzzified input features and the output targets (such as learning attitude and participation) to obtain a regression equation, and the regression equation is as follows:

[0036]

[0037] Among them, Y is the output result (learning attitude or participation), X is the input variable (learning behavior characteristics), and μ i (X) is the fuzzy membership function, and α i is the regression coefficient, and ε is the error term.

[0038] Step S14: Perform principal component analysis on the student behavior dataset, where redundant features are removed through dimensionality reduction and the core features affecting student behavior are extracted to obtain the reduced-dimensional learning behavior dataset;

[0039] It can be understood that in this step, redundant features in the student behavior data are removed through principal component analysis, thereby reducing the data dimension. Student behavior data often contains a large number of highly correlated features (such as learning duration, homework completion rate, classroom participation, etc.), and there may be a strong linear relationship between these features. The reduced-dimensional dataset reduces noise and unnecessary complexity, making subsequent analysis and modeling more efficient. When performing tasks such as student behavior analysis and learning performance prediction, the reduced-dimensional dataset can significantly improve the accuracy and generalization ability of the model. Because after dimensionality reduction, the most important features of the data are retained, while irrelevant noise is removed, avoiding the risk of overfitting.

[0040] Step S15: Based on a preset deep feature fusion algorithm, fuse the behavior features and learning features of students in the reduced-dimensional learning behavior dataset, and use the fused features as the reduced-dimensional learning behavior dataset.

[0041] It can be understood that this step performs fusion through a deep neural network model. Among them, the input layer of the model receives student behavior features and learning features. Usually, the behavior features and learning features are input through different input layers, or through the merged feature vectors. Through the deep neural network, these two types of features can be jointly modeled. The hidden layer performs multi-layer non-linear transformations through multiple neurons. At each layer, the network performs complex interaction processing based on the input behavior features and learning features, and gradually extracts deeper features that affect students' learning behaviors. The role of these layers is to extract the correlations between features, and through the weight adjustment learned by the network, optimize the feature fusion process. In the hidden layer, the ReLU or activation function is used to help the neural network capture non-linear relationships and prevent the model from being too simple to handle complex student behavior data. The output layer generates a fused feature set. These feature sets include the comprehensive information of student behavior features and learning features, and can provide effective input for subsequent learning behavior prediction or learning effect analysis. Through deep feature fusion, a new feature set is obtained, which contains the fusion results of behavior features and learning features. While retaining the original information, this feature set eliminates the redundancy and possible local biases between behavior and learning features, and provides a more comprehensive and efficient input for subsequent analysis.

[0042] Step S2: Extract the latent structure from the dimension-reduced learning behavior data set, analyze the behavior relationships between students through the extracted latent features, and classify the behavior patterns based on the analysis of the behavior relationships between students to obtain the behavior pattern classification results of the students;

[0043] It can be understood that this step can effectively extract and analyze the deep information in the student behavior data, reveal the similarities between students, and provide an accurate basis for classifying the student behavior patterns through latent feature extraction, graph convolutional network analysis, and K-means clustering. This process provides a solid data foundation for personalized education and accurate learning performance prediction. In this step, step S2 includes step S21, step S22, and step S23.

[0044] Step S21: Extract the latent features in the student learning behavior data based on a preset variational autoencoder to obtain a low-dimensional latent feature representation;

[0045] It can be understood that in this step, the variational autoencoder is used to extract the latent features in the student learning behavior data. Its basic structure consists of two parts: an encoder and a decoder.

[0046] Among them, the learning behavior data of students is first processed by an encoder. The encoder maps the high-dimensional behavior data to the latent space, where the high-dimensional behavior data is a low-dimensional latent feature representation. The encoder maps the input data through a neural network and outputs the mean and variance of the latent variables, which are used to represent the distribution of the latent variables. To ensure that the distribution of the latent variables conforms to the Gaussian distribution, the variational autoencoder adopts the reparameterization trick, that is, the mean and variance generate the latent variables through a standard normal distribution, so that the model can perform backpropagation training.

[0047] The latent features are decoded back to the data space through a decoder to obtain the reconstructed data closest to the input data. The purpose of the decoder is to try to reconstruct an output similar to the original data, which can help the model learn the distribution of the input data. In this step, the variational autoencoder optimizes the model parameters by simultaneously minimizing the reconstruction error and the KL divergence, so that the encoder can effectively extract the latent features of the student behavior data, and these latent features have good generative ability.

[0048] Step S22: Analyze the low-dimensional latent feature representation based on the graph convolutional network, where the behavior relationship between students is analyzed by constructing a student behavior relationship graph and an adjacency matrix to obtain a set of behavior relationships between students;

[0049] It can be understood that in this step, a student behavior relationship graph is constructed through node representation, edge representation, and adjacency matrix. Among them, node representation: The behavior pattern (low-dimensional latent feature) of each student will become a node in the graph. In this way, the behavior of each student is mapped to a node of the graph, and each node of the graph represents the behavior characteristics of a student. Edge representation: The behavior relationship between students is represented by the edges of the graph. The weight of the edge represents the similarity between students, usually calculated based on their distance or similarity in the behavior feature space. For example, if two students are very similar in terms of behavior dimensions such as homework submission and study duration, then these two students are connected in the graph and a larger edge weight is assigned; if their behaviors are quite different, the edge weight is smaller. Adjacency matrix: After constructing the graph, an adjacency matrix is used to represent the structure of the graph. The adjacency matrix represents the similarity or correlation degree between students and students. The non-zero elements in the matrix represent that there is a certain relationship between students, and the weight represents the strength of the relationship. This adjacency matrix is the input of the graph convolutional network, which is used to propagate information and extract the complex relationships between student behaviors.

[0050] The graph convolutional network in this step is a deep learning algorithm capable of processing graph-structured data. It can update the features of nodes through the neighboring nodes of the nodes, thereby capturing the dependencies and global information between the nodes. Among them, the core operation of graph convolution is to calculate the new feature representation of each node (student) through the adjacency matrix and the node feature matrix (the low-dimensional latent features of each student). Specifically, graph convolution can be calculated by the following formula:

[0051]

[0052] Among them, H (l) is the node feature matrix of the l-th layer, is the normalized adjacency matrix, W (l) is the weight matrix of the current layer, σ is the ReLU activation function, and H (l+1) is the node feature matrix of the (l + 1)-th layer.

[0053] After being processed by the graph convolutional network, the final feature representation of each student will not only contain its own behavior information but also the information transmitted from its neighbors (i.e., other students with similar behaviors). These final feature representations can reflect the behavioral similarities and group behavior patterns among students. Through graph convolution operations, the obtained set of student behavior features can not only reflect the direct behavioral similarities among students but also reveal some indirect behavioral connections.

[0054] Step S23: Cluster the set of behavioral relationships among students through the extended K-means clustering algorithm. Among them, the students are clustered by calculating the distance between the student behavior features and the cluster centers obtained by clustering using the Euclidean distance, and the classification result of the student behavior patterns is obtained.

[0055] It can be understood that through clustering in this step, similar behavior patterns among students can be discovered. For example, a group of students may show similar behavioral characteristics in terms of study duration, homework completion, etc. and can be grouped into one category; while another group of students may exhibit different behavior patterns. In this way, educators can more accurately identify the behavioral characteristics of different students and then take targeted educational measures.

[0056] Step S3: Perform seasonal hybrid extreme studentized residual algorithm and locally weighted regression processing on the classification result of the student behavior patterns to obtain the smoothed learning behavior trend data, and use the autoregressive conditional heteroskedasticity model to analyze the smoothed learning behavior trend data to obtain the volatility index of the students' learning performance;

[0057] It can be understood that through the seasonal hybrid extreme studentized residual algorithm and local weighted regression, outliers and local fluctuations in the learning behavior data are removed, making the data smoother and providing a robust basis for subsequent volatility analysis. Conducting volatility analysis on the smoothed data can accurately quantify the volatility of students' grades, providing a deeper perspective for the evaluation of students' learning achievements, helping educators identify students with large fluctuations in learning achievements, and thus implementing more targeted educational interventions. In this step, step S3 includes step S31, step S32, and step S33.

[0058] Step S31: According to the classification results of students' behavior patterns, use the seasonal hybrid extreme studentized residual algorithm to detect anomalies in the learning behavior data. Specifically, by constructing a seasonal hybrid model to analyze the time series of each student's behavior pattern, the classification results of students' behavior patterns are decomposed into seasonal components, trend components, and residual components. Then, through the seasonal hybrid extreme studentized residual algorithm, the learning behavior data with abnormal behavior marks is identified.

[0059] It can be understood that in this step, the learning behavior data is decomposed into three parts by using the STL decomposition method: Seasonal component: Captures the part of the periodic fluctuations in the data, such as the fluctuations in students' learning behavior in different semesters. Trend component: Represents the long-term change trend of students' behavior, which may reflect slow-changing behavioral characteristics such as students' learning attitudes and participation. Residual component: The remaining part, which contains abnormal fluctuations or noise in the data and is caused by some extreme or irregular events. Then, by combining these components of the time series data, a complete seasonal hybrid model is constructed, which can flexibly capture the seasonal changes and long-term trends of learning behavior.

[0060] After obtaining the complete seasonal hybrid model in this step, for the residual data at each time point, the studentized residual value is calculated. The studentized residual refers to the difference between the data point and its expected value. After standardization, the residuals of each data point have the same scale. Then, according to the value of the studentized residual, a preset threshold (such as 3 times the standard deviation) is used to determine whether this point is an abnormal data point. If the studentized residual of a certain point exceeds the threshold, then this point is regarded as an outlier and marked as abnormal behavior. Finally, the learning behavior data with abnormal behavior marks is obtained. By identifying and removing abnormal data in this step, the behavior data in the dataset is made more real and conforms to the norm, reducing the interference of noise, and thus improving the prediction and analysis accuracy of the entire learning behavior analysis system.

[0061] Step S32: Based on the local weighted regression algorithm, smooth the learning behavior data with abnormal behavior marks to obtain the smoothed learning behavior trend data.

[0062] It can be understood that in this step, for each data point with an abnormal behavior label, the locally weighted regression algorithm is used to assign weights to the neighboring data points around this data point. The weights are usually determined according to the relative position to this point. The data points closer have higher weights, and the data points farther away have lower weights. The weight calculation method in this step is the Gaussian kernel function. Through this function, the similarity between each data point and the target data point is calculated, and then the weights are calculated.

[0063] Use the weighted neighboring data points for local regression fitting. In this step, the method of linear regression is used to fit the data. Through this local fitting, the local trend of the data can be accurately captured, and the influence of abnormal data on the regression model can be reduced. By applying local regression to all data points in this step, the smoothed learning behavior data is obtained. These data can better reflect the long-term learning trend of students and remove the interference brought by abnormal behaviors.

[0064] The data smoothed by local weighted regression can reflect the learning trends of students at different time periods, including the changes in learning time and the fluctuations in learning participation.

[0065] Step S33: Process and model the smoothed learning behavior trend data based on the autoregressive model, and fit the established conditional heteroscedasticity model to the smoothed learning behavior trend data to obtain the volatility index of students' learning performance.

[0066] It can be understood that in this step, the smoothed learning behavior trend data is processed through the autoregressive model to obtain the time series data of students' learning behaviors. Through the analysis of the autoregressive model, the time series law in the learning behavior data is captured, and the possible future behavior trends are predicted, thus laying a foundation for the subsequent volatility analysis. Among them, the formula of the autoregressive model is as follows:

[0067] Y t =φ1Y t-1 +φ2Y t-2 +Λ+φ p Y t-p +τ

[0068] Among them, Y t is the learning behavior data at time point t, φ p is the model parameter with the model order of p, p is the model order, τ is the error term, and Y t-p is the learning behavior data with the model order of t - p.

[0069] After this step on the time series data of students' learning behaviors, the predicted values and real data of the autoregressive model are obtained, which are used as inputs to construct a conditional heteroscedasticity model. The maximum likelihood estimation method is used to estimate the parameters of the conditional heteroscedasticity model, and a volatility model most suitable for the smoothed learning behavior data is obtained. Through this process, the volatility of the data can be fitted and the future volatility can be predicted. Then, the conditional variance estimated by the conditional heteroscedasticity model can be used as a direct indicator of the volatility of learning performance. This indicator characterizes the volatility of students' learning performance under the condition of given historical learning behavior data.

[0070] It can be understood that through the combination of the autoregressive model and the conditional heteroscedasticity model in this step, the time series volatility in the learning behavior data can be effectively captured. Especially in the process of students' learning performance changes, it can reflect unstable and large-fluctuation learning patterns.

[0071] Step S4: Analyze and process the smoothed learning behavior trend data based on the bidirectional long short-term memory network and the random forest algorithm to obtain the relationship data between students' learning behaviors and grades, and generate a learning behavior analysis report on the relationship data and the volatility index of students' learning performance using the deep self-attention network.

[0072] It can be understood that by combining the bidirectional LSTM network, the random forest, and the deep self-attention network, the relationship between students' learning behaviors and grades can be comprehensively analyzed, thereby providing in-depth insights into the reasons for the fluctuations in learning performance and providing a basis for the formulation of personalized education and intervention strategies. In this step, step S4 includes step S41, step S42, step S43, step S44, and step S45.

[0073] Step S41: Use the smoothed learning behavior trend data and the volatility index of learning performance as input data to construct a time series data set, and the time series data set contains the behavioral historical information and grade volatility information of students;

[0074] It can be understood that in this step, when constructing the time series dataset, we combine the smoothed learning behavior trend data with the learning performance volatility index and arrange them in chronological order. Each time step in the time series dataset will contain behavioral history information and performance volatility information. Among them, the behavioral history information includes the learning behavior data of students, such as learning duration, online participation, homework submission frequency, etc. Such data reflects the learning activities and habits of students within a specific period. The performance volatility information includes the performance volatility of students within the same period, usually quantified by volatility indicators. Performance volatility reflects the degree of fluctuation of students' performance and may be affected by various factors such as learning attitude and learning time. For each student, a fixed time window (such as a semester, a month, etc.) is selected, and then the learning behavior data and performance volatility data of the student are collected within this window. The learning behavior and performance volatility data of each student within the time window are merged in chronological order to form a complete time series dataset. This dataset will be used to train the subsequent prediction model.

[0075] Then, according to the historical learning behavior and performance volatility of each student, the data is organized into a time series format. For each student, the behavioral data (such as learning duration, participation, etc.) and performance volatility (such as fluctuation range, fluctuation frequency, etc.) at each moment will be marked as a sample with a timestamp. If the selected time window is large, the sliding window method can be used to generate time series data to ensure that the model can learn the dynamic relationship between behavior and performance volatility.

[0076] Step S42: Encode the past learning behavior data and future learning behavior data through a bidirectional LSTM network to capture the temporal relationship between the learning behavior and performance of students;

[0077] It can be understood that in this step, the historical behavior data and performance volatility data of each student are used as time series inputs into a bidirectional LSTM network. The forward LSTM reads the data step by step from the start to the end of the time series, sequentially capturing the temporal dependencies between historical behaviors and performance. The backward LSTM processes the data step by step from the end to the start of the time series, capturing the impact of future behaviors on the current performance. This layer can provide context information from the future to the past. The bidirectional LSTM finally obtains a comprehensive understanding of the relationship between learning behaviors and performance volatility by merging the outputs of the forward and backward directions. This output will become a deep representation of the impact of student behaviors on performance volatility. Among them, the step method for temporal dependencies includes: the model processes each time series through a time window to analyze the comprehensive impact of past and future behaviors on performance volatility. Then, through learning with the bidirectional LSTM, the model can identify how learning behaviors gradually affect performance volatility and can determine which behavioral characteristics (such as frequent learning activities, early high-intensity learning, etc.) are most relevant to performance volatility. For example, a student's learning behaviors (such as continuous learning, frequent participation in discussions, etc.) in the past period may affect their future performance volatility, and current learning behaviors may also have a direct impact on short-term performance. By learning simultaneously from both ends of the time series, the bidirectional LSTM can not only capture the impact of short-term behaviors but also link long-term behavioral patterns with performance volatility, thus effectively revealing the multi-level impact of student behaviors on performance volatility.

[0078] The bidirectional LSTM can learn the temporal relationship of data from both the past and the future directions, enabling the model to comprehensively understand the dynamic connection between students' learning behaviors and performance volatility. By considering future information, the model avoids the limitations of unidirectional time series analysis and enhances the understanding of the complex interactions between student behaviors and performance.

[0079] Step S43: Construct multiple decision trees based on the temporal relationship between students' learning behaviors and performance. Each decision tree is trained on different subsets of the dataset, and the results of all decision trees are voted to obtain the final relationship data between learning behaviors and performance;

[0080] It is understandable that in this step, first, the learning behavior data and the performance volatility index of students are used as input features to establish the training dataset for each decision tree. In a random forest, the training data for each tree is obtained by randomly sampling the original dataset with replacement. This method ensures that the training datasets for each tree are different, thereby increasing the diversity of the model. Then, feature selection is performed. When each tree splits, instead of considering all features each time, a random subset of features is selected for splitting. This helps to avoid overfitting and makes the learning result of each tree not affected by a single feature. Furthermore, multiple trees are trained. At each node, the decision tree selects a feature to split the data, such that the variance of the subset after splitting is minimized, or equivalently, the classification error rate is minimized. Through training and voting, the random forest model finally obtains the relationship data between learning behavior and performance. These data not only reflect how students' learning behaviors affect their performance volatility but also reveal the specific contributions of different learning behaviors to performance. This dataset will be used for subsequent deep learning analysis to further guide the recommendation of personalized learning paths for students.

[0081] Step S44: Based on a deep self-attention network, perform feature analysis on the final relationship data and the learning performance volatility index, and extract the behavioral features that have the greatest impact on learning performance.

[0082] It is understandable that in this step, for the learning behavior data and the performance volatility data of each student in the deep self-attention network, the self-attention mechanism calculates the influence degree of each feature on the target prediction (such as the performance volatility index). This process is achieved by calculating the "correlation" between each pair of features, usually using the dot product or weighted sum method. Specifically, the network calculates the correlation between each feature (such as "extracurricular learning time") and other features (such as "classroom participation"). Through the calculated attention weights, the self-attention mechanism performs a weighted sum on each input feature. This process enables the network to emphasize the features that have a greater impact on students' performance volatility during the learning process, thereby reducing the attention to irrelevant or secondary features. The result after the weighted sum represents the contribution of each feature in the analysis of students' behaviors and performance volatility. The network automatically selects the most predictive behavioral features based on these weighted features. For example, it may be found that the "extracurricular reading time" or "online interaction frequency" has a relatively close relationship with students' performance volatility, while other factors such as "classroom performance" may have less impact on performance volatility. In this step, the deep self-attention network structure is a multi-layer stacked network structure. Each layer further performs a weighted sum and adjustment on the features of the previous layer, thereby gradually extracting more detailed feature relationships.

[0083] Step S45: Conduct grey relational analysis based on the behavioral characteristics that have the greatest impact on academic performance and the academic performance volatility index, and use the obtained correlation degree value as the performance impact degree value of the behavioral characteristics that have the greatest impact on academic performance. Summarize the behavioral characteristics that have the greatest impact on each student's academic performance and their corresponding impact degree values to obtain a learning behavior analysis report.

[0084] It can be understood that grey relational analysis can help accurately identify which behavioral characteristics have the most significant impact on academic performance volatility, thereby providing practically guiding intervention suggestions. Through a detailed analysis of each student's behavioral data and academic performance volatility, personalized learning behavior reports can be generated for students to help teachers, parents, or the students themselves understand and improve learning strategies.

[0085] Embodiment 2:

[0086] As Figure 2 shown, this embodiment provides a student learning behavior analysis system based on big data. Refer to Figure 2 The system includes an acquisition unit 701, a classification unit 702, an analysis unit 703, and a processing unit 704.

[0087] The acquisition unit 701 is used to acquire the learning behavior data of students, and perform data preprocessing and feature extraction on the learning behavior data of the students to obtain a reduced-dimensional learning behavior data set.

[0088] The classification unit 702 is used to extract the latent structure of the reduced-dimensional learning behavior data set, analyze the behavioral relationships between students through the extracted latent features, and classify the behavioral patterns based on the analysis of the behavioral relationships between students to obtain the behavioral pattern classification results of the students.

[0089] The analysis unit 703 is used to perform seasonal hybrid extreme studentized residual algorithm and locally weighted regression processing on the behavioral pattern classification results of the students to obtain smoothed learning behavior trend data, and analyze the smoothed learning behavior trend data using an autoregressive conditional heteroskedasticity model to obtain the academic performance volatility index of the students.

[0090] The processing unit 704 is used to analyze and process the smoothed learning behavior trend data based on a bidirectional long short-term memory network and a random forest algorithm to obtain the relationship data between the students' learning behaviors and their academic performance, and generate a learning behavior analysis report using a deep self-attention network for the relationship data and the academic performance volatility index of the students.

[0091] It should be noted that regarding the system in the above embodiments, the specific manners in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0092] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

[0093] As described above, these are only the specific implementation manners of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present invention, and all should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for analyzing students' learning behaviors based on big data, characterized in that, Including: Obtain the learning behavior data of students, and perform data preprocessing and feature extraction on the learning behavior data of the students to obtain a reduced-dimensional learning behavior data set; Perform latent structure extraction on the reduced-dimensional learning behavior data set, analyze the behavioral relationships between students through the extracted latent features, and classify behavioral patterns based on the analysis of the behavioral relationships between students to obtain the behavioral pattern classification results of the students; Perform seasonal hybrid extreme studentized residual algorithm and locally weighted regression processing on the student behavioral pattern classification results to obtain smoothed learning behavior trend data, and use the autoregressive conditional heteroskedasticity model to analyze the smoothed learning behavior trend data to obtain the volatility index of students' learning performance; Analyze and process the smoothed learning behavior trend data based on the bidirectional long short-term memory network and the random forest algorithm to obtain the relationship data between students' learning behaviors and grades, and use the deep self-attention network to generate a learning behavior analysis report for the relationship data and the volatility index of students' learning performance.

2. The method for analyzing students' learning behaviors based on big data according to claim 1, wherein , and perform data preprocessing and feature extraction on the learning behavior data of the students to obtain a reduced-dimensional learning behavior data set, including: Delete missing values, delete outliers and standardize the behavioral data of the students, calculate the similarity between pairwise data based on the standardized data, and retain the data with similarity greater than a preset threshold to obtain preprocessed behavioral data; Perform fuzzy C-means clustering on the preprocessed behavioral data to obtain a clustering result; Generate fuzzy rules based on the clustering result, use fuzzy logic regression to analyze the students' behavioral data, identify the learning attitudes and participation degrees of the students to obtain a student behavioral data set, and the student behavioral data set includes the behavioral characteristics and learning characteristics of the students; Perform principal component analysis on the student behavioral data set, where redundant features are removed through dimensionality reduction and the core features affecting students' behaviors are extracted to obtain a reduced-dimensional learning behavior data set; Fuse the behavioral characteristics and learning characteristics of the students in the reduced-dimensional learning behavior data set based on a preset deep feature fusion algorithm, and use the fused features as the reduced-dimensional learning behavior data set.

3. The method for analyzing students' learning behaviors based on big data according to claim 1, wherein , perform latent structure extraction on the reduced-dimensional learning behavior data set, analyze the behavioral relationships between students through the extracted latent features, and classify behavioral patterns based on the analysis of the behavioral relationships between students to obtain the behavioral pattern classification results of the students, including: Extract latent features in the students' learning behavior data based on a preset variational autoencoder to obtain a low-dimensional latent feature representation; Analyze the low-dimensional latent feature representation based on the graph convolutional network, where the behavioral relationships between students are analyzed by constructing a student behavior relationship graph and an adjacency matrix to obtain a set of behavioral relationships between students; Perform clustering processing on the set of behavioral relationships between students through an extended K-means clustering algorithm, where the students are clustered by calculating the distance between the students' behavioral characteristics and the clustering centers obtained by clustering using the Euclidean distance to obtain the behavioral pattern classification results of the students.

4. The method for analyzing students' learning behaviors based on big data according to claim 3, wherein ,The classified results of the student behavior patterns are processed by the seasonal hybrid extreme studentized residual algorithm and local weighted regression to obtain the smoothed learning behavior trend data, and the autoregressive conditional heteroskedasticity model is used to analyze the smoothed learning behavior trend data to obtain the volatility index of the student's academic performance, including: According to the classified results of the student behavior patterns, the seasonal hybrid extreme studentized residual algorithm is used to detect anomalies in the learning behavior data. Among them, by constructing a seasonal hybrid model to analyze the time series of each student behavior pattern, the classified results of the student behavior patterns are decomposed into seasonal components, trend components and residual components, and then the learning behavior data with anomaly behavior marks is identified through the seasonal hybrid extreme studentized residual algorithm; Based on the local weighted regression algorithm, the learning behavior data with anomaly behavior marks is smoothed to obtain the smoothed learning behavior trend data; Based on the autoregressive model, the smoothed learning behavior trend data is processed and modeled, and the established conditional heteroskedasticity model is fitted to the smoothed learning behavior trend data to obtain the volatility index of the student's academic performance.

5. The method for analyzing students' learning behaviors based on big data according to claim 1, wherein ,Based on the bidirectional long short-term memory network and the random forest algorithm, the smoothed learning behavior trend data is analyzed and processed to obtain the relationship data between the student's learning behavior and performance, and the deep self-attention network is used to generate a learning behavior analysis report for the relationship data and the volatility index of the student's academic performance, including: Taking the smoothed learning behavior trend data and the volatility index of the academic performance as input data, a time series data set is constructed, and the time series data set contains the behavior history information and performance fluctuation information of the students; The past learning behavior data and future learning behavior data are encoded by the bidirectional LSTM network to capture the temporal relationship between the student's learning behavior and performance; Based on the temporal relationship between the student's learning behavior and performance, multiple decision trees are constructed. Each decision tree is trained on different subsets of the data set, and the results of all decision trees are voted to obtain the final relationship data between the learning behavior and performance; Based on the deep self-attention network, feature analysis is performed on the final relationship data and the volatility index of the academic performance, and the behavior features that have the greatest impact on the academic performance are extracted; Based on the behavior features that have the greatest impact on the academic performance and the volatility index of the academic performance, grey relational analysis is performed, and the correlation degree value obtained from the analysis is used as the performance impact degree value of the behavior feature that has the greatest impact on the academic performance. The behavior features that have the greatest impact on the academic performance of each student and their corresponding impact degree values are summarized to obtain the learning behavior analysis report.

6. A student learning behavior analysis system based on big data, characterized in that, Including: An acquisition unit for acquiring the learning behavior data of the student, and performing data preprocessing and feature extraction on the learning behavior data of the student to obtain a reduced-dimensional learning behavior data set; A classification unit for extracting the latent structure of the reduced-dimensional learning behavior data set, analyzing the behavior relationship between students through the extracted latent features, and classifying the behavior patterns based on the analysis of the behavior relationship between students to obtain the classified results of the student behavior patterns; An analysis unit for performing seasonal hybrid extreme studentized residual algorithm and locally weighted regression processing on the classification result of the student behavior pattern to obtain smoothed learning behavior trend data, and using an autoregressive conditional heteroskedasticity model to analyze the smoothed learning behavior trend data to obtain a volatility index of the student's academic performance; A processing unit for analyzing and processing the smoothed learning behavior trend data based on a bidirectional long short-term memory network and a random forest algorithm to obtain relationship data between the student's learning behavior and performance, and using a deep self-attention network to generate a learning behavior analysis report for the relationship data and the volatility index of the student's academic performance.

7. The student learning behavior analysis system based on big data according to claim 6, characterized in that The acquisition unit includes: A first acquisition subunit for deleting missing values, deleting outliers, and normalizing the behavior data of the student, calculating the similarity between pairwise data based on the normalized data, and retaining the data with a similarity greater than a preset threshold to obtain preprocessed behavior data; A second acquisition subunit for performing fuzzy C-means clustering on the preprocessed behavior data to obtain a clustering result; A third acquisition subunit for generating fuzzy rules based on the clustering result, using fuzzy logic to analyze the student behavior data, and identifying the learning attitude and participation of the student to obtain a student behavior data set, where the student behavior data set includes the behavior characteristics and learning characteristics of the student; A fourth acquisition subunit for performing principal component analysis on the student behavior data set, where redundant features are removed through dimensionality reduction and the core features affecting the student behavior are extracted to obtain a learning behavior data set after dimensionality reduction; A fifth acquisition subunit for fusing the behavior characteristics and learning characteristics of the student in the learning behavior data set after dimensionality reduction based on a preset deep feature fusion algorithm, and using the fused features as the learning behavior data set after dimensionality reduction.

8. The student learning behavior analysis system based on big data according to claim 6, characterized in that The classification unit includes: A first classification subunit for extracting latent features in the student learning behavior data based on a preset variational autoencoder to obtain a low-dimensional latent feature representation; A second classification subunit for analyzing the low-dimensional latent feature representation based on a graph convolutional network, where the behavior relationship between students is analyzed by constructing a student behavior relationship graph and an adjacency matrix to obtain a set of behavior relationships between students; A third classification subunit for clustering the set of behavior relationships between students through an extended K-means clustering algorithm, where the students are clustered by calculating the distance between the student behavior characteristics and the clustering centers obtained by clustering using the Euclidean distance to obtain the classification result of the student behavior pattern.

9. The student learning behavior analysis system based on big data according to claim 6, characterized in that, The analysis unit includes: A first analysis unit for performing anomaly detection on the learning behavior data using a seasonal hybrid extreme studentized residual algorithm according to the classification result of the student behavior pattern, where the time series of each student behavior pattern is analyzed by constructing a seasonal hybrid model, and the classification result of the student behavior pattern is split into seasonal components, trend components, and residual components, and then the learning behavior data with abnormal behavior marks is identified through the seasonal hybrid extreme studentized residual algorithm; A second analysis unit, configured to perform smoothing processing on the learning behavior data containing abnormal behavior markers based on a locally weighted regression algorithm to obtain smoothed learning behavior trend data; A third analysis unit, configured to process and model the smoothed learning behavior trend data based on an autoregressive model, and fit the established conditional heteroskedasticity model to the smoothed learning behavior trend data to obtain a volatility index of the student's learning performance.

10. The system for analyzing students' learning behaviors based on big data according to claim 6, wherein The processing unit includes: A first processing subunit, configured to use the smoothed learning behavior trend data and the volatility index of the learning performance as input data to construct a time series data set, where the time series data set includes the student's behavior history information and performance volatility information; A second processing subunit, configured to encode the past learning behavior data and the future learning behavior data through a bidirectional LSTM network to capture the temporal relationship between the student's learning behavior and performance; A third processing subunit, configured to construct multiple decision trees based on the temporal relationship between the student's learning behavior and performance. Each decision tree is trained on a different subset of the data set, and the results of all decision trees are voted to obtain the final relationship data between the learning behavior and performance; A fourth processing subunit, configured to perform feature analysis on the final relationship data and the volatility index of the learning performance based on a deep self-attention network to extract the behavior features that have the greatest impact on the learning performance; A fifth processing subunit, configured to perform grey relational analysis on the behavior features that have the greatest impact on the learning performance and the volatility index of the learning performance, and use the analyzed correlation degree value as the performance impact degree value of the behavior features that have the greatest impact on the learning performance. The behavior features that have the greatest impact on each student's learning performance and their corresponding impact degree values are summarized to obtain a learning behavior analysis report.

Images