A Multi-Label Learning Behavioral Feature Selection Method and System Based on Shared Structure
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-14
- Publication Date
- 2026-08-14
AI Technical Summary
多标签特征选择分为：包装式方法，直接对特征子集进行迭代评估，虽然准确性有所提升，但计算复杂度较高且容易陷入局部最优情况，导致资源和成本的大量消耗；过滤式方法，通常基于信息论指标，比如基于因果判断、基于集群算法等，具有运算速度快的特点，但忽略了特征与目标任务之间的关联性；嵌入式方法，将特征选择过程与模型的训练过程进行融合，因其同时具有较好的性能和效率，近年来受到了广泛关注
[0024]第一，现有的多标签学习行为特征选择方法通常利用线性回归、自表达学习等技术进行特征选择，但这些方法在处理复杂的学习行为数据与学习目标时，通常选择将学习行为特征矩阵与标签矩阵分开进行运算，没考虑学习行为特征与学习目标间共享的结构信息。本发明结合非负矩阵分解（NMF）技术，同时对学习行为特征矩阵与标签矩阵进行分解，从而学习共享结构，进一步提升了所选学习行为特征的质量，并通过每个学习行为特征的权重区分各个学习行为特征。主要说明如下：
Smart Images
Figure CN122571031A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of smart teaching technology, and in particular relates to a method and system for selecting multi-label learning behavior features based on a shared structure. Background Technology
[0002] Learning behavior analysis, as an important research direction in education, aims to improve students' learning efficiency by collecting and analyzing various behavioral data during the learning process, thereby guiding teachers to optimize teaching strategies. In recent years, an increasing number of methods have attempted to improve the efficiency and performance of this behavior analysis. A common approach is to use mathematical modeling to treat multidimensional learning objectives as multi-labels, and then mine the correlation information between various student learning behavior data (such as learning time, interaction frequency, etc.) and these labels. However, with the ever-growing scale of data in modern society, while the collected high-dimensional data contains richer information, it also triggers the "curse of dimensionality," that is, the excessively high dimensionality of the data to be processed leads to a massive consumption of computing and storage resources, significantly increasing the cost of learning behavior analysis. Therefore, how to effectively address this issue and improve learning behavior analysis has become an urgent problem to be solved. If an efficient data dimensionality reduction method can be combined to guide the learning behavior analysis process, and if it can handle multi-dimensional and multi-category learning objectives while maintaining a certain level of efficiency and performance, then this problem can be addressed to some extent. Therefore, multi-label feature selection technology that can perform efficient dimensionality reduction without destroying the original data has become a research focus in recent years. Designing a feature selection model framework that can capture the potential relationship between student behavior and multiple learning objectives has become the key to further improving the effectiveness of learning behavior analysis.
[0003] To successfully integrate learning behavior analysis with multi-label feature selection and construct a framework for subsequent multi-label learning behavior feature selection methods, it is necessary to first model the original data. This involves treating each student as a sample, multiple learning objectives as labels, and student learning behavior data as features. Based on this, by selecting the most representative subset of features from the original learning behavior features, the behavioral features with the greatest impact on learning objectives can be effectively identified, thus assisting teachers in conducting learning behavior analysis. Multi-label feature selection can be categorized into: wrapper methods, which directly iteratively evaluate feature subsets, improving accuracy but exhibiting high computational complexity and a tendency to get trapped in local optima, leading to significant resource and cost consumption; filtering methods, typically based on information theory metrics such as causal judgments or cluster algorithms, offering fast computation but neglecting the correlation between features and the target task; and embedded methods, which integrate the feature selection process with the model training process, gaining widespread attention in recent years due to their combined good performance and efficiency. However, existing methods still have the following limitations: 1) They ignore the potential structural relationship between the sample space (also known as the feature space) and the label space, i.e., the potential relationship that similar labels should share similar features; 2) They directly obtain information from the original sample space and the original label matrix, ignoring the influence of redundant behavioral features, and the information utilization among multiple learning objectives is insufficient; 3) Traditional graph constraint methods cannot adapt well to the current large-scale behavioral features and learning objectives, which leads to an inaccurate subset of selected learning behavioral features, making it difficult for teachers to accurately locate the most critical behavioral features for the learning objectives. Therefore, there is an urgent need for a multi-label learning behavioral feature selection method that can effectively solve the above problems. Summary of the Invention
[0004] The purpose of this invention is to solve the problems existing in the prior art and to provide a method and system for selecting multi-label learning behavioral features based on a shared structure.
[0005] To achieve the above-mentioned objectives, the present invention specifically adopts the following technical solution:
[0006] In a first aspect, the present invention provides a multi-label learning behavior feature selection method based on a shared structure, comprising the following steps:
[0007] S1. Construct a learning behavior feature matrix from the learner's learning behavior characteristics and obtain a weighted learning behavior feature matrix through a feature weighting mechanism; construct a label matrix from the learner's learning objectives and initialize a pseudo-label matrix at the same time. Obtain the potential shared structure between the weighted learning behavior feature matrix and the pseudo-label matrix through non-negative matrix decomposition, and use the sum of the decomposition errors of the two matrices as the original optimization term. Maintain the structural consistency between the pseudo-label matrix and the label matrix through the pseudo-label dependency term.
[0008] S2. Construct a joint hypergraph regularization term to simultaneously consider the higher-order relationships in the weighted learning behavior feature matrix and the pseudo-label matrix; then linearly add the original optimization term, pseudo-label dependency term, joint hypergraph regularization term, and sparse constraint term as the objective function to be solved. After solving, obtain the weights of each learning behavior feature under multiple learning objectives, thereby completing the selection of learning behavior features.
[0009] Based on the above scheme, each step can be implemented in the following preferred manner.
[0010] As a preferred embodiment of the first aspect mentioned above, in S1, the pseudo-label dependency is obtained by subtracting two parts: the first part is the L1 norm of the sum matrix, which is the matrix obtained by adding the label matrix and the pseudo-label matrix; the second part is twice the L1 norm of the product matrix, which is the matrix obtained by multiplying the label matrix and the pseudo-label matrix element by element.
[0011] As a preferred embodiment of the first aspect, in S2, the method for constructing the joint hypergraph regularization term is as follows: first, the potential shared structure is represented by a matrix and its transpose is calculated; then, the transpose of the potential shared structure matrix is multiplied by the hypergraph Laplacian matrix to obtain an intermediate matrix, which is then multiplied by the potential shared structure matrix to obtain a square matrix. The trace of this square matrix is taken as the joint hypergraph regularization term, and the scalar value obtained after taking the trace is taken as the result of the joint hypergraph regularization term.
[0012] As a preferred embodiment of the first aspect above, the method for calculating the hypergraph Laplacian matrix is as follows: taking learners as nodes, an association matrix is constructed by simultaneously considering the learner similarity in the weighted learning behavior feature matrix and the pseudo-label matrix; hyperedges are established between two learners with similarity, and weights are assigned to each hyperedge to form a hyperedge weight matrix; finally, the hypergraph Laplacian matrix is calculated using non-normalization based on the association matrix and the hyperedge weight matrix.
[0013] As a preferred embodiment of the first aspect mentioned above, for two nodes corresponding to two learners, the Euclidean distance between the two nodes is calculated using the weighted learning behavior feature matrix and the pseudo-label matrix, respectively. Only when both calculated Euclidean distances are less than the corresponding similarity thresholds are the two learners considered to be similar learners.
[0014] As a preferred embodiment of the first aspect mentioned above, in S2, the sparse constraint term is the L21 norm of the learning behavior feature importance matrix, and the learning behavior feature importance matrix is the matrix formed by multiplying the two coefficient matrices generated during non-negative matrix decomposition.
[0015] Secondly, this invention provides a multi-label learning behavior feature selection system based on a shared structure, comprising:
[0016] The weight calculation module is used to preset one or more feedback text templates for each learning behavior feature, and obtain the weights of each learning behavior feature of the target learner under multiple learning objectives according to the multi-label learning behavior feature selection method based on shared structure described in any of the above-mentioned first aspect schemes.
[0017] The results output module is used to take the top K learning behavior features with the highest weights as driving features, and generate instruction information for optimizing teaching strategies or recommendation information for customizing learning programs from the feedback text template of the driving features, and then feed it back to the target learner.
[0018] Thirdly, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, can implement the multi-label learning behavior feature selection method based on a shared structure as described in any of the solutions in the first aspect above.
[0019] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the multi-label learning behavior feature selection method based on a shared structure as described in any of the solutions of the first aspect above.
[0020] Fifthly, the present invention provides a computer electronic device, which includes a memory and a processor;
[0021] The memory is used to store computer programs;
[0022] The processor is configured to, when executing the computer program, implement the multi-label learning behavior feature selection method based on a shared structure as described in any of the solutions of the first aspect above.
[0023] Compared with the prior art, the present invention has the following advantages:
[0024] First, existing multi-label learning behavior feature selection methods typically utilize techniques such as linear regression and self-expressive learning. However, when dealing with complex learning behavior data and learning objectives, these methods usually perform calculations on the learning behavior feature matrix and the label matrix separately, without considering the shared structural information between the learning behavior features and the learning objective. This invention combines non-negative matrix factorization (NMF) technology to simultaneously decompose the learning behavior feature matrix and the label matrix, thereby learning the shared structure and further improving the quality of the selected learning behavior features. Furthermore, it distinguishes each learning behavior feature by assigning a weight to each feature. The main points are explained below:
[0025] This invention combines the learning behavior analysis problem with the multi-label feature selection task, modeling learners' multiple learning objectives as a label matrix and learners' learning behavior features as a learning behavior feature matrix. The number of rows represents the number of learners, and the number of columns represents the number of learning behavior features. This modeling approach comprehensively considers the complex relationship between learning objectives and learning behavior features, facilitating the subsequent multi-label learning behavior feature selection process.
[0026] This invention utilizes nonnegative matrix factorization to learn the shared structure between learning objectives and learning behavior features, better capture the potential relationships in learning behavior data, and guide the subsequent feature selection process based on this to improve accuracy and robustness.
[0027] The main function of this invention is to provide an efficient multi-label learning behavior feature selection method. By selecting the most representative features with the highest relevance to learning objectives from learners' learning behavior data, it helps educators analyze learners' learning situations and assists in designing targeted teaching methods.
[0028] Second, existing methods for shared structure learning using traditional nonnegative matrix factorization typically use the label matrix and the learned behavior feature matrix directly. However, the learned behavior feature matrix contains redundant features, and the label matrix contains noisy labels, both of which negatively impact the overall performance of the method. To mitigate these effects, this invention transforms the learned behavior feature matrix into a weighted learned behavior feature matrix based on feature weights and converts the label matrix into a pseudo-label matrix for shared structure learning. This aims to more effectively reduce the interference of redundant features and noisy labels on the model, thereby further improving the feature selection process. The main points are explained below:
[0029] This invention utilizes feature weights to weight the learned behavior feature matrix, appropriately amplifying relevant features and reducing irrelevant features, thereby making the weighted learned behavior feature matrix structurally closer to the true distribution of the data. This method effectively reduces interference from redundant features, improves the stability of the learned shared structure, and guides the feature selection process.
[0030] This invention utilizes a pseudo-label matrix to replace the label matrix and achieves structural consistency between the two by setting pseudo-label dependencies. This transformation reduces the impact of noisy labels on shared structure learning, allowing each learning objective to be more accurately represented in the pseudo-label matrix, thereby improving the performance of the feature selection model.
[0031] This invention uses the low-dimensional space carrying the shared structure as the shared space, and utilizes the learner similarity in the weighted learning behavior feature matrix and the pseudo-label matrix to construct a joint hypergraph regularization term, thereby applying graph constraints to the shared space. Furthermore, by combining manifold learning with this joint hypergraph regularization term, the structural relationship between the pseudo-label space and the weighted sample space is preserved, further improving the stability of the shared space.
[0032] This invention effectively eliminates unnecessary features by introducing sparse constraint terms, improves the sparsity of the model, avoids overfitting problems in the feature selection process, and ensures the interpretability of the learning behavior analysis process. Attached Figure Description
[0033] Figure 1 This is a flowchart of the steps of the method of the present invention;
[0034] Figure 2 This is a flowchart illustrating the method for calculating the Laplacian matrix of a hypergraph in this invention.
[0035] Figure 3 This is a framework diagram for multi-label learning behavioral feature selection in the method of this invention;
[0036] Figure 4 This is a system block diagram of the present invention;
[0037] Figure 5 This is a schematic diagram of a computer electronic device provided by the present invention. Detailed Implementation
[0038] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.
[0039] In the description of this invention, it should be understood that the terms "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features.
[0040] When selecting learning behavior features, further exploring the potential shared structural information between learning behaviors and learning objectives can significantly improve performance and thus enhance the learning behavior analysis process. Inspired by this knowledge, this invention transforms learning behavior analysis into a special multi-label learning behavior feature selection task, proposing a multi-label learning behavior feature selection method based on shared structure. This method guides the selection process in multi-label scenarios by exploring the potential shared structural information between learning behaviors and learning objectives.
[0041] like Figure 1 As shown, in a preferred embodiment of the present invention, the multi-label learning behavior feature selection method based on shared structure includes the following S1-S2 steps. The specific implementation process of each step will be described in detail below.
[0042] S1. Construct a learning behavior feature matrix from the learner's learning behavior characteristics and obtain a weighted learning behavior feature matrix through a feature weighting mechanism; construct a label matrix from the learner's learning objectives and initialize a pseudo-label matrix at the same time. Obtain the potential shared structure between the weighted learning behavior feature matrix and the pseudo-label matrix through non-negative matrix decomposition, and use the sum of the decomposition errors of the two matrices as the original optimization term. Maintain the structural consistency between the pseudo-label matrix and the label matrix through the pseudo-label dependency term.
[0043] It should be noted that in S1 of this invention, by collecting and organizing the learning behavior characteristics of all learners, each learner is considered as a sample, and the learning behavior characteristics are used as the corresponding features of each sample, thereby constructing a learning behavior feature space. The aforementioned learning behavior characteristics refer to the behaviors taken by each learner during the learning process, such as learning duration, resource access methods, courseware access frequency, number of assignment submissions, number of assignment views, assignment completion time, number of class questions, quiz participation, and quiz scores. If there are d types of learner behaviors, then the number of learning behavior features is also d. After uniformly processing the learners' original behaviors, each behavior dimension corresponds to a learning behavior feature, thus representing the behavior taken by each learner as a d-dimensional feature vector. If the number of samples is n, indicating there are n learners, by uniformly modeling all learners, an n×d learning behavior feature matrix can be formed. This modeling approach lays the foundation for subsequent shared structure learning and feature selection space.
[0044] Similarly, this invention constructs a label space by treating learners' multiple learning objectives as label vectors in multi-label learning. Traditional learning objective modeling methods typically consider only a single value, while this invention fully considers the multi-learning objective characteristics of learners during the learning process, simultaneously incorporating multiple learning objectives or outcomes into the modeling process. For example, course completion, knowledge mastery, midterm or final exam scores, test grades, learning ability improvement, and learning participation levels are considered as different learning objectives, i.e., multiple labels. Furthermore, this invention labels the above learning objectives for each learner based on their actual learning performance, thus constructing a multi-label matrix representation. Specifically, when a learner's actual learning performance indicates that they have achieved the corresponding learning objective, this invention marks the label as 1; otherwise, it marks the label as 0. In this way, each learner can simultaneously correspond to multiple learning objectives, i.e., multiple label vectors, ultimately forming an n×c label matrix. Where c represents the number of learning objectives, i.e., the number of labels. This label space can comprehensively reflect the multidimensional learning objectives of each learner, providing effective support for subsequent selection of multi-label learning behavior features based on shared structure learning.
[0045] It should be noted that in S1 of this invention, after obtaining the learning behavior feature matrix and the label matrix, considering the consensus in feature selection that learners with similar labels should also share this similarity in related features, non-negative matrix factorization (NMF) can be used to obtain the potential shared structure between the learning behavior feature matrix and the label matrix. Specifically, during the decomposition, existing techniques first decompose the learning behavior feature matrix into a first basis matrix. and the first coefficient matrix The form of transpose multiplication, i.e. Similarly, the label matrix can also be decomposed into a second basis matrix. Second coefficient matrix The form of multiplication, that is .in, Indicates matrix transpose. This refers to the number of clusters. At this point, the first basis matrix can obtain the structural information of the learned behavior feature matrix, and the second basis matrix can obtain the structural information of the label matrix. Furthermore, by using the first basis matrix instead of the second basis matrix, the first basis matrix can obtain the potential shared structure between the learned behavior feature matrix and the label matrix; the first basis matrix is also called the shared space. Therefore, the following formula can be obtained:
[0046]
[0047] in, Let Frobenius norm be the square of the matrix.
[0048] Considering the potential presence of redundant features in the learning behavior feature matrix, this invention employs a feature weighting mechanism to adjust the learning behavior feature matrix and reduce its impact, thereby making it more closely approximate the real data distribution. Specifically, this is achieved by directly multiplying the learning behavior feature matrix by the feature weight vector. Relevant features are amplified by their corresponding weights, while irrelevant features are reduced by their corresponding weights, thus obtaining a weighted learning behavior feature matrix. The label matrix may contain some noise. To reduce the impact of this noise, this invention further adjusts the label matrix. Specifically, during the shared structure learning process, this invention uses a randomly initialized pseudo-label matrix. Instead of the label matrix, this invention uses the squared Frobenius norm of the matrix to measure the error of matrix decomposition. Therefore, the specific form of the original optimization term is:
[0049]
[0050] in, The decomposition error of the weighted learning behavior feature matrix; This represents the decomposition error of the pseudo-label matrix; Let Frobenius norm be the square of the matrix.
[0051] It should be noted that in S1 of this invention, the pseudo-label dependency is obtained by subtracting two parts: the first part is the L1 norm of the sum matrix, which is the matrix obtained by adding the label matrix and the pseudo-label matrix; the second part is twice the L1 norm of the product matrix, which is the matrix obtained by multiplying the label matrix and the pseudo-label matrix element by element.
[0052] Furthermore, the aforementioned pseudo-tag dependency can be represented as:
[0053]
[0054] in, For sum matrix; Let L1 norm be the sum matrix; This refers to element-wise multiplication of a matrix. It is a product matrix; It is twice the L1 norm of the product matrix; Let L be the L1 norm of the matrix.
[0055] S2. Construct a joint hypergraph regularization term to simultaneously consider the higher-order relationships in the weighted learning behavior feature matrix and the pseudo-label matrix; then linearly add the original optimization term, pseudo-label dependency term, joint hypergraph regularization term, and sparse constraint term as the objective function to be solved. After solving, obtain the weights of each learning behavior feature under multiple learning objectives, thereby completing the selection of learning behavior features.
[0056] It should be noted that in S2 of this invention, the construction method of the joint hypergraph regularization term is as follows: first, the potential shared structure is represented by a matrix and its transpose is calculated; then, the transpose of the potential shared structure matrix is multiplied by the hypergraph Laplace matrix to obtain an intermediate matrix, which is then multiplied by the potential shared structure matrix to obtain a square matrix. The trace of this square matrix is taken as the joint hypergraph regularization term, and the scalar value obtained after taking the trace is taken as the result of the joint hypergraph regularization term.
[0057] Furthermore, the aforementioned joint hypergraph regularization term aims to ensure the stability of the shared space, overcoming the limitation that general graph structures can only be modeled as dependencies, and can be expressed as follows: .in, The trace of the matrix; For potential shared structure matrices, It is the Laplacian matrix of the hypergraph.
[0058] Furthermore, such as Figure 2As shown, the calculation method of the hypergraph Laplacian matrix is as follows: taking learners as nodes, an association matrix is constructed by simultaneously considering the learner similarity in the weighted learning behavior feature matrix and the pseudo-label matrix; hyperedges are established between two learners with similarity, and weights are assigned to each hyperedge to form a hyperedge weight matrix; finally, the hypergraph Laplacian matrix is calculated using non-normalization based on the association matrix and the hyperedge weight matrix.
[0059] Furthermore, for two nodes corresponding to two learners, the Euclidean distance between the two nodes is calculated using the weighted learning behavior feature matrix and the pseudo-label matrix, respectively. Only when both calculated Euclidean distances are less than the corresponding similarity threshold are the two learners considered to be similar learners.
[0060] In this embodiment, an association matrix is constructed. In this invention, learners are treated as nodes and serve as the starting points of hyperedges. Learners are connected to the same hyperedge when they are close to each other, i.e., have high similarity. Unlike previous methods, to obtain accurate structural relationships, this invention considers both learner similarity in the weighted learning behavior feature matrix and learner similarity in the pseudo-label matrix. Only when two learners have similar learning behavior features and learning objectives are they allowed to be connected to the same hyperedge. Specifically, this invention first uses Euclidean distance as a similarity metric to calculate learner similarity in the weighted learning behavior feature matrix and the pseudo-label matrix, thereby constructing a learning behavior similarity matrix. Similarity matrix with learning objectives The position in the learning behavior similarity matrix is... element at For learners and learners The Euclidean distance between them represents the learners'... and learners Similarity in learning behaviors. Similarly, the position in the learning objective similarity matrix is... element at For learners and learners The Euclidean distance between them represents the learners'... and learners The similarity in learning objectives. Then, in the learning behavior similarity matrix, this invention will include learners... The average Euclidean distance of this learner to all other learners is used as the threshold for the similarity of their learning behaviors. Similarly, in the learning objective similarity matrix, this invention will include learners... The average Euclidean distance of all other learners is used as the similarity threshold for the learner's learning objectives. Through the above operations, each learner corresponds to a similarity threshold in both the learning behavior and learning goal dimensions. Therefore, the process of constructing the association matrix can be expressed as:
[0061]
[0062] in, For the position in the correlation matrix The element at the location is used to represent the learner. and learners Are they learners with similar characteristics? This indicates that the learner and learners They are similar, and vice versa; This is a conditional statement used to connect two parallel conditions.
[0063] In this embodiment, the hyperedge weight matrix The importance of each hyperedge was measured. In order to reduce the overhead of subsequent optimization process, the weight of each hyperedge was set to 1 when constructing the hyperedge weight matrix, and it remained unchanged during the optimization process.
[0064] Following the above process, this embodiment obtains the correlation matrix and the hyperedge weight matrix. Therefore, the hypergraph Laplacian matrix is further calculated using non-normalization. :
[0065]
[0066] in, Let be the vertex degree matrix; It is the hypermarginality matrix; It is the inverse of the matrix.
[0067] Based on the aforementioned hypergraph Laplacian matrix, this embodiment constructs a hypergraph by simultaneously considering learner similarity in the weighted learning behavior feature matrix and the pseudo-label matrix. A joint hypergraph regularization term is introduced, and the structural stability of the shared space is further strengthened by utilizing the characteristic that the hypergraph structure can capture higher-order dependencies.
[0068] It should be noted that in S2 of this invention, the sparse constraint term is the L21 norm of the learning behavior feature importance matrix, and the learning behavior feature importance matrix is the matrix formed by multiplying the two coefficient matrices generated during non-negative matrix decomposition.
[0069] In this embodiment, the L21 norm is introduced to improve robustness. Specifically, considering that two sparse matrices will appear during the shared structure learning process, namely... and The present invention will As a whole, sparsity constraints are applied. Specifically, this invention will... As a learning behavior feature importance matrix, each row of this matrix represents the importance of a learning behavior feature across different labels. When a learning behavior feature has a minimum value across all labels, it is considered an unimportant feature, and this row will gradually approach 0 as the objective function is optimized.
[0070] Therefore, this invention uses a weighted learning behavior feature matrix and a pseudo-label matrix to learn the shared structure, and further enhances the stability of the shared space by utilizing a joint hypergraph regularization term that can capture higher-order dependencies. The importance of each behavior feature is represented by a sparse constraint term. Finally, the objective function for multi-label learning behavior feature selection can be expressed as:
[0071]
[0072]
[0073] in, For sparse constraint terms; , and These are balancing parameters, used to balance the interactions between the terms in the objective function; These are constraints.
[0074] Based on the above objective function, since it has the L21 norm and is non-convex, it cannot be directly optimized. Therefore, it is necessary to transform the objective function, decompose it into multiple sub-problems, and solve them sequentially. Considering the constraints... Converted to the following Lagrange form :
[0075]
[0076] in, For Lagrange multipliers; It is a diagonal matrix, and the first element on its diagonal is... element Represented as:
[0077]
[0078] in, This represents the k-th row in the importance matrix of learning behavior features; This is to prevent extremely small numbers from being divisible by 0.
[0079] Based on the Lagrange formulas described above, update the first basis matrix sequentially. First coefficient matrix Second coefficient matrix and pseudo-label matrix It also requires dynamic updating of the hypergraph Laplacian matrix. We treat a complete update as one iteration and record the objective function value for each iteration. For example... Figure 3 As shown, once the objective function value reaches convergence and stability, the weights of each learning behavior feature can be obtained and used as the output, thus completing the importance measurement of different learning behavior features under multiple learning objectives.
[0080] The first basis matrix is discussed below. First coefficient matrix Second coefficient matrix and pseudo-label matrix The update process will be explained.
[0081] 1) Update the first basis matrix
[0082] By fixing the remaining variables and calculating the partial derivatives with respect to the first basis matrix, we can obtain its update method:
[0083]
[0084] in, For a given position in a matrix Elements at the location; and The first The second optimization and the first The first basis matrix during the next optimization.
[0085] 2) Update the first coefficient matrix
[0086] Similarly, by fixing the remaining variables and taking the partial derivative with respect to the first coefficient matrix, we can obtain its update method:
[0087]
[0088] in, and The first The second optimization and the first The first coefficient matrix during the second optimization.
[0089] 3) Update the second coefficient matrix
[0090] Similarly, by fixing the remaining variables and taking the partial derivative with respect to the second coefficient matrix, we can obtain its update method:
[0091]
[0092] in, and The first The second optimization and the first The second coefficient matrix during the next optimization.
[0093] 4) Update the pseudo-label matrix
[0094] Similarly, by fixing the remaining variables and taking the partial derivative with respect to the pseudo-label matrix, we can obtain its update method:
[0095]
[0096] in, and The first The second optimization and the first The pseudo-label matrix during the second optimization; Let be an auxiliary matrix, where the position is in The specific calculation method for the element at position is as follows:
[0097]
[0098] Through the above update process, the objective function value will eventually converge, resulting in a weight vector that reflects the importance of all learning behavior features. The weight vector contains the first... Value For the first The weights of each learning behavior feature.
[0099] Once this weight vector is obtained, the contribution of each learning behavior feature to the learning objective can be effectively assessed. By analyzing these weights, it is possible to accurately identify which learning behavior features promote learners' learning objectives and which inhibit them, thereby helping educators optimize subsequent teaching strategies. Furthermore, educators can design corresponding learning tasks and customize learning plans suitable for different learners based on the weights of their learning behavior features when facing different learning objectives.
[0100] Specifically, if three learning behavior characteristics have relatively high weights—study time, frequency of looking up in class, and frequency of question-and-answer interaction—then these characteristics are important and have a significant impact on learning objectives. Other learning behavior characteristics can be understood as having a smaller impact on learning objectives. In this case, educators can develop corresponding learning plans and use two teaching strategies—adjusting the time required for learners to study or adjusting the frequency of classroom questions—to improve learners' learning in a targeted and dynamic way. For learners who have longer study times but whose learning outcomes are not as expected, educators can provide targeted guidance to help them improve their learning methods. For learners who do not look up in class frequently, resulting in poor learning outcomes, educators can pay more attention to whether learners are looking up at the blackboard when explaining knowledge points. This approach, which combines learning behavior analysis with multi-label feature selection, provides educators with a data-driven analytical method. It aims to select different learning behavior features by analyzing and comparing multiple learning goal environments, thereby enabling individualized instruction, improving the accuracy of teaching decisions, allowing learners to learn more effectively, enhancing the interaction between educators and learners, and enabling learners to dynamically adjust their learning engagement when facing different learning goals, focusing their energy on appropriate learning behaviors, ultimately improving overall teaching quality and effectiveness.
[0101] The present invention will now demonstrate the application effect of the multi-label learning behavior feature selection method based on shared structure described in S1~S2 of the above embodiments on a specific dataset through a specific example, so as to facilitate understanding of the essence of the present invention.
[0102] Example
[0103] The specific implementation process of the multi-label learning behavior feature selection method based on shared structure used in this embodiment is as described above and will not be repeated here.
[0104] To more clearly demonstrate the technical effectiveness of the proposed method, the Education dataset from Yahoo! was selected for subsequent experiments. This dataset contains 5000 samples, 550 features, and 33 labels. The micro-average, most commonly used in the field of multi-label feature selection, was chosen as the evaluation metric. Fraction It calculates the overall precision and recall by performing global statistics on the prediction results for all categories, thus obtaining the final F1 score. The value ranges from 0 to 1, where 1 indicates that the model's predictions are perfect across all categories, with the largest proportion of true positives and no false positives or negatives. 0 indicates that the model's predictions are completely inaccurate, all predictions are wrong, or the model cannot correctly identify any positive samples.
[0105] Table 1 shows a comparison between the method of this invention and other classic methods for multi-label feature selection. These methods include MIFS, SCMFS, FLFS, and NCMFS, and their implementations are all existing technologies. For fairness, all methods selected the top 20% of features and conducted multiple experiments, averaging the results. This embodiment also records the standard deviation of multiple results.
[0106] Table 1. Micro-mean scores (mean) of different methods on the Education dataset (Standard deviation)
[0107] Experimental results show that the method of the present invention has a significant improvement over other multi-label feature selection methods. One major reason is that the method of the present invention obtains the potential shared structure between learning behavior and learning target by learning the shared structure based on the weighted learning behavior feature matrix and pseudo-label matrix, thereby effectively guiding the multi-label learning behavior feature selection process.
[0108] Furthermore, this embodiment also verified the complete objective function and the contribution of each function term through ablation experiments. That is, by setting multiple downgraded versions, the contributions of each function term can be better compared. The experimental results are shown in Table 2. The specific settings of the downgraded versions are as follows:
[0109] Version 1: When learning shared structures, the learning behavior feature matrix and label matrix are used directly, i.e. And because the label matrix is used directly, it is necessary to set... This means removing pseudo-tag dependencies while retaining the rest.
[0110] Version 2: Settings Remove the joint hypergraph regularization term, which corresponds to eliminating the constraint on the shared space, while the remaining terms are retained.
[0111] Version 3: By setting Remove the sparse constraint terms, which correspond to the shared space not being able to maintain good row sparsity, and retain the remaining terms.
[0112] Table 2. Ablation experimental results of the method of the present invention (average) (Standard deviation)
[0113] The results in Table 2 show that the complete objective function has good performance, and each function term contributes to the overall performance. The shared structure learning part based on the weighted learning behavior feature matrix and pseudo-label matrix is the key to improving the method's performance; the joint hypergraph regularization term and sparse constraint term also have a certain effect on improving the model's robustness and stability.
[0114] It should also be noted that the multi-label learning behavior feature selection method based on shared structure in the above embodiments can essentially be executed by a computer program or module. Therefore, similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a multi-label learning behavior feature selection system based on shared structure, corresponding to the multi-label learning behavior feature selection method based on shared structure provided in the above embodiments, such as... Figure 4 As shown, it includes:
[0115] The weight calculation module is used to preset one or more feedback text templates for each learning behavior feature, and obtain the weights of each learning behavior feature of the target learner under multiple learning objectives according to the multi-label learning behavior feature selection method based on shared structure described in the above embodiments.
[0116] The results output module is used to take the top K learning behavior features with the highest weights as driving features, and generate instruction information for optimizing teaching strategies or recommendation information for customizing learning programs from the feedback text template of the driving features, and then feed it back to the target learner.
[0117] It is understood that the multi-label learning behavior feature selection method based on shared structure described in S1~S2 above can essentially be implemented by a computer program. Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer program product corresponding to the multi-label learning behavior feature selection method based on shared structure provided in the above embodiments, which includes a computer program / instruction. When the computer program / instruction is executed by a processor, it can implement the multi-label learning behavior feature selection method based on shared structure as described in the above embodiments.
[0118] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer electronic device corresponding to the multi-label learning behavior feature selection method based on shared structure provided in the above embodiments, such as... Figure 5 As shown, it includes a memory and a processor;
[0119] The memory is used to store computer programs;
[0120] The processor is configured to implement the multi-label learning behavior feature selection method based on shared structure in the above embodiments when executing the computer program.
[0121] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.
[0122] Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer-readable storage medium corresponding to the multi-label learning behavior feature selection method based on shared structure provided in the above embodiments. The storage medium stores a computer program, which, when executed by a processor, can implement the multi-label learning behavior feature selection method based on shared structure in the above embodiments.
[0123] Specifically, in the computer-readable storage medium of the above three embodiments, the stored computer program is executed by a processor, which can perform the aforementioned steps S1 to S2.
[0124] It is understood that the aforementioned storage media may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Furthermore, the storage media may also be various media capable of storing program code, such as USB flash drives, external hard drives, magnetic disks, or optical discs.
[0125] It is understood that the processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.
[0126] It should also be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. In the embodiments provided in this application, the division of steps or modules in the system and method is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules or steps may be combined or integrated together, and a module or step may also be split.
[0127] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.
Claims
1. A method for selecting behavioral features in multi-label learning based on a shared structure, characterized in that, Includes the following steps: S1. Construct a learning behavior feature matrix from the learners' learning behavior characteristics and obtain a weighted learning behavior feature matrix through a feature weighting mechanism; A label matrix is constructed based on the learner's learning objectives, and a pseudo-label matrix is initialized simultaneously. The potential shared structure between the weighted learning behavior feature matrix and the pseudo-label matrix is obtained through non-negative matrix decomposition. The sum of the decomposition errors of the two matrices is used as the original optimization term. The structural consistency between the pseudo-label matrix and the label matrix is maintained through the pseudo-label dependency term. S2. Construct a joint hypergraph regularization term to simultaneously consider the higher-order relationships in the weighted learning behavior feature matrix and the pseudo-label matrix; then linearly add the original optimization term, pseudo-label dependency term, joint hypergraph regularization term, and sparse constraint term as the objective function to be solved. After solving, obtain the weights of each learning behavior feature under multiple learning objectives, thereby completing the selection of learning behavior features.
2. The multi-label learning behavior feature selection method based on shared structure as described in claim 1, characterized in that, In S1, the pseudo-label dependency is obtained by subtracting two parts: the first part is the L1 norm of the sum matrix, and the sum matrix is the matrix obtained by adding the label matrix and the pseudo-label matrix; The second part is twice the L1 norm of the product matrix, which is the matrix obtained by multiplying the label matrix and the pseudo-label matrix element by element.
3. The multi-label learning behavior feature selection method based on shared structure as described in claim 1, characterized in that, In S2, the joint hypergraph regularization term is constructed as follows: first, the potential shared structure is represented by a matrix and its transpose is calculated. Then, the transpose of the potential shared structure matrix is multiplied by the hypergraph Laplacian matrix to obtain an intermediate matrix, which is then multiplied by the potential shared structure matrix to obtain a square matrix. The trace of this square matrix is taken as the joint hypergraph regularization term, and the scalar value obtained after taking the trace is taken as the result of the joint hypergraph regularization term.
4. The multi-label learning behavior feature selection method based on shared structure as described in claim 3, characterized in that, The hypergraph Laplacian matrix is calculated as follows: learners are treated as nodes, and an association matrix is constructed by simultaneously considering the learner similarity in the weighted learning behavior feature matrix and the pseudo-label matrix; hyperedges are established between two similar learners, and weights are assigned to each hyperedge to form a hyperedge weight matrix; finally, the hypergraph Laplacian matrix is calculated using non-normalization based on the association matrix and the hyperedge weight matrix.
5. The multi-label learning behavior feature selection method based on shared structure as described in claim 4, characterized in that, For two nodes corresponding to two learners, the Euclidean distance between the two nodes is calculated using the weighted learning behavior feature matrix and the pseudo-label matrix, respectively. Only when both calculated Euclidean distances are less than the corresponding similarity thresholds are the two learners considered to be similar learners.
6. The multi-label learning behavior feature selection method based on shared structure as described in claim 1, characterized in that, In S2, the sparse constraint term is the L21 norm of the learning behavior feature importance matrix, which is the matrix formed by multiplying the two coefficient matrices generated during non-negative matrix decomposition.
7. A multi-label learning behavior feature selection system based on a shared structure, characterized in that, include: The weight calculation module is used to preset one or more feedback text templates for each learning behavior feature, and obtain the weights of each learning behavior feature of the target learner under multiple learning objectives according to the multi-label learning behavior feature selection method based on shared structure as described in any one of claims 1 to 6. The results output module is used to take the top K learning behavior features with the highest weights as driving features, and generate instruction information for optimizing teaching strategies or recommendation information for customizing learning programs from the feedback text template of the driving features, and then feed it back to the target learner.
8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it can implement the multi-label learning behavior feature selection method based on a shared structure as described in any one of claims 1 to 6.
9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the multi-label learning behavior feature selection method based on a shared structure as described in any one of claims 1 to 6.
10. A computer electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the multi-label learning behavior feature selection method based on a shared structure as described in any one of claims 1 to 6.