Dynamic grouping optimization method and device for large-scale online collaborative learning and medium

Abstract

The invention relates to a dynamic grouping optimization method and device for large-scale online collaborative learning and a medium. The method comprises the steps that questionnaire information, course information and learning behavior data are collected and stored in a student information database, a course information database and a learning history database respectively; based on student and course database data, constructing a knowledge graph reflecting student and course features; learning node features by using a graph embedding algorithm, and generating an embedded vector; generating a preliminary group by using a clustering algorithm; in combination with a clustering result and student data, a central intelligence set is used to accurately describe student ability, a learner portrait is generated, and students who accord with a standard are screened to take charge of a group leader; based on the matching degree of the group members and the group leader and the grouping rule, adopting an optimization algorithm to allocate the group members; and finally generating learning groups and dynamically adjusting the groups. According to the method, the learning process data of the students are collected in real time, the student features are modeled in a multi-dimensional mode, the cooperation requirements of different learning stages are met, and a globally optimal dynamic grouping scheme is provided.

Description

Technical Field

[0001] The present invention relates to the field of smart education technology, and in particular to a dynamic grouping optimization method, device and medium for large-scale online collaborative learning. Background Art

[0002] With the rapid development of information technology, online education has been widely used, especially Massive Open Online Courses (MOOC), which have become popular rapidly, providing flexible and efficient learning opportunities for learners around the world. Compared with traditional offline teaching, MOOC can accommodate tens of thousands of learners, which greatly expands the coverage of education. However, learners participating in large-scale online learning have diverse learning backgrounds, cultural backgrounds and personality characteristics, which brings great challenges to learning management and the improvement of learning effects.

[0003] In the MOOC environment, the construction of learning groups and the effectiveness of collaborative learning are crucial to improving students' learning outcomes. Through reasonable grouping, students can inspire each other and complement each other's strengths in collaborative learning, thereby improving learning efficiency. However, the following problems exist in the large-scale online learning environment, making the grouping of learning groups particularly complicated: (1) There are a large number of learners in a large-scale online learning environment. They are distributed in different countries or regions and have diverse cultural backgrounds, language abilities, and learning habits. The diversity of these characteristics makes it difficult to construct groupings that are suitable for learning needs. For example, students have significant differences in beliefs, language abilities, and information technology proficiency, and these characteristics may significantly affect the effectiveness of collaborative learning.

[0004] (2) Current learning group grouping methods often group students based on their basic information (such as knowledge level, learning style, etc.), which ignores the analysis and utilization of the multi-dimensional historical learning data accumulated by learners on online learning platforms. In particular, during the learning process, students’ learning performance and abilities will change dynamically as the course progresses. This variability is difficult to capture with traditional grouping methods, resulting in poor grouping results.

[0005] (3) Most existing grouping methods only perform static grouping for a certain stage of the course and lack the ability to make dynamic adjustments, making it difficult to meet the diverse needs of group building at different stages of the course. For example, in the early stages of the course, learners may need more homogeneous groups to reduce learning loneliness; while in the practice stage, heterogeneous groups may be needed to achieve complementary abilities. This change in demand has not been fully considered and addressed by existing methods.

[0006] (4) Current mainstream research and practice often focus on a single goal, such as improving students’ collaborative abilities or leveraging certain specific strengths of students, while ignoring the balance of other goals. For example, how to simultaneously reduce students’ sense of loneliness in learning, avoid conflicts between students, and stimulate students’ multi-faceted potential are still key issues that have not been fully resolved.

[0007] (5) Faced with the computational requirements of large-scale learner grouping, traditional algorithms are difficult to provide efficient optimization solutions within a limited time. In addition, there are many specific constraints that need to be met in online learning environments, such as time conflicts and language differences between learners, which further increases the complexity of the grouping problem.

[0008] At present, some studies have attempted to group students through clustering algorithms, optimization models and other technologies, but they are mainly suitable for small and medium-scale learning scenarios and are difficult to effectively handle the high complexity requirements of large-scale online learning environments. For example, traditional methods rarely consider the impact of students' diverse characteristics (such as beliefs, language skills, information technology proficiency, etc.) in large-scale online learning environments, and fail to fully tap the multidimensional historical data accumulated by students on the learning platform. In addition, many methods construct learning groups based on static grouping, lack dynamic adjustment capabilities, and fail to optimize grouping plans according to the needs of different learning stages of the course.

[0009] In summary, grouping learning groups in large-scale online learning environments still faces many challenges. How to effectively tap into the multi-dimensional characteristics of learners, dynamically adjust grouping schemes to meet the needs of different learning stages, and at the same time improve learning outcomes and reduce learning loneliness are issues that need to be addressed in the current online education field. Summary of the invention

[0010] The present invention provides a dynamic grouping optimization method, device and medium for large-scale online collaborative learning, which aims to solve the problems of low efficiency, poor effect and difficulty in dynamic adjustment of learning group grouping caused by the diversity of learner characteristics and dynamic changes in learning needs in a large-scale online learning environment.

[0011] To achieve the above object, the first aspect of the present invention provides a dynamic grouping optimization method for large-scale online collaborative learning, comprising the following steps: Collect questionnaire information filled out by students after logging in, course-related information, and students' behavior data during the learning process from the online learning platform, and store them in the student information database, course information database, and learning history database respectively; Based on the behavioral data in the student information database and the course information database, a knowledge graph reflecting the student characteristic information and the course characteristic information is constructed; Apply graph embedding algorithm to learn the features of nodes in the knowledge graph and generate node embedding vectors; Clustering the node embedding vectors using a clustering algorithm to generate preliminary results for learner grouping; Based on the clustering results and the data in the student information database and the learning history database, the single-valued neutral intelligence set is used to characterize and describe the learner's ability and generate a learner portrait for each learner; Select learners who are suitable to serve as group leaders based on the learner portrait and the preset group leader screening criteria; Based on the comprehensive matching degree between team members and team leaders and the grouping constraint rules, an optimization algorithm is used to allocate team members; Generate learning groups based on the optimized allocation results, and dynamically adjust the groupings during the learning process.

[0012] Furthermore, the knowledge graph construction method includes: Extracting students' characteristic information from the student information database; Extracting characteristic information of prerequisite courses from a course information database; Design an ontology model based on the characteristic information of students and prerequisite courses; According to the ontology model, students, courses and their characteristic information are organized into nodes of the knowledge graph, and the relationship between nodes is represented as edges to construct a preliminary knowledge graph; Based on the entities and their relationships in the preliminary knowledge graph, the knowledge graph is improved in combination with weight information to obtain the final knowledge graph.

[0013] Furthermore, the graph embedding algorithm is applied to learn the features of the nodes in the knowledge graph and generate node embedding vectors, including the following methods: Based on the constructed knowledge graph, determine the relationship weights between nodes in the graph, wherein the nodes include student characteristics, prerequisite courses, exercises, and knowledge points, and the relationship weights are used to represent the strength of association between nodes; Perform first-order and second-order similarity optimization on the node relationships in the knowledge graph. The first-order similarity optimizes the similarity of low-dimensional embedding vectors through directly connected nodes; the second-order similarity optimizes the similarity of low-dimensional embedding vectors through nodes with common neighbors. Based on the first-order similarity optimization and the second-order similarity optimization, the corresponding node embedding vectors are trained respectively, and the two node embedding vectors are fused to generate the final embedding vector of the node.

[0014] Furthermore, the method of clustering the node embedding vectors using a clustering algorithm and generating preliminary results of learner grouping includes: Based on the requirements of the clustering algorithm, the branch balance factor, leaf balance factor and spatial threshold of the clustering feature tree are set to control the number of branches of internal nodes, the number of clustering features of leaf nodes and the maximum value of the diameter or radius of each clustering feature in the leaf node. Insert the received node embedding vectors into the cluster feature tree one by one. If the cluster feature tree is empty, construct a leaf node and a subcluster, and insert the data point into it. If the cluster feature tree already has nodes, check whether the inserted data points exceed the spatial threshold, leaf balance factor, and branch balance factor one by one. If none of them exceed, complete the insertion operation. If any parameter value is exceeded, perform node splitting and cluster splitting operations according to the specific situation to create new nodes and clusters. Update the clustering feature information of each leaf node and non-leaf node in the clustering feature tree; after the node splits or the cluster splits, recursively update the newly generated node information upward to the parent node until the root node; The initial grouping of learners is completed through the construction of a clustering feature tree.

[0015] Furthermore, based on the clustering results combined with the data in the student information database and the learning history database, the learner's ability is characterized and described using a single-valued neutral intelligence set. The method of generating a learner portrait for each learner includes: Extract the learner node of each cluster from the learner clustering result set; According to the extracted learner node, the corresponding student feature information is obtained from the student information database, and the corresponding learning history record is obtained from the learning history database; According to the current course learning stage, determine the relevant data sources for calculating learner characteristics. If it is in the first stage, the questionnaire data and the learning records of the prerequisite courses are used first; if it is in the second to third stage, the questionnaire data and the learning records of the prerequisite courses are used first. q The second stage is based on the aforementioned data combined with the learning data of the current course; For each learner feature, relevant data is extracted from the student information database and classified, and a single-valued neutral intelligence number of each feature is generated according to the rules, wherein the single-valued neutral intelligence number includes a degree of truth, a degree of uncertainty, and a degree of distortion; For multi-dimensional data sources, a fusion strategy is used to integrate them and generate a comprehensive single-valued neutrino data of learner characteristics. The fusion strategy includes but is not limited to cloud model, weight distribution model and stage decay model. In different learning stages, the weight distribution of the single-valued neutrinometry of each feature is dynamically adjusted according to the availability of relevant data sources, and the final single-valued neutrinometry of the feature is calculated through the integration formula; The final single-valued neutrino data of learner characteristics are converted into structured portrait data, and the learner portrait based on the final single-valued neutrino data is output.

[0016] Further, methods for selecting learners suitable to serve as group leaders include: Receiving the learner portrait based on the final single-valued neutrino-intelligence number, extracting the feature value set of the students in each cluster; The feature value set is screened according to the preset threshold standard, and students who do not meet the threshold requirements are eliminated; Determine the characteristic indicators and their corresponding weights required for selecting group leaders based on the specific course scenario; Based on the characteristic values ​​of the screened students, the selected indicators and their weights, the competence of each student as a team leader is calculated; Sort the calculated competence of the group leader in descending order, and select the student with the highest competence in the sorted results as the group leader; The obtained group leader number is stored in the group leader set, and the remaining student numbers are stored in the member to be assigned set; Submit the group leader set and the member set to be assigned of all clusters to the group member assignment module.

[0017] Furthermore, based on the comprehensive matching degree between the team members and the team leaders and the grouping constraint rules, the allocation method of using the optimization algorithm to allocate the team members includes: Receive a set of team leaders and a set of members to be assigned; For the team leader and the member to be assigned in each cluster, the comprehensive matching degree is calculated based on the eigenvalues; According to the calculation results of the comprehensive matching degree, the grouping problem is formally modeled based on the E-CARGO model, and the objective function and constraints are set; Use the optimization algorithm to solve the objective function and obtain the best grouping result for each cluster while satisfying the constraints; Integrate the grouping results of all clusters into the final grouping solution; Provide feedback to students and teachers on the final grouping plan.

[0018] Furthermore, when the optimization algorithm is adjusted according to the dynamic grouping requirements, the following steps are also included: According to the teaching objectives of the current learning stage, dynamically adjust the grouping constraint rules and add or modify stage-specific constraints; Evaluate the grouping results of each learning stage. The evaluation indicators include the collaboration effect among group members, the quality of group task completion, and learners' satisfaction feedback; According to the evaluation results, optimize the parameters and weight allocation rules of the grouping model, and update the calculation formula of the comprehensive matching degree; The optimized grouping model is applied to the grouping adjustment process in the next stage.

[0019] To achieve the above object, the second aspect of the present invention provides a dynamic grouping optimization device for large-scale online collaborative learning, comprising the following modules: The data collection and processing module is used to collect questionnaire information, course-related information and students' learning behavior data from the online learning platform and store them in the database module, where the database module includes a student information database, a course information database and a learning history database, which are used to store the collected relevant data respectively; Graph construction and graph embedding training modules, where the graph construction module is used to construct a knowledge graph reflecting student feature information and course feature information based on the behavioral data in the student information database and the course information database; the graph embedding training module is used to perform feature learning on nodes in the knowledge graph and generate node embedding vectors; Learner clustering module, used to cluster node embedding vectors and generate preliminary results of learner grouping; The learner portrait module is used to characterize and describe the learner's ability based on the clustering results combined with the student information and learning history, and generate a learner portrait using a single-valued neutral intelligence set; The group leader selection module is used to select learners who are suitable to serve as group leaders based on learner portraits and group leader screening criteria; The team member allocation module is used to optimize the allocation of team members based on the comprehensive matching degree between team members and team leaders and the grouping constraint rules.

[0020] To achieve the above objectives, the third aspect of the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the dynamic grouping optimization method for large-scale online collaborative learning are executed.

[0021] Beneficial effects of the present invention: Compared with the prior art, the present invention provides a dynamic grouping optimization method, device and medium for large-scale online collaborative learning. By constructing a knowledge graph based on student characteristics and course characteristics, the multi-dimensional characteristics of students are comprehensively analyzed, and the feature information is learned by applying a graph embedding algorithm to generate efficient data representation to support subsequent grouping optimization. At the same time, the clustering algorithm is combined with the initial grouping and the use of single-valued neutral intelligence sets to characterize and describe the learner's ability, generate a more accurate learner portrait, and can more accurately characterize the learner's characteristics, select a group leader who meets the preset standards, and dynamically generate a learning group based on the comprehensive matching degree between the group members and the group leader and the grouping constraint rules. By dynamically adjusting the grouping strategy according to the progress of the course, the present invention solves the problems of low grouping efficiency and poor matching degree among group members in a large-scale online learning environment due to the diversity of learner characteristics and changes in course requirements, effectively reduces the sense of loneliness in learning, and improves the effect of collaborative learning and the overall quality of online learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments.

[0023] Figure 1 It is a framework diagram of a dynamic grouping optimization device for large-scale online collaborative learning disclosed in an embodiment of the present invention.

[0024] Figure 2 It is a workflow diagram of a group leader selection module disclosed in an embodiment of the present invention.

[0025] Figure numerals: 100, data collection and processing module; 200, graph construction and graph embedding training module; 300, learner clustering module; 400, learner portrait module; 500, team leader selection module; 600, team member allocation module. DETAILED DESCRIPTION

[0026] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0027] According to an embodiment of the present invention, it should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the following method, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0028] Before describing the method, it is necessary to define the basic sets and variables involved in student grouping and profiling in the online learning platform, which provides theoretical support for the subsequent method design and implementation. (hereinafter referred to as the current course) , For the number of students taking elective courses You need to learn some courses first, which are called courses. The prerequisite courses (hereinafter referred to as prerequisite courses) are extracted from students’ historical learning records in prerequisite courses and related course information, and then combined with learners’ characteristic information, the above data are used to profile students and optimize the learning group grouping. The learning process is divided into There are three learning stages, and each stage still needs to form a learning group for collaborative learning. The learning data generated in the process can be combined to dynamically divide and adjust the learning groups. is the set of students who filled in the questionnaire, among which Representing learners All questionnaire information filled in. A collection of students who have a record of studying in the current course. Similarly, the set of students representing the historical learning data of the prerequisite courses. If the student does not have the historical learning data of the prerequisite courses, then .

[0029] like Figure 1 As shown, the present invention provides a dynamic grouping optimization device for large-scale online collaborative learning, which is composed of the following modules: The data collection and processing module 100 is installed on the application server of the online learning platform and is used to collect the relevant questionnaire information filled out by students after logging into the online learning platform, such as learning style questionnaire, personality questionnaire, etc. At the same time, the learning-related data generated by students in the learning process of the online learning platform are collected, including learners' behavior data (such as courseware browsing times, video viewing records, etc.), task completion data (such as homework submission records, test participation, etc.), interactive communication data (such as posting, Q&A, etc.), platform usage data (such as learning time period, learning time interval, etc.), learning effect data (such as test scores, test scores, etc.), emotional data (such as learning satisfaction, emotional analysis results of a single course, etc.). And the course-related information provided by the teacher when creating the course, including exercises and knowledge points related to the course. At the beginning of the course, there are only questionnaire information and historical learning data generated by learners who have taken prerequisite courses on the learning platform. After collecting these data, the data collection and processing module will process these data and convert them into a unified standard format. And store the relevant feature information of the students in the student information database, the learning history data of the students in the learning history database, and the course-related information provided by the teacher in the course information database. After a period of time, all students will generate certain course learning data. The data collection and processing module will collect this data, convert it into a unified format, and store it in the corresponding database for subsequent group division and dynamic adjustment. The data collection and processing module will collect and process data according to a set time period (such as each learning stage of the course), and transmit the data to subsequent modules to adjust the grouping.

[0030] The graph construction and graph embedding training module 200 is used to extract student characteristics from the student information database, including the learner's religious beliefs, majors, ethnicity / race, region / time zone, learning goals and learning habits. Then, relevant information about the course, such as exercises and knowledge points, is extracted from the course information database. Next, these two types of information are combined to design the ontology model and construct the knowledge graph. Finally, the graph embedding algorithm is applied to learn the feature information of the nodes in the knowledge graph and obtain the embedding vectors of the nodes. These embedding vectors are submitted to the learner clustering module.

[0031] The learner clustering module 300 is used to receive the embedding vector data submitted by the graph construction and graph embedding training module, and cluster the learner nodes using a clustering algorithm to obtain a clustering result set, which is submitted to the learner portrait module.

[0032] The learner portrait module 400 is used to receive the clustering result set submitted by the learner clustering module, and perform the following processing on each small set in the clustering result set, that is, the cluster obtained by clustering. First, the student nodes in each cluster are taken out, and then the characteristic information of these students is extracted from the student information database, and the learning process data of these students is extracted from the learning history database. Combining these two types of information, a single-valued neutral intelligence set is used to characterize and describe the learner's ability, and the students in the cluster are profiled, and the portrait results are submitted to the group leader selection module.

[0033] The group leader selection module 500 is used to receive the portrait result information submitted by the learner portrait module. According to the set group leader evaluation criteria, the portrait result information is evaluated to find suitable candidates for group leaders to form a group leader set. The remaining students form a set of members to be assigned. The group leader set and the set of members to be assigned are submitted together to the group member assignment module.

[0034] The group member allocation module 600 is used to receive the group leader set and the member set information to be allocated submitted by the group leader selection module. First, the comprehensive matching degree between the member to be allocated and each group leader in the group leader set is calculated, and then the relevant constraints are taken out from the set constraint rule set. Then, the E-CARGO model is used to model and analyze the learning group grouping problem, and the IBM CPLEX optimization package is used to solve the grouping problem to obtain the final grouping optimization plan. Finally, the grouping optimization plan is fed back to students and teachers.

[0035] The following is a dynamic grouping optimization method for large-scale online collaborative learning to explain this solution in detail. The grouping optimization method includes the following steps: Step S100: Collect questionnaire information filled out by students after logging in, course-related information, and students' behavior data during the learning process from the online learning platform, and store them in the student information database, course information database, and learning history database respectively; Step S200: construct a knowledge graph reflecting student characteristic information and course characteristic information based on the behavior data in the student information database and the course information database; Step S300: Apply a graph embedding algorithm to perform feature learning on nodes in the knowledge graph to generate node embedding vectors; Step S400: clustering the node embedding vectors using a clustering algorithm to generate preliminary results of learner grouping; Step S500: Based on the clustering results and the data in the student information database and the learning history database, the single-valued neutral intelligence set is used to characterize and describe the learner's ability, and a learner portrait is generated for each learner; Step S600: Select learners who are suitable to serve as group leaders according to the learner portrait and the preset group leader screening criteria; Step S700: Based on the comprehensive matching degree between the group members and the group leaders and the grouping constraint rules, the group members are allocated using an optimization algorithm; Step S800: Generate a learning group according to the optimization allocation result, and dynamically adjust the grouping during the learning process.

[0036] In this embodiment, as described in the above steps S200 and S300, the working process of the graph construction and graph embedding training module 200 is divided into the following four steps: Step 1, graph construction and graph embedding training module 200 will determine which learning stage the current course learning process is in. If it is the first stage, it is necessary to extract the relevant feature information of the students from the student information database and the course information of the students' prerequisite courses from the course information database. Otherwise, directly use the previously constructed graph and trained graph embedding vector, jump to the learner clustering module, and submit the embedding vector result to the learner clustering module.

[0037] Step 2: Assumptions For Students Characteristics of students in large-scale online learning environments, such as religion, country, region / time zone, ethnicity / race, school, major, learning goals, and study habits. For Students Information about the prerequisite courses you have chosen, is the number of courses. Each course contains exercises and knowledge points. The number of exercises and knowledge points of different courses varies. Before building a knowledge graph, an ontology model needs to be designed. The ontology model mainly contains entities and relationships between entities. In this embodiment, students, characteristics of students, prerequisite courses, exercises included in prerequisite courses, and knowledge points involved in exercises are used as entities. Relational data can be further extracted based on these entities, such as student Features , which means that the student entity With feature entity If there is a relationship between ,feature and the relationship between the two A triple can be formed. According to the constructed ontology model, it can be further organized into a knowledge graph.

[0038] Step 3: Based on the knowledge graph constructed by the ontology model, apply the graph embedding algorithm to perform graph embedding training to learn the embedding vector representation of the node. Taking the classic LINE model as an example, the LINE model optimizes the calculation method of node similarity and proposes first-order similarity and second-order similarity. Among them, the first-order similarity means that when there is an edge connecting two vertices, the embedding vectors of the two nodes should be very similar in the low-dimensional space. The second-order similarity means that although there is no direct edge connecting the two vertices, they have the same neighbors, then the embedding vectors of the two nodes should also be very similar in the low-dimensional space. When calculating the first-order similarity, LINE first performs a similar operation on each undirected edge. , define the vertex and The joint probability distribution between is: (1) in, Vertex The low-dimensional embedding vector representation of represents a low-dimensional space, Vertex The transpose of the low-dimensional embedding vector representation of is a natural exponential function, which is used to convert the dot product into a probability value. At the same time, the empirical distribution is defined as: (2) in: in, is the empirical probability distribution, indicating the vertex and vertices The normalized value of the connection strength between is a set of vertices, Vertex and vertices The strength of the relationship between them, that is, the weight of the relationship, is the normalization coefficient, which represents the sum of all edge weights and is used to convert Convert to probability.

[0039] Then the optimization function defined by the LINE model is: (3) in, Represents the optimization objective function value of the LINE model. By minimizing this function, the model's embedded vector representation can be optimized. is the distance between two distributions, and KL divergence (Kullback-Leibler divergence) is often used as a distance metric. is the empirical probability distribution, Represents the joint probability distribution learned by the model, calculated by the embedding vector.

[0040] KL divergence is often used to measure the difference between two probability distributions. Therefore, the LINE model chooses to minimize the KL divergence of the two probability distributions and ignores the constant term to obtain the optimized objective function, as shown in formula (4): (4) By minimizing the objective function, the embedding vector of the vertex can be found. But the first-order similarity is only applicable to undirected graphs, not directed graphs. Therefore, the LINE model adds a second-order similarity, which is applicable not only to undirected graphs but also to directed graphs. The second-order similarity assumes that nodes with the same neighbors are similar. In this case, each vertex is regarded as a specific context information, so each vertex plays two roles, namely the vertex itself and the point as a context vertex for other vertices. For directed edges , defining a given vertex Generate context vertex The probability is: (5) in, Indicates that at a given vertex Under the condition of as the conditional probability of its context node, Represents a vertex A low-dimensional vector representation of the contextual role, Represents a vertex Embedded vector representation in low-dimensional space, Represents a vertex The context vector representation of is used for normalized calculation, representing all possible context nodes, is the number of contexts or vertices, Vertex The transpose of the low-dimensional embedding vector representation of Vertex The transpose of the context vector representation of . Similar to the first-order similarity, the objective function of the second-order similarity is: (6) in, The objective function value representing the second-order similarity is used to measure the difference between the conditional probability distribution learned by the model and the empirical distribution. For all vertices The sum is calculated, where represents a set of vertices, Used to indicate the importance of a node, which can be calculated by algorithms, such as PageRank. Represents the distance metric function between two conditional probability distributions, usually KL divergence, Represents the conditional probability distribution learned by the model, that is, at a given vertex Under the condition of , the probability distribution of other vertices as context (defined by formula (5)) is represents the empirical conditional probability distribution, that is, the conditional probability distribution calculated by normalizing the actual edge weights. The empirical distribution is defined as: (7) in, is the empirical conditional probability distribution, indicating that at a given vertex Under the condition of As the probability of the context node, is the edge weight, For Node The out-degree of , that is: in, For Node The neighbor set of . Also use KL divergence to replace the distance function, let , and ignoring the constant term, the objective function is: (8) Similar to the first-order similarity, the second-order similarity also needs to minimize the objective function. After training the LINE model using the first-order similarity and the second-order similarity respectively, the corresponding node embedding vector can be obtained. For example, the vertex The embedding vector obtained by first-order similarity calculation is , the embedding vector calculated by the second-order similarity is .Will and Concatenate them to get the final embedding vector of the node.

[0041] Step 4: Submit the obtained set of node embedding vectors to the learner clustering module 300.

[0042] In this embodiment, as described in step S400 above, the learner clustering module 300 receives the node embedding vector set submitted by the graph construction and graph embedding training module 200, and clusters the learners based on the embedding vector information using a clustering algorithm. Taking the BIRCH clustering algorithm as an example, the algorithm can use limited memory resources to complete high-quality clustering of large data sets. In BIRCH, a clustering feature (CF) tree is mainly constructed, and each node on the tree is composed of several clustering features. Each clustering feature CF can be summarized by a triple, that is, ,in is the number of data points. and The calculation formulas are as follows: (9) (10) in, express The linear sum of data points, Indicates data points, express The sum of squares of the data points.

[0043] In BIRCH, when two clusters are combined into one cluster, the clustering features can be added. For example and , when two clusters are merged, the CF of the large cluster is and The sum of Therefore, each cluster has its own CF. With CF, we can calculate the cluster centroid, cluster radius, cluster diameter, and the distance between different clusters.

[0044] The CF tree is mainly composed of three types of nodes: root nodes, internal nodes, and leaf nodes. There are three important parameters before constructing the CF tree, namely, branch balance factor , used to control the number of child nodes or branches contained in the internal node. The number of branches of all internal nodes shall not exceed Leaf balance factor , used to control the number of CFs contained in each leaf node, which cannot exceed Space threshold , used to control the maximum diameter or radius of each CF in the leaf node, which cannot exceed .

[0045] The execution process of the BIRCH algorithm is as follows: Initialize branch balance factor , leaf balance factor , spatial threshold .

[0046] Take data points from the database one by one and insert them into the CF tree. If the CF tree is empty, construct a leaf node and a sub-cluster. If the CF tree is not empty, check one by one whether the space threshold is exceeded after inserting the data point. , Leaf Balance Factor and the branch balance factor If the values ​​of the above three parameters are not exceeded, the insertion operation is performed. Otherwise, the cluster is split upward to create a new cluster to insert the data point. If the parent node is full, the cluster is split recursively until the root node.

[0047] Update the CF information of each leaf node and non-leaf node. If the node is split, insert a new tuple in the parent node and check the split until the root node.

[0048] After clustering by the clustering algorithm, the obtained clustering results are submitted to the learner portrait module 400 .

[0049] In this embodiment, as described in step S500 above, the learner portrait module 400 receives the clustering result set submitted by the learner clustering module 300. The learners in each cluster in the clustering results are portraited. The module first extracts the student nodes in each cluster, and then extracts the student's characteristic information from the student information database based on these student nodes, and extracts the student's historical learning records from the learning history database. Based on these two types of data, the learner's cognitive abilities, such as theoretical analysis ability and practical application ability, learning style, personality, leadership ability, social ability, learning investment, information technology proficiency, learning speed, etc. are portraited. Next, the module will determine which learning stage the current course learning process is in. If it is the first stage, the main data source used is the questionnaire information filled out by the learner and the historical learning records generated by his / her elective prerequisite courses. If it is the second- q stage, it will increase the number of learners in the course The historical learning data on the Internet is used as the data source. Next, we will explain in detail how to use this data to evaluate the characteristics of learners. First, use a single-valued neutral intelligence set to evaluate each student's characteristics, such as learning style, leadership ability, etc. The definition of a single-valued neutral intelligence set is as follows: set up is an object set, A single-valued neutrosophic set From 3 Function to a finite subset of the unit interval [0, 1] , and composition. It can be expressed as ,in, , and are three finite discrete value sets belonging to [0, 1], representing the degree of truth, degree of uncertainty, and degree of distortion, respectively, and satisfying , , and In particular, is a single-valued neutrosophic fuzzy number or a single-valued neutrosophic number. All subsequent expressions are expressed as single-valued neutrosophic numbers.

[0050] The following will explain in detail how to use single-valued neutroscopy to represent three types of data (questionnaire data, current course learning data, and prerequisite course historical data): (1) Regarding the questionnaire, students Filled out questionnaires, each of which can be used to calculate a learner characteristic, namely ,in, , for students Fill in The single-valued neutron number obtained from the questionnaire is , , For learners For features The degree of excellence, uncertainty and insufficiency of the learner's leadership. Taking leadership as an example, the degree of excellence means that the learner's answers to certain questions in the questionnaire can show his strong leadership, the degree of uncertainty means that the learner's answers cannot determine whether he has strong leadership, and the degree of insufficiency means that the learner's answers show his weak leadership. According to certain rules, the learner's answers on the questionnaire are divided into three categories (excellence, uncertainty and insufficiency), and the corresponding values ​​can be obtained by integrating the data of each type of answer.

[0051] (2) For the learning data of the current course, by calculating the learner The learning data generated in the current course can be used to obtain the single-valued neutroscopy number corresponding to the learner characteristics. ,in, represents the total number of learner characteristics; , which means calculating students The historical learning data of the current course is obtained The single-valued neutrino of a feature. Taking learning input as an example, learning input mainly includes six types of data: ① learning behavior data, mainly including courseware browsing records, course learning time, video viewing records, etc.; ② task completion data, mainly including homework submission records, test and examination participation, etc.; ③ interaction and communication data, mainly including forum posting records, question and answer records, etc.; ④ platform usage habit data, mainly including learning time period, learning interval, etc.; ⑤ learning effect data, mainly including various exams and test scores, etc.; ⑥ emotion and feedback data, mainly including learning satisfaction, emotional analysis results of students when learning courses, etc. Similar to the questionnaire, the data is divided into three categories according to certain rules, and the corresponding values ​​are calculated respectively.

[0052] (3) For the historical data of the prerequisite courses, similar to the learning data of the current course, The set of features calculated from the historical learning data of the elective prerequisite courses is ,in, Indicates the calculation of students The first Since there may be multiple prerequisite courses, assume that the total number of prerequisite courses is ,but Maybe indivual .

[0053] In order to make a more accurate student portrait, the single-valued neutral intelligence number of the learner characteristic index is obtained according to the cloud model theory. When there is only one prerequisite course, the current single-valued neutral intelligence number is the final single-valued neutral intelligence number; when there are multiple prerequisite courses, the cloud model is used to calculate the final single-valued neutral intelligence number. The cloud model is used to integrate the excellence, uncertainty and deficiency of multiple courses. The calculation formula is as follows: (11) in, , and For students In Features There are three numerical characteristics on it, namely expectation, entropy and super entropy. is the total number of prerequisite courses, Indicates students In Course Based on these three characteristics, the final excellence, uncertainty and insufficiency can be further calculated. The calculation formula is as follows: (12) in, Representing learners In the The value of the feature, for The influence coefficient ranges from 0.1 to 0.2.

[0054] Next, the feature evaluation values ​​calculated from the historical data are processed according to the determined learning stage. Before proceeding to the subsequent steps, the definitions of relevant symbols are given first. The learning data generated by the learner's learning of the prerequisite courses and the learning data generated by the current course are both learning history data, so historical data is used to represent the sum of these two types of data. The set of eigenvalues ​​calculated from historical learning data is ,in, is the number of learner characteristics; , which means that by calculating the learner The historical data obtained Since learner characteristics can be calculated from questionnaires and learning data, it is necessary to integrate these data to obtain the final single-valued neutron representation of learner characteristics. No. The final single-valued neutrino number of features is To express, that is .

[0055] In the first stage, if the learner has historical learning data of the prerequisite courses, .on the contrary In the 2nd q In the stage, if the learner has the historical learning data of the previous course, the single-valued neutrinometry of the features calculated by the two data is integrated using formula (13), otherwise = .

[0056] (13)

[0057] in, Representing learners In the The true degree of the single-valued neutron number on the features, Representing learners In the The uncertainty of the single-valued neutron number on the feature, Representing learners In the The degree of distortion of the single-valued neutron number on the feature, Indicates the calculation of students The first The single-valued neutrino number of features, Indicates the learning stage of the current course (such as stage 1, stage 2, etc.), represents the total number of stages of the current course learning, represents the single-valued neutron number of the feature after the learner integrates, The three components of are as follows: (14) Then, each feature can be calculated from the questionnaire information and historical learning data. Therefore, it is necessary to integrate the single-valued neutron numbers obtained from the questionnaire and the single-valued neutron numbers obtained from historical data. Different methods are used for integration according to the following situations: When the current course learning is in the first learning stage, if the learner only has questionnaire information and no historical learning data of the previous course, then the learner No. The final single-valued neutrino number of features and same, , , , .

[0058] If the learner has both questionnaire information and historical learning data of the prerequisite courses, the learner's No. The final single-valued neutrino number of features : (15) therefore, The three components of are: (16) in, and Represent different weights respectively, the default setting .

[0059] The current course is in the 2nd- q At this stage, learners No. The final single-valued neutron number of features is for: (17) (18) in, represents the initial weight, Represents the decay rate, which is a positive constant that determines the speed at which the weight decays. Indicates the current learning stage. represents the total number of learning stages of the current course. The three components of are: (19) Through the above calculations, the fuzzy measurement results of the learner in each feature can be obtained. All the calculation results are integrated and submitted to the group leader selection module 500.

[0060] In this embodiment, as described in step S600 above, the group leader selection module 500 receives the fuzzy measurement results of each learner on each feature submitted by the learner portrait module 400. The group leaders in each cluster are screened out according to the following steps. The following operations are performed for each cluster. The specific process is as follows: Figure 2 As shown: Step 1: Take out the eigenvalues ​​of the students in the cluster.

[0061] Step 2: Eliminate students who are not suitable to be group leaders according to the set threshold. Assume that the learner characteristic values ​​calculated so far are between 0 and 1. The thresholds set can be based on the characteristics that a group leader should have. For example, as a group leader, you should have certain leadership skills, so students with leadership skills below 0.5 are not suitable to be group leaders. At the same time, as a group leader, you also need to have a certain sense of responsibility, so students with conscientiousness below 0.5 are not suitable to be group leaders. By setting these threshold standards, students can be preliminarily screened and students who are not suitable to be group leaders can be eliminated. Teachers can set threshold standards according to course needs. After preliminary screening, students who meet the threshold requirements are further screened.

[0062] Step 3: There are many different course scenarios in the massive online learning environment. Take the practical application course scenario as an example. The courses in the practical application course scenario focus on examining students' practical application ability. Therefore, when selecting group leaders, learners' practical application ability should be the primary consideration. In addition, in this scenario, indicators such as learners' information technology proficiency and learning speed also need to be considered. Teachers can choose the indicators to be considered according to the specific course and set different weights.

[0063] Step 4: The competence of the remaining students as group leaders can be calculated based on the selected indicators and the set weights. The competence of the group leaders can be sorted in descending order to obtain the final group leaders.

[0064] Step 5: Put the obtained group leaders into the group leader set, and put the other students into the member set to be assigned. Each cluster will get the group leader set and the member set to be assigned, and the sets of all clusters are submitted to the group member assignment module 600 together.

[0065] In this embodiment, as described in the above steps S700 and S800, the working process of the member allocation module 600 is mainly divided into the following four steps.

[0066] Step 1: This module receives the group leader set and the set of members to be assigned submitted by the group leader selection module 500.

[0067] Step 2: Perform the following operations on each cluster to calculate the group leader in the cluster and members to be assigned The comprehensive matching degree is calculated as follows: (20) in, For the The weight value of the feature index, Indicates team leader In Features Unique-valued neutrosophic numbers on The scoring function of Representing learners In Features Unique-valued neutrosophic numbers on The scoring function is calculated as shown in formula (21). Representing learners No. The single-valued neutrino number of a feature.

[0068] (twenty one) (twenty two) After all the members to be assigned are calculated, the calculation result set (abbreviated as ) is passed to the next step.

[0069] Step 3: Receive the comprehensive matching calculation result set in the previous step , the E-CAEGO model is used to formally model and analyze the grouping problem. The E-CARGO model and objective function are shown as follows: (twenty three) (twenty four) in, It is a definition symbol, which indicates the definition or assignment of a mathematical object. Specifically, in a formula, it can indicate that the definition of a set is to contain elements , , , , , An ordered set of, in short, the meaning of this symbol is "defined as" or "assigned to", represents the problem of grouping learning groups, A set of classes representing abstract concepts, Representation and The associated object, represents a group of learning group members, represents the set of learners to be assigned, Represents a group of study groups. is the objective function value, To assign students To Group ,like , it means that the learner Assigned to study groups Otherwise, no allocation is made. is the number of learners, is the number of study groups. It is necessary to meet some general constraints, special constraints in the large-scale online learning environment, and constraints at different learning stages. These constraints are all derived from the constraint rule set. If the required constraints do not exist in the constraint rule set, you can add and supplement them yourself. The added constraints will be stored in the constraint rule set after review. General constraints include: each group has only one group leader; each learner can only be assigned to one group; there is no subjective conflict between learners; etc. Special constraints in the large-scale online learning environment include: learners have no conflicts in learning time; learners have little difference in language ability; etc. Constraints at different learning stages need to be set according to the requirements of each learning stage of the course. The calculation formulas of some constraints are shown in equations (25)-(29).

[0070] (25) (26) (27) (28) (29) in, is the number of learners, The number of study groups. Indicates The group leader of the group, formula (25) means that a group can only have one group leader, formula (26) means that each learner can only be in one group, and formula (27) means that there is no subjective conflict among learners assigned to the same group. It is a subjective conflict. Representing learners The set of available study time, Representing learners The set of languages ​​mastered, and Similarly, formula (28) When the intersection of the available time sets of learners assigned to the same group is empty, the learner and learners There will be conflicts in study time, i.e. , otherwise there is no conflict. When the intersection of the sets of languages ​​mastered by learners assigned to the same group is empty, that is, the learners and learners The language used is completely different, so learners and learners The large gap in language proficiency led to conflicts, i.e. , otherwise there is no conflict.

[0071] With the above objective function and related constraints, the IBM CPLEX optimization package can be used to solve the grouping problem. While maximizing the objective function, the corresponding constraints are met to obtain the best grouping optimization solution.

[0072] Step 4: Obtain the best grouping result for each cluster through calculation, integrate them together as the final grouping result, and feed back the obtained grouping result to students and teachers.

[0073] In view of the fact that the existing learning group construction method ignores the differences between learners in a large-scale online learning environment, does not make full use of the historical learning data accumulated by students on the online learning platform, and does not take into account that the importance of the data generated by learners in the current course will change as the course progresses, the present invention collects multiple historical learning data of students and difference data between learners, and classifies student characteristics according to the two goals in the method process of the present invention. Compared with the prior art, the present invention mines more differentiated feature data between learners in the online learning environment, such as information such as learners' beliefs and learning habits, and clusters students based on these feature data, which can ensure that there are certain similarities between learners, thereby reducing learners' learning loneliness. In addition, the present invention uses the historical data accumulated by learners on the online learning platform, such as learning time periods, learning intervals, etc. Through in-depth analysis of these data, learners' learning performance, such as learning ability and learning input, can be better displayed, thereby achieving better learning group grouping. At the same time, the present invention uses a function that changes over time to adjust the importance of the learning data generated by learners in the current course. By adjusting and updating the historical learning data, learners can be better modeled to provide support for the dynamic adjustment of learning groups. In summary, the present invention uses multi-dimensional student characteristics to profile students from different angles, and fully considers the changes in the importance of current course learning data, which can more accurately portray learners and provide strong support for forming efficient learning groups.

[0074] In order to achieve the goals of reducing learners' learning loneliness, giving full play to learners' abilities and advantages, and meeting the needs of different learning stages of the course, the present invention constructs a knowledge graph based on student feature information (such as beliefs, majors, schools, countries, time zones, learning goals and learning habits, etc.) and course information (such as exercises and knowledge points, etc.), and uses a clustering algorithm to cluster learners to ensure that there is a certain similarity between learners. Secondly, the present invention uses multiple student characteristics (such as learning style, learning investment, social skills, leadership skills, etc.) to profile students, which better portrays learners. Combining learner portraits and certain standards, the present invention selects the team leader of each group to lead the group to improve learning effects. Finally, according to the comprehensive matching degree between the team leader and the group members and the relevant constraints in the set constraint set (such as learning time, language ability constraints and constraints set at different learning stages, etc.), the group members are assigned to suitable groups. Compared with the prior art, it can not only reduce learners' learning loneliness and improve learning experience, but also fully stimulate the potential of each student, avoid conflicts between students, improve grouping effects, and provide students and teachers with better grouping optimization solutions. In addition, the present invention can adjust and update the grouping scheme according to the progress of the course to obtain an optimized grouping scheme, supporting and helping learners to efficiently complete tasks at different learning stages. Therefore, the grouping scheme provided by the present invention can improve students' learning effects, reduce teachers' teaching burden, and thus improve teaching quality and efficiency.

[0075] In order to better measure the learner's ability, the present invention uses a single-valued neutral intelligence set to measure the learner's ability. The single-valued neutral intelligence set comprehensively describes the learner's ability from three aspects: the degree of truth, the degree of uncertainty, and the degree of distortion. Compared with the prior art, the single-valued neutral intelligence set can better reflect the uncertainty and volatility of the learner's ability, which is consistent with the learner's ability level and its changes in the real learning process, thereby providing strong support for the formation of learning groups.

[0076] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0077] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0078] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0079] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it 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 all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.

[0080] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A dynamic grouping optimization method for large-scale online collaborative learning, characterized in that: The steps include: Collect questionnaire information filled out by students after logging in, course-related information, and students' behavior data during the learning process from the online learning platform, and store them in the student information database, course information database, and learning history database respectively; Based on the behavioral data in the student information database and the course information database, a knowledge graph reflecting the student characteristic information and the course characteristic information is constructed; Apply graph embedding algorithm to learn the features of nodes in the knowledge graph and generate node embedding vectors; Clustering the node embedding vectors using a clustering algorithm to generate preliminary results for learner grouping; Based on the clustering results and the data in the student information database and the learning history database, the single-valued neutral intelligence set is used to characterize and describe the learner's ability and generate a learner portrait for each learner; Select learners who are suitable to serve as group leaders based on the learner portrait and the preset group leader screening criteria; Based on the comprehensive matching degree between team members and team leaders and the grouping constraint rules, an optimization algorithm is used to allocate team members; Generate learning groups based on the optimized allocation results, and dynamically adjust the groupings during the learning process.

2. The dynamic grouping optimization method for large-scale online collaborative learning according to claim 1, characterized in that: The construction methods of knowledge graph include: Extracting students' characteristic information from the student information database; Extracting characteristic information of prerequisite courses from a course information database; Design an ontology model based on the characteristic information of students and prerequisite courses; According to the ontology model, students, courses and their characteristic information are organized into nodes of the knowledge graph, and the relationship between nodes is represented as edges to construct a preliminary knowledge graph; Based on the entities and their relationships in the preliminary knowledge graph, the knowledge graph is improved in combination with weight information to obtain the final knowledge graph.

3. The dynamic grouping optimization method for large-scale online collaborative learning according to claim 1, characterized in that: Apply graph embedding algorithms to learn features of nodes in knowledge graphs and generate node embedding vectors, including the following methods: Based on the constructed knowledge graph, determine the relationship weights between nodes in the graph, wherein the nodes include student characteristics, prerequisite courses, exercises, and knowledge points, and the relationship weights are used to represent the strength of association between nodes; Perform first-order and second-order similarity optimization on the node relationships in the knowledge graph. The first-order similarity optimizes the similarity of low-dimensional embedding vectors through directly connected nodes; the second-order similarity optimizes the similarity of low-dimensional embedding vectors through nodes with common neighbors. Based on the first-order similarity optimization and the second-order similarity optimization, the corresponding node embedding vectors are trained respectively, and the two node embedding vectors are fused to generate the final embedding vector of the node.

4. The dynamic grouping optimization method for large-scale online collaborative learning according to claim 1, characterized in that: Methods for clustering node embedding vectors using clustering algorithms and generating preliminary results for learner grouping include: Based on the requirements of the clustering algorithm, the branch balance factor, leaf balance factor and spatial threshold of the clustering feature tree are set to control the number of branches of internal nodes, the number of clustering features of leaf nodes and the maximum value of the diameter or radius of each clustering feature in the leaf node. Insert the received node embedding vectors into the cluster feature tree one by one. If the cluster feature tree is empty, construct a leaf node and a subcluster, and insert the data point into it. If the cluster feature tree already has nodes, check whether the inserted data points exceed the spatial threshold, leaf balance factor, and branch balance factor one by one. If none of them exceed, complete the insertion operation. If any parameter value is exceeded, perform node splitting and cluster splitting operations according to the specific situation to create new nodes and clusters. Update the clustering feature information of each leaf node and non-leaf node in the clustering feature tree; after the node splits or the cluster splits, recursively update the newly generated node information upward to the parent node until the root node; The initial grouping of learners is completed through the construction of a clustering feature tree.

5. The dynamic grouping optimization method for large-scale online collaborative learning according to claim 1, characterized in that: Based on the clustering results combined with the data in the student information database and the learning history database, the single-valued neutral intelligence set is used to characterize and describe the learner's ability. The method of generating a learner portrait for each learner includes: Extract the learner node of each cluster from the learner clustering result set; According to the extracted learner node, the corresponding student feature information is obtained from the student information database, and the corresponding learning history record is obtained from the learning history database; According to the current course learning stage, determine the relevant data sources for calculating learner characteristics. If it is in the first stage, the questionnaire data and the learning records of the prerequisite courses are used first; if it is in the second to third stage, the questionnaire data and the learning records of the prerequisite courses are used first. q The second stage is based on the aforementioned data combined with the learning data of the current course; For each learner feature, relevant data is extracted from the student information database and classified, and a single-valued neutral intelligence number of each feature is generated according to the rules, wherein the single-valued neutral intelligence number includes the degree of truth, the degree of uncertainty and the degree of distortion; For multi-dimensional data sources, a fusion strategy is used to integrate them and generate a comprehensive single-valued neutrino data of learner characteristics. The fusion strategy includes but is not limited to cloud model, weight distribution model and stage decay model. In different learning stages, the weight distribution of the single-valued neutrinometry of each feature is dynamically adjusted according to the availability of relevant data sources, and the final single-valued neutrinometry of the feature is calculated through the integration formula; The final single-valued neutrino data of learner characteristics are converted into structured portrait data, and the learner portrait based on the final single-valued neutrino data is output.

6. The dynamic grouping optimization method for large-scale online collaborative learning according to claim 5, characterized in that: Methods for selecting learners who are suitable to serve as group leaders include: Receiving the learner portrait based on the final single-valued neutrino-intelligence number, extracting the feature value set of the students in each cluster; The feature value set is screened according to the preset threshold standard, and students who do not meet the threshold requirements are eliminated; Determine the characteristic indicators and their corresponding weights required for selecting group leaders based on the specific course scenario; Based on the characteristic values ​​of the screened students, the selected indicators and their weights, the competence of each student as a team leader is calculated; Sort the calculated competence of the group leader in descending order, and select the student with the highest competence in the sorted results as the group leader; The obtained group leader number is stored in the group leader set, and the remaining student numbers are stored in the member to be assigned set; Submit the group leader set and the member set to be assigned of all clusters to the group member assignment module.

7. The dynamic grouping optimization method for large-scale online collaborative learning according to claim 1, characterized in that: Based on the comprehensive matching degree between team members and team leaders and the grouping constraint rules, the allocation method of using the optimization algorithm to allocate team members includes: Receive a set of team leaders and a set of members to be assigned; For the team leader and the member to be assigned in each cluster, the comprehensive matching degree is calculated based on the eigenvalues; According to the calculation results of the comprehensive matching degree, the grouping problem is formally modeled based on the E-CARGO model, and the objective function and constraints are set; Use the optimization algorithm to solve the objective function and obtain the best grouping result for each cluster while satisfying the constraints; Integrate the grouping results of all clusters into the final grouping solution; Provide feedback to students and teachers on the final grouping plan.

8. The dynamic grouping optimization method for large-scale online collaborative learning according to claim 1, characterized in that: When the optimization algorithm is adjusted according to the dynamic grouping requirements, the following steps are also included: According to the teaching objectives of the current learning stage, dynamically adjust the grouping constraint rules and add or modify stage-specific constraints; Evaluate the grouping results of each learning stage. The evaluation indicators include the collaboration effect among group members, the quality of group task completion, and learners' satisfaction feedback; According to the evaluation results, optimize the parameters and weight allocation rules of the grouping model, and update the calculation formula of the comprehensive matching degree; The optimized grouping model is applied to the grouping adjustment process in the next stage.

9. A dynamic grouping optimization device for large-scale online collaborative learning, characterized in that: Includes the following modules: The data collection and processing module is used to collect questionnaire information, course-related information and students' learning behavior data from the online learning platform and store them in the database module, where the database module includes a student information database, a course information database and a learning history database, which are used to store the collected relevant data respectively; Graph construction and graph embedding training modules, where the graph construction module is used to construct a knowledge graph reflecting student feature information and course feature information based on the behavioral data in the student information database and the course information database; the graph embedding training module is used to perform feature learning on nodes in the knowledge graph and generate node embedding vectors; Learner clustering module, used to cluster node embedding vectors and generate preliminary results of learner grouping; The learner portrait module is used to generate learner portraits based on clustering results, combined with student information and learning history, using single-valued neutral intelligence numbers to comprehensively characterize learner abilities from multiple angles; The group leader selection module is used to select learners who are suitable to serve as group leaders based on learner portraits and group leader screening criteria; The team member allocation module is used to optimize the allocation of team members based on the comprehensive matching degree between team members and team leaders and group constraint rules.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the dynamic grouping optimization method for large-scale online collaborative learning according to any one of claims 1 to 8 are executed.

Images