Job arrangement-oriented ai agent collaboration method and system

Abstract

The present application relates to the technical field of intelligent education, and more particularly to an AI intelligent agent collaboration method and system for homework arrangement. The method comprises: acquiring multi-modal learning behavior data of a target learner, and constructing a weighted three-part association graph containing a learner, a knowledge point and a homework content node; configuring multiple collaborative artificial intelligence agents to perform similarity propagation on the corresponding subgraph to generate an association strength matrix; obtaining a comprehensive association strength matrix through linear weighted fusion by a collaborative aggregation module; sorting and recommending entries in a homework library based on the comprehensive matrix, and updating the graph by writing back the interaction feedback. The method improves the personalization and accuracy of homework recommendation.

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, and in particular to an AI intelligent agent collaborative method and system for homework assignment. Background Technology

[0002] In the current field of intelligent education, personalized homework assignment recommendations are typically based on learners' historical grades, knowledge mastery levels, or behavioral log data, employing methods such as collaborative filtering, knowledge graph reasoning, or deep learning models. The conventional approach is to construct a single-dimensional learner-assignment association model, such as a recommendation system based on matrix factorization or graph attention networks, which extracts learner features and assignment attributes to calculate similarity or predict scores. This single-task model often projects different factors such as knowledge point mastery, assignment difficulty, and learner preferences into the same feature space, relying on large-scale labeled data or predefined rule templates for recall and ranking. Furthermore, some systems introduce knowledge graphs, associating learners and assignment nodes through knowledge points; however, the edge weights in the graph are often determined by static expert annotations or simple statistical frequencies, lacking the ability to adaptively adjust in real-time to interactive feedback.

[0003] These existing solutions have two significant shortcomings. First, multimodal learning behavior data is not fully integrated and utilized. Traditional approaches typically only process single-modal information such as clickstreams, answer accuracy, or assignment completion time, ignoring the potential correlations in unstructured data such as voice, eye tracking, and note logs. This results in an insufficiently refined characterization of the match between learners' cognitive states and assignments. Second, the recommendation process lacks a multi-dimensional collaborative mechanism. Conventional knowledge graph recommendation systems often only consider the propagation of single relationships between learners and knowledge points or learners and assignments, failing to capture the three-way interaction between learners and knowledge points, knowledge points and assignments, and learners and assignments simultaneously. This single-path similarity calculation is prone to the cold-start problem. When new assignments or knowledge points are added, the recommendation results highly depend on existing sparse interaction records, making it difficult to meet the needs of personalized dynamic adjustments. When faced with continuous updates to behavioral data within the learning cycle, existing methods usually require offline retraining of the entire model, making it impossible to incrementally optimize the recommendation sequence online. This causes the system to lag in responding to learners' real-time states, ultimately affecting the relevance of assignments and learning effectiveness. Summary of the Invention

[0004] The embodiments of the present invention provide an AI intelligent agent collaboration method and system for task deployment, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides an AI agent collaboration method for job assignment, comprising: Acquire multimodal learning behavior data generated by the target learner within a preset learning period, perform feature encoding and relation extraction processing on the multimodal learning behavior data, construct a three-part association graph containing learner nodes, knowledge point nodes and assignment content nodes, and calculate the weight of the association edges between each node based on historical interaction data to form a weighted three-part association graph. Configure at least three collaborative AI agents corresponding to the learner-knowledge point association dimension, the knowledge point-assignment content association dimension, and the learner-assignment content association dimension, respectively. Each collaborative AI agent performs similarity propagation calculation on the corresponding subgraph of the weighted tripartite association graph and independently generates the association strength matrix of the corresponding dimension. The association strength matrix generated by each collaborative AI agent is input into the collaborative aggregation module. The linear weighted aggregation mechanism is used to perform element-wise weighted fusion of each association strength matrix to generate a comprehensive association strength matrix. Based on the comprehensive association strength matrix, a comprehensive association score is calculated for each candidate task in the pre-constructed task content library. Candidate task items that meet the preset association threshold are selected, and personalized task recommendation sequences are generated by sorting them in descending order of comprehensive association score. The interactive feedback data of the target learner on the recommendation sequence is written back to the weighted tripartite association graph to update the association edge weights.

[0006] The multimodal learning behavior data is processed by feature encoding and relation extraction to construct a three-part association graph containing learner nodes, knowledge point nodes, and assignment content nodes. Based on historical interaction data, the weights of the association edges between nodes are calculated to form a weighted three-part association graph, including: Learner identifiers, interacted knowledge point identifiers, and accessed assignment content identifiers are extracted from the multimodal learning behavior data to construct learner node sets, knowledge point node sets, and assignment content node sets. Establish learner-knowledge point association edges based on the learner's historical interaction records with knowledge points; establish knowledge point-assignment content association edges based on the coverage mapping relationship of knowledge points in assignment content; and establish learner-assignment content association edges based on the learner's historical completion records with assignment content. The first weight of the learner-knowledge point association edge is calculated based on the interaction frequency and interaction accuracy. The second weight of the knowledge point-assignment content association edge is calculated based on the coverage ratio of the knowledge point in the assignment content. The third weight of the learner-assignment content association edge is calculated based on the assignment completion quality score. The set of learner nodes, the set of knowledge point nodes, the set of assignment content nodes, and the three types of associated edges carrying the first weight, the second weight, and the third weight respectively are summarized to form the weighted ternary association graph.

[0007] Each collaborative AI agent performs similarity propagation calculations on the corresponding subgraph of the weighted tripartite association graph, independently generating an association strength matrix for the corresponding dimension, including: The first collaborative AI agent corresponding to the learner-knowledge point association dimension takes the bipartite subgraph formed by the learner node set and the knowledge point node set in the weighted tripartite association graph as the operation object, and uses the learner-knowledge point association edge weight as the initial propagation value to perform similarity propagation calculation and generate a learner-knowledge point association strength matrix. The second collaborative AI agent, corresponding to the knowledge point-task content association dimension, takes the bipartite subgraph formed by the knowledge point node set and the task content node set in the weighted tripartite association graph as the operation object, and uses the knowledge point-task content association edge weight as the initial propagation value to perform similarity propagation calculation and generate a knowledge point-task content association strength matrix. The third collaborative AI agent corresponding to the learner-assignment content association dimension takes the bipartite subgraph formed by the learner node set and the assignment content node set in the weighted tripartite association graph as the operation object, and uses the learner-assignment content association edge weight as the initial propagation value to perform similarity propagation calculation and generate a learner-assignment content association strength matrix.

[0008] The similarity propagation calculation includes: The edge weight matrix of the corresponding bipartite subgraph is used as the initial association strength matrix for propagation, and the similarity propagation iterative process is initiated. In each round of propagation iteration, the association strength matrix of the current round is used as the left multiplication matrix, and the edge weight matrix of the corresponding bipartite subgraph is used as the right multiplication matrix. Matrix multiplication is performed so that the association strength of the current round is propagated to the adjacent nodes along the node association direction represented by the edge weights in the bipartite graph, and the association strength matrix of the next round is obtained. Perform row normalization on each row of the next round association strength matrix to normalize the sum of the values ​​of each row to a preset constant value, and obtain the row normalized next round association strength matrix; The association strength matrix of the next round after row normalization is weighted and superimposed with the association strength matrix of the current round according to a preset retention ratio to obtain the propagation result matrix of this round. The preset retention ratio controls the information retention weight of the association strength of the current round in the superimposed result. Calculate the matrix norm difference between the current propagation result matrix and the previous propagation result matrix. When the matrix norm difference is lower than a preset convergence threshold, terminate the propagation iteration and output the current propagation result matrix as the final state association strength matrix of the corresponding dimension of the collaborative AI agent.

[0009] The correlation strength matrices are fused element-wise using a linear weighted aggregation mechanism to generate a comprehensive correlation strength matrix, including: Obtain a preset set of dimension weight coefficients, which includes a first weight coefficient, a second weight coefficient, and a third weight coefficient corresponding to the learner-knowledge point association dimension, the knowledge point-assignment content association dimension, and the learner-assignment content association dimension, respectively. Each dimension weight coefficient is non-negative and the sum of the three is a preset normalization constant. Perform matrix dimension alignment processing on each association strength matrix to make the learner-knowledge point association strength matrix, the knowledge point-assignment content association strength matrix, and the learner-assignment content association strength matrix have the same number of matrix rows and columns; For each element position in the three association strength matrices that have the same row index and column index, the corresponding element in the learner-knowledge point association strength matrix is ​​weighted by the first weight coefficient, the corresponding element in the knowledge point-task content association strength matrix is ​​weighted by the second weight coefficient, and the corresponding element in the learner-task content association strength matrix is ​​weighted by the third weight coefficient. The three weighted results are summed to obtain the element value of that position in the comprehensive association strength matrix. The comprehensive correlation strength matrix is ​​normalized element-wise to map all element values ​​to a preset value range, resulting in a normalized comprehensive correlation strength matrix.

[0010] A comprehensive association score is calculated for each candidate assignment in the pre-built assignment content library. Candidate assignments that meet a preset association threshold are selected, and personalized assignment recommendation sequences are generated by sorting them in descending order of their comprehensive association scores. The interactive feedback data of the target learner on the recommended sequences is then written back to the weighted tripartite association graph to update the association edge weights, including: For each candidate task in the task content library, extract all association strength values ​​of the candidate task in the row corresponding to the target learner from the comprehensive association strength matrix, and perform a weighted summation operation on all extracted association strength values ​​to obtain the comprehensive association score of the candidate task. All candidate job items are sorted in descending order of comprehensive correlation score. A preset number of candidate job items with comprehensive correlation scores higher than a preset cutoff threshold are selected to generate the personalized job recommendation sequence. Obtain the completion rate and accuracy data of the target learner for each recommended task in the personalized task recommendation sequence, and calculate the real-time weight adjustment amount for this interaction based on the completion rate and accuracy data; The instantaneous weight adjustment amount and the historical weight of the corresponding associated edge in the weighted three-part association graph are weighted and merged according to a preset attenuation coefficient, and the weight of the corresponding associated edge is updated with the weighted merging result.

[0011] The instantaneous weight adjustment amount and the historical weights of the corresponding associated edges in the weighted three-part association graph are weighted and merged according to a preset attenuation coefficient, and the weights of the corresponding associated edges are updated with the weighted merging result, including: The instant weight adjustment amount is calculated and is proportional to the product of the completion data and the accuracy data. When the completion data is lower than a preset completion threshold or the accuracy data is lower than a preset accuracy threshold, the instant weight adjustment amount is negative. Obtain the time decay factor corresponding to the associated edge. The time decay factor is determined by a preset decay function based on the time interval between the most recent update time of the associated edge and the current time. The larger the time interval, the closer the value of the time decay factor is to zero, so that the weight of the associated edge that has not been updated for a long time will naturally decay over time. Multiply the historical weight by the time decay factor to obtain the decayed historical weight of the associated edge; The historical weights after decay are summed with the instantaneous weight adjustment to obtain the updated associated edge weights. The updated associated edge weights are then written back to the weight storage location of the associated edge in the weighted three-part association graph, thus completing this feedback-driven closed-loop update of edge weights.

[0012] A second aspect of the present invention provides an AI intelligent agent collaborative system for task deployment, comprising: The graph building unit is used to acquire multimodal learning behavior data generated by the target learner within a preset learning period, perform feature encoding and relation extraction processing on the multimodal learning behavior data, construct a three-part association graph containing learner nodes, knowledge point nodes and assignment content nodes, and calculate the weight of the association edges between each node based on historical interaction data to form a weighted three-part association graph. The collaborative computing unit is configured with at least three collaborative AI agents corresponding to the learner-knowledge point association dimension, the knowledge point-assignment content association dimension, and the learner-assignment content association dimension, respectively. Each collaborative AI agent performs similarity propagation calculation on the corresponding subgraph of the weighted tripartite association graph and independently generates the association strength matrix of the corresponding dimension. The weighted aggregation unit is used to input the association strength matrix generated by each collaborative AI agent into the collaborative aggregation module, and perform element-wise weighted fusion of each association strength matrix through a linear weighted aggregation mechanism to generate a comprehensive association strength matrix. The recommendation update unit is used to calculate the comprehensive association score of each candidate task in the pre-constructed task content library based on the comprehensive association strength matrix, filter the candidate task tasks that meet the preset association threshold, generate a personalized task recommendation sequence by sorting them in descending order of comprehensive association score, and write back the interactive feedback data of the target learner to the weighted tripartite association graph to update the association edge weights.

[0013] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] By fully leveraging multimodal learning behavior data to construct a weighted tripartite association graph, the complex relationships between learners, knowledge points, and assignment content are presented in a quantitative form. Edge weights are calculated based on historical interaction data, ensuring the graph structure accurately reflects the actual learning status and providing a high-fidelity information foundation for subsequent personalized recommendations. This multi-dimensional association modeling method significantly improves the targeting of assignment recommendations, effectively identifying learners' knowledge weaknesses and strengths, and avoiding recommendation biases caused by traditional single data sources.

[0016] Three collaborative AI agents independently handle the similarity propagation calculations for the learner-knowledge point, knowledge point-assignment content, and learner-assignment content subgraphs. Each agent focuses on its own association dimension, generating its own association strength matrix through parallel computation to avoid interference and confusion between different dimensions. This division of labor and collaboration mechanism significantly enhances the robustness of the recommendation results; even when data is sparse in one dimension, information from other dimensions can still ensure the stability and reliability of the recommendations.

[0017] The collaborative aggregation module fuses the three association strength matrices element-wise using a linear weighting mechanism to form a comprehensive strength matrix reflecting global associations. This weighted fusion process allows for flexible adjustment of the contribution of different dimensions, adapting to various learning scenarios and individual differences. Assignment items are selected based on the comprehensive association strength matrix, ensuring that the recommended sequence balances knowledge coverage with matching the learner's current ability level, achieving truly personalized assignments.

[0018] Interactive feedback data is written back to the weighted association graph in real time and the edge weights are dynamically updated, enabling the recommendation system to have self-learning and adaptive capabilities. As the learning process progresses, the system continuously optimizes its understanding of the learner's cognitive trajectory, and the recommendation effect exhibits a spiral evolution of continuous improvement. This method completely changes the traditional fixed pattern of homework assignment, dynamically generating the most suitable combination of assignments for each learner, significantly improving learning efficiency and depth of knowledge mastery. Attached Figure Description

[0019] Figure 1A flowchart illustrating the collaborative method of AI agents for task assignment. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0022] Figure 1 This is a flowchart illustrating the AI ​​agent collaboration method for task assignment according to an embodiment of the present invention.

[0023] AI agent collaboration methods for task assignment include: Acquire multimodal learning behavior data generated by the target learner within a preset learning period, perform feature encoding and relation extraction processing on the multimodal learning behavior data, construct a three-part association graph containing learner nodes, knowledge point nodes and assignment content nodes, and calculate the weight of the association edges between each node based on historical interaction data to form a weighted three-part association graph. Configure at least three collaborative AI agents corresponding to the learner-knowledge point association dimension, the knowledge point-assignment content association dimension, and the learner-assignment content association dimension, respectively. Each collaborative AI agent performs similarity propagation calculation on the corresponding subgraph of the weighted tripartite association graph and independently generates the association strength matrix of the corresponding dimension. The association strength matrix generated by each collaborative AI agent is input into the collaborative aggregation module. The linear weighted aggregation mechanism is used to perform element-wise weighted fusion of each association strength matrix to generate a comprehensive association strength matrix. Based on the comprehensive association strength matrix, a comprehensive association score is calculated for each candidate task in the pre-constructed task content library. Candidate task items that meet the preset association threshold are selected, and personalized task recommendation sequences are generated by sorting them in descending order of comprehensive association score. The interactive feedback data of the target learner on the recommendation sequence is written back to the weighted tripartite association graph to update the association edge weights.

[0024] The multimodal learning behavior data is processed by feature encoding and relation extraction to construct a three-part association graph containing learner nodes, knowledge point nodes, and assignment content nodes. Based on historical interaction data, the weights of the association edges between nodes are calculated to form a weighted three-part association graph, including: Learner identifiers, interacted knowledge point identifiers, and accessed assignment content identifiers are extracted from the multimodal learning behavior data to construct learner node sets, knowledge point node sets, and assignment content node sets. Establish learner-knowledge point association edges based on the learner's historical interaction records with knowledge points; establish knowledge point-assignment content association edges based on the coverage mapping relationship of knowledge points in assignment content; and establish learner-assignment content association edges based on the learner's historical completion records with assignment content. The first weight of the learner-knowledge point association edge is calculated based on the interaction frequency and interaction accuracy. The second weight of the knowledge point-assignment content association edge is calculated based on the coverage ratio of the knowledge point in the assignment content. The third weight of the learner-assignment content association edge is calculated based on the assignment completion quality score. The set of learner nodes, the set of knowledge point nodes, the set of assignment content nodes, and the three types of associated edges carrying the first weight, the second weight, and the third weight respectively are summarized to form the weighted ternary association graph.

[0025] After obtaining the raw multimodal learning behavior data, it needs to be structured to construct a ternary association graph that supports collaborative agent computation. Multimodal learning behavior data typically includes various activity records of learners on online learning platforms, such as video viewing time, exercise answer records, knowledge point browsing paths, assignment submission timestamps, and discussion forum interactions. This data is scattered across different tables and needs to be uniformly organized using data fusion techniques.

[0026] In the data preprocessing stage, the collected multimodal learning behavior data is first cleaned, removing records with more than 30% missing values. The timestamp format is standardized, and key features of unstructured text content (such as discussion thread content) are extracted using natural language processing techniques. For learner identification extraction, a unique user ID is read from the user login log table, and static attributes such as grade and subject preference are obtained by linking it to the user basic information table. For the extraction of interacted knowledge point identifiers, the learner's video viewing records and exercise answer records are traversed. The video viewing records contain knowledge point tags corresponding to the videos, and the exercise answer records contain knowledge point metadata associated with the questions. By deduplicating the knowledge point fields of these records, the set of knowledge points actually encountered by each learner within the preset learning cycle can be obtained. For the extraction of accessed assignment content identifiers, the assignment numbers completed by the learner are read from the assignment submission record table, and assignment numbers that the learner viewed but did not submit are extracted from the assignment browsing log. The two sets are merged and deduplicated to form the learner's assignment access set.

[0027] After extracting the identifiers, three types of node sets are constructed. The learner node set is denoted as... ,in This represents the number of active learners in the system. The set of knowledge point nodes is denoted as... ,in This represents the total number of knowledge points covered in the curriculum. The set of assignment content nodes is denoted as... ,in This indicates the number of all candidate assignments in the assignment content library. In addition to carrying a basic identifier, each node also needs to be attached with an attribute vector. For example, learner nodes are attached with an assessment value of their learning ability level, knowledge point nodes are attached with a difficulty coefficient label, and assignment content nodes are attached with metadata such as question type and estimated completion time.

[0028] After the node set is constructed, three types of connection edges need to be established based on different data sources and association logic. For the establishment of learner-knowledge point connection edges, the historical interaction record table is traversed to find each learner's learning behavior records for each knowledge point. These records include the completion rate of the video corresponding to the knowledge point, the number of times related practice questions were answered, and the number of correct answers. If the learner... Knowledge points If there is at least one interaction, then an association edge is established between the two. To establish the association between knowledge points and assignment content, it is necessary to analyze the knowledge point coverage structure of the assignment content. Each assignment typically consists of multiple questions, and each question is associated with a specific knowledge point tag. By parsing the question-knowledge point mapping relationship in the assignment metadata, the knowledge points can be statistically analyzed. In the content of the assignment The number of questions covered. If knowledge points... Homework If at least one question in the problem is related, then an association edge is established. To establish the learner-assignment content association, query the learner's assignment completion history table and extract the learner's information. Homework The submission history. A link is established only when a learner has actually completed and submitted an assignment. No direct association edges will be established for browsing behavior that has not been submitted.

[0029] After establishing the association edges, the weight value of each edge needs to be calculated to quantify the association strength. For the first weight calculation of learner-knowledge point association edges, two dimensions are considered: interaction frequency and interaction accuracy. Interaction frequency reflects the learner's level of engagement in learning the knowledge point, while interaction accuracy reflects the learner's mastery level of the knowledge point. Let the learner... Knowledge points The total number of interactions is The correct number of interactions is The accuracy rate is then calculated as follows: To balance the influence of both factors, the first weight is calculated using a weighted harmonic mean. ,in The maximum number of interactions among all learners with all knowledge points, used to normalize the frequency metric. and These are hyperparameters, typically set and It places more emphasis on the degree of mastery.

[0030] The calculation of the second weight for the edge linking knowledge points and assignment content is mainly based on the coverage percentage of the knowledge point in the assignment. Let the assignment... The total number of questions included is The knowledge points involved The number of questions is Then the knowledge points In the homework The coverage ratio in Considering the varying importance of knowledge points within the curriculum, a knowledge point importance coefficient is introduced. As a moderating factor, this coefficient can be obtained through the hierarchical position of the knowledge point in the course syllabus or the teacher's annotation. The final second weight is calculated as follows: This ensures that the edge weights corresponding to the core knowledge points are increased.

[0031] The calculation of the third weight for the learner-assignment content association edge relies on the assignment completion quality score. The quality score is a comprehensive indicator, including factors such as assignment score, submission timeliness, and answer format. Let the learner... Complete the assignment The original score obtained is (Maximum score is 100), the number of days before the deadline for submission is: The timeliness reward coefficient is then calculated as follows: Early submissions will receive up to a 7% bonus. The correctness of the answer format will be scored by the machine-graded module. (Values ​​range from 0 to 1) are quantified. Taking into account the above factors, the third weight is calculated as follows: The higher the value, the better the learner's completion of the assignment, which reflects the degree of matching between the learner and the assignment content.

[0032] After calculating all edge weights, the three types of node sets will be... , , and each carrying weight , , The three types of edge sets are integrated into a data structure. A weighted ternary association graph is constructed using adjacency lists or sparse matrices. ,in , , These represent the three types of edge sets, respectively. In actual storage, three independent sparse matrices are typically used to represent the three types of edge relationships: the learner-knowledge point association matrix. for 3D matrix, knowledge point-assignment content association matrix for 3D matrix, learner-assignment content association matrix for A three-dimensional matrix. The index of a non-zero element in the matrix indicates the existence of an associated edge, and the element value is the weight of that edge. To facilitate efficient computation by the collaborative agent, a fast lookup structure with row and column indices is established for each of the three matrices, supporting the rapid location of all associated nodes and their weights by dimension (learner, knowledge point, or assignment content).

[0033] After constructing the three-part association graph, the graph structure needs to be validated for rationality. This involves checking for isolated nodes (nodes without any connecting edges). If isolated nodes exist, it indicates missing data collection or that the node did not generate effective interaction within the preset learning period. Based on business rules, it's necessary to decide whether to remove or add default weights. Next, check if the distribution of edge weights meets expectations. Calculate the mean and standard deviation of the three types of edge weights. If an extreme skewed distribution of edge weights occurs, it may be necessary to adjust the parameter configuration in the weight calculation formula. Finally, use graph topology analysis algorithms to calculate the graph's connectivity index, ensuring the graph structure can support subsequent similarity propagation calculations and avoiding numerous broken subgraphs that could prevent effective collaboration between the agents.

[0034] Each collaborative AI agent performs similarity propagation calculations on the corresponding subgraph of the weighted tripartite association graph, independently generating an association strength matrix for the corresponding dimension, including: The first collaborative AI agent corresponding to the learner-knowledge point association dimension takes the bipartite subgraph formed by the learner node set and the knowledge point node set in the weighted tripartite association graph as the operation object, and uses the learner-knowledge point association edge weight as the initial propagation value to perform similarity propagation calculation and generate a learner-knowledge point association strength matrix. The second collaborative AI agent, corresponding to the knowledge point-task content association dimension, takes the bipartite subgraph formed by the knowledge point node set and the task content node set in the weighted tripartite association graph as the operation object, and uses the knowledge point-task content association edge weight as the initial propagation value to perform similarity propagation calculation and generate a knowledge point-task content association strength matrix. The third collaborative AI agent corresponding to the learner-assignment content association dimension takes the bipartite subgraph formed by the learner node set and the assignment content node set in the weighted tripartite association graph as the operation object, and uses the learner-assignment content association edge weight as the initial propagation value to perform similarity propagation calculation and generate a learner-assignment content association strength matrix.

[0035] After obtaining the weighted tripartite graph, the three collaborative AI agents each perform similarity propagation calculations on their respective bipartite subgraphs, mining implicit relationships between nodes from different association dimensions. Each collaborative AI agent adopts a graph propagation-based collaborative filtering mechanism to diffuse explicit edge weight information to indirectly associated node pairs through multi-hop propagation paths, thereby overcoming the limitations of sparse direct connections and generating more comprehensive association strength assessment results.

[0036] The first collaborative AI agent is responsible for handling the learner-knowledge point association dimension. This agent extracts a bipartite subgraph from the weighted tripartite association graph, consisting of the learner node set and the knowledge point node set. This subgraph retains only the edges between learner nodes and knowledge point nodes, removing all connections related to task content nodes. The extracted bipartite subgraph can be represented as follows: ,in Represents the set of learner nodes. Represents a set of knowledge point nodes. This represents the set of edges linking learners and knowledge points. Each edge... Carrying weight As the initial propagation value for similarity propagation, this weight has taken into account information from both the frequency of interaction and the accuracy of interaction.

[0037] When performing similarity propagation calculations on the bipartite subgraph, the first collaborative AI agent constructs a propagation mechanism based on the idea of ​​collaborative filtering. For any learner node... and knowledge point nodes Even if there is no direct connection between them, a relationship can be established through the propagation path of the intermediate node. Specifically, learners Other learners If multiple knowledge points exhibit similar interaction patterns, then they share similar learning preferences; knowledge points Other knowledge points If learners frequently interact with each other within the same group, their content attributes are likely related. This two-way communication mechanism allows for the prediction of learner dynamics. Knowledge points that have not been directly interacted with The degree of potential mastery.

[0038] The propagation calculation is performed using a matrix iteration approach. First, an initial correlation strength matrix is ​​constructed. , of which elements equal to edge weight If an edge exists, the value is zero; otherwise, it is zero. Then, an iterative formula is used. To carry out dissemination and updates, among which The similarity matrix between learners This is a similarity matrix between knowledge points. and For the propagation coefficient, For each iteration round. Elements of the learner similarity matrix. By calculating learners and The overlap of the interactive knowledge point sets is determined using Jaccard similarity or cosine similarity as a metric. The elements of the knowledge point similarity matrix are... By calculating knowledge points and The overlap of the learner set is determined. After multiple iterations until convergence or a preset iteration limit is reached, the final learner-knowledge point association strength matrix is ​​obtained. Each element of this matrix reflects the predicted strength of the learner's overall mastery of the corresponding knowledge point.

[0039] The second collaborative AI agent is responsible for handling the knowledge point-task content association dimension. This agent extracts a bipartite subgraph from the weighted ternary association graph, consisting of the knowledge point node set and the task content node set, denoted as... ,in Represents the set of task content nodes. This represents the set of edges that link knowledge points to assignment content. Each edge... weight As the initial propagation value, this weight already reflects the coverage and importance of knowledge points in the assignment.

[0040] When performing similarity propagation calculations on the bipartite subgraph, the second collaborative AI agent employs a similar collaborative filtering propagation mechanism. For any knowledge point... and homework content Even if there is no direct overlap between them, a relationship can be established through co-occurrence patterns. (If knowledge points...) With knowledge points If assignments frequently appear in the same set of assignments, then they have a symbiotic relationship in terms of content structure; if assignments With homework If two or more knowledge points are distributed similarly, they are similar in their assessment objectives. This dissemination mechanism can provide richer evidence of knowledge point connections for recommending assignment content.

[0041] The propagation calculation also employs a matrix iteration method. An initial correlation strength matrix is ​​constructed. , of which elements equal to edge weight If an edge exists, the value is zero; otherwise, it is zero. This is determined using an iterative formula. To carry out dissemination and updates, among which The similarity matrix between knowledge points (which can be compared with the similarity matrix in the first intelligent agent) (Use the same calculation method or recalculate based on the job co-occurrence pattern) This is a similarity matrix between the assignment content. and The propagation coefficient. Elements of the similarity matrix between task content. By calculating the job and The overlap of the included knowledge point sets is determined, and the comprehensive similarity is calculated by combining multi-dimensional features such as question difficulty distribution and question type distribution. After iterative convergence, the final knowledge point-assignment content association strength matrix is ​​obtained. Each element of this matrix reflects the overall correlation strength between the corresponding knowledge point and the corresponding homework content.

[0042] The third collaborative AI agent is responsible for handling the learner-task content association dimension. This agent extracts a bipartite subgraph from the weighted tripartite association graph, consisting of the learner node set and the task content node set, denoted as […]. ,in This represents the set of edges that associate learners with their assignments. Each edge... weight As the initial propagation value, this weight has integrated information from multiple dimensions, including historical assignment scores, submission timeliness, and answer format.

[0043] When performing similarity propagation calculations on this bipartite subgraph, the third collaborative AI agent also employs a collaborative filtering propagation mechanism. For any learner and homework content Even learners No homework completed Predictions can also be made using the completion records of similar learners. If learners With learners If completed assignments exhibit similar scoring patterns and completion quality, they are similar in learning ability and style; if assignments With homework If a task is completed by the same group of learners with similar performance, then the task is correlated in terms of difficulty level and suitability. Through this propagation mechanism, it is possible to directly predict the learner's suitability for the incomplete task.

[0044] Propagation calculation to construct the initial correlation strength matrix , of which elements equal to edge weight If an edge exists, the value is zero; otherwise, it is zero. This is determined using an iterative formula. To carry out dissemination and updates, among which The similarity matrix between learners (which can be compared with the similarity matrix of the first agent) (Shared or recalculated based on job completion mode) The similarity matrix between task content (can be compared with the similarity matrix in the second agent) (Shared or recalculated based on learner performance patterns) and This represents the propagation coefficient. Learner similarity in this dimension can be calculated by combining factors such as the Pearson correlation coefficient of the task score sequence and the similarity of the completion time distribution. Task content similarity can be calculated by combining statistical characteristics such as the score distribution and accuracy distribution after learners complete the task. After iterative convergence, the final learner-task content association strength matrix is ​​obtained. Each element of this matrix directly reflects the corresponding learner's predicted completion quality or fit for the corresponding task content.

[0045] The propagation computation processes of the three collaborative AI agents are executed independently and in parallel, each operating on its own dedicated bipartite subgraph to avoid direct confusion between information from different dimensions. The propagation coefficients are optimized based on the sparsity and noise level of the data in each dimension. Dimensions with sparser interactive data use larger propagation coefficients to enhance information diffusion, while dimensions with higher data quality use larger initial value retention coefficients to maintain the reliability of the original information. Each agent sets a convergence criterion during the propagation process: when the matrix change in two consecutive iterations is less than a preset threshold... (For example Iteration can stop when the maximum number of iterations (e.g., 50) is reached.

[0046] The three generated correlation strength matrices , and Capture related information from different dimensions. Reflects the learner's mastery of each knowledge point. It reflects the coverage and correlation between knowledge points and homework content. These three matrices directly reflect the quality of learners' historical interactions with assignment content. They provide multi-perspective input for the subsequent collaborative aggregation module, and by integrating the correlation strengths of different dimensions, they can generate more accurate and robust comprehensive correlation evaluation results, thereby supporting the accuracy of personalized assignment recommendations.

[0047] The similarity propagation calculation includes: The edge weight matrix of the corresponding bipartite subgraph is used as the initial association strength matrix for propagation, and the similarity propagation iterative process is initiated. In each round of propagation iteration, the association strength matrix of the current round is used as the left multiplication matrix, and the edge weight matrix of the corresponding bipartite subgraph is used as the right multiplication matrix. Matrix multiplication is performed so that the association strength of the current round is propagated to the adjacent nodes along the node association direction represented by the edge weights in the bipartite graph, and the association strength matrix of the next round is obtained. Perform row normalization on each row of the next round association strength matrix to normalize the sum of the values ​​of each row to a preset constant value, and obtain the row normalized next round association strength matrix; The association strength matrix of the next round after row normalization is weighted and superimposed with the association strength matrix of the current round according to a preset retention ratio to obtain the propagation result matrix of this round. The preset retention ratio controls the information retention weight of the association strength of the current round in the superimposed result. Calculate the matrix norm difference between the current propagation result matrix and the previous propagation result matrix. When the matrix norm difference is lower than a preset convergence threshold, terminate the propagation iteration and output the current propagation result matrix as the final state association strength matrix of the corresponding dimension of the collaborative AI agent.

[0048] Similarity propagation calculation is based on an iterative update mechanism and is executed separately within each collaborative AI agent. A detailed explanation is provided using the first collaborative AI agent in the learner-knowledge point association dimension as an example. This agent receives the constructed learner-knowledge point bipartite graph subgraph and its corresponding edge weight matrix. The weight matrix has dimensions of , of which Line number Column elements The calculations have already been completed in the aforementioned process.

[0049] Before starting the propagation iteration, set the edge weight matrix. Directly copy as the propagation initial state matrix This matrix serves as the initial distribution of association strength in round 0. A maximum upper limit is set for the number of iteration rounds. The value is typically between 10 and 50 to avoid excessive consumption of computing resources. Initialize the round counter. .

[0050] Entering the During round propagation iteration, the correlation strength matrix of the current round is read first. The dimensions of this matrix are maintained at 1. Perform a left matrix multiplication operation to calculate the intermediate propagation result. The physical meaning of this operation lies in the fact that the current association strength of a learner node with each knowledge point is propagated to neighboring knowledge point nodes through the association path represented by the edge weights. The specific calculation of matrix multiplication is as follows: ;in Indicate learner In the Wheelchair knowledge points The strength of the association, Representing knowledge points With knowledge points The second-order association weight between knowledge points. This weight captures the implicit similarity between knowledge points when they are frequently interacted with by multiple learners.

[0051] Obtain the intermediate propagation result matrix Then, normalization is performed on each row of the matrix. For the first row... Calculate the sum of all elements in the row. .like This indicates that the learner has no associated propagation signal in the current iteration, and at this time, all elements in this row are set to 0. This ensures the stability of the normalized values. Then for the first Scaling is performed on each element in the row: After row normalization, the matrix The sum of each row of elements is equal to the preset constant value 1. This process eliminates the numerical bias caused by the difference in the total correlation strength of different learners, making the propagation signal comparable across learner dimensions.

[0052] Next, a weighted overlay operation is performed, incorporating a preset retention ratio parameter. The value ranges from 0 to 1, and is usually set to 0.2 to 0.5. Calculate the propagation result matrix for this round: This weighted overlay mechanism achieves information smoothing, parameters Controls the degree to which the original association strength of the current round is preserved. Higher... The lower value makes the algorithm more inclined to retain historical information and reduce the iteration step size; Values ​​accelerate the propagation of correlated signals, but may introduce numerical oscillations. Matrix That is, the first The correlation strength matrix of the round is used as the input for the next iteration.

[0053] After completing this round of iterations, it is necessary to determine whether the algorithm has reached convergence. Calculate the Frobenius norm difference between the current round's propagation result matrix and the previous round's propagation result matrix: The norm difference reflects the global magnitude of the changes in matrix elements between two iterations. The Frobenius norm is calculated as follows: ;Will With preset convergence threshold Comparison, usually The range of values ​​is to .like Once the propagation process has converged to a steady state, the iteration loop is terminated, and the matrix is... Output as the final state association strength matrix of the learner-knowledge point association dimension. If And the current round , then let Return to the starting point of the iteration and continue to execute the next round of propagation calculation.

[0054] The second collaborative AI agent, representing the knowledge point-assignment content association dimension, and the third collaborative AI agent, representing the learner-assignment content association dimension, both employ the same propagation iteration framework. The second agent uses an edge weight matrix. As the initial matrix The matrix dimension is Intermediate variables during propagation are represented as , and The retention ratio parameter can be set independently. The third agent uses an edge weight matrix. As the initial matrix The matrix dimension is The corresponding variable during the propagation process is , and The retention ratio parameter is set to .

[0055] The three agents can perform propagation iterative calculations in parallel, each independently maintaining its iterative state variables without interfering with the others. In practical deployment, the maximum number of iteration rounds and convergence thresholds can be set for each agent based on the data sparsity of different association dimensions. For dimensions with high sparsity of the edge weight matrix, such as the learner-task content association dimension, the maximum number of iteration rounds can be appropriately increased to over 30 rounds, and the convergence threshold can be relaxed to [missing information]. This ensures that the propagation signal spreads sufficiently to the sparse region. For dimensions with relatively dense edge weight distribution, such as the knowledge point-assignment content association dimension, a stricter convergence threshold can be used. This is to obtain a more refined estimate of the association strength.

[0056] A numerical stability guarantee mechanism exists during the propagation iteration process. After matrix multiplication, if the intermediate result matrix... The number of elements in the data exceeds the threshold. In this case, perform a global scaling on the entire matrix, dividing all elements by . This prevents floating-point overflow during subsequent normalization calculations. In the weighted overlay step, if all elements in a row are zero, that row is automatically filled with the values ​​of the corresponding row from the previous round after overlay, avoiding information loss caused by rows with zero values.

[0057] After each collaborative AI agent completes its propagation iteration, the output final-state correlation strength matrix has a clear physical interpretation. Learner-Knowledge Point Final-State Matrix In, elements Indicate learner Knowledge points The comprehensive mastery assessment value not only integrates direct interaction records but also incorporates learners' performance information on similar knowledge points through a propagation mechanism. Knowledge Point-Assignment Content Final State Matrix In, elements Quantifying knowledge points With homework The degree of internal correlation reflects the extent to which the assignment covers the knowledge point. Learner-Assignment Content Final State Matrix In, elements Predicting learners Homework The completion fit is based on the learner's actual performance on similar historical assignments.

[0058] The three final-state association strength matrices, as independent association dimension estimation results, are fed into the subsequent collaborative aggregation module for multi-dimensional fusion processing, ultimately generating a comprehensive association strength matrix for personalized job recommendation.

[0059] The correlation strength matrices are fused element-wise using a linear weighted aggregation mechanism to generate a comprehensive correlation strength matrix, including: Obtain a preset set of dimension weight coefficients, which includes a first weight coefficient, a second weight coefficient, and a third weight coefficient corresponding to the learner-knowledge point association dimension, the knowledge point-assignment content association dimension, and the learner-assignment content association dimension, respectively. Each dimension weight coefficient is non-negative and the sum of the three is a preset normalization constant. Perform matrix dimension alignment processing on each association strength matrix to make the learner-knowledge point association strength matrix, the knowledge point-assignment content association strength matrix, and the learner-assignment content association strength matrix have the same number of matrix rows and columns; For each element position in the three association strength matrices that have the same row index and column index, the corresponding element in the learner-knowledge point association strength matrix is ​​weighted by the first weight coefficient, the corresponding element in the knowledge point-task content association strength matrix is ​​weighted by the second weight coefficient, and the corresponding element in the learner-task content association strength matrix is ​​weighted by the third weight coefficient. The three weighted results are summed to obtain the element value of that position in the comprehensive association strength matrix. The comprehensive correlation strength matrix is ​​normalized element-wise to map all element values ​​to a preset value range, resulting in a normalized comprehensive correlation strength matrix.

[0060] After completing the independent calculations of each collaborative AI agent, the association strength matrices generated from the three association dimensions need to be fused to obtain a unified representation that comprehensively reflects the complex relationships between learners, knowledge points, and assignment content. The collaborative aggregation module first retrieves a preset set of dimension weight coefficients from the system configuration file. This set contains three core parameters: the first weight coefficient... This is used to adjust the contribution of the learner-knowledge point association dimension to the final fusion result, and its numerical range is set to... Second weighting coefficient The influence weights used to adjust the correlation between knowledge points and homework content are also constrained by... Within the interval; third weighting coefficient This is used to adjust the strength of the learner-assignment content association dimension, and its value range is also [missing information]. To ensure the mathematical rigor of the fusion process, the weight coefficients of the three dimensions must satisfy the normalization constraint. This normalization constant is always 1 in practical applications. Different subject areas and teaching scenarios can adjust the weight allocation strategy according to actual needs; for example, in basic subjects that emphasize the integrity of the knowledge system, the weight can be increased. The value of [the value] can be improved in application-oriented courses that emphasize the cultivation of personalized abilities. and Weight configuration.

[0061] Before performing matrix fusion, the inconsistency in the dimensions of the three association strength matrices must be addressed. (Learner-Knowledge Point Association Strength Matrix) The dimension is ,in Indicates the number of learners. Represents the number of knowledge points; Knowledge point-assignment content correlation strength matrix The dimension is ,in Indicates the number of assignments; Learner-Assignment Content Association Strength Matrix The dimension is Since the nodes described by the three matrices are associated with different objects, directly performing element-wise operations is mathematically infeasible. Therefore, a matrix dimension alignment mechanism is introduced to extend the three matrices to the same dimension by constructing a unified index mapping relationship.

[0062] The specific matrix dimension alignment strategy is implemented using a tensor expansion method. The three two-dimensional matrices are expanded into a unified three-dimensional tensor space, with its dimension specification set as follows: The index of each element of this tensor Corresponding learners Knowledge Points and homework content The triples formed. For the learner-knowledge point association strength matrix. By performing a broadcast copy operation along the third dimension, it can be replicated. Element value at position Copy to extended tensor All values ​​that satisfy the first two dimensions of the index are The position, that is, for any All Similarly, the correlation strength matrix between knowledge points and homework content... Perform a broadcast operation along the first dimension, Element value at position Extend to tensors All values ​​satisfying the last two-dimensional index are The position, that is, for any All For the learner-task content association strength matrix Perform broadcast replication along the second dimension, Element value at position Extend to tensors All first-dimensional and third-dimensional indices are The position, that is, for any All After the above expansion operations, the three different-dimensional association strength matrices are uniformly mapped to the same... 3D tensor space.

[0063] After dimensional alignment is completed, for each location in the extended tensor space Perform a weighted fusion computation. Extract the element values ​​of the three extended tensors at that position. , and Multiply by the corresponding dimension weight coefficients to calculate the fusion value. This fusion value comprehensively reflects the learner's response across three dimensions of association. Complete the knowledge points Homework This assessment of the suitability of the learning task involves traversing all elements in the tensor space. At each position, perform element-wise weighted summation to obtain the preliminary comprehensive correlation strength tensor. .

[0064] Considering that the core association in practical assignment recommendation scenarios is the matching relationship between learners and assignment content, it is necessary to integrate the three-dimensional fusion tensor. Dimension reduction is performed. An aggregation operation is executed on the fusion tensor along the knowledge point dimension (the second dimension), by calculating each... Location in all knowledge point index The weighted average of the values ​​above yields the learner-assignment content comprehensive correlation strength matrix. The specific calculation method is as follows: for learners and homework Comprehensive correlation strength value This aggregation process compresses multi-dimensional information from the knowledge point dimension into a single scalar value, achieving a projection mapping from a three-dimensional association space to a two-dimensional matching space. The aggregated matrix... Dimensions The first in the matrix Line number Column elements This indicates that the system is for the learner Recommended homework A comprehensive fit assessment.

[0065] Because the weighted fusion process may cause the overall correlation strength value to exceed a reasonable range, the matrix needs to be adjusted. Perform normalization. First, traverse the matrix. Find the maximum value among all elements in the array. and minimum element value Calculate the numerical span For each element in the matrix Applying linear normalization transformation Map all element values ​​to The standard interval. The normalized matrix. This process preserves the relative magnitudes of the original elements while eliminating computational biases caused by differences in numerical scale. Special handling is required during the normalization process. In the case of degradation, when all comprehensive correlation strength values ​​are exactly the same, the normalization result is uniformly set to the preset default value of 0.5 to avoid numerical anomalies caused by division by zero.

[0066] Normalized composite correlation strength matrix Its elements serve as the core basis for subsequent assignment recommendation and ranking. This directly reflects the system's impact on learners. Recommended homework The confidence level is relatively high. Value indicates the task It engages with learners across multiple dimensions, including knowledge coverage, difficulty matching, and historical interaction preferences. The current learning state of the data is highly consistent with the data and should be prioritized at the beginning of the recommended sequence. To further improve the discriminative power of the fusion results, a nonlinear activation transformation can be introduced after normalization, such as applying the hyperbolic tangent function. ,in This is the amplification factor, adjusted by... The value of can enhance the differentiation between high and low correlation strength values, making the recommendation ranking more distinct.

[0067] In practical engineering implementations, to address the computational efficiency challenges in scenarios with large-scale learner and job libraries, a sparse matrix storage structure can be used to optimize the tensor expansion and fusion computation process. Leveraging the sparsity of the three association strength matrices, broadcast copying and weighted summation operations are performed only on non-zero elements, avoiding redundant computation on a large number of zero-value elements. Simultaneously, a parallel computing framework is introduced, partitioning the tensor space according to learner or job dimensions, and performing fusion computation in parallel on multiple computing nodes, significantly reducing overall processing latency. For online recommendation scenarios requiring real-time response, the normalized comprehensive association strength matrix can be pre-cached. When a new recommendation request arrives, the cached result is queried directly. The matrix recalculation process is only triggered when the learner generates a new interaction that causes the weight of the associated edge to be updated, thus achieving millisecond-level response latency while ensuring recommendation accuracy.

[0068] A comprehensive association score is calculated for each candidate assignment in the pre-built assignment content library. Candidate assignments that meet a preset association threshold are selected, and personalized assignment recommendation sequences are generated by sorting them in descending order of their comprehensive association scores. The interactive feedback data of the target learner on the recommended sequences is then written back to the weighted tripartite association graph to update the association edge weights, including: For each candidate task in the task content library, extract all association strength values ​​of the candidate task in the row corresponding to the target learner from the comprehensive association strength matrix, and perform a weighted summation operation on all extracted association strength values ​​to obtain the comprehensive association score of the candidate task. All candidate job items are sorted in descending order of comprehensive correlation score. A preset number of candidate job items with comprehensive correlation scores higher than a preset cutoff threshold are selected to generate the personalized job recommendation sequence. Obtain the completion rate and accuracy data of the target learner for each recommended task in the personalized task recommendation sequence, and calculate the real-time weight adjustment amount for this interaction based on the completion rate and accuracy data; The instantaneous weight adjustment amount and the historical weight of the corresponding associated edge in the weighted three-part association graph are weighted and merged according to a preset attenuation coefficient, and the weight of the corresponding associated edge is updated with the weighted merging result.

[0069] After the comprehensive correlation strength matrix is ​​generated, personalized recommendation ranking is performed for each candidate job item in the job content library. Assume there are a total of... Each candidate task must be for the current target learner. Generate recommended sequences. This is achieved from the normalized and nonlinearly activated composite association strength matrix. Extract the first one from the middle. All elements of a row, the row vector is denoted as Each element Indicate learner With candidate jobs The comprehensive correlation strength value after multi-dimensional fusion and non-linear enhancement.

[0070] For candidate jobs Its comprehensive correlation score The calculation requires integrating multiple information sources. First, consider the basic correlation strength of the task with the current learner's corresponding row. This value integrates multi-dimensional information such as learners' mastery of knowledge points, the relationship between knowledge points and assignments, and historical completion records. However, relying on a single value may not fully reflect the recommended value of assignments; therefore, an assignment difficulty suitability coefficient is introduced. and the timeliness coefficient of the work As an auxiliary scoring factor.

[0071] Difficulty adaptability coefficient According to the assignment The difficulty level of the annotation and the learner The historical average difficulty level of the assignment is matched. Assume the assignment... The difficulty level is marked as (The range of values ​​is) ),learner The set of completed assignments in the history of assignments is denoted as . Calculate the learner's historical average completion difficulty for: ;in The cardinality of the set of completed tasks. For history assignments The difficulty level. The difficulty suitability coefficient is calculated using a Gaussian kernel function: ;in This is the difficulty-matching tolerance hyperparameter, typically set to 0.8. When the difficulty of the candidate task is close to the learner's historical adaptation difficulty, The closer to 1; the greater the difference, the better. The smaller the value, the less likely it is to suppress the recommendation weight of tasks that are too difficult or too easy.

[0072] Job timeliness coefficient The decay is based on the interval between the job posting time and the current time. Assume the job... The publication timestamp is The current timestamp is The time interval is (Unit: days). The timeliness coefficient adopts an exponential decay model: ;in The time decay rate hyperparameter, typically set to 0.05, causes the timeliness coefficient of assignments published more than 20 days ago to decay to approximately 0.37. This mechanism ensures that newly published assignments receive higher recommendation priority, aligning with the time-series requirements of the teaching schedule.

[0073] Taking into account the above factors, candidate jobs The formula for calculating the comprehensive correlation score is as follows: ;in Assignment type weighting factors: Different weights are assigned based on the assignment type (e.g., basic exercises, comprehensive applications, extended exploration). Basic exercises are weighted differently. The recommended value is 1.0 for comprehensive application type, 1.2 for comprehensive application type, and 0.9 for extended exploration type, in order to balance the recommended proportions of different types of assignments.

[0074] After calculating the comprehensive correlation score for all candidate jobs, a score vector is constructed. Perform a descending sort operation on the vector to obtain the sorted job index sequence. ,in Set a preset cutoff threshold. This threshold can be dynamically adjusted based on historical recommendation acceptance rates, with a typical initial value set at 0.65. This filters out those that meet the criteria. The number of recommended tasks is limited to a preset limit. (Typical values ​​range from 8 to 12), and the final generated personalized job recommendation sequence is denoted as... .

[0075] After the recommended assignments are generated and pushed to learners, the system continuously monitors learners' interactions with each recommended assignment. This monitoring targets the assignments within the recommended assignments. Acquire learners Completion and accuracy data. Completion data is collected through the job submission interface and recorded as follows. value range Where 0 indicates not started and 1 indicates fully submitted. Accuracy data is obtained through an automated grading system and denoted as... Similarly, the range of values This indicates the percentage of correct answers submitted.

[0076] Based on completion and accuracy data, calculate the immediate weight adjustment for this interaction. This adjustment requires simultaneously updating the weights of the three types of association edges in the weighted tripartite graph. For the learner-knowledge point association edge... Its real-time weight adjustment amount According to the assignment Knowledge points involved Performance calculation. Assume the assignment. Related to knowledge points The set of questions is The learner's accuracy rate in this set is ,but: ;in The number of questions. A reward and punishment mechanism is established (a positive adjustment is made when the accuracy rate is higher than 0.5, and a negative adjustment is made when it is lower than 0.5). This is the adjustment strength coefficient for the learner-knowledge point edge, typically set to 0.15. This formula ensures that interactions with high completion and high accuracy significantly increase the edge weight, while interactions with low completion or poor accuracy weaken the edge weight.

[0077] For the knowledge point-assignment content related edge Its real-time weight adjustment amount This mainly reflects the actual importance of this knowledge point in the assignment. The calculation formula is: ;in Indicates completion of homework A collection of learners The numerator represents the number of learners, and the numerator represents the total number of learners in each knowledge point. Average accuracy rate on related questions This is the adjustment strength coefficient for the knowledge point-assignment pair, typically set to 0.10. This mechanism uses group performance data to calibrate the correlation strength between knowledge points and assignments; if a knowledge point is generally performed poorly on questions, its weight in that assignment is reduced.

[0078] For learner-assignment content association edges Its real-time weight adjustment amount Taking into account completion rate, accuracy rate, and timely submission: ;in For learners Submit assignment The number of days until the deadline (a negative value if the submission is late). The allowed completion time for the task (in days). This is the adjustment strength coefficient for the learner-assignment side, typically set to 0.20. The formula assigns extra weight to timely and high-quality completed assignments, while penalizing delayed or low-quality completion.

[0079] After calculating the real-time weight adjustments for each type of edge, a weighted average needs to be performed with the historical weights to update the associated edge weights. An exponential moving average strategy is employed, incorporating a preset decay coefficient. This coefficient controls the proportion of historical weights retained, typically set to 0.85. For learner-knowledge point association edges, the updated weights... The calculation is as follows: ;in This represents the current historical weight. This formula ensures smooth weight updates, avoiding drastic weight fluctuations caused by a single interaction. Similarly, the weight updates for the knowledge point-assignment content association edges and the learner-assignment content association edges are as follows: ; ; The weight update operation is triggered immediately after each learner submits a recommendation assignment, and the updated weight values ​​are written back to the corresponding edge storage structure of the weighted ternary association graph. To prevent numerical overflow caused by weight accumulation, an upper limit constraint is imposed on all edge weights. At that time, forced truncation is (Typical value is 10.0). A lower limit protection mechanism is also set; when the weight is lower than... (When the typical value is 0.01), maintain To prevent edge associations from completely disappearing.

[0080] Through the aforementioned dynamic weight update mechanism, the weighted tripartite association graph can continuously learn about changes in learners' actual abilities and feedback on assignment quality. This allows subsequent recommendation sequences to be generated based on the latest association states, forming a closed-loop adaptive system of "recommendation-feedback-update-re-recommendation." This real-time update strategy significantly improves the recommendation system's sensitivity to individual learner differences and dynamic changes in knowledge mastery, ensuring that assignment recommendations always maintain a high degree of matching with learners' current learning status.

[0081] The instantaneous weight adjustment amount and the historical weights of the corresponding associated edges in the weighted three-part association graph are weighted and merged according to a preset attenuation coefficient, and the weights of the corresponding associated edges are updated with the weighted merging result, including: The instant weight adjustment amount is calculated and is proportional to the product of the completion data and the accuracy data. When the completion data is lower than a preset completion threshold or the accuracy data is lower than a preset accuracy threshold, the instant weight adjustment amount is negative. Obtain the time decay factor corresponding to the associated edge. The time decay factor is determined by a preset decay function based on the time interval between the most recent update time of the associated edge and the current time. The larger the time interval, the closer the value of the time decay factor is to zero, so that the weight of the associated edge that has not been updated for a long time will naturally decay over time. Multiply the historical weight by the time decay factor to obtain the decayed historical weight of the associated edge; The historical weights after decay are summed with the instantaneous weight adjustment to obtain the updated associated edge weights. The updated associated edge weights are then written back to the weight storage location of the associated edge in the weighted three-part association graph, thus completing this feedback-driven closed-loop update of edge weights.

[0082] Once a learner completes an assignment in the recommended sequence, the resulting interactive feedback data must be immediately reflected in the edge weight update of the weighted tripartite graph. This closed-loop update mechanism is achieved through a decaying weighted merging strategy, which absorbs the latest learning behavior features while retaining the accumulated knowledge from historical behaviors. Simultaneously, it applies natural decay to edges that have not generated interaction for a long time, ensuring the timeliness and accuracy of the graph structure.

[0083] First, determine how to calculate the instantaneous weight adjustment. For learners... With knowledge points The associated edges between Its instantaneous adjustment amount The calculation requires comprehensive consideration of two core indicators: completion rate and accuracy rate. It reflects the learner's overall engagement with the assignment, while accuracy reflects the learner's overall engagement with the assignment. This reflects their actual level of mastery of specific knowledge points. When calculating, first extract the homework... Related knowledge points All questions constitute a subset of questions Statistical learners' accuracy rate on this subset Then completeness With accuracy Multiply by, then multiply by the preset adjustment strength coefficient This yields the basic adjustment amount. However, a conditional judgment mechanism needs to be introduced: when the completion rate... Below the preset completion threshold If the accuracy rate is low, it indicates that the learner has not invested enough effort in the assignment. In this case, regardless of the accuracy rate, the adjustment amount should be negative to reflect the learner's declining mastery of the relevant knowledge points; if the accuracy rate is high... Below the preset accuracy threshold Even if the completion rate is high, it still indicates that the learner has a comprehension barrier regarding that knowledge point, and in this case, the adjustment amount is also negative. This negative adjustment mechanism is implemented by adding a sign determination function before the basic adjustment amount, and this sign function outputs... or The calculation depends on whether both the completion rate and accuracy rate simultaneously meet their respective thresholds. The complete calculation process is as follows: first calculate the product term. Then judge and Whether both conditions are met simultaneously, and if so, the sign function is taken. Otherwise take Finally obtained ,in This is the sign determination function mentioned above.

[0084] For knowledge points With the content of the assignment The associated edges between Its instantaneous adjustment amount The calculation logic differs. Since this edge describes the coverage relationship of knowledge points in the assignment, its adjustment is based on the collective performance of all learners who completed the assignment on the corresponding knowledge points. Specifically, the process involves first obtaining the completed assignment data... learner set For each learner in the set Calculate its knowledge points Accuracy on related questions Then, the average of all correct rates is taken to obtain the average correct rate for that knowledge point in this assignment. The overall completion rate of the assignment (defined as a set) Average completion rate of learners ) and average accuracy Multiply by, then multiply by the adjusted strength coefficient This yields the basic adjustment amount. Similarly, a dual threshold determination is introduced: when... or When the adjustment value is negative, it indicates that the assignment is not effective in testing the corresponding knowledge point, and the weight of the associated edge needs to be reduced to decrease the probability of recommending the assignment in the future.

[0085] learner With the content of the assignment The associated edges between Its instantaneous adjustment amount The calculation directly depends on the individual learner's overall performance on the assignment. Specifically: the learner's performance on the assignment... completion level With overall accuracy Multiply the results, then consider the submission timeliness factor. Define the submission timeliness score as the assignment score. Allowed completion time The timeliness score is higher as the submission time is closer to the publication time than the actual time taken by the learner. The timeliness score is calculated by multiplying completion rate, accuracy, and timeliness score together, and then multiplying by an adjustment factor. The basic adjustment amount is obtained. The double threshold determination rule is: when... or When the time is right, adjust the quantity to a negative value.

[0086] After calculating the immediate adjustment amount, a time decay mechanism needs to be introduced to handle historical weights. Each associated edge maintains a timestamp of its most recent update in the system. Record the moment when the weight of this edge was last updated due to interactive feedback. The current moment is denoted as... Calculate the time interval Time decay factor Calculated using the exponential decay function, the function form is: ,in This is the decay rate hyperparameter, and its value determines the decay rate. It is usually set according to the learning period, for example, in a learning scenario measured in weeks. Possible values to In between, the edge weights that haven't been updated for a month naturally decay back to their original values. to Between. The properties of the exponential function guarantee the time interval. The larger the attenuation factor, the greater the attenuation factor. The closer it gets to zero, the more gradually it forgets edges that have not been interacted with for a long time.

[0087] Get the historical weight of the associated edge Then, it is combined with the time decay factor. Multiply to obtain the decayed historical weights. The physical significance of this step lies in the fact that the learner's knowledge acquisition information contained in the historical weights becomes invalid over time, especially when the learner has not interacted with the knowledge point or assignment for a long time. Their past mastery level can no longer accurately reflect their current true state, therefore, it needs to be reduced using a decay factor. The decayed historical weights. It represents the state of prior knowledge after time-sensitivity correction.

[0088] Decayed historical weights With real-time weight adjustment Summing yields the updated associated edge weights. This weighted merging operation integrates historical data with the latest feedback: the decayed historical weights serve as a baseline, and the immediate adjustment amount acts as an incremental correction; the two are combined to form a new weight estimate. It's important to note that the updated weights may exceed a preset reasonable range, therefore boundary pruning is necessary. Specifically: if... Then force it to be set to ;like Then force it to be set to This pruning operation ensures that all edge weights remain constant. Within the interval, to prevent the weight values ​​from diverging in extreme cases, which could cause subsequent similarity propagation calculations to fail.

[0089] After completing the calculation and pruning of the weight values, the updated associated edge weights will be... Write back to the corresponding storage location of the weighted ternary association graph. Specifically, for the learner-knowledge point edge... ,Will Write back to edge weight matrix The Line number Column; for knowledge point - homework content side ,Will Write back to edge weight matrix The Line number List; For learners - assignment content side ,Will Write back to edge weight matrix The Line number The column also updates the timestamp of the most recent update for that edge. For the current moment This is to ensure that the time interval is calculated correctly during the next update.

[0090] This closed-loop update process is automatically triggered each time a learner completes a recommended assignment and generates interactive feedback. If the assignment completed by a learner is related to multiple knowledge points, then all relevant knowledge points need to be traversed, and the above update process needs to be executed for each learner-knowledge point edge. This is because the assignment recommendation sequence... The graph contains multiple candidate tasks. As learners complete the tasks sequentially, the system continuously receives feedback data and updates the weights of relevant edges in the graph structure. This continuous update mechanism enables the weighted tripartite graph to dynamically track changes in learners' knowledge mastery, providing a more accurate basis for calculating the association strength for the next round of task recommendations.

[0091] To ensure numerical stability during the update process, an outlier detection mechanism needs to be introduced in actual deployment. For example, if the weight of an edge fluctuates drastically within a short period of time (e.g., the weight change exceeds a threshold in two consecutive updates), an outlier detection mechanism can be introduced. When an anomaly is triggered, an alert is issued, automatic updates for that edge are suspended, and manual review or a more conservative update strategy is adopted. Furthermore, for edges in the early stages of the system that have not yet accumulated sufficient historical interaction data, their time decay factor can be appropriately increased. The value of is chosen to reduce the decay rate of historical weights and avoid excessive weight decay caused by data sparsity.

[0092] Through the coordinated operation of the aforementioned steps, including real-time adjustment calculation, introduction of time decay factor, decay-weighted merging, and boundary pruning and write-back, a closed-loop update of the edge weights in the weighted tripartite association graph is achieved. This update mechanism organically integrates the learners' latest learning behavior characteristics with their historical knowledge mastery status. Simultaneously, it utilizes the time decay function to achieve natural forgetting of outdated information, ensuring that the graph structure always reflects the current learning status of the learner group. This provides high-quality weight input for subsequent similarity propagation calculations by collaborative agents, ultimately supporting the continuous optimization of personalized assignment recommendations.

[0093] A second aspect of the present invention provides an AI intelligent agent collaborative system for task deployment, comprising: The graph building unit is used to acquire multimodal learning behavior data generated by the target learner within a preset learning period, perform feature encoding and relation extraction processing on the multimodal learning behavior data, construct a three-part association graph containing learner nodes, knowledge point nodes and assignment content nodes, and calculate the weight of the association edges between each node based on historical interaction data to form a weighted three-part association graph. The collaborative computing unit is configured with at least three collaborative AI agents corresponding to the learner-knowledge point association dimension, the knowledge point-assignment content association dimension, and the learner-assignment content association dimension, respectively. Each collaborative AI agent performs similarity propagation calculation on the corresponding subgraph of the weighted tripartite association graph and independently generates the association strength matrix of the corresponding dimension. The weighted aggregation unit is used to input the association strength matrix generated by each collaborative AI agent into the collaborative aggregation module, and perform element-wise weighted fusion of each association strength matrix through a linear weighted aggregation mechanism to generate a comprehensive association strength matrix. The recommendation update unit is used to calculate the comprehensive association score of each candidate task in the pre-constructed task content library based on the comprehensive association strength matrix, filter the candidate task tasks that meet the preset association threshold, generate a personalized task recommendation sequence by sorting them in descending order of comprehensive association score, and write back the interactive feedback data of the target learner to the weighted tripartite association graph to update the association edge weights.

[0094] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0095] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0096] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI agent collaborative method for task assignment, characterized in that, include: Acquire multimodal learning behavior data generated by the target learner within a preset learning period, perform feature encoding and relation extraction processing on the multimodal learning behavior data, construct a three-part association graph containing learner nodes, knowledge point nodes and assignment content nodes, and calculate the weight of the association edges between each node based on historical interaction data to form a weighted three-part association graph. Configure at least three collaborative AI agents corresponding to the learner-knowledge point association dimension, the knowledge point-assignment content association dimension, and the learner-assignment content association dimension, respectively. Each collaborative AI agent performs similarity propagation calculation on the corresponding subgraph of the weighted tripartite association graph and independently generates the association strength matrix of the corresponding dimension. The association strength matrix generated by each collaborative AI agent is input into the collaborative aggregation module. The linear weighted aggregation mechanism is used to perform element-wise weighted fusion of each association strength matrix to generate a comprehensive association strength matrix. Based on the comprehensive association strength matrix, a comprehensive association score is calculated for each candidate task in the pre-constructed task content library. Candidate task items that meet the preset association threshold are selected, and personalized task recommendation sequences are generated by sorting them in descending order of comprehensive association score. The interactive feedback data of the target learner on the recommendation sequence is written back to the weighted tripartite association graph to update the association edge weights.

2. The method according to claim 1, characterized in that, The multimodal learning behavior data is processed by feature encoding and relation extraction to construct a three-part association graph containing learner nodes, knowledge point nodes, and assignment content nodes. Based on historical interaction data, the weights of the association edges between nodes are calculated to form a weighted three-part association graph, including: Learner identifiers, interacted knowledge point identifiers, and accessed assignment content identifiers are extracted from the multimodal learning behavior data to construct learner node sets, knowledge point node sets, and assignment content node sets. Establish learner-knowledge point association edges based on the learner's historical interaction records with knowledge points; establish knowledge point-assignment content association edges based on the coverage mapping relationship of knowledge points in assignment content; and establish learner-assignment content association edges based on the learner's historical completion records with assignment content. The first weight of the learner-knowledge point association edge is calculated based on the interaction frequency and interaction accuracy. The second weight of the knowledge point-assignment content association edge is calculated based on the coverage ratio of the knowledge point in the assignment content. The third weight of the learner-assignment content association edge is calculated based on the assignment completion quality score. The set of learner nodes, the set of knowledge point nodes, the set of assignment content nodes, and the three types of associated edges carrying the first weight, the second weight, and the third weight respectively are summarized to form the weighted ternary association graph.

3. The method according to claim 1, characterized in that, Each collaborative AI agent performs similarity propagation calculations on the corresponding subgraph of the weighted tripartite association graph, independently generating an association strength matrix for the corresponding dimension, including: The first collaborative AI agent corresponding to the learner-knowledge point association dimension takes the bipartite subgraph formed by the learner node set and the knowledge point node set in the weighted tripartite association graph as the operation object, and uses the learner-knowledge point association edge weight as the initial propagation value to perform similarity propagation calculation and generate a learner-knowledge point association strength matrix. The second collaborative AI agent, corresponding to the knowledge point-task content association dimension, takes the bipartite subgraph formed by the knowledge point node set and the task content node set in the weighted tripartite association graph as the operation object, and uses the knowledge point-task content association edge weight as the initial propagation value to perform similarity propagation calculation and generate a knowledge point-task content association strength matrix. The third collaborative AI agent corresponding to the learner-assignment content association dimension takes the bipartite subgraph formed by the learner node set and the assignment content node set in the weighted tripartite association graph as the operation object, and uses the learner-assignment content association edge weight as the initial propagation value to perform similarity propagation calculation and generate a learner-assignment content association strength matrix.

4. The method according to claim 3, characterized in that, The similarity propagation calculation includes: The edge weight matrix of the corresponding bipartite subgraph is used as the initial association strength matrix for propagation, and the similarity propagation iterative process is initiated. In each round of propagation iteration, the association strength matrix of the current round is used as the left multiplication matrix, and the edge weight matrix of the corresponding bipartite subgraph is used as the right multiplication matrix. Matrix multiplication is performed so that the association strength of the current round is propagated to the adjacent nodes along the node association direction represented by the edge weights in the bipartite graph, and the association strength matrix of the next round is obtained. Perform row normalization on each row of the next round association strength matrix to normalize the sum of the values ​​of each row to a preset constant value, and obtain the row normalized next round association strength matrix; The association strength matrix of the next round after row normalization is weighted and superimposed with the association strength matrix of the current round according to a preset retention ratio to obtain the propagation result matrix of this round. The preset retention ratio controls the information retention weight of the association strength of the current round in the superimposed result. Calculate the matrix norm difference between the current propagation result matrix and the previous propagation result matrix. When the matrix norm difference is lower than a preset convergence threshold, terminate the propagation iteration and output the current propagation result matrix as the final state association strength matrix of the corresponding dimension of the collaborative AI agent.

5. The method according to claim 1, characterized in that, The correlation strength matrices are fused element-wise using a linear weighted aggregation mechanism to generate a comprehensive correlation strength matrix, including: Obtain a preset set of dimension weight coefficients, which includes a first weight coefficient, a second weight coefficient, and a third weight coefficient corresponding to the learner-knowledge point association dimension, the knowledge point-assignment content association dimension, and the learner-assignment content association dimension, respectively. Each dimension weight coefficient is non-negative and the sum of the three is a preset normalization constant. Perform matrix dimension alignment processing on each association strength matrix to make the learner-knowledge point association strength matrix, the knowledge point-assignment content association strength matrix, and the learner-assignment content association strength matrix have the same number of matrix rows and columns; For each element position in the three association strength matrices that have the same row index and column index, the corresponding element in the learner-knowledge point association strength matrix is ​​weighted by the first weight coefficient, the corresponding element in the knowledge point-task content association strength matrix is ​​weighted by the second weight coefficient, and the corresponding element in the learner-task content association strength matrix is ​​weighted by the third weight coefficient. The three weighted results are summed to obtain the element value of that position in the comprehensive association strength matrix. The comprehensive correlation strength matrix is ​​normalized element-wise to map all element values ​​to a preset value range, resulting in a normalized comprehensive correlation strength matrix.

6. The method according to claim 1, characterized in that, A comprehensive association score is calculated for each candidate assignment in the pre-built assignment content library. Candidate assignments that meet a preset association threshold are selected, and personalized assignment recommendation sequences are generated by sorting them in descending order of their comprehensive association scores. The interactive feedback data of the target learner on the recommended sequences is then written back to the weighted tripartite association graph to update the association edge weights, including: For each candidate task in the task content library, extract all association strength values ​​of the candidate task in the row corresponding to the target learner from the comprehensive association strength matrix, and perform a weighted summation operation on all extracted association strength values ​​to obtain the comprehensive association score of the candidate task. All candidate job items are sorted in descending order of comprehensive correlation score. A preset number of candidate job items with comprehensive correlation scores higher than a preset cutoff threshold are selected to generate the personalized job recommendation sequence. Obtain the completion rate and accuracy data of the target learner for each recommended task in the personalized task recommendation sequence, and calculate the real-time weight adjustment amount for this interaction based on the completion rate and accuracy data; The instantaneous weight adjustment amount and the historical weight of the corresponding associated edge in the weighted three-part association graph are weighted and merged according to a preset attenuation coefficient, and the weight of the corresponding associated edge is updated with the weighted merging result.

7. The method according to claim 6, characterized in that, The instantaneous weight adjustment amount and the historical weights of the corresponding associated edges in the weighted three-part association graph are weighted and merged according to a preset attenuation coefficient, and the weights of the corresponding associated edges are updated with the weighted merging result, including: The instant weight adjustment amount is calculated and is proportional to the product of the completion data and the accuracy data. When the completion data is lower than a preset completion threshold or the accuracy data is lower than a preset accuracy threshold, the instant weight adjustment amount is negative. Obtain the time decay factor corresponding to the associated edge. The time decay factor is determined by a preset decay function based on the time interval between the most recent update time of the associated edge and the current time. The larger the time interval, the closer the value of the time decay factor is to zero, so that the weight of the associated edge that has not been updated for a long time will naturally decay over time. Multiply the historical weight by the time decay factor to obtain the decayed historical weight of the associated edge; The historical weights after decay are summed with the instantaneous weight adjustment to obtain the updated associated edge weights. The updated associated edge weights are then written back to the weight storage location of the associated edge in the weighted three-part association graph, thus completing this feedback-driven closed-loop update of edge weights.

8. An AI-powered collaborative system for task assignment, used to implement the method as described in any one of claims 1-7, characterized in that, include: The graph building unit is used to acquire multimodal learning behavior data generated by the target learner within a preset learning period, perform feature encoding and relation extraction processing on the multimodal learning behavior data, construct a three-part association graph containing learner nodes, knowledge point nodes and assignment content nodes, and calculate the weight of the association edges between each node based on historical interaction data to form a weighted three-part association graph. The collaborative computing unit is configured with at least three collaborative AI agents corresponding to the learner-knowledge point association dimension, the knowledge point-assignment content association dimension, and the learner-assignment content association dimension, respectively. Each collaborative AI agent performs similarity propagation calculation on the corresponding subgraph of the weighted tripartite association graph and independently generates the association strength matrix of the corresponding dimension. The weighted aggregation unit is used to input the association strength matrix generated by each collaborative AI agent into the collaborative aggregation module, and perform element-wise weighted fusion of each association strength matrix through a linear weighted aggregation mechanism to generate a comprehensive association strength matrix. The recommendation update unit is used to calculate the comprehensive association score of each candidate task in the pre-constructed task content library based on the comprehensive association strength matrix, filter the candidate task tasks that meet the preset association threshold, generate a personalized task recommendation sequence by sorting them in descending order of comprehensive association score, and write back the interactive feedback data of the target learner to the weighted tripartite association graph to update the association edge weights.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images