A student growth portrait generation system based on a graph neural network

Abstract

This invention discloses a student growth profile generation system based on graph neural networks, comprising: a data acquisition module for acquiring task data and multi-source evaluation data to construct a student feature sequence; a mapping and labeling module for fitting the student feature sequence and generating an ability mapping sequence; a relationship extraction module for constructing a student growth relationship graph based on the student feature sequence and the ability mapping sequence; a feature interaction module for generating a node feature matrix; an aggregation and update module for performing feature aggregation on the student growth relationship graph to obtain node representation vectors; a profile generation module for generating a growth profile sequence based on the node representation vectors; and a granularity partitioning module for granularizing the growth profile sequence. This invention utilizes graph neural networks and other technologies, possessing advantages such as strong relationship modeling capabilities, high accuracy in growth analysis, and high profile generation efficiency.

Description

Technical Field

[0001] This invention relates to the field of educational data analysis technology, and in particular to a student growth profile generation system based on graph neural networks. Background Technology

[0002] With the continuous advancement of educational informatization, various teaching platforms have gradually begun to collect and manage students' behavioral data, task outcome data, and evaluation data during the learning process. By recording students' performance in course learning, project practice, and comprehensive tasks, a considerable set of growth data can be formed, which can then be used to analyze and evaluate students' abilities. To more intuitively display students' development across different ability dimensions, some educational platforms use data statistics and visualization technologies to organize student evaluation results and display student growth in the form of ability radar charts, performance curves, etc., thereby assisting teachers in conducting teaching analysis and providing learning feedback to students.

[0003] In existing technologies, student development profiles are typically obtained based on weighted calculations of course grade statistics or evaluation data. Ability scores are derived by linearly weighting different evaluation indicators, and then a visual ability map is generated based on the score results. However, these methods mainly rely on simple statistical models and lack the ability to effectively model the relationships between students across different tasks. Furthermore, student development data usually comes from multiple sources such as teacher evaluations, peer evaluations, and industry mentor evaluations. These multi-source data exhibit complex relationship structures, making it difficult for traditional methods to effectively integrate and analyze these relationships. Consequently, the generated development profiles fail to accurately reflect the true state of students' ability development.

[0004] Furthermore, there are often complex relationships between task data, ability indicators, and evaluation results in a student's development process. For example, a single task may simultaneously affect multiple ability indicators, and there may also be ability correlations between different tasks. Existing technologies typically treat these data as independent features, lacking a holistic modeling method for the relationship structure between students, tasks, and ability indicators. Therefore, it is difficult to extract key features from a student's development process in complex data environments, affecting the accuracy of the development profile generation.

[0005] Therefore, how to provide a student growth profile generation system based on graph neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a student growth profile generation system based on graph neural networks. This invention constructs a student growth relationship graph containing student nodes, task nodes, and ability indicator nodes by structuring student task data and multi-source evaluation data. It then utilizes graph neural networks to perform neighborhood aggregation and feature propagation calculations on the node features in the relationship graph, thereby extracting student ability features and generating a growth profile sequence. Simultaneously, it combines an ability space search method to evaluate and optimize the distribution of student abilities. This system enables automatic analysis of student ability structures and generation of growth profiles, possessing advantages such as strong relationship modeling capabilities, high accuracy in growth analysis, and high efficiency in profile generation.

[0007] A student growth profile generation system based on graph neural networks according to an embodiment of the present invention includes: The data acquisition module is used to acquire and preprocess student task data and multi-source evaluation data to construct student feature sequences; The mapping and labeling module is used to perform nonlinear fitting on the evaluation dimension features in the student feature sequence through the KAN network, and perform ability attribution labeling operation on each evaluation dimension based on the fitting results to generate an ability mapping sequence. The relationship extraction module is used to extract a set of nodes based on the student feature sequence and ability mapping sequence. It uses the Node2Vec algorithm to perform random walk sampling in the node set, performs embedding training operation on the sampling results through the Skip-Gram model to extract the connection relationship, and constructs a student growth relationship graph by combining the node set. The feature interaction module is used to construct the FusionNet model to perform feature fusion operations on student feature sequences, perform feature compression and mapping operations on the fusion results, generate node feature vectors, write the node feature vectors into the corresponding nodes of the student growth relationship graph, and construct the node feature matrix. The aggregation and update module is used to perform heterogeneous graph attention calculation operations on the student growth relationship graph in the HGT model based on the node feature matrix. It performs weighted aggregation of the features of the neighboring nodes of each node according to the attention results, and performs feature propagation and node update operations to obtain the node representation vector. The profile generation module is used to construct a sequence of capability indicators based on node representation vectors, and to perform capability space search and fitness evaluation operations using the MAP-Elites algorithm, and to construct a growth profile sequence based on the evaluation results. The granularity segmentation module is used to perform granularity segmentation operations on the growth profile sequence to generate growth profile data of different granularities.

[0008] Optionally, the task data includes task structure, task evidence, evidence binding information, and key task markers; the multi-source evaluation data represents teacher scores, teacher comments, student self-evaluations, peer evaluations, and industry mentor evaluations stored according to grading standards from teachers, students, and industry mentors; the preprocessing includes field alignment, missing value imputation, outlier removal, and time alignment; the node set includes three types of nodes: student, task, and ability indicator; the HGT model represents a heterogeneous graph neural network model; and the granularity includes student perspective, teacher perspective, and management perspective.

[0009] Optionally, the mapping tag module includes: The evaluation dimension features are extracted from the student feature sequence, and the evaluation dimension features are rearranged according to the order of their appearance in the task data. At the same time, the numerical range of each evaluation dimension feature is normalized to construct the evaluation dimension feature set. In the node function structure of the KAN network, piecewise calculation of node functions and continuous curve fitting are performed on the features of each evaluation dimension, and inter-layer accumulation and nonlinear mapping are performed to generate a dimension fitting sequence. Based on the set of capability indicators in the preset capability map template, capability indicator matching calculation is performed on the dimension fitting sequence. The correspondence between evaluation dimensions and capability indicators is established according to the matching relationship between the fitting results of each dimension and the set of capability indicators, and a capability mapping set is generated. Based on the ability mapping set, the ability attribution labeling operation is performed on each evaluation dimension feature in the student feature sequence. The evaluation dimension features and ability indicators are jointly labeled and aggregated according to the student identifier to generate an ability mapping sequence. The student identifier represents a unique information label used to record the student's identity.

[0010] Optionally, the piecewise calculation of the node function and the continuous curve fitting operation specifically include: Based on the numerical range of each evaluation dimension feature in the evaluation dimension feature set, an interval partitioning operation is performed on the evaluation dimension features to divide the numerical interval corresponding to each evaluation dimension feature into multiple continuous segmented intervals, and a node function interval sequence is established according to the order of the segmented intervals. In the node function structure of the KAN network, the node function is segmented and calculated for each evaluation dimension feature according to the node function interval sequence. The evaluation dimension feature within each segment interval is mapped according to the preset node function, and the node function calculation results are recorded in the order of the segment intervals to generate a function result sequence. A set of fitted curves is constructed based on the sequence of function results, and the curvature detection operation is performed using the Laplace operator to extract the change feature sequence corresponding to each evaluation dimension. A dimensional feature set is constructed based on the function result sequence and the change feature sequence, and the SimSiam method is used to perform comparative learning training on the dimensional feature set to extract a stable dimensional representation sequence. Based on the dimensional representation sequence, perform inter-layer accumulation and nonlinear transformation operations on the function result sequence to generate a dimensional fitting sequence.

[0011] Optionally, the curvature detection operation and contrastive learning training specifically include: Based on the order of each fitted curve in the sequence, adjacent curve point pairs are established, and the difference between the numerical changes between the curve points is calculated to generate a sequence of change amplitudes. The Laplace operator is applied to the fitted curve set based on the change amplitude sequence. The second-order difference operation is performed between each curve point and its corresponding preceding and following curve points to obtain the curvature value sequence of the corresponding curve points. Based on the curvature value sequence, the set of fitted curves is judged for curvature changes. Positions with curvature values ​​greater than a preset curvature threshold are marked as curvature change nodes, and a curvature marking sequence is generated based on the marking results. Extract the changing nodes from the fitted curve set based on the curvature marker sequence, and construct the changing node set according to the arrangement order in the fitted curve set; Based on the set of changing nodes, perform a trend correlation operation on the features of each evaluation dimension to generate a sequence of changing features; A set of dimensional features is constructed based on the sequence of function results and the sequence of change features. A feature pairing sequence is established according to the order of the evaluation dimensional features in the set of dimensional features. The function results and change features corresponding to the same evaluation dimensional feature are combined into feature representation pairs. Feature enhancement operations are performed on the dimensional feature set based on the feature pairing sequence. Random masking and numerical perturbation are applied to the function result sequence and the changed feature sequence, respectively, and a comparison feature sequence is generated according to the order of the evaluation dimensional features. The SimSiam method is used to perform mapping calculations on the contrast feature sequences to generate the representation feature sequence and the predicted representation sequence, respectively. Perform cosine similarity calculation on the representation feature sequence and the predicted representation sequence, and generate a consistency score sequence based on the cosine similarity; Based on the consistency score sequence, a stable representation filtering operation is performed on the representation feature sequence. Representation features with a consistency score higher than a preset consistency threshold are used as stable dimension representations, and a dimension representation sequence is generated.

[0012] Optionally, the relationship extraction module includes: Extract a node set based on the student feature sequence and ability mapping sequence, and perform node identification encoding and sequential numbering operations on the student nodes, task nodes and ability indicator nodes in the node set to generate a node sequence; The Node2Vec algorithm is used to perform random walk sampling on the node sequence, and the visited nodes are recorded in the order of the random walk path to generate the node walk sequence; Input the node walking sequence into the Skip-Gram model, extract each node from the node walking sequence according to the access order, and establish a node training sequence with each node as the center node. Based on the positional relationship of nodes in the node walking sequence, a fixed-size neighborhood range is determined, and a set of neighborhood nodes corresponding to each central node is generated. The number of times nodes co-occur is calculated based on the set of each central node and its neighboring nodes, and a co-occurrence matrix is ​​constructed. Frequency statistics are then performed on the co-occurrence matrix to generate a co-occurrence frequency sequence. The embedding vector is initialized for each node based on the co-occurrence frequency sequence, and the vector mapping operation between the center node and the neighboring nodes is performed based on the co-occurrence matrix to generate the node representation sequence. Based on the node representation sequence, perform vector update operations on the center node and neighboring nodes, and repeat the co-occurrence statistics, vector mapping and update operations according to the order of arrangement in the node walk sequence to generate the embedding representation sequence; Calculate the embedding similarity relationship between nodes based on the embedding representation sequence and extract the node connection relationship to generate the edge relationship sequence; Construct a student growth relationship graph based on the node set and the node edge relationship sequence.

[0013] Optionally, the feature interaction module includes: A FusionNet model containing a parallel feature extraction structure is constructed. In the parallel feature extraction structure, convolutional mapping and activation operations are performed on the task data and multi-source evaluation data in the student feature sequence, respectively, to generate multiple sets of branch feature sequences. Based on the arrangement order of the feature sequences of each branch in the FusionNet model, a weighted fusion calculation is performed on multiple sets of branch feature sequences, and the fusion result is subjected to residual superposition processing to generate a fused feature sequence. The fused feature sequence is subjected to feature compression. Fully connected mapping and dimension compression are performed according to the feature dimension arrangement order in the fused feature sequence, and the compressed feature sequence is obtained through normalization. Each feature in the compressed feature sequence is mapped and assigned according to the node identifiers of student nodes, task nodes, and ability indicator nodes to generate node feature vectors. Write the node feature vectors into the corresponding student nodes, task nodes, and ability indicator nodes in the student growth relationship graph, and perform matrix arrangement operation on all node feature vectors in the order of the node set to construct the node feature matrix.

[0014] Optionally, the process of generating the branch feature sequence specifically includes: Obtain the student feature sequence, and perform a reordering operation on each feature item in the student feature sequence according to the student identifier, task identifier and evaluation dimension identifier. Based on the feature source, divide the student feature sequence into task feature sequence and evaluation feature sequence. The task feature sequence and the evaluation feature sequence are respectively input into the parallel feature extraction structure in the FusionNet model. In each feature extraction channel, the convolution window sliding calculation is performed according to the feature arrangement order. Convolution mapping calculation is performed in each convolution window through the convolution kernel of preset size to generate convolution feature sequence. Perform the ReLU activation operation on the convolutional feature sequence to generate an activated feature sequence; Two-dimensional empirical mode decomposition is performed on each activation feature in the activation feature sequence. Upper and lower envelope curves are constructed based on the local extrema of each activation feature, and intrinsic mode functions are calculated based on the upper and lower envelope curves to generate a mode sequence. Based on the modal sequence, perform mode screening and mode combination operations on each intrinsic mode function, and perform feature reconstruction calculation on the combination results; The reconstruction results are organized according to the branch arrangement order in the FusionNet model to generate a branch feature sequence.

[0015] Optionally, the aggregation update module includes: Read the node feature matrix and the student growth relationship graph, perform node type classification operation on the node feature matrix according to the node identifiers of student nodes, task nodes and ability indicator nodes, and establish a node type sequence according to the arrangement order in the node set; Extract the set of neighboring nodes of each node in the student growth relationship graph based on the node type sequence, and establish an adjacency relationship sequence based on the connection relationship between each node. In the HGT model, heterogeneous graph attention calculation is performed on each node according to the adjacency sequence, attention weights are initialized for the connection relationships between different types of nodes according to the node type sequence, and weight allocation is performed on each neighboring node according to the adjacency sequence to generate an attention weight sequence. Based on the attention weight sequence, a weighted aggregation calculation is performed on the node features in the neighborhood node set to generate an aggregated feature sequence; Based on the aggregated feature sequence and the node feature matrix, a feature propagation operation is performed. The aggregated feature sequence is written into the corresponding node according to the arrangement order in the node set. Feature superposition and feature update operations are performed on each node to generate an updated node feature sequence. Based on the updated node feature sequence, multi-layer feature propagation calculation is performed in the HGT model. By repeatedly performing a preset number of rounds of neighbor node extraction, attention weight calculation, weighted aggregation and node update operations, a stable node feature sequence is generated. Based on a stable sequence of node features, node representation extraction is performed on student nodes, task nodes, and ability indicator nodes to construct node representation vectors.

[0016] Optionally, the image generation module includes: Extract student node representations, task node representations, and ability indicator node representations according to the order of the node representation vectors in the node set, and perform node type separation operation to generate student representation sequence, task representation sequence, and ability representation sequence; A set of capability indicators is established based on the capability representation sequence. According to the capability dimension arrangement order in the preset capability map template, the capability dimension classification operation is performed on the capability representation sequence. Each capability indicator in the capability representation sequence is grouped and organized according to the capability dimension to generate a capability indicator sequence. Perform association mapping calculations on the student representation sequence, task representation sequence, and ability indicator sequence, and construct an association feature sequence based on the association mapping results; The MAP-Elites algorithm is used to construct a capability space grid based on the capability index sequence, and the associated feature sequence is mapped to the corresponding capability space grid to generate a capability distribution sequence. Based on the capability distribution sequence, a fitness evaluation operation is performed on each capability spatial location, and a fitness calculation is performed on each capability spatial location based on the associated feature sequence to generate a fitness sequence; Based on the fitness sequence, a capability space search operation is performed on the capability distribution sequence. The associated features with the highest fitness values ​​are selected in the capability space grid to construct a growth profile sequence.

[0017] The beneficial effects of this invention are: First, this invention acquires student task data and multi-source evaluation data, establishes a mapping relationship between evaluation dimensions and ability indicators, and constructs a student growth relationship graph containing student nodes, task nodes, and ability indicator nodes. Based on this, a graph neural network is used to perform neighborhood information aggregation and feature propagation processing on the node features in the relationship graph, thereby realizing the modeling of complex relationships in student growth data. Compared with the existing technology that only uses statistical methods or linear weighting to calculate abilities, this invention can extract richer ability features from structured relationship data and improve the accuracy of student growth profile generation.

[0018] Secondly, this invention generates a growth profile sequence that reflects the student's ability structure by performing feature fusion and embedding representation learning on the student's feature sequence and using the ability space search method to perform fitness assessment and space search processing on the student's ability indicators. This allows the student's development in different ability dimensions to be expressed through a unified data structure. Compared with the traditional growth profile generated based on a single evaluation indicator, this invention can comprehensively analyze multi-source evaluation data, improving the objectivity and completeness of the student ability analysis results.

[0019] Finally, based on the generated growth profile sequence, this invention performs granular segmentation processing on the growth profile sequence, generating growth profile data of different granularities according to the student's perspective, teacher's perspective, and management perspective. This allows different roles to view the corresponding level of ability analysis results according to their own needs, thereby providing students with intuitive growth feedback, teachers with a basis for learning analysis, and managers with overall ability structure analysis data. It has the advantages of strong growth data analysis capabilities, rich profile expression forms, and wide applicability. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a module structure diagram of a student growth profile generation system based on graph neural networks proposed in this invention; Figure 2 This invention presents a flowchart of the student growth relationship graph embedding training process for a student growth profile generation system based on graph neural networks. Figure 3 This is a flowchart of the growth profile generation and ability sequence construction of a student growth profile generation system based on graph neural networks proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figures 1-3 A student growth profile generation system based on graph neural networks includes: The data acquisition module is used to acquire and preprocess student task data and multi-source evaluation data to construct student feature sequences; The mapping and labeling module is used to perform nonlinear fitting on the evaluation dimension features in the student feature sequence through the KAN network, and perform ability attribution labeling operation on each evaluation dimension based on the fitting results to generate an ability mapping sequence. The relationship extraction module is used to extract a set of nodes based on the student feature sequence and ability mapping sequence. It uses the Node2Vec algorithm to perform random walk sampling in the node set, performs embedding training operation on the sampling results through the Skip-Gram model to extract the connection relationship, and constructs a student growth relationship graph by combining the node set. The feature interaction module is used to construct the FusionNet model to perform feature fusion operations on student feature sequences, perform feature compression and mapping operations on the fusion results, generate node feature vectors, write the node feature vectors into the corresponding nodes of the student growth relationship graph, and construct the node feature matrix. The aggregation and update module is used to perform heterogeneous graph attention calculation operations on the student growth relationship graph in the HGT model based on the node feature matrix. It performs weighted aggregation of the features of the neighboring nodes of each node according to the attention results, and performs feature propagation and node update operations to obtain the node representation vector. The profile generation module is used to construct a sequence of capability indicators based on node representation vectors, and to perform capability space search and fitness evaluation operations using the MAP-Elites algorithm, and to construct a growth profile sequence based on the evaluation results. The granularity segmentation module is used to perform granularity segmentation operations on the growth profile sequence to generate growth profile data of different granularities.

[0023] In this embodiment, the task data includes task structure, task evidence, evidence binding information, and key task markers. The multi-source evaluation data represents teacher scores, teacher comments, student self-evaluations, peer evaluations, and industry mentor evaluations stored according to grading standards. Preprocessing includes field alignment, missing value imputation, outlier removal, and time alignment. The node set includes three types of nodes: student, task, and ability indicator. The HGT model represents a heterogeneous graph neural network model with granularity including student perspective, teacher perspective, and management perspective.

[0024] In this embodiment, the mapping tag module includes: The evaluation dimension features are extracted from the student feature sequence, and the evaluation dimension features are rearranged according to the order of their appearance in the task data. At the same time, the numerical range of each evaluation dimension feature is normalized to construct the evaluation dimension feature set. In the node function structure of the KAN network, piecewise calculation of node functions and continuous curve fitting are performed on the features of each evaluation dimension, and inter-layer accumulation and nonlinear mapping are performed to generate a dimension fitting sequence. Based on the set of capability indicators in the preset capability map template, capability indicator matching calculation is performed on the dimension fitting sequence. The correspondence between evaluation dimensions and capability indicators is established according to the matching relationship between the fitting results of each dimension and the set of capability indicators, and a capability mapping set is generated. Based on the ability mapping set, the ability attribution labeling operation is performed on each evaluation dimension feature in the student feature sequence. The evaluation dimension features and ability indicators are jointly labeled and aggregated according to the student identifier to generate the ability mapping sequence. Here, the student identifier represents a unique information label used to record the student's identity.

[0025] In this embodiment, the piecewise calculation of node functions and the continuous curve fitting operation specifically include: Based on the numerical range of each evaluation dimension feature in the evaluation dimension feature set, an interval partitioning operation is performed on the evaluation dimension features to divide the numerical interval corresponding to each evaluation dimension feature into multiple continuous segmented intervals, and a node function interval sequence is established according to the order of the segmented intervals. In the node function structure of the KAN network, the node function is segmented and calculated for each evaluation dimension feature according to the node function interval sequence. The evaluation dimension feature within each segment interval is mapped according to the preset node function, and the node function calculation results are recorded in the order of the segment intervals to generate a function result sequence. A set of fitted curves is constructed based on the sequence of function results, and the curvature detection operation is performed using the Laplace operator to extract the change feature sequence corresponding to each evaluation dimension. A dimensional feature set is constructed based on the function result sequence and the change feature sequence, and the SimSiam method is used to perform comparative learning training on the dimensional feature set to extract a stable dimensional representation sequence. Based on the dimensional representation sequence, perform inter-layer accumulation and nonlinear transformation operations on the function result sequence to generate a dimensional fitting sequence.

[0026] In this embodiment, the capability index matching calculation specifically includes: Read the set of capability indicators from the preset capability map template and establish a capability indicator sequence according to the order of the capability indicators in the capability map template; A similarity assessment operation is performed on the dimensional fitting results and the capability index sequence. The Mahalanobis distance is calculated as the similarity score based on the numerical difference between the dimensional fitting results and each capability index. The similarity scores are then sorted from high to low to generate the index matching sequence. Based on the indicator matching sequence, perform a correspondence construction operation on the features of each evaluation dimension, establish an association between each evaluation dimension feature and the capability indicators in the corresponding indicator matching sequence whose similarity scores are higher than the preset reference score threshold, and generate a capability mapping set between evaluation dimensions and capability indicators. The ability index weights are assigned to each evaluation dimension based on the ability mapping set, and the evaluation dimensions are sorted according to their order in the student feature sequence to obtain the ability index matching results.

[0027] In this embodiment, the curvature detection operation and comparative learning training specifically include: Based on the order of each fitted curve in the sequence, adjacent curve point pairs are established, and the difference between the numerical changes between the curve points is calculated to generate a sequence of change amplitudes. The Laplace operator is applied to the fitted curve set based on the change amplitude sequence. The second-order difference operation is performed between each curve point and its corresponding preceding and following curve points to obtain the curvature value sequence of the corresponding curve points. Based on the curvature value sequence, the set of fitted curves is judged for curvature changes. Positions with curvature values ​​greater than a preset curvature threshold are marked as curvature change nodes, and a curvature marking sequence is generated based on the marking results. Extract the changing nodes from the fitted curve set based on the curvature marker sequence, and construct the changing node set according to the arrangement order in the fitted curve set; Based on the set of changing nodes, perform a trend correlation operation on the features of each evaluation dimension to generate a sequence of changing features; A set of dimensional features is constructed based on the sequence of function results and the sequence of change features. A feature pairing sequence is established according to the order of the evaluation dimensional features in the set of dimensional features. The function results and change features corresponding to the same evaluation dimensional feature are combined into feature representation pairs. Feature enhancement operations are performed on the dimensional feature set based on the feature pairing sequence. Random masking and numerical perturbation are applied to the function result sequence and the changed feature sequence, respectively, and a comparison feature sequence is generated according to the order of the evaluation dimensional features. The SimSiam method is used to perform mapping calculations on the contrast feature sequences to generate the representation feature sequence and the predicted representation sequence, respectively. Perform cosine similarity calculation on the representation feature sequence and the predicted representation sequence, and generate a consistency score sequence based on the cosine similarity; Based on the consistency score sequence, a stable representation filtering operation is performed on the representation feature sequence. Representation features with a consistency score higher than a preset consistency threshold are used as stable dimension representations, and a dimension representation sequence is generated.

[0028] In this embodiment, the relationship extraction module includes: Extract a node set based on the student feature sequence and ability mapping sequence, and perform node identification encoding and sequential numbering operations on the student nodes, task nodes and ability indicator nodes in the node set to generate a node sequence; The Node2Vec algorithm is used to perform random walk sampling on the node sequence, and the visited nodes are recorded in the order of the random walk path to generate the node walk sequence; Input the node walking sequence into the Skip-Gram model, extract each node from the node walking sequence according to the access order, and establish a node training sequence with each node as the center node. Based on the positional relationship of nodes in the node walking sequence, a fixed-size neighborhood range is determined, and a set of neighborhood nodes corresponding to each central node is generated. The number of times nodes co-occur is calculated based on the set of each central node and its neighboring nodes, and a co-occurrence matrix is ​​constructed. Frequency statistics are then performed on the co-occurrence matrix to generate a co-occurrence frequency sequence. The embedding vector is initialized for each node based on the co-occurrence frequency sequence, and the vector mapping operation between the center node and the neighboring nodes is performed based on the co-occurrence matrix to generate the node representation sequence. Based on the node representation sequence, perform vector update operations on the center node and neighboring nodes, and repeat the co-occurrence statistics, vector mapping and update operations according to the order of arrangement in the node walk sequence to generate the embedding representation sequence; Calculate the embedding similarity relationship between nodes based on the embedding representation sequence and extract the node connection relationship to generate the edge relationship sequence; Construct a student growth relationship graph based on the node set and the node edge relationship sequence.

[0029] In this embodiment, the feature interaction module includes: A FusionNet model containing a parallel feature extraction structure is constructed. In the parallel feature extraction structure, convolutional mapping and activation operations are performed on the task data and multi-source evaluation data in the student feature sequence, respectively, to generate multiple sets of branch feature sequences. Based on the arrangement order of the feature sequences of each branch in the FusionNet model, a weighted fusion calculation is performed on multiple sets of branch feature sequences, and the fusion result is subjected to residual superposition processing to generate a fused feature sequence. The fused feature sequence is subjected to feature compression. Fully connected mapping and dimension compression are performed according to the feature dimension arrangement order in the fused feature sequence, and the compressed feature sequence is obtained through normalization. Each feature in the compressed feature sequence is mapped and assigned according to the node identifiers of student nodes, task nodes, and ability indicator nodes to generate node feature vectors. Write the node feature vectors into the corresponding student nodes, task nodes, and ability indicator nodes in the student growth relationship graph, and perform matrix arrangement operation on all node feature vectors in the order of the node set to construct the node feature matrix.

[0030] In this embodiment, the process of generating the branch feature sequence specifically includes: Obtain the student feature sequence, and perform a reordering operation on each feature item in the student feature sequence according to the student identifier, task identifier and evaluation dimension identifier. Based on the feature source, divide the student feature sequence into task feature sequence and evaluation feature sequence. The task feature sequence and the evaluation feature sequence are respectively input into the parallel feature extraction structure in the FusionNet model. In each feature extraction channel, the convolution window sliding calculation is performed according to the feature arrangement order. Convolution mapping calculation is performed in each convolution window through the convolution kernel of preset size to generate convolution feature sequence. Perform the ReLU activation operation on the convolutional feature sequence to generate an activated feature sequence; Two-dimensional empirical mode decomposition is performed on each activation feature in the activation feature sequence. Upper and lower envelope curves are constructed based on the local extrema of each activation feature, and intrinsic mode functions are calculated based on the upper and lower envelope curves to generate a mode sequence. Based on the modal sequence, perform mode screening and mode combination operations on each intrinsic mode function, and perform feature reconstruction calculation on the combination results; The reconstruction results are organized according to the branch arrangement order in the FusionNet model to generate a branch feature sequence.

[0031] In this embodiment, the aggregation update module includes: Read the node feature matrix and the student growth relationship graph, perform node type classification operation on the node feature matrix according to the node identifiers of student nodes, task nodes and ability indicator nodes, and establish a node type sequence according to the arrangement order in the node set; Extract the set of neighboring nodes of each node in the student growth relationship graph based on the node type sequence, and establish an adjacency relationship sequence based on the connection relationship between each node. In the HGT model, heterogeneous graph attention calculation is performed on each node according to the adjacency sequence, attention weights are initialized for the connection relationships between different types of nodes according to the node type sequence, and weight allocation is performed on each neighboring node according to the adjacency sequence to generate an attention weight sequence. Based on the attention weight sequence, a weighted aggregation calculation is performed on the node features in the neighborhood node set to generate an aggregated feature sequence; Based on the aggregated feature sequence and the node feature matrix, a feature propagation operation is performed. The aggregated feature sequence is written into the corresponding node according to the arrangement order in the node set. Feature superposition and feature update operations are performed on each node to generate an updated node feature sequence. Based on the updated node feature sequence, multi-layer feature propagation calculation is performed in the HGT model. By repeatedly performing a preset number of rounds of neighbor node extraction, attention weight calculation, weighted aggregation and node update operations, a stable node feature sequence is generated. Based on a stable sequence of node features, node representation extraction is performed on student nodes, task nodes, and ability indicator nodes to construct node representation vectors.

[0032] In this embodiment, the image generation module includes: Extract student node representations, task node representations, and ability indicator node representations according to the order of the node representation vectors in the node set, and perform node type separation operation to generate student representation sequence, task representation sequence, and ability representation sequence; A set of capability indicators is established based on the capability representation sequence. According to the capability dimension arrangement order in the preset capability map template, the capability dimension classification operation is performed on the capability representation sequence. Each capability indicator in the capability representation sequence is grouped and organized according to the capability dimension to generate a capability indicator sequence. Perform association mapping calculations on the student representation sequence, task representation sequence, and ability indicator sequence, and construct an association feature sequence based on the association mapping results; The MAP-Elites algorithm is used to construct a capability space grid based on the capability index sequence, and the associated feature sequence is mapped to the corresponding capability space grid to generate a capability distribution sequence. Based on the capability distribution sequence, a fitness evaluation operation is performed on each capability spatial location, and a fitness calculation is performed on each capability spatial location based on the associated feature sequence to generate a fitness sequence; Based on the fitness sequence, a capability space search operation is performed on the capability distribution sequence. The associated features with the highest fitness values ​​are selected in the capability space grid to construct a growth profile sequence.

[0033] In this embodiment, the process of generating the capability distribution sequence specifically includes: A capability space dimension set is established based on the order of capability dimensions in the capability index sequence, and the capability space is divided into intervals according to the value range of each capability dimension, thereby constructing a capability space grid composed of multiple capability space units in the capability space. Extract each associated feature from the associated feature sequence, and perform a capability dimension mapping operation on each associated feature according to the capability dimension correspondence in the capability index sequence to generate a capability feature sequence; Based on the set of capability space dimensions, spatial location calculation is performed on each capability feature in the capability feature sequence, and the spatial location of the capability feature in the capability space grid is determined by combining the interval division results corresponding to each capability dimension. Perform a grid writing operation on each capability feature according to its spatial location, and record the spatial location number corresponding to each capability feature according to the arrangement order in the capability space grid to generate a capability location sequence. Based on the capability location sequence, feature aggregation operation is performed on each capability space unit to combine and organize the associated features mapped to the same capability space unit, generating a capability distribution sequence.

[0034] In this embodiment, the granularity division module includes: Extract the capability indicator data corresponding to each capability dimension according to the order of the growth profile sequence in the capability dimension, and perform a granularity identifier matching operation on the growth profile sequence according to the preset granularity template to generate a granularity identifier sequence. The growth profile sequence is classified according to the granularity identifier sequence, and student perspective set, teacher perspective set and management perspective set are established according to the classification rules in the granularity template. Based on the student perspective set, the ability indicator data in the growth profile sequence are reorganized into ability dimensions, and the ability indicator data corresponding to each ability dimension are extracted according to the order of the ability dimensions to generate the student perspective profile sequence. Based on the teacher's perspective set, the ability indicator data in the growth profile sequence are reorganized into a capability diagnosis, and the capability indicator data corresponding to each capability dimension are extracted in the order of capability dimension arrangement to generate the teacher's perspective profile sequence. Based on the management perspective set, the group capability structure reorganization operation is performed on the capability indicator data in the growth profile sequence, and the capability indicator data corresponding to each capability dimension is extracted according to the order of capability dimensions to generate the management perspective profile sequence. Growth profile data of different granularities are constructed based on the growth profile sequence from three perspectives.

[0035] Example 1: To verify the feasibility of this invention in practice, it was applied to a student growth data analysis scenario within a comprehensive teaching and training environment. In this environment, students continuously complete various tasks and activities during their learning process, including course tasks, practical tasks, and comprehensive ability training tasks. During these tasks, students submit document deliverables, presentation materials, code outputs, video materials, and research reports, and receive multi-source evaluations from teachers, peers, and industry mentors. As the learning process progresses, students generate a large amount of task and evaluation data. However, in traditional data analysis methods, this data is usually scattered across different recording systems. When analyzing changes in student abilities, teachers mainly rely on single-score statistics or simple average ratings, making it difficult to comprehensively analyze student performance across different tasks or clearly reflect the correlation between student abilities, thus hindering the accurate depiction of student growth.

[0036] In this scenario, the student growth profile generation system based on graph neural networks of this invention is deployed to uniformly collect and process task data and multi-source evaluation data of students during the learning process. The system first acquires task structure information, task achievement evidence, and evaluation data for students in various learning tasks, and performs field alignment, outlier removal, and time alignment processing on the data to construct a student feature sequence. Subsequently, a mapping mechanism is used to establish a mapping relationship between evaluation dimensions and ability indicators. For example, evaluation dimensions such as expression quality, teamwork, problem-solving ability, and technical implementation ability are mapped to ability indicators such as expression ability, collaboration ability, professional ability, and innovation ability. After completing the mapping process, the system extracts student nodes, task nodes, and ability indicator nodes based on the student feature sequence and ability mapping sequence, and constructs a student growth relationship graph among the node sets.

[0037] After the growth relationship graph is constructed, the system uses an embedding training method to learn node relationships. It obtains potential association paths between nodes through random walk sampling and extracts node connectivity using an embedding representation learning method. Based on this, the system performs feature fusion processing on student task data and multi-source evaluation data, writing the fused feature vectors into the corresponding nodes in the growth relationship graph to form a node feature matrix. Subsequently, a heterogeneous graph neural network is used to perform neighborhood information aggregation and feature propagation processing on the node features in the relationship graph, enabling student nodes to receive information from task nodes and ability indicator nodes, thereby obtaining stable node representation vectors.

[0038] After obtaining the node representation vectors, the system constructs a sequence of ability indicators based on the student node representations and the ability indicator node representations. It then performs an ability distribution assessment in the ability space, selecting the optimal combination of ability distributions through an ability space search method, thereby generating a sequence of student growth profiles. This growth profile sequence clearly reflects a student's comprehensive performance across multiple ability dimensions, including expression, collaboration, professional skills, innovative problem-solving, and professional ethics. Subsequently, the system performs granularity partitioning of the growth profile sequence according to preset granularity rules, generating growth profile data from student, teacher, and management perspectives, thus meeting the needs of different roles in student growth analysis. Students can view their own ability changes through the student-perspective growth profile, teachers can analyze student ability structure and learning problems through the teacher-perspective growth profile, and administrators can analyze the overall ability distribution through the management-perspective growth profile.

[0039] In practical applications, this invention analyzes and processes a certain scale of student growth data. The system collects and processes task data and evaluation data from multiple students across multiple learning tasks. By constructing a student growth relationship graph and utilizing graph neural networks for feature learning, it generates student growth profile data. To verify the effectiveness of this invention, the student growth profile results generated by the system were compared with the ability analysis results generated by traditional statistical methods. Indicators such as ability analysis accuracy, ability structure recognition rate, and profile generation efficiency were recorded. Specific data are shown in the table below: Table 1. Statistical Table of Student Growth Profile Generation System Operation Performance

[0040]

[0041] As shown in Table 1, with the increasing number of student tasks and the scale of evaluation data, the growth profile generated by this invention significantly outperforms traditional statistical methods in terms of ability recognition accuracy, increasing the average ability recognition accuracy from approximately 72% to approximately 90%. Furthermore, in terms of ability structure recognition, this invention, by constructing a student growth relationship graph and utilizing graph neural networks for feature propagation processing, can more accurately identify the relationships between students across different ability dimensions, increasing the ability structure recognition rate from approximately 64% to approximately 88%. Regarding profile generation efficiency, this invention, through graph structure feature learning and node feature aggregation calculation, reduces the average profile generation time from approximately 5.5 seconds to approximately 2.3 seconds, indicating that this invention not only improves the accuracy of ability analysis but also significantly enhances the efficiency of growth profile generation.

[0042] In summary, practical application in student growth data analysis scenarios demonstrates that this invention can effectively integrate student task data with multi-source evaluation data, model the relationship between students, tasks, and ability indicators through graph neural networks, thereby generating a more accurate student growth profile, and provide corresponding levels of growth analysis data for different roles through a granularity partitioning mechanism, thus possessing high practical application value.

[0043] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A student growth profile generation system based on graph neural networks, characterized in that, include: The data acquisition module is used to acquire and preprocess student task data and multi-source evaluation data to construct student feature sequences; The mapping and labeling module is used to perform nonlinear fitting on the evaluation dimension features in the student feature sequence through the KAN network, and perform ability attribution labeling operation on each evaluation dimension based on the fitting results to generate an ability mapping sequence. The relationship extraction module is used to extract a set of nodes based on the student feature sequence and ability mapping sequence. It uses the Node2Vec algorithm to perform random walk sampling in the node set, performs embedding training operation on the sampling results through the Skip-Gram model to extract the connection relationship, and constructs a student growth relationship graph by combining the node set. The feature interaction module is used to construct the FusionNet model to perform feature fusion operations on student feature sequences, perform feature compression and mapping operations on the fusion results, generate node feature vectors, write the node feature vectors into the corresponding nodes of the student growth relationship graph, and construct the node feature matrix. The aggregation and update module is used to perform heterogeneous graph attention calculation operations on the student growth relationship graph in the HGT model based on the node feature matrix. It performs weighted aggregation of the features of the neighboring nodes of each node according to the attention results, and performs feature propagation and node update operations to obtain the node representation vector. The profile generation module is used to construct a sequence of capability indicators based on node representation vectors, and to perform capability space search and fitness evaluation operations using the MAP-Elites algorithm, and to construct a growth profile sequence based on the evaluation results. The granularity segmentation module is used to perform granularity segmentation operations on the growth profile sequence to generate growth profile data of different granularities.

2. The student growth profile generation system based on graph neural networks according to claim 1, characterized in that, The task data includes task structure, task evidence, evidence binding information, and key task markers. The multi-source evaluation data represents teacher scores, teacher comments, student self-evaluations, peer evaluations, and industry mentor evaluations stored according to grading standards from teachers, students, and industry mentors. The preprocessing includes field alignment, missing value imputation, outlier removal, and time alignment. The node set contains three types of nodes: student, task, and ability indicator. The HGT model represents a heterogeneous graph neural network model. The granularity includes student perspective, teacher perspective, and management perspective.

3. The student growth profile generation system based on graph neural networks according to claim 1, characterized in that, The mapping tag module includes: The evaluation dimension features are extracted from the student feature sequence, and the evaluation dimension features are rearranged according to the order of their appearance in the task data. At the same time, the numerical range of each evaluation dimension feature is normalized to construct the evaluation dimension feature set. In the node function structure of the KAN network, piecewise calculation of node functions and continuous curve fitting are performed on the features of each evaluation dimension, and inter-layer accumulation and nonlinear mapping are performed to generate a dimension fitting sequence. Based on the set of capability indicators in the preset capability map template, capability indicator matching calculation is performed on the dimension fitting sequence. The correspondence between evaluation dimensions and capability indicators is established according to the matching relationship between the fitting results of each dimension and the set of capability indicators, and a capability mapping set is generated. Based on the ability mapping set, the ability attribution labeling operation is performed on each evaluation dimension feature in the student feature sequence. The evaluation dimension features and ability indicators are jointly labeled and aggregated according to the student identifier to generate an ability mapping sequence. The student identifier represents a unique information label used to record the student's identity.

4. The student growth profile generation system based on graph neural networks according to claim 3, characterized in that, The piecewise calculation of node functions and continuous curve fitting operations specifically include: Based on the numerical range of each evaluation dimension feature in the evaluation dimension feature set, an interval partitioning operation is performed on the evaluation dimension features to divide the numerical interval corresponding to each evaluation dimension feature into multiple continuous segmented intervals, and a node function interval sequence is established according to the order of the segmented intervals. In the node function structure of the KAN network, the node function is segmented and calculated for each evaluation dimension feature according to the node function interval sequence. The evaluation dimension feature within each segment interval is mapped according to the preset node function, and the node function calculation results are recorded in the order of the segment intervals to generate a function result sequence. A set of fitted curves is constructed based on the sequence of function results, and the curvature detection operation is performed using the Laplace operator to extract the change feature sequence corresponding to each evaluation dimension. A dimensional feature set is constructed based on the function result sequence and the change feature sequence, and the SimSiam method is used to perform comparative learning training on the dimensional feature set to extract a stable dimensional representation sequence. Based on the dimensional representation sequence, perform inter-layer accumulation and nonlinear transformation operations on the function result sequence to generate a dimensional fitting sequence.

5. The student growth profile generation system based on graph neural networks according to claim 4, characterized in that, The curvature detection operation and comparative learning training specifically include: Based on the order of each fitted curve in the sequence, adjacent curve point pairs are established, and the difference between the numerical changes between the curve points is calculated to generate a sequence of change amplitudes. The Laplace operator is applied to the fitted curve set based on the change amplitude sequence. The second-order difference operation is performed between each curve point and its corresponding preceding and following curve points to obtain the curvature value sequence of the corresponding curve points. Based on the curvature value sequence, the set of fitted curves is judged for curvature changes. Positions with curvature values ​​greater than a preset curvature threshold are marked as curvature change nodes, and a curvature marking sequence is generated based on the marking results. Extract the changing nodes from the fitted curve set based on the curvature marker sequence, and construct the changing node set according to the arrangement order in the fitted curve set; Based on the set of changing nodes, perform a trend correlation operation on the features of each evaluation dimension to generate a sequence of changing features; A set of dimensional features is constructed based on the sequence of function results and the sequence of change features. A feature pairing sequence is established according to the order of the evaluation dimensional features in the set of dimensional features. The function results and change features corresponding to the same evaluation dimensional feature are combined into feature representation pairs. Feature enhancement operations are performed on the dimensional feature set based on the feature pairing sequence. Random masking and numerical perturbation are applied to the function result sequence and the changed feature sequence, respectively, and a comparison feature sequence is generated according to the order of the evaluation dimensional features. The SimSiam method is used to perform mapping calculations on the contrast feature sequences to generate the representation feature sequence and the predicted representation sequence, respectively. Perform cosine similarity calculation on the representation feature sequence and the predicted representation sequence, and generate a consistency score sequence based on the cosine similarity; Based on the consistency score sequence, a stable representation filtering operation is performed on the representation feature sequence. Representation features with a consistency score higher than a preset consistency threshold are used as stable dimension representations, and a dimension representation sequence is generated.

6. The student growth profile generation system based on graph neural networks according to claim 1, characterized in that, The relationship extraction module includes: Extract a node set based on the student feature sequence and ability mapping sequence, and perform node identification encoding and sequential numbering operations on the student nodes, task nodes and ability indicator nodes in the node set to generate a node sequence; The Node2Vec algorithm is used to perform random walk sampling on the node sequence, and the visited nodes are recorded in the order of the random walk path to generate the node walk sequence; Input the node walking sequence into the Skip-Gram model, extract each node from the node walking sequence according to the access order, and establish a node training sequence with each node as the center node. Based on the positional relationship of nodes in the node walking sequence, a fixed-size neighborhood range is determined, and a set of neighborhood nodes corresponding to each central node is generated. The number of times nodes co-occur is calculated based on the set of each central node and its neighboring nodes, and a co-occurrence matrix is ​​constructed. Frequency statistics are then performed on the co-occurrence matrix to generate a co-occurrence frequency sequence. The embedding vector is initialized for each node based on the co-occurrence frequency sequence, and the vector mapping operation between the center node and the neighboring nodes is performed based on the co-occurrence matrix to generate the node representation sequence. Based on the node representation sequence, perform vector update operations on the center node and neighboring nodes, and repeat the co-occurrence statistics, vector mapping and update operations according to the order of arrangement in the node walk sequence to generate the embedding representation sequence; Calculate the embedding similarity relationship between nodes based on the embedding representation sequence and extract the node connection relationship to generate the edge relationship sequence; Construct a student growth relationship graph based on the node set and the node edge relationship sequence.

7. The student growth profile generation system based on graph neural networks according to claim 1, characterized in that, The feature interaction module includes: A FusionNet model containing a parallel feature extraction structure is constructed. In the parallel feature extraction structure, convolutional mapping and activation operations are performed on the task data and multi-source evaluation data in the student feature sequence, respectively, to generate multiple sets of branch feature sequences. Based on the arrangement order of the feature sequences of each branch in the FusionNet model, a weighted fusion calculation is performed on multiple sets of branch feature sequences, and the fusion result is subjected to residual superposition processing to generate a fused feature sequence. The fused feature sequence is subjected to feature compression. Fully connected mapping and dimension compression are performed according to the feature dimension arrangement order in the fused feature sequence, and the compressed feature sequence is obtained through normalization. Each feature in the compressed feature sequence is mapped and assigned according to the node identifiers of student nodes, task nodes, and ability indicator nodes to generate node feature vectors. Write the node feature vectors into the corresponding student nodes, task nodes, and ability indicator nodes in the student growth relationship graph, and perform matrix arrangement operation on all node feature vectors in the order of the node set to construct the node feature matrix.

8. The student growth profile generation system based on graph neural networks according to claim 7, characterized in that, The process of generating the branch feature sequence specifically includes: Obtain the student feature sequence, and perform a reordering operation on each feature item in the student feature sequence according to the student identifier, task identifier and evaluation dimension identifier. Based on the feature source, divide the student feature sequence into task feature sequence and evaluation feature sequence. The task feature sequence and the evaluation feature sequence are respectively input into the parallel feature extraction structure in the FusionNet model. In each feature extraction channel, the convolution window sliding calculation is performed according to the feature arrangement order. Convolution mapping calculation is performed in each convolution window through the convolution kernel of preset size to generate convolution feature sequence. Perform the ReLU activation operation on the convolutional feature sequence to generate an activated feature sequence; Two-dimensional empirical mode decomposition is performed on each activation feature in the activation feature sequence. Upper and lower envelope curves are constructed based on the local extrema of each activation feature, and intrinsic mode functions are calculated based on the upper and lower envelope curves to generate a mode sequence. Based on the modal sequence, perform mode screening and mode combination operations on each intrinsic mode function, and perform feature reconstruction calculation on the combination results; The reconstruction results are organized according to the branch arrangement order in the FusionNet model to generate a branch feature sequence.

9. A student growth profile generation system based on graph neural networks according to claim 1, characterized in that, The aggregation update module includes: Read the node feature matrix and the student growth relationship graph, perform node type classification operation on the node feature matrix according to the node identifiers of student nodes, task nodes and ability indicator nodes, and establish a node type sequence according to the arrangement order in the node set; Extract the set of neighboring nodes of each node in the student growth relationship graph based on the node type sequence, and establish an adjacency relationship sequence based on the connection relationship between each node. In the HGT model, heterogeneous graph attention calculation is performed on each node according to the adjacency sequence, attention weights are initialized for the connection relationships between different types of nodes according to the node type sequence, and weight allocation is performed on each neighboring node according to the adjacency sequence to generate an attention weight sequence. Based on the attention weight sequence, a weighted aggregation calculation is performed on the node features in the neighborhood node set to generate an aggregated feature sequence; Based on the aggregated feature sequence and the node feature matrix, a feature propagation operation is performed. The aggregated feature sequence is written into the corresponding node according to the arrangement order in the node set. Feature superposition and feature update operations are performed on each node to generate an updated node feature sequence. Based on the updated node feature sequence, multi-layer feature propagation calculation is performed in the HGT model. By repeatedly performing a preset number of rounds of neighbor node extraction, attention weight calculation, weighted aggregation and node update operations, a stable node feature sequence is generated. Based on a stable sequence of node features, node representation extraction is performed on student nodes, task nodes, and ability indicator nodes to construct node representation vectors.

10. A student growth profile generation system based on graph neural networks according to claim 1, characterized in that, The image generation module includes: Extract student node representations, task node representations, and ability indicator node representations according to the order of the node representation vectors in the node set, and perform node type separation operation to generate student representation sequence, task representation sequence, and ability representation sequence; A set of capability indicators is established based on the capability representation sequence. According to the capability dimension arrangement order in the preset capability map template, the capability dimension classification operation is performed on the capability representation sequence. Each capability indicator in the capability representation sequence is grouped and organized according to the capability dimension to generate a capability indicator sequence. Perform association mapping calculations on the student representation sequence, task representation sequence, and ability indicator sequence, and construct an association feature sequence based on the association mapping results; The MAP-Elites algorithm is used to construct a capability space grid based on the capability index sequence, and the associated feature sequence is mapped to the corresponding capability space grid to generate a capability distribution sequence. Based on the capability distribution sequence, a fitness evaluation operation is performed on each capability spatial location, and a fitness calculation is performed on each capability spatial location based on the associated feature sequence to generate a fitness sequence; Based on the fitness sequence, a capability space search operation is performed on the capability distribution sequence. The associated features with the highest fitness values ​​are selected in the capability space grid to construct a growth profile sequence.

Images