Cognitive Diagnostic Methods and Systems Based on Heterogeneous Hypergraph Neural Networks

Abstract

This invention discloses a cognitive diagnosis method and system based on a heterogeneous hypergraph neural network, belonging to the field of intelligent education technology. The invention first maps knowledge concepts to nodes and exercises to hyperedges, constructing a heterogeneous hypergraph, and initializes a learnable dense vector for each user. Then, it trains a cognitive diagnosis model based on the heterogeneous hypergraph neural network for user cognitive diagnosis. This invention preserves high-order combinatorial relationships of knowledge concepts in exercises through the hypergraph, avoiding semantic loss caused by binary connections. It introduces Cellular Sheaf theory, utilizing learnable Sheaf-restricted mapping matrices and Sheaf Dirichlet Energy regularization to reasonably handle heterogeneity in educational data, preventing feature collapse and improving diagnostic accuracy and interpretability.

Description

Technical Field

[0001] This invention belongs to the field of intelligent education technology, and in particular relates to a cognitive diagnosis method and system based on heterogeneous hypergraph neural networks. Background Technology

[0002] Cognitive diagnosis is a core task in intelligent education systems and adaptive learning platforms. Its goal is not merely to assess student scores, but to identify students' specific mastery of different knowledge concepts. As online education platforms accumulate massive amounts of response data and question banks continue to expand, the knowledge structures involved in the questions are becoming increasingly complex. Many comprehensive questions often test multiple cross-disciplinary knowledge points simultaneously. Traditional educational measurement methods often struggle to fully capture these complex relationships, making it difficult for teachers to identify students' weak knowledge points in a timely and accurate manner, and to provide appropriate remedial learning paths. Therefore, how to accurately extract students' cognitive states and the structural correlations behind the questions from response data has become a crucial issue in the field of intelligent education.

[0003] In recent years, numerous methods have been proposed for cognitive diagnosis. Existing methods can be broadly categorized into two types:

[0004] (1) Methods based on traditional educational measurement. For example, Item Response Theory (IRT) and DINA models. These methods usually rely on Q matrices constructed by experts and explicit probability assumptions or logical rules to estimate students’ mastery of knowledge points, and have strong interpretability.

[0005] (2) Deep learning-based methods. For example, NeuralCDM, GNN-CDM and other models. These methods use the representation learning ability of neural networks to learn more complex feature relationships from the interaction data of students, exercises and knowledge points, so as to improve diagnostic accuracy.

[0006] Despite the progress made by the aforementioned methods, existing cognitive diagnostic methods still have significant shortcomings when dealing with complex educational scenarios. First, most existing methods are still based on a simple binary association between questions and knowledge points, making it difficult to express the higher-order combinations of knowledge concepts involved in a single question, thus easily losing information about the overall structure of the question. Second, deep learning, especially graph neural network methods, often implicitly assume isomorphism, meaning connected nodes should have similar features. However, comprehensive questions in educational data often connect heterogeneous knowledge points with significant differences; simple aggregation can easily lead to inappropriate feature smoothing, weakening the original independence of different knowledge points. Third, existing methods typically treat question difficulty as a simple superposition of the difficulties of several knowledge points, lacking modeling of the complementary effects and structural difficulty resulting from combinations of different knowledge points, thus failing to accurately explain the true reasons for students' incorrect answers.

[0007] Therefore, the key issues in current cognitive diagnostic tasks can be summarized as follows:

[0008] First, existing methods are insufficient in modeling the higher-order relationships between questions and knowledge points, making it difficult to depict the structural characteristics of multiple knowledge concepts working together within comprehensive questions.

[0009] Second, existing graph learning methods are prone to over-smoothing of features when processing heterogeneous relationships in educational data, leading to the loss of information on differences in knowledge points.

[0010] Third, existing models lack effective mining of the structural difficulty formed by the combination of knowledge points, making it difficult to achieve more accurate and fine-grained cognitive state diagnosis.

[0011] In conclusion, to improve the accuracy and reliability of cognitive diagnosis, a novel cognitive diagnosis method needs to be designed that can simultaneously characterize the higher-order relationships between questions and knowledge points, adapt to heterogeneous knowledge structures while preserving knowledge differences, and uncover the structural difficulty of knowledge combinations. This would allow for a more accurate assessment of students' mastery of knowledge points and provide more effective support for personalized teaching. Summary of the Invention

[0012] The purpose of this invention is to solve the problems existing in the prior art and to provide a cognitive diagnosis method and system based on heterogeneous hypergraph neural networks.

[0013] To achieve the above-mentioned objectives, the present invention specifically adopts the following technical solution:

[0014] In a first aspect, the present invention provides a cognitive diagnosis method based on a heterogeneous hypergraph neural network, comprising the following steps:

[0015] S1. Perform structured mapping on the set of knowledge concepts and the set of exercises obtained from the online learning platform, instantiate each knowledge concept as a node, instantiate each exercise as a hyperedge, construct a heterogeneous hypergraph, and initialize a learnable low-dimensional dense vector for each user as the initial representation of the user's potential cognitive state.

[0016] S2. Train a cognitive diagnostic model based on a heterogeneous hypergraph neural network on the heterogeneous hypergraph for user cognitive diagnosis. In the model training, firstly, a Sheaf-restricted mapping matrix is ​​generated based on a multilayer perceptron and the heterogeneous hypergraph is processed by a Sheaf-Laplacian-based heterogeneous feature aggregation mechanism to simultaneously encode the ontological features and structural difference features of the knowledge concept to obtain the final representation of the knowledge concept. Then, the low-dimensional dense vector and the final representation of the knowledge concept are interactively calculated under monotonicity constraints to output a dual result containing a cognitive state profile and a response prediction matrix.

[0017] Based on the above scheme, each step can be implemented in the following preferred manner.

[0018] As a preferred embodiment of the first aspect mentioned above, in S2, the generation process of the Sheaf-restricted mapping matrix is ​​as follows: First, the first multilayer perceptron obtains the initial feature vectors of the nodes from the heterogeneous hypergraph, and the second multilayer perceptron obtains the initial feature vectors of the hyperedges from the heterogeneous hypergraph. The two initial feature vectors with correlation are concatenated to form an integrated feature vector, which is used as the input of the nonlinear mapping network. Finally, the feature vector output by the nonlinear mapping network is processed by a nonlinear activation function and reshaped to the target size, thereby obtaining the Sheaf-restricted mapping matrix.

[0019] As a preferred embodiment of the first aspect mentioned above, in S2, the specific process of obtaining the final representation of the knowledge concept through the heterogeneous feature aggregation mechanism is as follows: obtain a heterogeneous hypergraph neural network formed by the sequential cascading of multiple feature aggregation layers, use the heterogeneous hypergraph as the initial knowledge point feature matrix and input it into the heterogeneous hypergraph neural network, and use the knowledge point feature matrix output by the last feature aggregation layer of the network as the final representation of the knowledge concept.

[0020] As a preferred embodiment of the first aspect mentioned above, the specific processing flow in each feature aggregation layer is as follows: First, the knowledge point feature matrix input from the previous feature aggregation layer is weighted using a weight matrix that preserves ontology information to obtain the first weighted knowledge point feature matrix; simultaneously, the knowledge point feature matrix input from the previous feature aggregation layer is processed by the normalized Sheaf-Laplacian operator, and then the processed knowledge point feature matrix is ​​weighted using a weight matrix that aggregates heterogeneous information to obtain the second weighted knowledge point feature matrix; the error between the two weighted knowledge point feature matrices is calculated and processed by a nonlinear activation function to obtain the knowledge point feature matrix output from this feature aggregation layer, which serves as the input to the next feature aggregation layer.

[0021] As a preferred embodiment of the first aspect mentioned above, in S2, the specific process of interactive computation between the low-dimensional dense vector and the final representation of the knowledge concept is as follows:

[0022] S21. First, perform a vector inner product operation on the user's low-dimensional dense vector and the final representation of the knowledge concept. Then, use the Sigmoid function to compress and activate the inner product operation result, and output the user's knowledge mastery vector as a cognitive state profile to represent the user's knowledge acquisition status for different knowledge concepts.

[0023] S22. Subsequently, based on the hyperedge representation that aggregates heterogeneous information, the difficulty parameter and discrimination parameter of the exercise are generated, and the difficulty parameter, discrimination parameter and the knowledge mastery vector are input together into the third multilayer perceptron to predict the answer probability, forming the answer prediction matrix, while applying positive value constraints to the weight parameters of the third multilayer perceptron.

[0024] As a preferred embodiment of the first aspect mentioned above, for each exercise's corresponding hyperedge, the final representation vectors of the knowledge concepts associated within the hyperedge are normalized and aggregated to obtain the hyperedge representation; each row of the hyperedge representation corresponds to the representation vector of an exercise; the difficulty parameter is obtained by performing a linear transformation on the exercise's representation vector and then mapping it to the (0,1) interval using the Sigmoid function; the discrimination parameter is obtained by performing another set of linear transformations on the representation vector of the same exercise and then activating the output using the Softplus function.

[0025] As a preferred embodiment of the first aspect mentioned above, during training, the cognitive diagnostic model uses a weighted sum of the binary cross-entropy loss and the Sheaf heteromatch regularization loss as the total loss, and updates the cognitive diagnostic model parameters based on minimizing the total loss. The binary cross-entropy loss is calculated based on the predicted answer probability and the actual answer label.

[0026] Secondly, the present invention provides a cognitive diagnostic system based on a heterogeneous hypergraph neural network, comprising:

[0027] The data acquisition module is used to use a large-scale online intelligent question bank as the input source of exercises. Whenever a user completes a set of exercises on the online learning platform, it acquires the user's answer log and the association data between the exercises and knowledge concepts, thereby constructing a heterogeneous hypergraph.

[0028] The cognitive diagnosis module is used to obtain the trained cognitive diagnosis model in the cognitive diagnosis method based on heterogeneous hypergraph neural network described in the first aspect above, and then input the newly constructed heterogeneous hypergraph into the cognitive diagnosis model. Using the Sheaf restricted mapping matrix, the knowledge concept features are dynamically projected onto the local context space of the current exercise. Forward propagation is performed through the heterogeneous message passing mechanism to obtain the final representation of the knowledge concept that integrates high-order heterogeneous information in the current answering process. Then, the user's low-dimensional dense vector and the final representation of the knowledge concept are subjected to monotonicity-constrained interactive calculation. The output includes a dual result containing a cognitive state profile and an answer prediction matrix, and is pushed to the user on the online learning platform according to a preset method.

[0029] Thirdly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the cognitive diagnosis method based on heterogeneous hypergraph neural network as described in any of the solutions of the first aspect above.

[0030] Fourthly, the present invention provides a computer electronic device, which includes a memory and a processor;

[0031] The memory is used to store computer programs;

[0032] The processor is configured to, when executing the computer program, implement the cognitive diagnostic method based on heterogeneous hypergraph neural network as described in any of the first aspects above.

[0033] Compared with the prior art, the present invention has the following advantages:

[0034] First, traditional cognitive diagnostic models or basic graph neural networks, when assessing user cognitive states, are often limited by the simple "node-edge" binary connection assumption, failing to consider the high-order semantic integrity widely present in educational scenarios. For example, a comprehensive exercise is not a simple linear accumulation of knowledge concepts, but an organic combination of multiple knowledge concepts in a specific context; similarly, the user's answering behavior is also controlled by the combined influence of these multi-dimensional knowledge concepts. Existing methods tend to decompose this complex "many-to-one" inclusion relationship into scattered "one-to-one" edges, destroying the structural information of the exercise as a high-order whole, leading to semantic loss when the model handles complex interactions. Hypergraphs have significant advantages in representing deep interaction relationships between multiple nodes when dealing with complex systems. Therefore, this invention reconstructs the cognitive diagnostic task into a structured representation learning task based on hypergraph networks, proposing a cognitive diagnostic framework that can preserve the inherent high-order topological features of educational data:

[0035] First, this invention uses hypergraph theory to reshape the physical representation of data. Knowledge concepts are directly mapped to nodes in a heterogeneous hypergraph, and each exercise is abstracted as a hyperedge connecting multiple knowledge concept nodes. This modeling approach fundamentally breaks through the dependence of traditional graph structures on binary relations, enabling the model to fully preserve the high-order combinatorial relationships between knowledge concepts within an exercise. For example, an exercise that examines three knowledge concepts simultaneously is represented in the hypergraph as a single hyperedge containing three nodes, rather than three independent edges.

[0036] Secondly, to address the inference challenges posed by sparse response data, this invention designs a co-evolutionary node and hyperedge representation learning framework. The model no longer involves unidirectional feature extraction but instead establishes a bidirectional, mutually feeding information flow field: when updating a node representation, the model aggregates the comprehensive contextual information of all hyperedges involved in that node; while when updating a hyperedge representation, the model non-linearly aggregates all the knowledge points it contains.

[0037] Finally, to ensure the diagnostic results are interpretable from an educational perspective, this invention constructs a unified diagnostic layer that satisfies the monotonicity assumption. That is, the predicted probability of answering should increase as the user's proficiency with the knowledge concept improves. The model utilizes the learned high-order knowledge concept representations and the user's low-dimensional dense vectors to generate answer probabilities and knowledge mastery vectors through constrained interactive computation. This design allows the model to minimize prediction errors while inferring the implicit difficulty and discrimination parameters of each exercise.

[0038] Second, traditional hypergraph convolution is essentially based on the isomatch assumption, which smooths the features of nodes within the hyperedge. However, connected nodes often have very different attributes. Directly using traditional hypergraph neural networks for feature aggregation cannot reasonably handle the dismatch features commonly found in educational data. This leads to knowledge concepts with vastly different properties (such as "algebra" and "geometry") being forcibly assimilated in the same comprehensive exercise, causing feature collapse. Ultimately, the model cannot distinguish which knowledge concept the user has mastered. Therefore, this invention introduces the Cellular Sheaf theory from algebraic topology and designs a dismatch feature aggregation mechanism that can tolerate local inconsistencies:

[0039] First, this invention uses Cellular Sheaf theory to construct a high-dimensional geometric structure, treating features as something independent of static vectors attached to nodes. Specifically, each node and each hyperedge is assigned an independent vector space, and Sheaf constraint mapping matrices are introduced into the node feature space and hyperedge feature space to map the ontological features of knowledge concepts to the local contextual features of exercises. Therefore, each knowledge concept can play different roles in different exercises, and its different performances under different question types can be effectively captured. Utilizing the learnability of the Sheaf constraint mapping matrix, the model can dynamically adjust the weight matrix of each knowledge concept according to the specific content of the exercise. For exercises containing vastly different knowledge concepts, the Sheaf structure can learn higher structural difficulty weights through non-linear constraint mapping, thus accurately simulating the real educational principle that "interdisciplinary problems are often more difficult than exercises that only test a single knowledge concept," compensating for the shortcomings of traditional models in mining structural difficulty.

[0040] Secondly, this invention proposes a decoupled normalized Sheaf Laplacian operator. This operator first uses a learnable Sheaf-restricted mapping matrix to explicitly calculate the differences in node features projected onto the hyperedge space, and then compares the distances of nodes within the same hyperedge in the projection space. Therefore, it can objectively measure the complementarity or conflict of knowledge concepts. More importantly, the model strictly preserves the differences between knowledge concepts when aggregating information, avoiding simple feature averaging and truly respecting the independence of each knowledge concept.

[0041] Finally, this invention introduces a novel regularization objective: the Sheaf Dirichlet Energy, which is used as the Sheaf heteropairing regularization loss. Unlike general Laplace smoothing, it minimizes the difference between the node features after the restricted mapping and the hyperedge features as an energy function. Therefore, it mathematically reasonably derives the conclusion that "the heteropairing knowledge concepts should achieve logical consistency within the problem, rather than being close to each other in the original feature space." Attached Figure Description

[0042] Figure 1 This is a flowchart of the steps of the present invention;

[0043] Figure 2 This is a schematic diagram illustrating the process of constructing a heterogamous hypergraph in this invention;

[0044] Figure 3 A schematic diagram for obtaining the final representation of the knowledge concept and the hyperedge representation;

[0045] Figure 4 A schematic diagram illustrating the interactive computation between low-dimensional dense vectors and the final representation of knowledge concepts;

[0046] Figure 5 This is a system block diagram of the present invention;

[0047] Figure 6 This is a schematic diagram of a computer electronic device provided by the present invention. Detailed Implementation

[0048] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.

[0049] In the description of this invention, it should be understood that the terms "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features.

[0050] like Figure 1 As shown, in a preferred embodiment of the present invention, the cognitive diagnosis method based on heterogeneous hypergraph neural network includes the following steps S1-S2. The specific implementation process of each step will be described in detail below.

[0051] S1. Perform structured mapping on the set of knowledge concepts and the set of exercises obtained from the online learning platform, instantiate each knowledge concept as a node, instantiate each exercise as a hyperedge, construct a heterogeneous hypergraph, and initialize a learnable low-dimensional dense vector for each user as the initial representation of the user's potential cognitive state.

[0052] It should be noted that in step S1 of the present invention, as Figure 2 As shown, the set of knowledge concepts is defined as , which includes There are 1 knowledge concept, each denoted as _ . , For knowledge concept indexing, The total number of knowledge concepts; the set of exercises is defined as... , which includes There are 10 exercises, each denoted as 1. , For the exercise index, This represents the total number of exercises. After obtaining the above two sets, this invention uses an expert-calibrated Q-matrix (association matrix). Constructing a heterogeneous hypergraph , A set of nodes; This is a set of hyperedges. Furthermore, each element in the Q matrix above takes the value 0 or 1, representing the relationship between exercises and knowledge points. For example, the first element in the Q matrix... Line 1 Column elements represent the first Is the first exercise related to the first...? These knowledge concepts are related, when When, it means the first The first exercise and the first There are connections between knowledge concepts, and vice versa.

[0053] It should also be noted that in step S1 of this invention, based on the user's previous answers to each exercise, the prior art typically generates a corresponding answer matrix first. , This represents the total number of users. Each row in the answer matrix corresponds to a user's answer to each question. Each element in the answer matrix takes the values ​​0, 1, or null, representing a question answered correctly, answered incorrectly, or not answered, respectively. For example, the first row in the answer matrix... Line 1 Non-empty elements in the column This indicates that the student In the The invention presents the answers to individual exercises. To address the sparsity of the answer data, this invention does not directly construct a graph connection based on the answer matrix, but instead... (The sentence is incomplete and requires more context to be translated accurately.) Initialize a learnable low-dimensional dense vector This vector, serving as the initial representation of the student's potential cognitive state, will be iteratively updated during the training of the cognitive diagnostic model using the gradient of the problem difficulty parameter via backpropagation. Let be the set of real numbers. Let be the feature dimension of this initial representation.

[0054] S2. Train a cognitive diagnostic model based on a heterogeneous hypergraph neural network on the heterogeneous hypergraph for user cognitive diagnosis. In the model training, firstly, a Sheaf-restricted mapping matrix is ​​generated based on a multilayer perceptron and the heterogeneous hypergraph is processed by a Sheaf-Laplacian-based heterogeneous feature aggregation mechanism to simultaneously encode the ontological features and structural difference features of the knowledge concept to obtain the final representation of the knowledge concept. Then, the low-dimensional dense vector and the final representation of the knowledge concept are interactively calculated under monotonicity constraints to output a dual result containing a cognitive state profile and a response prediction matrix.

[0055] It should be noted that in step S2 of this invention, the representation of knowledge concepts in different exercises has a dynamic characteristic, that is, the same knowledge concept will present different aspects in different exercise contexts. In order to capture this fine-grained semantic difference and construct a geometric structure, it is first necessary to establish a constraint mapping between knowledge concept nodes and exercise hyperedges, that is, a Sheaf constraint mapping matrix.

[0056] In this invention, the Sheaf constraint mapping matrix generation process is as follows: First, the first multilayer perceptron obtains the initial feature vectors of the nodes from the heterogeneous hypergraph. The initial feature vectors of hyperedges are obtained from the heterogeneous hypergraph by the second multilayer perceptron. The two initial feature vectors that are related are concatenated to form an integrated feature vector, which is then used as the input to the nonlinear mapping network. Finally, the feature vector output by the nonlinear mapping network is processed by a nonlinear activation function and reshaped to the target size, thereby obtaining the Sheaf constraint mapping matrix.

[0057] Furthermore, the nonlinear mapping network employs a multilayer perceptron.

[0058] In this embodiment, for each pair of node-hyperedge pairs with an association relationship in the heterogeneous hypergraph This step uses a parameter-shared multilayer perceptron as a mapping generator to dynamically predict the Sheaf-restricted mapping matrix after projecting node features onto the hyperedge local space. Specifically, the initial feature vectors of the nodes and the initial feature vectors of the hyperedges are integrated through a concatenation operation, and this is used as the input to a nonlinear mapping network (MLP).

[0059]

[0060] in, for A Sheaf-restricted mapping matrix of dimension , whose subscripts This indicates a node. Mapping to hyperedge superior; is the feature dimension of nodes in the heterogeneous hypergraph; For shape reshaping operations; It is a non-linear activation function; It is a nonlinear mapping network; Indicates a splicing operation; These are the trainable parameters for the nonlinear mapping network. The final Sheaf-restricted mapping matrix quantifies the projection state of a knowledge concept under a specific exercise, and serves as the geometric basis for subsequent heterogeneous difference calculations.

[0061] It should be noted that in step S2 of this invention, traditional hypergraph neural networks often ignore the heterogeneity between nodes when performing convolution operations, resulting in overly smooth features. In order to make full use of the geometric structure constructed by the constraint mapping, this step proposes a heterogeneous feature aggregation mechanism based on Sheaf-Laplacian to simultaneously encode ontological features and structural difference features of knowledge points.

[0062] In this invention, the specific process of obtaining the final representation of knowledge concepts through the heterogeneous feature aggregation mechanism is as follows: Obtain a heterogeneous hypergraph neural network formed by the sequential cascading of multiple feature aggregation layers; use the heterogeneous hypergraph as the initial knowledge point feature matrix and input it into the heterogeneous hypergraph neural network; use the knowledge point feature matrix output by the last feature aggregation layer of the network as the final representation of knowledge concepts.

[0063] In each feature aggregation layer, the specific processing flow is as follows: First, the knowledge point feature matrix input from the previous feature aggregation layer is weighted using a weight matrix that preserves ontology information to obtain the first weighted knowledge point feature matrix; simultaneously, the knowledge point feature matrix input from the previous feature aggregation layer is processed by the normalized Sheaf-Laplacian operator, and then the processed knowledge point feature matrix is ​​weighted using a weight matrix that aggregates heterogeneous information to obtain the second weighted knowledge point feature matrix; the error between the two weighted knowledge point feature matrices is calculated and processed by a nonlinear activation function to obtain the knowledge point feature matrix output from this feature aggregation layer, which serves as the input to the next feature aggregation layer.

[0064] In this embodiment, for the first For feature aggregation at each layer, the model does not directly average neighbor features; instead, it calculates the diffusion error. The specific formula is as follows:

[0065]

[0066] in, and The first Layer, First The knowledge point feature matrix of the layer; and These are the weight matrices for preserving ontology information and aggregating heterogeneous information, respectively; For the normalized Sheaf Laplacian operator, its value at nodes The specific calculations are as follows:

[0067]

[0068] in, For nodes in the processed knowledge point feature matrix Features of the location; Indicates the presence of nodes super-edge ; For super-edge The degree of the hyperedge, i.e. Total number of nodes included; superscript Indicates transpose; Indicates inclusion within a hyperedge On, except for nodes Other nodes, That is, with nodes Associated nodes; For from the first Nodes obtained from the layer knowledge point feature matrix Features; For nodes Mapping to hyperedge When above, the corresponding Sheaf constraint mapping matrix; For from the first Nodes obtained from the layer knowledge point feature matrix The characteristics of the formula. Explicitly captures the current knowledge concept node within the hyperedge of the same exercise. Other related knowledge concept nodes Feature misalignment in the projection space. When this aggregation logic is applied iteratively to multiple levels, such as... Figure 3 As shown, the output knowledge point feature matrix of the last layer is the final representation of the knowledge concept that integrates higher-order heterogeneous structure information. Each row corresponds to the final representation vector of a knowledge concept.

[0069] Furthermore, the number of feature aggregation layers in the above-mentioned heterogeneous hypergraph neural network can be 2 or 4 layers, or can be selected by those skilled in the art according to actual needs, and is not limited in this invention.

[0070] It should be noted that in step S2 of this invention, in order to obtain cognitive diagnostic results with educational interpretability, it is necessary to map the learned knowledge concept representations to the user's cognitive state space and generate answer predictions. Considering the monotonicity assumption in educational measurement, that is, the stronger the user's ability, the higher the probability of answering questions correctly, this step designs a cognitive diagnostic prediction strategy that satisfies the monotonicity constraint.

[0071] In this invention, such as Figure 4 As shown, the specific process of interactive computation between low-dimensional dense vectors and the final representation of knowledge concepts is as follows:

[0072] S21. First, perform a vector inner product operation between the user's low-dimensional dense vector and the final representation of the knowledge concept. Then, use the Sigmoid function to numerically compress and activate the result of the inner product operation, and output the user's knowledge mastery vector as a cognitive state profile to represent the user's knowledge acquisition status for different knowledge concepts.

[0073]

[0074] in, For users Knowledge mastery vector; The Each element represents a user. For the The probability of mastering a knowledge concept.

[0075] S22. Subsequently, based on the hyperedge representation that aggregates heterogeneous information, the difficulty parameter and discrimination parameter of the exercise are generated, and the difficulty parameter, discrimination parameter, and the knowledge mastery vector are input together into the third multilayer perceptron to predict the probability of answering. This forms the response prediction matrix, and simultaneously adjusts the weight parameters of the third multilayer perceptron. Apply positive constraints ( ):

[0076]

[0077] In this embodiment, the specific method for obtaining the hyperedge representation is as follows:

[0078] For the The hyperedge corresponding to each exercise The final representation vectors of the knowledge concepts associated within the hyperedge are normalized and aggregated to obtain the hyperedge representation. :

[0079]

[0080] in, and These are trainable weight and bias parameters; To be with the super-edge A collection of related knowledge concepts; for The number of knowledge concepts in it; For the first The final representation vector of each knowledge concept. This is obtained from... This is a hyperedge representation that integrates high-order knowledge combination relations and heterogeneous structural information.

[0081] In this embodiment, each row of the hyperedge representation corresponds to a representation vector of a problem. The difficulty parameter is obtained by performing a linear transformation on the representation vector of the problem and then mapping it to the (0,1) interval using the Sigmoid function. The discrimination parameter is obtained by performing another set of linear transformations on the representation vector of the same problem and then activating the output using the Softplus function.

[0082] Specifically, no. Difficulty parameters of the exercises With the Discrimination parameters of the exercises Represented as:

[0083]

[0084]

[0085] in, and All of these are trainable weight parameters; and All of these are trainable bias parameters; For the first The representation vector of each exercise is derived from a row in the hyperedge representation. For the difficulty parameter, this invention employs a Sigmoid function constraint. This ensures that the probability of a user mastering each knowledge concept falls within the same numerical range; for the discrimination parameter, this invention employs a Softplus function constraint. This is to ensure that the discrimination parameter meets the non-negativity requirement in educational measurement.

[0086] Therefore, the knowledge mastery vector generated by this invention As a core diagnostic indicator, it is a dimension of The probability vector. This vector precisely quantifies the probability of a user's mastery of each knowledge concept, ranging from (0,1). Probability of answering. As a validation and recommendation metric, it represents the model's prediction of the user's correct answer rate on each exercise question.

[0087] The primary objective of this invention is to accurately determine each user's level of understanding of knowledge concepts. Unlike traditional methods, this invention outputs... This is derived from inferences based on heterogeneous hypergraphs, meaning that the mastery probability not only reflects a user's memory of a single knowledge concept, but also implicitly reflects the user's comprehensive ability to handle cross-domain knowledge combinations and complex structured exercises. Furthermore, based on... With its high-precision prediction, this invention can further generate personalized remedial practice paths for users (i.e., recommend exercises that users currently have a low level of mastery of but whose predicted pass rate is in the "zone of proximal development").

[0088] In this invention, during training, the cognitive diagnostic model uses a weighted sum of binary cross-entropy loss and Sheaf heteromatch regularization loss as the total loss, and updates the cognitive diagnostic model parameters based on minimizing the total loss. The binary cross-entropy loss is calculated based on the predicted answer probability and the actual answer label.

[0089] In this embodiment, after obtaining the predicted response probability from the cognitive diagnostic model and the feature differences calculated based on the Sheaf structure, the training objective of the model is to simultaneously minimize the prediction error and the geometric structure inconsistency. Furthermore, this embodiment utilizes binary cross-entropy loss to optimize prediction performance. By comparing the consistency between the predicted response probability and the actual response label, the following loss term is constructed:

[0090]

[0091] in, To compare the predicted probability of response with the actual response label Calculated binary cross-entropy loss; The set of observed user-exercise interaction pairs.

[0092] To ensure that the Sheaf constraint mapping matrix effectively captures the geometric structure on the data manifold and prevents it from degenerating into trivial solutions, this embodiment introduces Sheaf Dirichlet energy as the Sheaf heterogeneous regularization loss, which is defined as the sum of the differences in the features of all hyperedge interior nodes in the projection space:

[0093]

[0094] Minimizing this loss can help features achieve consistency in their local contexts, enhancing the model's robustness to heterogeneous structures.

[0095] Therefore, the total loss of this invention can be expressed as:

[0096]

[0097] in, The hyperparameter is used to balance the prediction task with structure regularization. In each training iteration, the total loss can be jointly updated using the backpropagation algorithm and stochastic gradient descent (SGD) or the Adam optimizer to update the low-dimensional dense vectors, multilayer perceptron parameters, and convolutional weights in the model.

[0098] The present invention will now demonstrate the application effect of the cognitive diagnosis method based on heterogeneous hypergraph neural network described in S1~S2 of the above embodiments on a specific dataset through a specific example, so as to facilitate understanding of the essence of the present invention.

[0099] Example

[0100] The specific implementation process of the cognitive diagnosis method based on heterogeneous hypergraph neural network used in this embodiment is as described above and will not be repeated here.

[0101] To demonstrate the technical effectiveness of the proposed method, it was validated in a real-world teaching scenario. The dataset used in this invention is an educational dataset collected from a real, computer-supported large-scale online learning scenario (such as the ASSISTments question bank). In this dataset, each node represents a knowledge concept in the large-scale online learning scenario, and each hyperedge represents an exercise containing multiple knowledge concepts, forming a "knowledge concept-exercise" heterogeneous hypergraph. During the course, this invention uses exercises covering multiple knowledge points in combination, and experts pre-label the knowledge concepts (i.e., the Q-matrix) contained in each exercise during the teaching and research phase. In class, students' individual answer records are collected as input for model training. Students' cognitive states are initialized as learnable low-dimensional dense vectors, and student interactions with exercises are recorded as answer logs for real-time data collection. Furthermore, to verify the effectiveness of the method in handling heterogeneous data, this embodiment specifically selects exercises with strong heterogeneity (i.e., containing knowledge concepts with significant differences in nature) that cross chapters or disciplines.

[0102] The evaluation metrics used are the prediction accuracy (ACC) and the area under the ROC curve (AUC), which are most commonly used in the field of cognitive diagnosis. Both values ​​range from 0 to 1; a higher value indicates a more accurate prediction of student performance, and thus reflects a more reliable diagnosis of cognitive state. Table 1 shows a comparison between the proposed method and current mainstream cognitive diagnosis benchmarks (such as DINA, NeuralCDM, and HyperCDM). For fairness, the experiments were conducted under the same data split ratio (e.g., 80% training, 20% testing).

[0103] Table 1. Comparison of prediction performance of different methods on educational datasets

[0104] Experimental results show that the method proposed in this invention has a significant improvement over the benchmark method. One major reason is that the method of this invention considers the heterogeneity of knowledge points within each exercise instance from the perspective of the whole and the structure.

[0105] In addition, this embodiment also conducted an ablation study to better examine the contribution of the proposed innovations. The experimental results are shown in Table 2, and the specific details of each module are as follows:

[0106] Backbone: Uses a basic neurocognitive diagnostic framework (such as NeuralCDM) to perform simple interactive predictions using only the student's low-dimensional dense vectors and the ID features of the exercises, without including any graph structure information.

[0107] W / Hyper: Based on the Backbone, a regular hypergraph structure is constructed according to the Q matrix, and a conventional hypergraph convolutional neural network (HGNN) is used to aggregate knowledge concept features.

[0108] W / Sheaf: Based on W / Hyper, it introduces the heterogeneous feature aggregation mechanism, the core of this invention, to explicitly model the structural difficulty within the exercises and the heterogeneous relationship between knowledge concepts, corresponding to the complete model of this invention.

[0109] Table 2. Ablation Experiment Results of the Innovative Module of this Invention

[0110] The results of the ablation experiment show that the two innovative modules proposed in this invention are very effective. The introduction of the heterogeneous hypergraph structure is conducive to encoding the higher-order relationship of "multiple knowledge concepts constitute a problem", while the heterogeneous feature aggregation mechanism solves the problems of non-smooth connection and heterogeneous conflict between knowledge points, thus facilitating the diagnosis of students' cognitive state.

[0111] It should also be noted that the cognitive diagnosis method based on heterogeneous hypergraph neural networks in the above embodiments can essentially be executed by a computer program or module. Therefore, similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a cognitive diagnosis system based on heterogeneous hypergraph neural networks, corresponding to the cognitive diagnosis method based on heterogeneous hypergraph neural networks provided in the above embodiments, such as... Figure 5 As shown, it includes:

[0112] The data acquisition module is used to use a large-scale online intelligent question bank as the input source of exercises. Whenever a user completes a set of exercises on the online learning platform, it acquires the user's answer log and the association data between the exercises and knowledge concepts, thereby constructing a heterogeneous hypergraph.

[0113] The cognitive diagnosis module is used to obtain the trained cognitive diagnosis model in the cognitive diagnosis method based on heterogeneous hypergraph neural network described in the above embodiments, and then input the newly constructed heterogeneous hypergraph into the cognitive diagnosis model. Using the Sheaf restricted mapping matrix, the knowledge concept features are dynamically projected onto the local context space of the current exercise. Forward propagation is performed through the heterogeneous message passing mechanism to obtain the final representation of the knowledge concept that integrates high-order heterogeneous information in the current answering process. Then, the user's low-dimensional dense vector and the final representation of the knowledge concept are subjected to monotonicity-constrained interactive calculation. The output includes a dual result containing a cognitive state profile and an answer prediction matrix, and is pushed to the user on the online learning platform according to a preset method.

[0114] It is understood that the cognitive diagnosis method based on heterogeneous hypergraph neural networks described in S1-S2 above can essentially be implemented by a computer program. Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer electronic device corresponding to the cognitive diagnosis method based on heterogeneous hypergraph neural networks provided in the above embodiments, such as... Figure 6 As shown, it includes a memory and a processor;

[0115] The memory is used to store computer programs;

[0116] The processor is configured to implement the cognitive diagnosis method based on heterogeneous hypergraph neural network in the above embodiments when executing the computer program.

[0117] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0118] Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer-readable storage medium corresponding to the cognitive diagnosis method based on heterogeneous hypergraph neural network provided in the above embodiments. The storage medium stores a computer program, which, when executed by a processor, can realize the cognitive diagnosis method based on heterogeneous hypergraph neural network in the above embodiments.

[0119] It is understood that the aforementioned storage media may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Furthermore, the storage media may also be various media capable of storing program code, such as USB flash drives, external hard drives, magnetic disks, or optical discs.

[0120] It is understood that the processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0121] It should also be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. In the embodiments provided in this application, the division of steps or modules in the system and method is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules or steps may be combined or integrated together, and a module or step may also be split.

[0122] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A cognitive diagnostic method based on a heterogeneous hypergraph neural network, characterized in that, Includes the following steps: S1. Perform structured mapping on the set of knowledge concepts and the set of exercises obtained from the online learning platform, instantiate each knowledge concept as a node, instantiate each exercise as a hyperedge, construct a heterogeneous hypergraph, and initialize a learnable low-dimensional dense vector for each user as the initial representation of the user's potential cognitive state. S2. Train a cognitive diagnostic model based on a heterogeneous hypergraph neural network on the heterogeneous hypergraph for user cognitive diagnosis. In the model training, firstly, a Sheaf-restricted mapping matrix is ​​generated based on a multilayer perceptron and the heterogeneous hypergraph is processed by a Sheaf-Laplacian-based heterogeneous feature aggregation mechanism to simultaneously encode the ontological features and structural difference features of the knowledge concept to obtain the final representation of the knowledge concept. Then, the low-dimensional dense vector and the final representation of the knowledge concept are interactively calculated under monotonicity constraints to output a dual result containing a cognitive state profile and a response prediction matrix.

2. The cognitive diagnosis method based on heterogeneous hypergraph neural network as described in claim 1, characterized in that, In S2, the Sheaf constraint mapping matrix is ​​generated as follows: First, the first multilayer perceptron obtains the initial feature vectors of the nodes from the heterogeneous hypergraph, and the second multilayer perceptron obtains the initial feature vectors of the hyperedges from the heterogeneous hypergraph. The two initial feature vectors with correlation are concatenated to form an integrated feature vector, which is used as the input of the nonlinear mapping network. Finally, the feature vector output by the nonlinear mapping network is processed by a nonlinear activation function and reshaped to the target size, thereby obtaining the Sheaf constraint mapping matrix.

3. The cognitive diagnosis method based on heterogeneous hypergraph neural network as described in claim 1, characterized in that, In S2, the specific process of obtaining the final representation of knowledge concepts through the heterogeneous feature aggregation mechanism is as follows: Obtain the heterogeneous hypergraph neural network formed by the sequential cascading of multiple feature aggregation layers, use the heterogeneous hypergraph as the initial knowledge point feature matrix and input it into the heterogeneous hypergraph neural network, and use the knowledge point feature matrix output by the last feature aggregation layer of the network as the final representation of knowledge concepts.

4. The cognitive diagnosis method based on heterogeneous hypergraph neural network as described in claim 3, characterized in that, In each feature aggregation layer, the specific processing flow is as follows: First, the knowledge point feature matrix input from the previous feature aggregation layer is weighted using a weight matrix that preserves ontology information to obtain the first weighted knowledge point feature matrix; simultaneously, the knowledge point feature matrix input from the previous feature aggregation layer is processed by the normalized Sheaf-Laplacian operator, and then the processed knowledge point feature matrix is ​​weighted using a weight matrix that aggregates heterogeneous information to obtain the second weighted knowledge point feature matrix; the error between the two weighted knowledge point feature matrices is calculated and processed by a nonlinear activation function to obtain the knowledge point feature matrix output from this feature aggregation layer, which serves as the input to the next feature aggregation layer.

5. The cognitive diagnosis method based on heterogeneous hypergraph neural network as described in claim 1, characterized in that, In S2, the specific process of interactive computation between the low-dimensional dense vector and the final representation of the knowledge concept is as follows: S21. First, perform a vector inner product operation on the user's low-dimensional dense vector and the final representation of the knowledge concept. Then, use the Sigmoid function to compress and activate the inner product operation result, and output the user's knowledge mastery vector as a cognitive state profile to represent the user's knowledge acquisition status for different knowledge concepts. S22. Subsequently, based on the hyperedge representation that aggregates heterogeneous information, the difficulty parameter and discrimination parameter of the exercise are generated, and the difficulty parameter, discrimination parameter and the knowledge mastery vector are input together into the third multilayer perceptron to predict the answer probability, forming the answer prediction matrix, while applying positive value constraints to the weight parameters of the third multilayer perceptron.

6. The cognitive diagnosis method based on heterogeneous hypergraph neural network as described in claim 5, characterized in that, For each exercise's corresponding hyperedge, the final representation vectors of the knowledge concepts associated within the hyperedge are normalized and aggregated to obtain the hyperedge representation. Each row of the hyperedge representation corresponds to the representation vector of an exercise. The difficulty parameter is obtained by performing a linear transformation on the exercise's representation vector and then mapping it to the (0,1) interval using the Sigmoid function. The discrimination parameter is obtained by performing another set of linear transformations on the representation vector of the same exercise and then activating the output using the Softplus function.

7. The cognitive diagnosis method based on heterogeneous hypergraph neural network as described in claim 1, characterized in that, During training, the cognitive diagnostic model uses a weighted sum of binary cross-entropy loss and Sheaf heteromatch regularization loss as the total loss, and updates the cognitive diagnostic model parameters based on minimizing the total loss. The binary cross-entropy loss is calculated based on the predicted answer probability and the actual answer label.

8. A cognitive diagnostic system based on a heterogeneous hypergraph neural network, characterized in that, include: The data acquisition module is used to use a large-scale online intelligent question bank as the input source of exercises. Whenever a user completes a set of exercises on the online learning platform, it acquires the user's answer log and the association data between the exercises and knowledge concepts, thereby constructing a heterogeneous hypergraph. The cognitive diagnosis module is used to obtain the trained cognitive diagnosis model in the cognitive diagnosis method based on heterogeneous hypergraph neural network as described in any one of claims 1 to 7, then input the newly constructed heterogeneous hypergraph into the cognitive diagnosis model, use the Sheaf restricted mapping matrix to dynamically project the knowledge concept features to the local context space of the current exercise, perform forward propagation through the heterogeneous message passing mechanism to obtain the final representation of the knowledge concept that integrates high-order heterogeneous information in the current answering process, and then perform interactive calculations on the user's low-dimensional dense vector and the final representation of the knowledge concept under monotonicity constraints, outputting a dual result containing a cognitive state profile and an answer prediction matrix, and pushing it to the user on the online learning platform according to a preset method.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the cognitive diagnosis method based on heterogeneous hypergraph neural networks as described in any one of claims 1 to 7.

10. A computer electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to implement the cognitive diagnosis method based on heterogeneous hypergraph neural network as described in any one of claims 1 to 7 when executing the computer program.

Images