Online education cognition diagnosis method based on heterogeneous conceptual graph construction and modeling enhancement

Abstract

The invention discloses an online education cognition diagnosis method based on heterogeneous conceptual graph construction and modeling enhancement, and the method comprises the following steps: extracting practice records of a plurality of students, and extracting a plurality of knowledge concepts in the practice records; selecting knowledge concept pairs of which the similarity is higher than a threshold value, identifying the relation between knowledge concepts, and constructing an accurate heterogeneous conceptual graph through triple generation and triple evaluation; dividing the heterogeneous concept graph into three sub-graphs according to the relationship type; basic representations of the three sub-graphs are obtained through a sub-graph encoder; the important sub-graph features are dynamically identified through a self-adaptive selector; updating knowledge concepts and practice features through a graphic information aggregator; a diagnostic prediction of the student's knowledge grasp level is generated based on the response records and the graphical information. According to the method, the complex relation between knowledge concepts is mined, the accurate inference on the knowledge mastering level is improved, the basic sub-graph features are adaptively recognized, and graph representation is integrated.

Description

Technical Field

[0001] The present invention belongs to the technical field of online education, and in particular relates to an online education cognitive diagnosis method based on heterogeneous concept graph construction and modeling enhancement. Background Art

[0002] As a key component of online education systems, cognitive diagnosis (CD) is the basis for subsequent educational tasks, such as computer adaptive testing (CAT). The purpose of intelligence tests is to assess learners' knowledge level based on their current performance in answering questions. A good example can be found in Figure 1 In the dataset, students A and B responded to several exercises (e.g., {e_i}, i = 1, 2, 3, 4). Furthermore, an expert-constructed Q-matrix represents the relationship between the exercises and the knowledge concepts. Cognitive diagnostics infers the students' mastery of the knowledge concepts based on the response records and the Q-matrix. This mastery level not only reflects the students' proficiency in the knowledge concepts but also helps provide targeted recommendations for their subsequent learning.

[0003] Originally developed from psychometrics, there are several representative CD models, such as Item Response Theory (IRT) and Multidimensional Item Response Theory (MIRT). IRT and its variants are scalable mathematical frameworks. While easy to implement, their scalability and interpretability remain limited. With advances in deep learning, cognitive diagnostics has also incorporated deep learning to improve accuracy and interpretability, such as NCDM, KaNCD, and ORCDF. These deep learning-based methods primarily utilize multilayer perceptrons to improve the accuracy of CD based on response records and the relationship between students and exercises.

[0004] However, most CD methods ignore the dependencies between knowledge concepts, which are valuable for assessing mastery levels. For example, students must master matrix multiplication as a necessary foundation to master linear algebra. Therefore, the mastery level of the parent concept will affect the mastery of the child concept. This dependency between knowledge concepts is important information for inferring the degree of students' mastery of the knowledge concepts. To this end, some CD methods use prerequisite relationships to describe the dependencies between knowledge concepts. These works model this relationship through a hierarchical structure, however, this oversimplifies the relationship between knowledge concepts. For example, although there is no precedence relationship between matrix row operations and matrix row echelons, these two knowledge concepts can enhance each other's understanding, indicating the existence of other relationship types, such as synergistic relationships. Summary of the Invention

[0005] Although some studies have proposed characterizing the precedence relationships between knowledge concepts, this modeling approach remains limited. Existing methods primarily model the interaction between students and exercises, largely ignoring the relationships between underlying knowledge concepts and therefore failing to fully simulate students' mastery levels. To address these issues, this application proposes a heterogeneous concept graph-enhanced cognitive diagnosis model (HCGCDM). Specifically, this application first designs a heterogeneous concept graph construction module to construct a comprehensive heterogeneous concept graph. This graph first identifies pairs of knowledge concepts with a high probability of forming triples. Next, this application uses retrieval-augmented generation (RAG) to retrieve and evaluate triples under three representative relation types: precedence, parallel, and synergy. By constructing a heterogeneous concept graph, the interconnections between knowledge concepts are strengthened, overcoming the oversimplified relational representations in previous CD methods. To effectively capture the heterogeneous graph structure, this application proposes a heterogeneous concept graph modeling and aggregation module, which involves three key operations: a subgraph encoder, an adaptive selector, and a graph information aggregator. Taking into account the different interactions between knowledge concepts, the three subgraphs are decomposed based on the three relationships. The subgraph encoder then generates an expressive representation of each subgraph using a suitable graph neural network. For example, a graph attention network (GAT) is used to model precedence relations, as children's knowledge concepts at different levels contribute differently to parent knowledge concepts. To address the different precedence relations of knowledge concepts in exercises, an adaptive selector for dynamic subgraph embedding integration is developed. Secondly, a graph information aggregator is designed to merge graph information with the original graph representation with weights. The last step of HCGCDM is to jointly utilize response records and heterogeneous concept graph information to infer students' mastery level.

[0006] To achieve the above objectives, the online education cognitive diagnosis method based on heterogeneous concept map construction and modeling enhancement disclosed in this application includes the following steps:

[0007] Extracting multiple students' practice records and extracting multiple knowledge concepts from the practice records;

[0008] Select knowledge concept pairs with similarity higher than a threshold, identify the connections between knowledge concepts, and construct a heterogeneous concept graph;

[0009] The heterogeneous concept graph is divided into three subgraphs according to the relationship type;

[0010] Obtain basic representations for the three sub-graphs through the sub-graph encoder;

[0011] Dynamically identify important subgraph features through adaptive selectors;

[0012] Update knowledge concepts and practice features through a graphical information aggregator;

[0013] Generate diagnostic predictions of students' knowledge levels based on response records and graphical information.

[0014] Furthermore, in the heterogeneous concept graph, knowledge concepts are represented as nodes, and the relationships between knowledge concepts are described as edges; the relationships between the knowledge concepts include:

[0015] If mastering knowledge concept a is a necessary condition for mastering knowledge concept b, then there is a precedence relationship between a and b;

[0016] The parallel relationship means that two knowledge concepts exist at the same level;

[0017] The characteristic of a synergistic relationship is that the two knowledge concepts reinforce each other's understanding and application.

[0018] Furthermore, for multiple students S = {s1,s2,...,s N}, N is the total number of students, there exists an exercise E={e1,e2,...,e M}, M is the total number of exercises, knowledge concept C={c1,c2,...,c K}, K is the total number of knowledge concepts, recorded as the response record set, where y is the answer score, that is, 1 means that the student answered the exercise e correctly, otherwise it is 0; the Q matrix represents the relationship between the exercise and the knowledge concept, defined as Q = (q ij ) M×K , if exercise i contains knowledge concept j, then q ij =1, otherwise q ij =0.

[0019] Furthermore, a large language model is used to generate triples of relationships between knowledge concepts:

[0020] Concept Pair Extraction: Pairs of knowledge concepts are fed as input to a large language model; the knowledge concepts are embedded into vectors and their similarity is calculated; for concept pairs whose similarity exceeds a threshold, they are identified as potential candidates for relationship detection:

[0021] Pairs={(c i ,c j )|sim(c i ,c j )>θ

[0022] c i and c j are the i-th and j-th concepts respectively;

[0023] Triple generation: Utilize retrieval-enhanced generation to improve the reliability of generated content by leveraging external knowledge to obtain triple T1;

[0024] Triple verification: Use a large language model to verify the validity of the triple; in this step, the large language model is required to use retrieval-enhanced generation to analyze whether the triple in T1 is correct, so as to further enhance the credibility of the heterogeneous concept graph. The final triple is represented as T2 = {(c i ,r p ,c j )}|c i ,c j ∈C,r p ∈R}, R is the edge type set, r p is the edge type;

[0025] Each triple in T2 is regarded as an edge in the heterogeneous concept graph G, thereby transforming T2 into G; given the differences in relationship type information, G is decomposed into three relationship-specific subgraphs, denoted as G1, G2, and G3, involving pre-order relations, parallel relations, and collaborative relations.

[0026] Furthermore, the triple generation specifically includes:

[0027] Establishing an external knowledge base EK, wherein the external knowledge base is derived from knowledge concepts in web pages and mathematical knowledge base;

[0028] Using the extracted pairs, retrieve the top k relevant external knowledge P in EK by embedding similarity KB ;

[0029] Using the retrieved external knowledge P KB , and the extracted pairs and the original prompt Pgen form the final prompt P, which is fed to the large language model:

[0030]

[0031] After the instruction P is sent to the large language model, the large language model processes it and generates the corresponding triples, which are recorded as T1 = {(c i ,r p ,c j )}|c i ,c j ∈C,r p ∈R}.

[0032] Furthermore, the sub-image encoder comprises:

[0033] Use the pre-trained language model to capture the semantic features of knowledge concepts as the initial representation of nodes:

[0034] x i =PLM(text(c i ))

[0035] PLM represents the pre-trained language model;

[0036] In subgraph G1, the precedence relation requires modeling the different contributions of child concepts to their parent concepts; therefore, the representation of a node in subgraph G1 is calculated as:

[0037]

[0038] α ij is the attention weight, W1 is the trainable parameter matrix, and N(i) is the set of neighbors of node i;

[0039] In subgraph G2, parallel relations benefit from the neighborhood aggregation of graph convolutional networks to capture the local structural similarities between knowledge concepts; therefore, the representation of nodes in subgraph G2 uses:

[0040]

[0041] in, yes The degree matrix of is the sum of the original adjacency matrix and the unit matrix, W2 is the trainable parameter matrix;

[0042] In subgraph G3, for the collaborative relationship, CoAttention-GNN is used to characterize how two knowledge concepts enhance each other’s understanding; therefore, the representation of the nodes in subgraph G3 is:

[0043]

[0044] Among them, W 31 and W 32 are the trainable parameter matrices respectively.

[0045] Furthermore, the adaptive selector automatically assigns weights to subgraphs to enhance the overall performance of the exercise, specifically including:

[0046] Get three subgraphs α 1:3 The attention weights are as follows:

[0047]

[0048] Among them, W k is the trainable parameter matrix, d is the vector dimension, W q is a trainable parameter matrix, h v is the feature vector of node v;

[0049] Multiply the subgraph features with their corresponding weights to get the representation of the node:

[0050]

[0051] Among them, W vis a trainable parameter matrix.

[0052] Furthermore, the graph information aggregator organically integrates graph information into knowledge concepts and exercises, specifically including:

[0053] Assign weight μ to graph information i :

[0054] μ i =σ(f1·([h',h e ]))

[0055] Among them, h e is the feature representation of the problem, and f1 is the linear layer;

[0056] Get a representation with graph information:

[0057] h' e =h e +μ i e h'.

[0058] Furthermore, an interaction function based on deep learning is used to predict students’ responses:

[0059]

[0060] in, is the predicted value, h s is the student's representation, h′ diff is the characteristic representation of the difficulty of the exercise, h′ disc is the feature representation of the discriminability of the exercises, and f2 is a multi-layer perceptron;

[0061] The loss function L is the cross entropy loss between the output y and the true label y:

[0062]

[0063] Among them, y i is the true label.

[0064] In summary, the beneficial effects of this application are as follows:

[0065] This application is the first work to model knowledge concepts as heterogeneous concept graphs in cognitive diagnosis tasks to improve the assessment of students' knowledge mastery level.

[0066] This application proposes a heterogeneous concept map construction module to explore the complex connections between knowledge concepts and improve the accurate inference of knowledge mastery level.

[0067] This application proposes a heterogeneous concept graph modeling and aggregation module that can adaptively identify essential subgraph features and integrate graph representations. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 Examples of student practice responses;

[0069] Figure 2 The application processing process;

[0070] Figure 3 Examples of misunderstandings caused by directly generating triples;

[0071] Figure 4 Use the retrieved external knowledge PKB, the extracted pairs and the original prompt Pgen to form the final prompt P, which is fed to the LLM;

[0072] Figure 5 Experimental results. DETAILED DESCRIPTION

[0073] The present invention will be further described below with reference to the accompanying drawings, but the present invention is not limited in any way. Any changes or substitutions made based on the teachings of the present invention fall within the scope of protection of the present invention.

[0074] Artificial intelligence (AI) refers to the theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI also involves studying the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0075] Natural language processing (NLP) is a key area of ​​research in computer science and artificial intelligence. It studies the theories and methods that enable effective communication between humans and computers using natural language. Natural language processing (NLP) integrates linguistics, computer science, and mathematics. Therefore, research in this field involves natural language—the language we use in everyday life—and is closely linked to the study of linguistics. Natural language processing technologies typically include text processing, semantic understanding, machine translation, robotic question answering, and knowledge graphs.

[0076] The technical solutions provided in the embodiments of this application involve technologies such as machine learning and natural language processing of artificial intelligence, and are specifically introduced and explained through the following embodiments.

[0077] The task overview of this application assumes that there are students S={s1,s2,...,s N}, Exercise E={e1,e2,...,e M}, knowledge concept C={c1,c2,...,c K}. Let it be the response record set, where y is the answer score, that is, 1 means the student answered the exercise e correctly, otherwise it is 0. The Q matrix represents the relationship between the exercise and the knowledge concept, defined as Q = (q ij ) M×K , if exercise i contains knowledge concept j, then q ij =1, otherwise q ij =0.

[0078] Heterogeneous concept graph This application defines knowledge concepts and their relationships as heterogeneous concept graphs. Heterogeneous concept graphs are composed of finite triples T = {(c i ,r i ,c j )}|c i ,c j ∈C,r i ∈R}, where R is the set of relationships between knowledge concepts. A heterogeneous concept graph is also denoted as a directed graph G, where G = (C, D), C is the node set, and D is the edge set. Next, this application divides G into three subgraphs based on different edge types R, denoted as G1, G2, and G3.

[0079] Problem Definition Given a student's answer record Res, a matrix Q∈R and a heterogeneous concept graph G, the goal of this application is to obtain the student's mastery of each knowledge concept.

[0080] refer to Figure 2 , the HCGCDM of this application consists of two parts. In the heterogeneous concept graph construction module, first, knowledge concept pairs with high similarity are selected, the connections between knowledge concepts are identified, and an accurate heterogeneous concept graph is constructed through triple generation and triple evaluation. Then, the heterogeneous concept graph is divided into three subgraphs according to the relationship type. The heterogeneous concept graph modeling and aggregation module includes a subgraph encoder, an adaptive selector, and a graph information aggregator. This application first models the three subgraphs based on different graph neural networks, and obtains the basic representation through the subgraph encoder. Then, this application develops an adaptive selector to dynamically identify important subgraph features. Through the graph information aggregator, the knowledge concepts and practice features are updated. HCGCDM finally generates a diagnostic prediction of the student's knowledge mastery level based on the response records and graph information.

[0081] In one embodiment, heterogeneous concept graph construction includes:

[0082] To enhance the interconnectedness between knowledge concepts, this application constructs a heterogeneous graph based on knowledge concepts, called a heterogeneous concept graph. This heterogeneous concept graph not only enriches the interactions between knowledge concepts, but also visualizes the relationships between knowledge concepts, providing valuable insights for subsequent student performance analysis. In the heterogeneous concept graph, knowledge concepts are represented as nodes, and the relationships between them are described as edges. Because the various types of relationships between knowledge concepts are difficult to automatically obtain, based on existing literature, this application mainly models three representative relationships in this work:

[0083] -If mastering knowledge concept a is a necessary condition for mastering knowledge concept b, then there is a precedence relationship between a and b.

[0084] -Parallel relationship means that two knowledge concepts exist at the same level, such as "addition" and "subtraction".

[0085] -Synergistic relationships are characterized by two knowledge concepts reinforcing each other's understanding and application.

[0086] While deep learning-based methods exist to explore pre-order relations, they are unable to obtain triples involving pre-order relations without a training set. Furthermore, current methods lack an automated framework for extracting triples from parallel and collaborative relations. Inspired by the powerful capabilities of large language models (LLMs), this application proposes to leverage LLMs to generate triples from three types of relations.

[0087] Concept Pair Extraction. To accurately capture the relationships between knowledge concepts, this application feeds pairs of knowledge concepts as input to the LLM. Considering that not all pairs are useful, and taking into account the computational efficiency of the LLM when processing pairs, this application designs pair extraction to select pairs with a higher probability of forming meaningful triples. This application embeds knowledge concepts into vectors and calculates their similarity. For concept pairs whose similarity exceeds a threshold, this application identifies them as potential candidates for relationship detection:

[0088] Pairs={(c i ,c j )|sim(c i ,c j )>θ

[0089] Since knowledge concepts have different semantics, if LLM directly generates triples using the prompt Pgen and the extracted pairs Pair, it may lead to misunderstandings, such as Figure 3As shown. Therefore, the present application utilizes retrieval-augmented generation (RAG) to improve the reliability of generated content by effectively utilizing external knowledge. For example, "average" has many meanings, but using external knowledge of mathematics, it is obvious that "average" generally refers to the average. Specifically, the present application first establishes an external knowledge base EK to promote more accurate RAG, which is derived from knowledge concepts in web pages and mathematical knowledge bases. Then, the present application uses the extracted pairs to retrieve the top k relevant external knowledge PKB in EK by embedding similarity. Finally, the present application uses the retrieved external knowledge PKB, as well as the extracted pairs and the original prompt Pgen to form the final prompt P, which is fed to the LLM, as shown Figure 4 As shown:

[0090]

[0091] After the instruction P is sent to the LLM, the LLM processes it and generates the corresponding triplet, which is recorded as T1 = {(c i ,r p ,c j )}|c i ,c j ∈C,r p ∈R}.

[0092] The quality of the triples generated by triple verification directly affects the construction of the heterogeneous concept graph. In order to ensure the accuracy of the triples generated by triple generation, this application uses LLM again to verify the validity of the triples. In this step, this application requires LLM to use RAG to analyze whether the triples in T1 are correct to further enhance the credibility of the heterogeneous concept graph. RAG uses a similar process to triple generation. In order to save space, this application omits RAG. The final triple is represented as T2={(c i ,r p ,c j )}|c i ,c j ∈C,r p ∈R}.

[0093] This application regards each triple in T2 as an edge in the heterogeneous concept graph G, thereby transforming T2 into G. In view of the differences in relationship type information, this application decomposes G into three relationship-specific subgraphs, denoted as G1, G2, and G3, involving pre-order relations, parallel relations, and collaborative relations.

[0094] Heterogeneous concept graph modeling

[0095] Subgraph Encoder Due to the different interaction patterns between knowledge concepts, this application uses a dedicated GNN architecture for each subgraph. First, this application uses a pre-trained language model to capture the semantic features of knowledge concepts as the initial representation of nodes:

[0096] x i =PLM(text(c i ))

[0097] In G1, the precedence relation requires modeling the different contributions of child concepts to their parent concepts, which is naturally handled by two layers of GATs. Therefore, the representation of a node in G1 can be calculated as:

[0098]

[0099] In G2, parallel relations benefit from the neighborhood aggregation of graph convolutional networks to capture the local structural similarities between knowledge concepts. Therefore, the representation of nodes in G2 uses:

[0100]

[0101] In G3, for collaborative relationships, this application uses CoAttention-GNN to characterize how two knowledge concepts enhance each other's understanding. Therefore, the representation of nodes in G3 can be achieved through:

[0102]

[0103] Adaptive Selector Intuitively, different exercises focus on different types of relationships. For example, when students solve an algebra problem, the prerequisite relationships between knowledge concepts play a greater role in solving the problem. In other words, different exercises show different reliance on specific subgraph structures. Therefore, this application develops an adaptive selector to automatically assign weights to subgraphs, enhancing the overall performance of the exercises. This application first obtains the attention weights of the three subgraphs α1:3 as follows:

[0104]

[0105] Then, we multiply the subgraph features with their corresponding weights to get the representation of the node:

[0106]

[0107] In order to organically integrate graph information into knowledge concepts and exercises, this application designs a graph information aggregator. It first assigns a weight μ to the graph information. i :

[0108] μ i =σ(f1·([h',h e ]))

[0109] Then, we get a representation with graph information:

[0110] h' e =h e +μ i eh'

[0111] Model training:

[0112] This application uses deep learning-based interaction functions to predict student responses:

[0113]

[0114] The loss function L of HCGCDM is the cross entropy loss between the output y and the true label y:

[0115]

[0116] We conducted experiments on three real-world datasets: Assist2009-2010, Junyi, and Math1. For Junyi, we removed students with fewer than 15 responses to ensure sufficient interaction. We then randomly sampled 10,000 students and retained the 666 concepts contained in their responses.

[0117] Baseline Model. To fully verify the performance of HCGCDM, this application selected four baselines for comparison:

[0118] MIRT is a representative model of traditional IRT.

[0119] NCDM is a work that uses neural networks to model interaction functions.

[0120] KANCD considers knowledge concept associations to improve students' prediction performance.

[0121] HyperCDM proposes a hypergraph cognitive diagnosis model to capture homogeneous effects and alleviate over-smoothing.

[0122] In constructing the modular heterogeneous concept graph, this application used GPT-4o for the LLM for triple generation and verification. For model training, this application set epoch = 1, learning rate = 0.002, and dim = 20. All experiments were conducted on PyTorch 1.12.0 and a GeForce RTX 3090 with 24GB of RAM.

[0123] For the prediction performance task, this application uses common metrics to evaluate the performance of the CD method, including AUC, ACC, and MSE. Next, this application adopts the student performance prediction task, which aims to predict the student's response to the exercises and verify the performance of the CD method.

[0124] As shown in Table 1, HCGCDM outperforms other baselines on the Assist2009-2010 and Math1 datasets. Heterogeneous concept graphs enhance the interconnections between knowledge concepts and the fusion of knowledge concepts and exercise representations, thereby improving the accuracy of diagnosis. HyperCDM outperforms HCGCDM on the Junyi dataset. This application attributes this to two reasons: (1) In the Junyi dataset, knowledge concepts and exercises are exactly the same, which leads to interference in graph construction and results in suboptimal performance. For example, the text descriptions of knowledge concepts and exercises are both counted. (2) HyperCDM uses hypergraphs to reveal more relationships, thereby enriching information. However, HyperCDM has a key problem, which is that it is too time-consuming, which hinders its practical application in online education. For other baselines, the psychometrics-based method achieved the worst results. In general, the deep learning-based method performed better than the psychometrics-based method (MIRT). This may be because MIRT's modeling is too simple and it is difficult to fully capture the characteristics of students, exercises, and knowledge concepts.

[0125] Table 1 Experimental results on Assist2009-2010, Junyi and Math1 datasets.

[0126]

[0127] In order to study the impact of the number of knowledge concepts on diagnostic performance, this application compares the enhancements of AUC, ACC, and MSE of our proposal with those of other baselines. For the three datasets, this application chooses a method based on psychometrics (MIRT) for comparison. Figure 5 As shown, the improvements in AUC, ACC, and MSE are most pronounced in Math1 (Math), followed by Assist2009-2010 (Assist) and Junyi. Compared to the other two datasets, the Math1 dataset contains fewer knowledge concepts and data, resulting in limited information availability. Therefore, the construction of a heterogeneous concept graph enriches interactions, enabling the model to more effectively utilize sparse data and improve reasoning accuracy. In contrast, the Junyi dataset suffers from semantic overlap between knowledge concepts and exercise descriptions, leading to information redundancy during training. This fundamental data limitation leads to limited improvements on the Junyi dataset.

[0128] This application proposes HCGCDM, which improves the accuracy of CD by modeling knowledge concepts as heterogeneous concept graphs to explore deep and complex connections. Specifically, this application designs a heterogeneous concept graph construction module to identify complex connections between knowledge concepts. Through heterogeneous concept graph modeling and aggregation, knowledge concepts and exercise representations are combined with graph information. This application demonstrates the effectiveness of HCGCDM through several experiments. Exploring more relationships between knowledge concepts and further enhancing the interpretability of CD is a promising future research direction.

[0129] In summary, the beneficial effects of this application are as follows:

[0130] This application is the first work to model knowledge concepts as heterogeneous concept graphs in cognitive diagnosis tasks to improve the assessment of students' knowledge mastery level.

[0131] This application proposes a heterogeneous concept map construction module to explore the complex connections between knowledge concepts and improve the accurate inference of knowledge mastery level.

[0132] This application proposes a heterogeneous concept graph modeling and aggregation module that can adaptively identify essential subgraph features and integrate graph representations.

[0133] As used herein, the word "preferred" is intended to serve as an example, instance, or illustration. Any aspect or design described herein as "preferred" is not necessarily to be construed as advantageous over other aspects or designs. Rather, the use of the word "preferred" is intended to present concepts in a concrete manner. As used in this application, the term "or" is intended to mean an inclusive "or" rather than an exclusive "or." That is, unless otherwise specified or clear from the context, "X employs A or B" is intended to mean any of the naturally inclusive permutations. That is, if X employs A; X employs B; or X employs both A and B, then "X employs A or B" is satisfied in any of the foregoing examples.

[0134] Moreover, although the present disclosure has been shown and described with respect to one or implementation, those skilled in the art will think of equivalent variations and modifications based on reading and understanding of this specification and the accompanying drawings. The present disclosure includes all such modifications and variations and is limited only by the scope of the appended claims. In particular, with respect to the various functions performed by the above-mentioned components (such as elements, etc.), the terms used to describe such components are intended to correspond to any component (unless otherwise indicated) that performs the specified function of the component (such as it is functionally equivalent), even if structurally different from the disclosed structure that performs the function in the exemplary implementation of the present disclosure shown herein. In addition, although the specific features of the present disclosure have been disclosed with respect to only one of several implementations, such features can be combined with one or other features of other implementations that can be desired and advantageous for a given or specific application. Moreover, insofar as the terms "including", "having", "containing" or their variations are used in specific embodiments or claims, such terms are intended to be included in a manner similar to the term "comprising".

[0135] The functional units in the embodiments of the present invention may be integrated into a single processing module, or each unit may exist physically separately, or multiple or more units may be integrated into a single module. The aforementioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium. The aforementioned storage medium may be a read-only memory, a magnetic disk, or an optical disk, etc. The aforementioned devices or systems may execute the storage method in the corresponding method embodiment.

[0136] In summary, the above embodiment is one implementation method of the present invention, but the implementation method of the present invention is not limited to the described embodiment. Any other changes, modifications, substitutions, combinations, and simplifications that deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. An online education cognitive diagnosis method based on heterogeneous concept map construction and modeling enhancement, characterized by: The following steps are involved: Extracting multiple students' practice records and extracting multiple knowledge concepts from the practice records; Select knowledge concept pairs with similarity higher than a threshold, identify the connections between knowledge concepts, and construct a heterogeneous concept graph; The heterogeneous concept graph is divided into three subgraphs according to the relationship type; Obtain basic representations for the three sub-graphs through the sub-graph encoder; Dynamically identify important subgraph features through adaptive selectors; Update knowledge concepts and practice features through a graphical information aggregator; Generate diagnostic predictions of students' knowledge levels based on response records and graphical information.

2. The online education cognitive diagnosis method based on heterogeneous concept map construction and modeling enhancement according to claim 1 is characterized in that: In the heterogeneous concept graph, knowledge concepts are represented as nodes, and the relationships between knowledge concepts are described as edges. The relationships between the knowledge concepts include: If mastering knowledge concept a is a necessary condition for mastering knowledge concept b, then there is a precedence relationship between a and b; The parallel relationship indicates that two knowledge concepts exist at the same level; A synergistic relationship is one in which two knowledge concepts reinforce each other's understanding and application.

3. The online education cognitive diagnosis method based on heterogeneous concept map construction and modeling enhancement according to claim 1 is characterized in that: For multiple students S={s1,s2,...,s N }, N is the total number of students, there exists an exercise E={e1,e2,...,e M }, M is the total number of exercises, knowledge concept C={c1,c2,...,c K }, K is the total number of knowledge concepts, recorded as the response record set, where y is the answer score, that is, 1 means that the student answered exercise e correctly, otherwise it is 0; the Q matrix represents the relationship between exercises and knowledge concepts, defined as Q = (q ij ) M×K , if exercise i contains knowledge concept j, then q ij =1, otherwise q ij =0.

4. The online education cognitive diagnosis method based on heterogeneous concept map construction and modeling enhancement according to claim 3 is characterized in that: Use a large language model to generate triples based on the relationship between knowledge concepts: Concept pair extraction: Pairs of knowledge concepts are fed as input to a large language model; the knowledge concepts are embedded into vectors and similarity is calculated; for concept pairs whose similarity exceeds a threshold, they are identified as potential candidates for relationship detection: Pairs={(c i ,c j )|sim(c i ,c j )>θ c i and c j are the i-th and j-th concepts respectively; Triple generation: Utilize retrieval-enhanced generation to improve the reliability of generated content by leveraging external knowledge to obtain triple T1; Triple verification: Use the large language model to verify the validity of the triple; in this step, the large language model uses retrieval-enhanced generation to analyze whether the triple in T1 is correct to further enhance the credibility of the heterogeneous concept graph. The final triple is represented as T2 = {(c i ,r p ,c j )}|c i ,c j ∈C,r p ∈R}, R is the edge type set, r p is the edge type; Each triple in T2 is regarded as an edge in the heterogeneous concept graph G, thereby transforming T2 into G; given the differences in relationship type information, G is decomposed into three relationship-specific subgraphs, denoted as G1, G2, and G3, involving pre-order relations, parallel relations, and collaborative relations.

5. The online education cognitive diagnosis method based on heterogeneous concept map construction and modeling enhancement according to claim 4 is characterized in that: The triple generation specifically includes: Establishing an external knowledge base EK, wherein the external knowledge base is derived from knowledge concepts in web pages and mathematical knowledge base; Using the extracted pairs, retrieve the top k relevant external knowledge P in EK by embedding similarity KB ; Use the retrieved external knowledge P KB , and the extracted pairs and the original prompt Pgen form the final prompt P, which is fed to the large language model: After the instruction P is sent to the large language model, the large language model processes it and generates the corresponding triples, which are recorded as T1 = {(c i ,r p ,c j )}|c i ,c j ∈C,r p ∈R}.

6. The online education cognitive diagnosis method based on heterogeneous concept map construction and modeling enhancement according to claim 5 is characterized in that: The sub-image encoder comprises: Use the pre-trained language model to capture the semantic features of knowledge concepts as the initial representation of nodes: x i =PLM(text(c i )) PLM represents the pre-trained language model; In subgraph G1, the precedence relation requires modeling the different contributions of child concepts to their parent concepts; therefore, the representation of a node in subgraph G1 is calculated as: α ij is the attention weight, W1 is the trainable parameter matrix, and N(i) is the set of neighbors of node i; In subgraph G2, parallel relations benefit from the neighborhood aggregation of graph convolutional networks to capture the local structural similarities between knowledge concepts; therefore, the representation of nodes in subgraph G2 uses: in, yes The degree matrix of is the sum of the original adjacency matrix and the unit matrix, W2 is the trainable parameter matrix; In subgraph G3, for the collaborative relationship, CoAttention-GNN is used to characterize how two knowledge concepts enhance each other’s understanding; therefore, the representation of the nodes in subgraph G3 is: Among them, W 31 and W 32 are the trainable parameter matrices respectively.

7. The online education cognitive diagnosis method based on heterogeneous concept map construction and modeling enhancement according to claim 6 is characterized in that: The adaptive selector automatically assigns weights to subgraphs to enhance the overall performance of the exercise, specifically: Get three subgraphs α 1:3 The attention weights are as follows: Among them, W k is the trainable parameter matrix, d is the vector dimension, W q is a trainable parameter matrix, h v is the feature vector of node v; Multiply the subgraph features with their corresponding weights to get the representation of the node: Among them, W v is a trainable parameter matrix.

8. The online education cognitive diagnosis method based on heterogeneous concept map construction and modeling enhancement according to claim 7 is characterized in that: The graph information aggregator organically integrates graph information into knowledge concepts and exercises, specifically including: Assign weight μ to graph information i : μ i =σ(f1·([h',h e ])) Among them, h e is the feature representation of the problem, and f1 is the linear layer; Get a representation with graph information: h’ e =h e +μ i e h’。 9. The online education cognitive diagnosis method based on heterogeneous concept map construction and modeling enhancement according to claim 8 is characterized in that: Use deep learning-based interaction functions to predict student responses: in, is the predicted value, h s is the student's representation, h′ diff is the characteristic representation of the difficulty of the exercise, h′ disc is the feature representation of the discriminability of the exercises, and f2 is a multi-layer perceptron; The loss function L is the cross entropy loss between the output y and the true label y: Among them, y i is the true label.

Images