Multiview graph transformer cognitive assessment system, method, and storage medium with fused explicit features

Abstract

The application discloses a multi-view graph Transformer cognitive assessment system and method fusing explicit features and a storage medium. The system comprises a feature embedding module, which is used for generating basic representations for three types of entities, namely learners, questions and knowledge concepts, and fusing difficulty coefficients of questions and knowledge concepts into their representations respectively; an intra-view information aggregation module, which is used for constructing three heterogeneous bipartite graphs, namely learner-question, learner-knowledge concept and question-knowledge concept, and updating node representations based on graph Transformer; a multi-view feature fusion module, which is used for fusing representations of each entity in different views to generate specialized representations; and a cognitive assessment module, which is used for mapping the specialized representations of learners to a vector space with the number of knowledge concepts as the dimension, and the obtained vector is the cognitive assessment result and is used for predicting the probability of answering any question correctly. The application fuses explicit difficulty, models multi-view graph structure and interacts with graph Transformer in depth, thereby significantly improving the accuracy and interpretability of cognitive assessment.

Description

Technical Field

[0001] This application relates to the field of cognitive state assessment and prediction technology, specifically to a multi-view Transformer cognitive assessment system, method, and storage medium that integrates explicit features. Background Technology

[0002] Online learning platforms are entering a critical phase of expansion and technological advancement, with the education sector generating and accumulating an unprecedented scale of learner behavior data. This wealth of behavioral data offers limitless possibilities for cognitive assessment and personalized learning guidance.

[0003] Learners, knowledge concepts, and questions can be considered as three independent entities. The process of learners solving questions after learning knowledge concepts can be viewed as a comprehensive interaction. Breaking down this comprehensive interaction reveals the pairwise interactions between these independent entities: learners learning knowledge concepts, learners answering questions, and questions containing knowledge concepts. By deconstructing these interactions, we can characterize and infer the learner's cognitive state, depict the inherent attributes and relevance of knowledge concepts, and represent the knowledge assessment requirements and difficulty characteristics of the questions.

[0004] Early cognitive diagnostic models relied on statistical methods, using mathematical approaches such as logistic functions and parameter estimation for reasoning, with typical methods including Item Response Theory (OPT) and the DINA method. While these early methods offered good interpretability, their limited modeling capabilities made it difficult to describe the complex interactions of the real world, thus constraining diagnostic accuracy and generalization. With the intervention of deep learning, cognitive diagnostic models can delve deeper into the intrinsic relationships between entities. Although deep learning can characterize these relationships, the GNN aggregators currently used in deep learning models are insufficient in expressing the high-frequency, dynamic, and complex dependencies in knowledge interactions during the information aggregation stage. Furthermore, in terms of feature utilization, existing deep models primarily rely on implicit representations learned from learners' answer records, failing to incorporate explicit representations such as question difficulty. From an educational perspective, explicit features have clear educational significance; the inability to incorporate explicit features limits the richness of the model's representations. Summary of the Invention

[0005] In the first aspect, this application provides a multi-view graph Transformer cognitive assessment system that integrates explicit features.

[0006] The technical solution of this application is as follows: a multi-view graph Transformer cognitive assessment system integrating explicit features, including:

[0007] The feature embedding module is used to define the basic representation of entities in the form of a trainable matrix in a vector space, the entities including learners, questions, and knowledge concepts, and is used to define the question difficulty coefficient and the knowledge concept difficulty coefficient, and map the question difficulty coefficient and the knowledge concept difficulty coefficient to the basic representation of the question and the knowledge concept, respectively.

[0008] The in-view information aggregation module is used to build a heterogeneous bipartite graph and update the node representations in the heterogeneous bipartite graph based on the graph Transformer; the heterogeneous bipartite graph includes a learner-topic graph, a learner-knowledge concept graph, and a topic-knowledge concept graph; in the heterogeneous bipartite graph, the entities are treated as nodes, and the interaction relationships between entities are treated as edges;

[0009] A multi-view feature fusion module is used to fuse the representations of the entity in each heterogeneous bipartite graph and output the specialized representation of each entity.

[0010] The cognitive assessment module maps the specialized representations of each entity to a vector space with the number of knowledge concepts as the dimension. The learner's representation after mapping is used as the learner's cognitive assessment result, and the probability of the learner's correct prediction of the questions already answered is used as the output of the cognitive assessment module.

[0011] Furthermore, a triplet set is used to represent the interaction record between learners, questions, and knowledge concepts, and a Q-matrix is ​​used to represent the knowledge concepts associated with the questions. In the triplet set, real label values ​​are set for learners who answer questions correctly and incorrectly.

[0012] The difficulty level of a question is defined as the difference between 1 and the correct answer rate of all learners for that question in the interaction record.

[0013] The difficulty coefficient of a knowledge concept is defined as 1, which is the difference between the correct answer rate of all learners on questions related to that knowledge concept in the interaction record.

[0014] Furthermore, in the feature embedding module, the difficulty coefficient of the question and the difficulty coefficient of the knowledge concept are mapped to the basic representations of the question and the knowledge concept respectively through a dedicated linear transformation layer, as follows:

[0015] ;

[0016] ;

[0017] In the formula, , These represent the question difficulty coefficient and the knowledge concept difficulty coefficient, respectively. , These represent the difficulty level of the question and the difficulty level of the knowledge concept, respectively. , These represent the weight matrix parameters and bias parameters in the embedding model of the question difficulty coefficient, respectively. , These represent the weight matrix parameters and bias parameters in the knowledge concept difficulty coefficient embedding model, respectively.

[0018] but:

[0019] The learner's initial representation is its basic representation; the question's initial representation is the sum of its basic representation and the question's difficulty coefficient embedding; and the knowledge concept's initial representation is the sum of its basic representation and the knowledge concept's difficulty coefficient embedding.

[0020] Furthermore, in the multi-view feature fusion module, two sub-representations of a single entity in different heterogeneous bipartite graphs are stacked into a view matrix. The view matrix is ​​input into the self-attention module, which calculates the similarity score between the sub-representations and assigns attention weights to each sub-representation based on the similarity score. The weighted aggregate representation of the entity is calculated based on the attention weights assigned in the sub-table.

[0021] The weighted aggregate representation is averaged along the view matrix dimension, and the entity's specialized representation is obtained through layer normalization.

[0022] Furthermore, in the cognitive assessment module, the specialized representations of entities are mapped to a vector space with the number of knowledge concepts as the dimension through independent linear layers, and the representation of questions is enhanced through a knowledge enhancement mechanism.

[0023] Furthermore, the knowledge enhancement mechanism is configured as follows:

[0024] The mapped set of knowledge concepts associated with the question is averaged to obtain the knowledge background vector of the question.

[0025] By concatenating the mapped question representation with the knowledge background vector, we obtain the knowledge-enhanced question representation.

[0026] Furthermore, in the cognitive assessment module, the inner vector product of the learner's representation after mapping and the question representation after knowledge enhancement is calculated, and the predicted probability of the learner answering the question correctly is output through function mapping.

[0027] Secondly, this application provides a multi-view graph Transformer cognitive evaluation method that integrates explicit features. The method is implemented based on a multi-view graph Transformer cognitive evaluation system that integrates explicit features, and specifically includes the following steps:

[0028] S1. Collect the learner's answers to the questions, obtain the Q matrix corresponding to the answers, and assign an initial learning embedding vector to the learner.

[0029] S2. Construct learner-topic diagram, learner-knowledge concept diagram, and topic-knowledge concept diagram;

[0030] S3. Run the trained multi-view graph Transformer cognitive evaluation system with fused explicit features on three heterogeneous bipartite graphs and update all nodes;

[0031] S4. Integrate learners' representations in learner-topic diagrams and learner-knowledge concept diagrams, and output specialized representations of learners;

[0032] S5. Input the learner's specialized representation into the cognitive assessment module for mapping, and output the learner's cognitive assessment results through the cognitive assessment module.

[0033] Furthermore, the method is used to predict the probability of a learner answering an unanswered new question correctly, and includes the following steps: obtaining the question representation of the unanswered new question and its associated knowledge concept representation; inputting the learner's specialized representation, the question representation of the unanswered new question and its associated knowledge concept representation into the cognitive assessment module to output the predicted probability of the unanswered new question.

[0034] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned multi-view graph Transformer cognitive assessment method that integrates explicit features.

[0035] Due to the adoption of the above technical solution, the beneficial effects of this application are as follows:

[0036] 1. The features of this application are richer. Specifically, addressing the problem of insufficient utilization of explicit features in the prior art as pointed out in the background section, this application combines explicit prior knowledge with the basic representation of entities based on the difficulty coefficient of the questions and knowledge concepts involved in the interactive data, thereby providing the model with richer basic representations.

[0037] 2. This application enables deep interaction modeling. Specifically, it innovatively employs a graph Transformer architecture, constructing three heterogeneous bipartite graphs in pairs using interaction relationships of different properties between entities. These three heterogeneous bipartite graphs are independent of each other, and information is aggregated in each graph through the multi-head self-attention mechanism of the graph Transformer. This approach achieves profound modeling of complex dependencies in cognitive interactions by dynamically evaluating the importance of neighboring nodes.

[0038] 3. In the multi-view feature fusion module and the cognitive evaluation module, this application uses a self-attention module to weightedly integrate the representations of various heterogeneous bipartite graphs, generate the embedding of globally unified nodes, and then project them onto an interpretable knowledge space to predict the final answer performance based on the matching degree between the learner's state and the question requirements. Attached Figure Description

[0039] The accompanying drawings, which are provided to further illustrate this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application.

[0040] Figure 1 The architecture diagram of the Transformer cognitive assessment system, which integrates explicit features, is provided for this application. Detailed Implementation

[0041] Based on the background art, the cognitive assessment in this application can be formally described as a given learner's historical response record. and The matrix infers learners' performance on new questions and generates an interpretable cognitive state vector, where each dimension corresponds to the learner's mastery of the relevant knowledge concept.

[0042] For a knowledge concept, it is a definite entity. The teaching syllabus and examination outline both provide clear knowledge concepts, and the examiners create questions based on these concepts. The correspondence between a question and a knowledge concept is a definite entity, not determined subjectively. When a question tests a knowledge concept, it can be assumed that the question is related to or connected with that knowledge concept.

[0043] In this application, the set of entities is first defined, and the set of learners is represented as follows: The set of questions is The set of knowledge concepts is ,in , , The set represents the number of learners, questions, and knowledge concepts, respectively, with each element representing an entity. Furthermore, the interaction between learners and questions is recorded as a response log. These interaction records are represented as a set of triples. ,in learner For the question The answer result, This indicates a correct answer; otherwise, it indicates an incorrect answer. The knowledge concepts associated with the question are represented by the Q-matrix. This means that the Q matrix consists of 0s and 1s. Indicates the topic Examined the concept of knowledge Otherwise, there would be no examination.

[0044] To incorporate prior information with clear educational significance into the model, it is necessary to define the explicit difficulty characteristics of questions and knowledge concepts. Based on a given learner's historical response record, a difficulty coefficient is calculated for each question and knowledge concept. For each question... First, define a subset of all historical interaction records related to the topic. The difficulty level of a question is defined as the difference between 1 and the total correct answer rate of all learners for that question in the interaction log.

[0045] ;

[0046] Similarly, a micro-averaging strategy is used to calculate the difficulty coefficient of a knowledge concept. The difficulty coefficient of a knowledge concept is defined as the difference between 1 and the correct answer rate of all learners on questions related to that knowledge concept in the interaction record:

[0047] Using the Q-matrix, we first define the concept of knowledge. The relevant subset of interaction records is Then knowledge concept The difficulty level is defined as follows:

[0048] .

[0049] Specifically, for "cold start" questions or knowledge concepts that have not appeared in historical data, their difficulty coefficient is replaced by the average of questions or knowledge concepts with known difficulty coefficients.

[0050] As attached Figure 1 As shown, this application provides a multi-view graph Transformer cognitive assessment system that integrates explicit features, including:

[0051] The feature embedding module is used to define the basic representation of entities in the form of a trainable matrix in a vector space, the entities including learners, questions, and knowledge concepts; and to define the question difficulty coefficient and the knowledge concept difficulty coefficient, and map the question difficulty coefficient and the knowledge concept difficulty coefficient to the basic representation of the question and the knowledge concept, respectively.

[0052] To represent learners, questions, and knowledge concepts in vector space, we first define trainable embedding matrices for each of these three types of entities:

[0053] ;

[0054] in, , , Embedding matrices representing learners, questions, and knowledge concepts, respectively; Indicates the embedding dimension.

[0055] For learners ,topic and knowledge concepts ,That The fundamental representation of dimensions , , It can be obtained from its one-hot encoding vector and the corresponding embedding matrix:

[0056] ;

[0057] in, , , Representing learners ,topic Knowledge concepts The one-hot encoded vector.

[0058] In the feature embedding module, the question difficulty coefficient and the knowledge concept difficulty coefficient are mapped to the basic representations of the question and knowledge concept respectively through a dedicated linear transformation layer, as follows:

[0059] ;

[0060] ;

[0061] In the formula, , These represent the question difficulty coefficient and the knowledge concept difficulty coefficient, respectively. , These represent the difficulty level of the question and the difficulty level of the knowledge concept, respectively. , These represent the weight matrix parameters and bias parameters in the embedding model of the question difficulty coefficient, respectively. , These represent the weight matrix parameters and bias parameters in the knowledge concept difficulty coefficient embedding model, respectively.

[0062] The learner's initial representation is its basic representation; the initial representation of a question is the sum of its basic representation and the question's difficulty coefficient embedding; the initial representation of a knowledge concept is the sum of its basic representation and the knowledge concept's difficulty coefficient embedding, i.e.:

[0063] ;

[0064] in, , , These represent the initial representations of the learner, the topic, and the knowledge concept, respectively.

[0065] To specialize in modeling interactions of different natures, three independent heterogeneous bipartite graphs are established using an in-view information aggregation module, forming a learner-problem graph. Learner-Knowledge Concept Map And the topic - knowledge concept map The multi-view graph construction method can effectively reduce information interference that may occur due to the mixed modeling of different logical relationships, laying the foundation for subsequent specialized graph Transformer information aggregation.

[0066] In an independent bipartite graph, a Graph Transformer (GT) is applied to update the nodes. In a heterogeneous bipartite graph, the entities are treated as nodes, and the interactions between entities are treated as edges.

[0067] Learner nodes In the learner-topic diagram Learner-Knowledge Concept Map Taking the update in the middle as an example, after the first The update process of a layer graph Transformer can be represented as follows:

[0068] ;

[0069] In the formula, and For the first l Layered learners and the topic The representation, if , That is, respectively, the features embedded in the feature embedding module. ; and For the first l Layered learners and knowledge concepts The representation, if , That is, respectively obtained by the feature embedding module .

[0070] Learner - Learner in the Problem Diagram and its neighboring problem nodes Taking the interaction as an example, the first l The calculation process of the layer diagram Transformer is as follows:

[0071] In the l The first layer of the network In each attention head, the query vector for each head is first obtained through a linear transformation. Key vector Value vector :

[0072] ;

[0073] in, These are trainable weight matrices used to obtain the query vector, key vector, and value vector, respectively. This is an embedded representation of the corresponding student and question.

[0074] This represents the dimension of the corresponding attention head, with each head dimension being uniformly defined as follows: , This indicates the number of attention heads.

[0075] Subsequently, the weights between the learner and the questions they have answered are calculated based on scaled dot product attention. These weights are then multiplied by the value vector of the corresponding question to obtain the attention output vector for each learner, which is then aggregated. The attention results of the heads are processed through a linear mapping to obtain the final aggregate vector. :

[0076] ;

[0077] in, , This represents the trainable weight matrix in a linear mapping model.

[0078] Multi-head attention output Learner representations that aggregate question information, aggregate vectors The data is fed into a feedforward neural network, which contains a residual connection and a normalization layer before and after it. The calculation process is as follows:

[0079] ;

[0080] in, and These represent the intermediate features of the above calculation process. For learners to go through the learner-topic diagram for the first time The updated representation after the layer diagram Transformer; and The weight matrix is ​​a learnable matrix for the computation process; for brevity, the bias term is omitted from the formula. The updated representation of the problem can also be obtained through the graph Transformer mechanism of this bipartite graph. Similarly, in the learner-knowledge concept bipartite diagram, we can obtain the learner representation of the aggregated knowledge concepts. Aggregate learners' knowledge concept representation Furthermore, the question representation that aggregates knowledge concepts can be obtained from the question-knowledge concept map. Knowledge concept representation that aggregates question information .

[0081] In this embodiment, a single-layer graph Transformer is used. Ultimately, learner and question representations are obtained in the learner-question graph. and In the learner-knowledge concept map, we obtain the learner and knowledge concept representations. and The problem and knowledge concept representations are obtained from the problem-knowledge concept map. and .

[0082] After aggregating information from various views, each entity obtains multiple independent representations in different bipartite graphs. To form a global and unified understanding of the entities, the multi-view feature fusion module effectively merges these representations from different views and outputs specialized representations for each entity.

[0083] Taking the learner node as an example, in the previous module, the learner Representations were obtained from the learner-topic map and the learner-knowledge concept map, respectively. In the multi-view feature fusion module, two sub-representations of a single entity in different heterogeneous bipartite graphs are stacked into a view matrix. The view matrix is ​​input into the self-attention module, which calculates similarity scores between sub-representations and assigns attention weights to each sub-representation based on these scores. Based on this, the model can determine whether the learner's overall representation is formed through their problem-solving behavior (from...). More important is its knowledge structure (from...) More importantly, the attention weights assigned to the sub-tables are used for calculation. :

[0084] ;

[0085] In the formula, The weighted aggregate representation of the learner entity. , , These represent the weight matrices that the above models can be trained.

[0086] Finally, the weighted aggregate representation is averaged along the view matrix dimension, and specialized representations of the entities are obtained through layer normalization:

[0087] ;

[0088] in, Representing the learner's specialized representation, Representation layer normalization, This indicates that the first dimension is averaged.

[0089] Similarly, questions and knowledge concepts also acquire their respective specialized representations through self-attention mechanisms and average pooling. and .

[0090] To further suppress numerical instability caused by excessively large representation ranges, and to enhance the nonlinear expressive power of question embeddings, a tanh activation function is introduced after the question fusion representation to improve training stability and representation convergence.

[0091] ;

[0092] in, express.

[0093] To enhance the interpretability of the model and align it with the general paradigm of cognitive diagnosis, the specialized representations of the three types of entities are first projected uniformly onto a model based on the number of knowledge concepts through independent linear layers. In a dimensional space with clear cognitive meaning, before projection, the dimension is meaningless, and its value may be 8, 16, 32, etc. After projection, the dimension represents the number of knowledge concepts, and one dimension value can represent the measurement value of one knowledge concept. The specific transformation process is as follows:

[0094] ;

[0095] ;

[0096] ;

[0097] in, , , These are respectively represented as the learner's cognitive state vector after projection, the knowledge requirement vector after question projection, and the feature vector after knowledge concept projection. and These are represented as the linear projection matrix and bias term in the process of transforming learners, topics, and knowledge concepts, respectively.

[0098] In the cognitive assessment module, specialized representations of entities are mapped to a vector space with the number of knowledge concepts as the dimension through independent linear layers, and the representation of questions is enhanced through a knowledge enhancement mechanism.

[0099] The knowledge enhancement mechanism is configured as follows:

[0100] The mapped set of knowledge concepts associated with the question is averaged to obtain the knowledge background vector of the question:

[0101] For the question To obtain the set of knowledge concepts associated with it. The averaged representations of knowledge concepts after projection yield:

[0102] ;

[0103] in, This represents a general knowledge background vector for the corresponding question. Represented as related to the title The number of related knowledge concepts, This represents the feature vector after the previous knowledge concept is projected.

[0104] By concatenating the mapped question representation with the knowledge background vector, we obtain the knowledge-enhanced question representation:

[0105] Projection representation of the problem itself Its knowledge background vector The data is then concatenated and fused through a final linear layer to achieve dimensionality reduction, resulting in a knowledge-enhanced question representation. :

[0106] ;

[0107] in, and These are the trainable weight matrix and the bias, respectively.

[0108] Finally, learners In the title The learner's performance on the test is modeled as a matching process between the learner's cognitive state and the knowledge requirements of the question. The alignment is measured using a vector inner product, and the result is then mapped to the probability of a correct answer using a sigmoid function.

[0109]

[0110] in, For the Sigmoid function, This represents the dot product of vectors.

[0111] The probability of answering correctly is the final output of the system, while the final representation generated by the system for the learner is... This represents the student's level of mastery over various knowledge concepts, i.e., the final cognitive diagnostic vector. Each dimension of the cognitive state vector corresponds to the learner's level of mastery of the corresponding knowledge concept. Ultimately, As a trainable variable within the system, it is optimized through continuous iteration.

[0112] In practice, the learner's answer (correct or incorrect) is a typical binary classification problem. The binary cross-entropy loss function, widely used and robust in classification tasks, is employed, as shown in the following formula:

[0113] ;

[0114] in, Indicate learner In the title The real answer tag on the screen This indicates that the system predicts learners. Answer the question correctly. The probability, express log function, This represents the loss function.

[0115] The cognitive diagnostic module generates interpretable assessment results of learners' mastery of knowledge concepts. Taking the fused learner representation as input, the module transforms it into a cognitive state vector aligned with the knowledge concept system through a mapping function (such as a linear layer). Each dimension of this vector corresponds to a knowledge concept, and the numerical value represents the learner's mastery of that knowledge concept, serving as the final cognitive assessment result.

[0116] To ensure the validity of the cognitive assessment results, the module also predicts the learner's probability of answering already answered questions based on this cognitive state and compares the prediction with the actual answer labels. During model training, by minimizing the error between the predicted probability and the actual answer, the mapping function and the upstream representation generation process are optimized inversely, thereby driving the cognitive state vector to continuously approximate the learner's true knowledge structure. Therefore, the prediction accuracy for already answered questions is not the module's final output target, but a key mechanism for supervising and verifying the rationality of the cognitive assessment results.

[0117] As a verification of the scheme in this application:

[0118] To comprehensively and from multiple perspectives evaluate the effectiveness of the system proposed in this application, five real-world datasets were selected for experimentation. These datasets cover different subject areas and data scales, and each is equipped with a Q-matrix defined by domain experts to clarify the relationship between questions and knowledge concepts. Table 1 lists detailed statistical information for the five datasets, including the number of learners, the number of questions, the number of knowledge concepts, and the total number of interaction records, as well as several key indicators. Sparsity refers to the sparsity of the dataset, calculated by dividing the total number of interaction records by the product of the number of learners and the number of questions. Average Correct Rate represents the average score of all learners' responses in the dataset. Concepts covered per Learner represents the average number of knowledge concepts covered in the questions answered by learners. Proportion of Concepts covered per Learner represents the proportion of knowledge concepts covered by learners out of all knowledge concepts. Specific descriptions of each dataset are as follows:

[0119] Assist2017: This dataset originates from the 2017 ASSISTments Data Mining Competition and was collected by the online education platform ASSISTments. The data underwent preprocessing: for students who answered the same question repeatedly, only their first answer was retained. To ensure the validity of the interactive data, only students who answered more than 15 questions were retained.

[0120] EdNet-1: This dataset is a simplified form of EdNet, a large-scale hierarchical dataset in education. It describes a learner's learning process as a series of "question-answer" pairs. The data was collected by Santa, a multi-platform artificial intelligence tutoring system in South Korea, and primarily involves English listening and reading tests. To ensure direct comparison with baseline models and the reproducibility of the experiment, a preprocessed version from the paper "Inductive cognitive diagnosis for fast student learning in web-based intelligent education systems" was used. Specifically, the data contained in its publicly available codebase was directly adopted.

[0121] FrcSub: This dataset originates from a fraction addition and subtraction test conducted by a primary school. As a widely used public dataset, the FrcSub dataset has a sparsity of 1, meaning that every student answered all the test questions.

[0122] Math1 & Math2: These two datasets are derived from two final math exams at a high school. The questions are divided into objective and subjective questions. Similar to FrcSub, the sparsity of these two datasets is also 1. In data processing, the scores for objective questions were binary-coded: answers with a normalized score of 1 (i.e., full marks) were considered correct (denoted as 1), while answers with a score less than 1 were considered incorrect (denoted as 0).

[0123] Table 1. Dataset Statistics

[0124]

[0125] The specific experimental setup in this application is as follows:

[0126] Baseline model:

[0127] MIRT: As a multidimensional extension of Item Response Theory (IRT), MIRT is a classic model in the field of cognitive diagnostics. Unlike IRT, which uses a single ability value, MIRT represents both the learner's cognitive state and the characteristics of the items as multidimensional vectors in the same latent space. Each dimension of these vectors corresponds to a specific knowledge skill or latent trait.

[0128] DINA: DINA is a classic non-compensatory cognitive diagnostic model. Its core assumption is based on AND gate logic, which states that learners must master all the knowledge and skills required to answer a question correctly. To explain the uncertainty in actual answers, DINA introduces two noise parameters for each question: error rate (the probability of answering incorrectly despite having the ability) and guess rate (the probability of answering correctly despite lacking the ability).

[0129] NCDM: NCDM is a pioneering deep learning cognitive diagnostic framework designed to overcome the limitations of traditional models that rely on manually designed interaction functions. Its core idea is to utilize multi-layered neural networks to automatically learn and model the complex non-linear interactions between learners and questions. To ensure the interpretability of the neural network model, NCDM introduces the monotonicity assumption, ensuring that the probability of a learner correctly predicting a question does not decrease as their proficiency in relevant knowledge and skills improves.

[0130] RCD: RCD is a graph-based framework whose core is to model the complete hierarchical relationships between learners, questions, and knowledge concepts as a unified relational graph. This graph aims to simultaneously capture inter-layer interactions, such as the interaction between learners and questions, and intra-layer structural relationships, such as dependencies between knowledge concepts. Finally, the model utilizes a multi-layered attention network to aggregate this relational information for diagnostic purposes.

[0131] ICDM: ICDM is an inductive model that enables rapid diagnosis of new learners by avoiding the time-consuming retraining required by traditional transductive models. Its core mechanism is to construct a student-centered graph (SCG) and infer a new learner's knowledge level by aggregating information about the new learner's neighboring nodes in the graph (such as the questions they have answered).

[0132] CD-IDF: ID-CDF is a framework designed to address the issues of unidentifiable and interpretability overfitting in cognitive diagnostic models through a novel "Response-Proficiency-Response" paradigm. Inspired by encoder-decoder models, it employs a diagnostic module to inductively map responses to capability features, and a prediction module to reconstruct responses from these features.

[0133] Evaluation indicators:

[0134] To ensure a comprehensive and robust evaluation of model performance, the model's prediction results are evaluated from two perspectives: classification and regression. For classification tasks, which predict whether the learner's answer is correct or incorrect, accuracy (ACC) and area under the ROC curve (AUC) are used. For regression tasks, which measure the accuracy of the model's predicted probabilities, root mean square error (RMSE) is used.

[0135] Accuracy (ACC): As a primary classification metric, ACC measures the proportion of correctly predicted responses. To calculate ACC, the continuous probability values ​​output by the model are first converted into binary class labels using a threshold of 0.5. A higher ACC value indicates better predictive performance. Its formal definition is as follows:

[0136]

[0137] in This represents the set of all learner-question interaction pairs in the test set. This is an indicator function; its value is 1 when its internal condition is true, and 0 otherwise.

[0138] The Area Under the ROC Curve (AUC) evaluates a model's ability to correctly distinguish between positive and negative classes (correct answers). Its value ranges from 0.5 to 1.0, where 0.5 represents the performance of random guessing, and 1.0 represents perfectly correct classification. A key advantage of AUC is its robustness in handling imbalanced datasets, as its evaluation of predicted ranking quality does not depend on any specific classification threshold. A higher AUC value indicates better model performance.

[0139] Root Mean Square Error (RMSE): From a regression perspective, RMSE quantifies the average error between the model's predicted probability and the actual response. A lower RMSE value indicates a smaller deviation between the predicted and actual values, suggesting a more accurate and reliable model. Because its calculation includes a squared term for the error, this metric is particularly sensitive to larger prediction errors and incurs a higher penalty. Therefore, RMSE effectively reflects the model's performance in the worst-case scenario and is a commonly used indicator for measuring the model's predictive stability and robustness. Its calculation method is as follows:

[0140]

[0141] in, It is an interactive pair The true score, That is the corresponding predicted probability.

[0142] Implementation details:

[0143] Hardware environment:

[0144] All experiments were conducted using Python 3.8 and CUDA 11.6. For the system model proposed in this application, PyTorch 1.13.0 was used, and the graph neural network module was implemented using the DGL library (version 1.1.2+cu116). The hardware environment included two NVIDIA GeForce RTX 4090 GPUs (each with 24GB of VRAM) and a 16-core CPU (3.35 GHz). The operating system was Ubuntu 22.04.5 LTS.

[0145] Parameter settings:

[0146] In the system model, the embedding dimension of all entities is set to 64. The model architecture employs a single-layer graph Transformer with two attention heads, and its internal feedforward network has a hidden dimension of 128. The Adam optimizer is used for model training. The batch size is adjusted for each dataset: 64 for Assist2017 and FrcSub, 128 for EdNet-1, and 32 for Math1 and Math2. To mitigate overfitting, a Dropout ratio of 0.2 is applied. All learnable parameters are initialized using the Xavier normalization method.

[0147] Results analysis:

[0148] In cognitive diagnosis, learners' true cognitive state cannot be directly observed. Therefore, following the standard practice in most existing studies, the effectiveness of the model is indirectly evaluated by predicting learners' responses to questions. To comprehensively validate the performance of the system model, two main aspects are addressed: first, the proposed model is compared with multiple baseline models; second, a series of ablation experiments are conducted to deeply analyze the roles of key modules in the model, specifically the impact of explicit feature fusion and the graph Transformer aggregator on the model.

[0149] Baseline model performance comparison:

[0150] As shown in Table 2, overall, the system model proposed in this application achieves state-of-the-art or highly competitive performance across all datasets and metrics, demonstrating its universal effectiveness and robustness. Several key observations can be drawn from the results. First, all deep learning-based models—including NCDM, RCD, ICDM, CD-IDF, and the model in this application—generally outperform classic models such as MIRT and DINA. This performance gap is particularly pronounced on large-scale datasets, highlighting the superior ability of neural networks to capture the inherent complex nonlinear interactions during the learning process.

[0151] Secondly, on the large-scale, highly sparse datasets ASSIST2017 and EdNet-1, the deep learning-based model consistently and significantly outperformed the classical model. This demonstrates the effectiveness of deep neural networks in capturing complex nonlinear interactions in large-scale data, with the model in this application exhibiting unique advantages and achieving state-of-the-art performance in these challenging scenarios. Although ICDM achieved a slight lead over EdNet-1 in the RMSE metric, the difference was minimal. Conversely, on smaller, denser datasets, namely math1, math2, and FrcSub, the performance gap between the deep learning model and the classical model narrowed. This is because classical models are designed and tested on small, complete datasets similar to offline exams, allowing them to remain competitive on such datasets. However, the model in this application still maintained state-of-the-art performance on these datasets, highlighting its robustness across different data types. The fact that the model in this application can achieve state-of-the-art results on datasets with vastly different sizes and sparsities strongly demonstrates its excellent generalization ability.

[0152] Table 2 Performance comparison results between the model in this application and the baseline model

[0153]

[0154] Model ablation experiment:

[0155] To verify the contribution of the key innovations of the model in this application and to demonstrate the rationality of its architecture selection, a series of ablation experiments were conducted. Several model variants were designed and compared by removing key components or replacing the graph aggregator with other mainstream and representative graph neural networks (GNNs). These variants are defined as follows:

[0156] w / o Difficulty: Removed the model version of the explicit difficulty blending module.

[0157] w / o Aggregation: The model version of the Graph Transformer aggregation module (i.e., module two) has been completely removed. Since it no longer generates representations of multiple views, the subsequent multi-view fusion module (i.e., module three) has also been removed.

[0158] GCN, GAT, GraphSAGE, GIN: Replace the graph Transformer in the model of this application with the model version of the corresponding GNN aggregator.

[0159] The performance comparisons of these variants with the full model (EF-MGT) are shown in Table 3. A detailed breakdown of the results in the table follows:

[0160] First, the contribution of the explicit difficulty fusion module was evaluated. By comparing the full model and the w / o Difficulty variant, it was found that removing this module resulted in a significant decrease in model performance across all metrics on all five datasets, with an average decrease in AUC of 0.55%. This indicates that incorporating the explicit difficulty of knowledge concepts and questions as features into the model provides valuable prior knowledge, thereby significantly improving the model's prediction accuracy.

[0161] Secondly, the overall impact of the graph aggregation framework was analyzed by comparing the complete model with the w / o Aggregation variant. The results show that removing the graph aggregation and fusion modules resulted in a sustained and significant performance decrease across all datasets. For example, the AUC metric decreased by an average of 0.82% across the five datasets. This finding also confirms that the multi-view graph structure is the cornerstone of the model presented in this application, and that using only the initial embeddings is insufficient to capture the complex relationships required for accurate diagnosis.

[0162] Finally, to demonstrate the rationale for choosing the Graph Transformer, its performance was compared with several other classic GNN aggregators. As shown in Table 3, the model using the Graph Transformer outperforms variants using GCN, GAT, GraphSAGE, and GIN in most cases. Although GAT and GCN also show strong competitiveness on some small and dense datasets (such as math1-AUC and FrcSub-AUC), with results comparable to the model in this application, the Graph Transformer remains the optimal choice in terms of overall performance.

[0163] Table 3 Ablation Experiment Results

[0164]

[0165] Furthermore, the assessment method for new learners is based on the Transformer cognitive assessment system, which integrates explicit features and includes the following steps:

[0166] S1. Collect the learner's answers to the questions, obtain the Q matrix corresponding to the answers, and assign an initial learning embedding vector to the learner.

[0167] S2. Construct learner-topic diagram, learner-knowledge concept diagram, and topic-knowledge concept diagram;

[0168] S3. Run the trained multi-view graph Transformer cognitive evaluation system with fused explicit features on three heterogeneous bipartite graphs and update all nodes;

[0169] S4. Integrate learners' representations in learner-topic diagrams and learner-knowledge concept diagrams, and output specialized representations of learners;

[0170] S5. Input the learner's specialized representation into the cognitive assessment module for mapping, and output the learner's cognitive assessment results through the cognitive assessment module.

[0171] The above method can be used to predict the probability of a learner answering a new unanswered question correctly. It includes the following steps: obtaining the question representation of the new unanswered question and its associated knowledge concept representation; inputting the learner's specialized representation, the question representation of the new unanswered question and its associated knowledge concept representation into the cognitive assessment module to output the predicted probability of the new unanswered question.

[0172] The model's prediction of the accuracy of new questions that learners have not answered is consistent with the model's prediction logic on the test set: during the training phase, the model has completed the construction of embeddings that integrate explicit difficulty and global feature learning for all questions (including new questions), generating fixed question representations, and simultaneously obtaining the embedding representations of knowledge concepts and learners. By using the trained new question representation along with the corresponding knowledge concepts and student embedding representations as input to the model, the model can output the probability that the learner answers the new question correctly. The new question's own representation is not updated specifically because there are no corresponding student-question interaction records.

[0173] This application also provides a multi-view graph Transformer cognitive evaluation apparatus that integrates explicit features, including a processor and a memory. Optionally, the apparatus may further include a communication interface and a bus. The processor, communication interface, and memory can communicate with each other via the bus. The communication interface can be used for information transmission. The processor can call logical instructions in the memory to execute the multi-view graph Transformer cognitive evaluation method integrating explicit features described in the above embodiments.

[0174] 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.

[0175] Memory, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as program instructions / modules corresponding to the methods in the embodiments of this disclosure. The processor executes functional applications and data processing by running the program instructions / modules stored in the memory, thereby implementing the multi-view graph Transformer cognitive assessment method that integrates explicit features as described in the above embodiments.

[0176] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory may include high-speed random access memory and may also include non-volatile memory.

[0177] This disclosure provides a computer-readable storage medium storing computer-executable instructions configured to execute the above-described multi-view graph Transformer cognitive assessment method based on fused explicit features.

[0178] The aforementioned computer-readable storage medium may be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.

[0179] The technical solutions of this disclosure can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more 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 method described in this disclosure. The aforementioned storage medium can be a non-transitory storage medium, including: a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, and other media capable of storing program code; it can also be a transient storage medium.

[0180] This application was supported by the National Natural Science Foundation of China (72171136, 72571161) and the Natural Science Foundation of Shandong Province (ZR2024QG014).

[0181] For any parts not mentioned in this application, existing technologies may be used or referenced.

[0182] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A multi-view Transformer cognitive assessment system integrating explicit features, characterized in that, include: The feature embedding module defines the basic representations of entities in a vector space in the form of trainable matrices. These entities include learners, questions, and knowledge concepts. The module also defines question difficulty coefficients and knowledge concept difficulty coefficients, mapping them to their respective basic representations. Within the feature embedding module, the question difficulty coefficients and knowledge concept difficulty coefficients are mapped to their respective basic representations through dedicated linear transformation layers, as detailed below: ； ； In the formula, , These represent the question difficulty coefficient and the knowledge concept difficulty coefficient, respectively. , These represent the difficulty level of the question and the difficulty level of the knowledge concept, respectively. , These represent the weight matrix parameters and bias parameters embedded in the question difficulty coefficient model, respectively. , These represent the weight matrix parameters and bias parameters in the knowledge concept difficulty coefficient embedding model, respectively. but: The learner's initial representation is its basic representation; the question's initial representation is the sum of its basic representation and the question's difficulty coefficient embedding; the knowledge concept's initial representation is the sum of its basic representation and the knowledge concept's difficulty coefficient embedding. The in-view information aggregation module is used to build a heterogeneous bipartite graph and update the node representations in the heterogeneous bipartite graph based on the graph Transformer; the heterogeneous bipartite graph includes a learner-topic graph, a learner-knowledge concept graph, and a topic-knowledge concept graph; in the heterogeneous bipartite graph, the entities are treated as nodes, and the interaction relationships between entities are treated as edges; The multi-view feature fusion module is used to fuse the representations of the entity in each heterogeneous bipartite graph and output the specialized representation of each entity. In the multi-view feature fusion module, the two sub-representations of a single entity in different heterogeneous bipartite graphs are stacked into a view matrix. The view matrix is ​​input to the self-attention module, which calculates the similarity score between the sub-representations and assigns attention weights to each sub-representation based on the similarity score. The weighted aggregate representation of the entity is calculated based on the attention weights assigned in the sub-table. The weighted aggregate representation is averaged along the view matrix dimension, and the specialized representation of the entity is obtained through layer normalization. The cognitive assessment module maps the specialized representations of each entity to a vector space with the number of knowledge concepts as the dimension. The learner's representation after mapping is used as the learner's cognitive assessment result, and the probability of the learner's correct prediction of the questions already answered is used as the output of the cognitive assessment module.

2. The multi-view Transformer cognitive assessment system integrating explicit features according to claim 1, characterized in that, The interaction records between learners, questions, and knowledge concepts are represented by a set of triples, and the knowledge concepts associated with the questions are represented by a Q matrix. In the set of triples, real label values ​​are set for the questions that learners answer correctly and incorrectly. The difficulty level of a question is defined as the difference between 1 and the correct answer rate of all learners for that question in the interaction record. The difficulty coefficient of a knowledge concept is defined as 1, which is the difference between the correct answer rate of all learners on questions related to that knowledge concept in the interaction record.

3. The multi-view Transformer cognitive assessment system integrating explicit features according to claim 2, characterized in that, In the cognitive assessment module, specialized representations of entities are mapped to a vector space with the number of knowledge concepts as the dimension through independent linear layers, and the representation of questions is enhanced through a knowledge enhancement mechanism.

4. The multi-view Transformer cognitive assessment system integrating explicit features according to claim 3, characterized in that, The knowledge enhancement mechanism is configured as follows: The mapped set of knowledge concepts associated with the question is averaged to obtain the knowledge background vector of the question. By concatenating the mapped question representation with the knowledge background vector, we obtain the knowledge-enhanced question representation.

5. The multi-view Transformer cognitive assessment system integrating explicit features according to claim 4, characterized in that, In the cognitive assessment module, the vector inner product of the learner's representation after mapping and the question representation after knowledge enhancement is calculated, and the predicted probability of the learner answering the question correctly is output through function mapping.

6. A multi-view Transformer cognitive assessment method integrating explicit features, characterized in that: The method is implemented based on the multi-view graph Transformer cognitive assessment system that integrates explicit features as described in any one of claims 1 to 5, and specifically includes the following steps: S1. Collect the learner's answers to the questions, obtain the Q matrix corresponding to the answers, and assign an initial learning embedding vector to the learner. S2. Construct learner-topic diagram, learner-knowledge concept diagram, and topic-knowledge concept diagram; S3. Run the trained multi-view graph Transformer cognitive evaluation system with fused explicit features on three heterogeneous bipartite graphs and update all nodes; S4. Integrate learners' representations in learner-topic diagrams and learner-knowledge concept diagrams, and output specialized representations of learners; S5. Input the learner's specialized representation into the cognitive assessment module for mapping, and output the learner's cognitive assessment results through the cognitive assessment module.

7. The multi-view graph Transformer cognitive assessment method integrating explicit features according to claim 6, characterized in that, The method is used to predict the probability of a learner answering an unanswered new question correctly, and includes the following steps: obtaining the question representation of the unanswered new question and its associated knowledge concept representation; inputting the learner's specialized representation, the question representation of the unanswered new question and its associated knowledge concept representation into the cognitive assessment module to output the predicted probability of the unanswered new question.

8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the method described in any one of claims 6-7.

Images