Cognitive diagnosis method for multi-level interaction constraint perception

Abstract

The invention discloses a cognitive diagnosis method based on multi-level interaction constraint perception, relates to the technical field of cognitive diagnosis, and aims to more accurately perform quantitative analysis on cognitive states of students, firstly, based on historical performance data of the students, a hierarchical relation graph of knowledge concepts, practice and student abilities is constructed, and the knowledge concepts, the practice and the student abilities are obtained; the hierarchy and relevance among the three are comprehensively described, then a multi-round iteration hierarchical perception neural network is designed, through gradually associating knowledge concepts with practice hierarchies and practice and student ability hierarchies, features of different hierarchies are fused in multi-round iteration, question difficulty and student ability features with more expressive force are generated, and finally, the question difficulty and student ability features are extracted. In the diagnosis stage, the framework further combines student ability characteristics and personalized knowledge proficiency of multi-level interaction constraint perception to realize accurate evaluation of the student knowledge state, and the question difficulty and the multi-level characteristics thereof are fused to provide more comprehensive support for cognitive diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of cognitive diagnosis, and more particularly to a cognitive diagnosis method of multi-level interactive constraint perception. Background Art

[0002] Cognitive diagnosis (CD) plays a key role in educational data mining. It aims to reveal students' mastery of specific concepts through in-depth analysis of their response records. With the increasing adoption of intelligent educational technologies, the demand for cognitive diagnosis to improve individual learning outcomes and support personalized learning is growing. By analyzing data on student-task interactions, cognitive diagnosis not only effectively models students' knowledge level but also assesses their current learning status, helping them better understand their learning progress and improve learning efficiency.

[0003] Cognitive diagnostic models (CDMs) have undergone multiple stages of development, and their performance is highly dependent on the design of interactive mechanisms to characterize the complex relationships between students and knowledge points. While traditional item response theory (IRT) and its extended version, multidimensional item response theory (MIRT), provide a theoretical foundation for CDM research, their inherent assumptions present certain limitations in practical applications. As a classic form of CDM, the DINA model can deeply analyze students' mastery of knowledge points, thereby supporting the development of personalized learning paths. In recent years, with the integration of deep neural networks and graph neural network technologies, the performance of these models has been significantly improved, opening up new directions for cognitive diagnostic research.

[0004] Existing cognitive diagnostic frameworks have improved diagnostic accuracy to a certain extent by analyzing students' explicit response records (such as exercise difficulty and related knowledge concepts) and specific behaviors (such as guessing). However, these approaches primarily focus on automatically inferring students' knowledge proficiency through cognitive diagnostic frameworks, failing to fully consider the hierarchical relationships between different explicit records and, therefore, failing to more reasonably estimate students' true knowledge status. Summary of the Invention

[0005] The purpose of the present invention is to provide a cognitive diagnosis method with multi-level interactive constraint perception, which combines the student ability characteristics and personalized knowledge proficiency of multi-level interactive constraint perception to achieve accurate assessment of students' knowledge status, and integrates the difficulty of questions with their multi-level characteristics to provide more comprehensive support for cognitive diagnosis.

[0006] In order to achieve the above-mentioned purpose, the technical solution provided by one embodiment of the present invention is as follows:

[0007] A multi-level interactive constraint perception cognitive diagnosis method includes the following steps:

[0008] S1. First, identify and construct the hierarchical relationship of explicit records in real educational scenarios;

[0009] S2. Secondly, we deeply explore the feature interactions between layers and achieve more reasonable knowledge state modeling through iterative constraint-aware strategies;

[0010] S3. Finally, the integration of multi-level interactive perception features and personalized features significantly improves the expressive power of individual features and hierarchical features.

[0011] As a further improvement of the present invention, multiple core sets are defined, specifically including:

[0012] Student set S={s1,s2,...,s n}, contains n students;

[0013] Practice set E={e1,e2,...,e m}Includes m different exercises;

[0014] Knowledge point set KC = {kc1, kc2, ..., kc k} contains k different knowledge points; in addition, it also includes multiple level sets:

[0015] Knowledge point hierarchy set CH = {ch1,ch2,...,ch h}, describing the hierarchical structure of knowledge points;

[0016] Exercise level set EH = {eh1,eh2,...,eh h}, indicating the hierarchical structure of the exercise questions;

[0017] Student level set SH={sh1,sh2,...,sh h}, indicating students’ ability performance at different levels.

[0018] Here, h represents the number of levels involved in the framework, which is set according to the knowledge points, exercises, and students' cognitive development level, and the number of levels h is the same for all level sets;

[0019] Finally, the answer response set R = {0, 1} is used to represent the student's answer situation, where 1 indicates a completely correct answer and 0 indicates an incorrect answer.

[0020] As a further improvement of the present invention, the student's test record T is defined as a five-tuple set (s, e, ch, eh, sh), where s∈S represents a student in the student set, e∈E represents a practice question in the practice set, ch∈CH represents the level to which the knowledge concept belongs, eh∈EH represents the level to which the practice question belongs, and sh∈SH represents the student's ability level. In order to describe the association between the practice and the knowledge point, a predetermined Q matrix is introduced, denoted as Q=Q ij m×k , where m is the number of exercises and k is the number of knowledge points, if and only if exercise e i and knowledge kc j When associated, Q ij =1; otherwise, Q ij =0.

[0021] As a further improvement of the present invention, the student's personalized knowledge proficiency vector θ sp is a vector of length k, which is used to characterize the student's proficiency in each knowledge point. Its calculation process is composed of the student's unique hot encoding vector x s Multiply it with a trainable knowledge proficiency matrix S and map it through the sigmoid activation function, which is specifically expressed as follows:

[0022] θ sp =σ(x s ×S),

[0023] θ sp ∈(0,1) 1×k represents the student’s personalized proficiency in each knowledge point, x s ∈{0,1} 1×n is the student’s unique hot encoding, used to identify the specific student, S∈R n×k Is a trainable parameter matrix that captures the proficiency relationship between students and knowledge points. Function σ(·) represents the sigmoid activation function, which is used to normalize the output value. The one-hot encoding of the exercise x e Used to characterize the characteristics of the exercise, the difficulty of the knowledge point test h diff p ∈(0,1) 1×K and practice discrimination h disc ∈(0,1) further characterizes the difficulty and discrimination ability of the exercise, and its calculation formulas are:

[0024] h diff p =σ(x e ×A),

[0025] h disc e =σ(x e×B), where

[0026] x e ∈{0,1} 1×m Represents the one-hot encoding of exercise e, A∈R M×K and B∈R M×1 is a trainable parameter matrix. At the same time, the Q matrix is used to describe the relationship between exercises and knowledge points. The calculation formula of its correlation vector is:

[0027] Q e =x e ×Q,

[0028] Among them, Q e ∈{0,1} 1×k It reflects the degree of relevance between the exercise and the knowledge point;

[0029] The performance of students in completing the exercises can be described by the following model: r = M(θ,η e ), where r represents the student's response (such as the correct answer or score), θ represents the student's knowledge level, and η e is a parameter related to practice, and M is based on a neurocognitive diagnostic model that divides the student’s overall knowledge proficiency θ into two components:

[0030] θ=S(θ sh ,θ sp ),

[0031] h diff =D(h diff p ,h diff h ), where θ sh represents the capability characteristics constrained by the hierarchical structure, and θ sp Represents the individual ability characteristics of students, h diff h ∈(0,1)1×K represents the difficulty of the questions constrained by the hierarchical structure, while functions S and D combine the two to reflect the multi-dimensional influence of students' knowledge mastery level and question difficulty.

[0032] As a further improvement of the present invention, through the relationship matrix R(e i ,k j ) and R(s i ,e j ) represent the correlation between exercises and knowledge points, and between students and exercises respectively. The specific expressions are:

[0033] ε ek ={(e i , k j ), (k j , e i)|R(e i , k j )=1},

[0034] ε se ={(s i , e j ), (e j , s i )|R(s i , e j )=1},

[0035] Among them, R(e i ,k j )=1 and R(s i ,e j )=1 respectively means constructing bidirectional edges to represent the relationship between exercises and knowledge points, and between students and exercises, and expressing the hierarchical relationship of all nodes through a unified hierarchical identification calculation function: H(v i )←h(v i ),v i ∈{e i ,c j ,s i}, where h(v i ) is a unified level identification calculation function, which calculates the level identification of different types of nodes respectively, specifically:

[0036]

[0037] By integrating the hierarchical identification of nodes and the relationship between edges, we construct two graph structures representing different hierarchical relationships: the hierarchical graph G of knowledge points and exercises. ec and exercises with students' hierarchical graph G se , which is expressed as:

[0038] G ec =(H ec ,E ec ),G se =(H se ,E se ),

[0039] Among them, H ec and H se Indicates the hierarchical identity of the corresponding node, E ec and E se The edge set of the graph depicts the complex hierarchical interactive relationships between knowledge points and exercises, and between exercises and students.

[0040] As a further improvement of the present invention, the embedding of knowledge concept level (t) k,First, calculate the attention correlation score e between node k and its neighbor node d (t) kd , the score is calculated using the following formula:

[0041] e (t) kd =ξ(w T a [Wch (t) k ∥Weh (t) d ]),

[0042] Where ξ(·) represents the LeakyReLU activation function, eh (t) d is the embedding representation at the topic level, w a is a learnable weight vector, W is the weight matrix of the linear transformation, and the symbol ∥ represents the vector concatenation operation;

[0043] Next, the calculated e (t) kd After normalization, the neighbor attention weight α is obtained (t) kd , and its calculation formula is:

[0044]

[0045] Among them, G^ ec Indicates the current node k in the knowledge point and exercise level graph G ec The set of neighbors in ;

[0046] Then, by the attention weight α (t) kd The information of its related topic-level neighbor nodes is weightedly aggregated to obtain the aggregated knowledge concept hierarchical embedding ch^ (t) k :

[0047]

[0048] Where φ(·) is the Sigmoid activation function;

[0049] Finally, the knowledge concept hierarchy is updated through multiple rounds of iterative embedding using the following formula:

[0050] ch k (t+1) =ch k (t) +β k (t) ch^ k (t) ,

[0051] Among them, the fusion weight β k (t) Calculated by the following formula:

[0052]

[0053] Among them, f k is the score calculated by linear transformation, w b and b are learnable parameters.

[0054] As a further improvement of the present invention, through the topic-knowledge point hierarchy graph G ec and topic-student level graph G se It is iteratively updated, and the specific calculation process is as follows:

[0055] First, the topic level embedding is updated through the knowledge concept level, and the calculation formula is:

[0056]

[0057] Among them, ch k (t) Represents the hierarchical embedding of the current knowledge concept, G ec is the set of neighborhood knowledge points of topic d, α dk (t) is the attention weight, representing the contribution of knowledge point k to question d, and φ(·) represents the activation function;

[0058] Then, the question embedding is updated at the student level using the following formula:

[0059]

[0060] Among them, sh z (t) represents the ability embedding of student z, α dz (t) is the attention weight, G se is the neighborhood student set of question d;

[0061] Next, the above two update results are weighted and fused to obtain the final title embedding update:

[0062]

[0063] Among them, the weight parameter β1 (t) and β2 (t) Adaptive calculation, as follows:

[0064]

[0065] Among them, qd (t) Indicates the current feature representation of question d, q~ d (k，t) and q~ d (s，t) are the updated embedding representations of knowledge points and students respectively, [·∥·] is the feature concatenation operation, w c is a trainable weight parameter

[0066] As a further improvement of the present invention, the embedding of student ability levels is completed through message aggregation of question level information and dynamic adjustment of itself. The specific calculation is as follows:

[0067] First, based on the student's neighborhood question set G se ,

[0068]

[0069] Among them, α zd (t) is the attention weight, which measures the contribution of question d to the embedding update of student z, and φ(·) is the activation function;

[0070] Based on message aggregation, the update formula for embedding the student ability level is:

[0071] sh z (t+1) =sh z (t) +sh~ z (t) , among which, sh z (t) is the embedding representation of student z at the current level, sh~ (t) is the update amount obtained by aggregating the topics.

[0072] As a further improvement of the present invention, the level of difficulty h of the calculation question is diff h and the hierarchical characteristics of students’ knowledge mastery θ sh , which is calculated as follows:

[0073] h diff h =σ(W1·(eh (t+1) d ∥ch (t+1) k +b1),

[0074] θsh=σ(W2·(eh (t+1) d ∥sh (t+1) z +b2),

[0075] Among them, W1 and W2 are trainable weight matrices, b1 and b2 are bias terms, and the activation function σ(·) is the Sigmoid function.

[0076] As a further improvement of the present invention, by designing an adaptively optimized personalized weighting strategy, hierarchical features are effectively combined with personalized features to generate a comprehensive status representation of the student:

[0077] θ=α p ·θ sh +(1-α p )·θ sp ,

[0078] h diff =β p ·h diff h +(1-β p )·h diff p

[0079] Among them, α p and β p is an adjustable personalized weight coefficient, which is used to control the weight ratio of students' knowledge mastery characteristics and question difficulty characteristics in the fusion process. The basic paradigm of the cognitive diagnosis model is used to predict students' response to exercises. The specific formula is:

[0080] y=Q e o(θ-h diff )×h disc ;

[0081] The cross entropy loss function is used to measure the difference between the predicted value y and the student’s actual answer label r. The specific form of the loss function is as follows:

[0082]

[0083] Among them, r i represents the student’s actual answer label, y i represents the predicted probability of the model.

[0084] Compared with the prior art, the advantages of the present invention are:

[0085] The present invention achieves accurate assessment of students' knowledge status by combining students' ability characteristics and personalized knowledge proficiency with multi-level interactive constraint perception, and integrates the difficulty of questions with their multi-level characteristics to provide more comprehensive support for cognitive diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 This is a cognitive diagnosis framework diagram of the present invention;

[0087] Figure 2 The kernel density plots show the performance of the multi-level interactive constraints of the present invention on two typical data sets (PISA-Read and PISA-Math);

[0088] Figure 3 This is a graph showing the experimental results of the present invention with the number of layers ranging from 4 to 30;

[0089] Figure 4 The CEKT model of the present invention tracks the evolution of students' knowledge status over time in the process of solving three programming problems;

[0090] Figure 5 This is a diagram of the MHCD cognitive diagnosis results of the present invention;

[0091] Figure 6 is a table of key statistical features of each data set of the present invention;

[0092] Figure 7 This is a table of experimental results of the MHCD framework of the present invention;

[0093] Figure 8 This is a table of experimental results of various classic cognitive diagnostic models after combining the present invention with MHCD. DETAILED DESCRIPTION

[0094] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention; it is obvious that the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0095] Example 1:

[0096] To validate the performance of the multi-level interactive constraint-aware cognitive diagnostic framework, experiments were conducted based on the PISA 2015 public dataset. This dataset contains over 500,000 student records from 73 countries and regions, and sub-datasets in the areas of science, reading, and mathematics were selected for research. During the preprocessing stage, responses marked as "Full credit" were considered correct (labeled 1), while responses marked as "No credit" and "Partial credit" were classified as incorrect (labeled 0), and unanswered data were directly ignored. Furthermore, to ensure statistical significance of the sample, only student data with at least 30 response records were retained in the Science and Reading datasets, while this threshold was set at 20 for the Math dataset. Figure 6Key statistical characteristics of each dataset are summarized. To ensure the credibility of the experimental results, all datasets were randomly divided into a training set (70%), a validation set (10%), and a test set (20%) to support model construction and performance evaluation. This strict division provides a solid foundation for subsequent experiments.

[0097] In this embodiment, the model parameters are first initialized using the Xavier method, and training is performed through 5-fold cross validation, and the final evaluation results are based on the average of five rounds of validation. The experiment is run on a 64-bit Ubuntu 20.04.5 LTS server equipped with a 2.30 GHz Intel Xeon Gold 5218 processor and a 32 GB Tesla V100 graphics card. The computing framework uses PyTorch for model training and optimization. In order to comprehensively evaluate the effect of the multi-level interactive constraint perception cognitive diagnosis model, a variety of classic benchmark models are selected for comparative experiments. All models are reproduced based on the public code library, and their performance is ensured to reach the best level through meticulous hyperparameter optimization, thereby ensuring the fairness and reliability of the results. The experimental design uses the area under the curve (AUC), accuracy (ACC) and root mean square error (RMSE) as evaluation criteria, and comprehensively analyzes the model performance in multiple dimensions.

[0098] IRT (Lord, 2012): This method is based on item response theory. It models students’ performance in answering questions to estimate their ability level and the difficulty of the test questions. It is suitable for measuring a single potential ability.

[0099] MIRT (Ackerman, 2014): MIRT is an extension of the IRT model that supports multidimensional latent ability modeling. This method captures student responses across multiple dimensions, simultaneously estimating both student ability and test item parameters, making it suitable for assessing complex cognitive abilities.

[0100] DINA (DeLaTorre, 2009): The DINA model assumes that each question involves multiple knowledge points and infers a student's mastery of these knowledge points based on their responses. The model describes the status of knowledge points as a binary classification (mastered or not mastered), making it suitable for refined cognitive diagnosis.

[0101] NCDM (Wang et al., 2020): NCDM combines neural networks to learn the complex relationship between students and questions. Through multi-layer network modeling, this method can efficiently perform cognitive diagnosis and provide interpretable predictions of student ability.

[0102] RCD (Gao et al., 2021): The RCD framework improves the accuracy of cognitive diagnosis and representation learning ability by modeling the relationship between students, questions and knowledge points.

[0103] GCDM (Su et al., 2022): GCDM uses heterogeneous cognitive graphs to model the interactions between students, skills, and questions, and combines performance propagators with attention-based aggregators to evaluate skill mastery levels.

[0104] SCD (Wang et al., 2023): SCD assists cognitive diagnosis through self-supervised learning and graph obfuscation technology, providing more accurate diagnosis for students with sparse interaction records.

[0105] DCD (Chen et al., 2024): The DCD model separates conceptual factors through students' answer records and limited annotation information, and aligns these factors to achieve efficient cognitive diagnosis.

[0106] ICDM (Liu et al., 2024): ICDM uses a student-centered graph to infer the knowledge mastery level of new students without retraining the model while maintaining high diagnostic performance.

[0107] The inability to directly measure students' true knowledge makes evaluating the performance of cognitive diagnostic models challenging. While improving the model's predictive accuracy is not the ultimate goal, using student performance prediction as an indirect measure can effectively verify the model's rationality. This analysis further demonstrates the unique value of the Multi-Level Interactive Constraint Perception (MHCD) framework in capturing students' knowledge. Figure 7 The experimental results of the MHCD framework were presented, revealing the following key points: First, the MHCD model achieved optimal performance in all three main evaluation indicators, indicating that multi-level joint modeling of knowledge concepts, exercises, and student abilities can significantly improve the performance of the model. Second, compared with traditional cognitive diagnosis methods, models based on neurocognitive diagnosis perform better. For example, models such as NCDM, RCD, and DCD are significantly better than traditional models such as IRT and MIRT. Finally, whether on the PISA-Science dataset with a larger data volume or the PISA-Read dataset with a smaller data scale, the MHCD model showed excellent performance, which not only demonstrated its good universality, but also further verified its strong adaptability and evaluation capabilities in large-scale data scenarios.

[0108] In addition, ablation experiments were designed to verify the effectiveness of the three key relational modules (CHR, EHR, and SHR) in the MHCD framework. CHR represents the knowledge concept hierarchy relation, EHR represents the exercise hierarchy relation, and SHR represents the student ability hierarchy relation. In the experiments, each of these three modules was replaced with a standard aggregation layer, which simply aggregates the input without learning feature relationships, while the other modules remained unchanged. Figure 7The experimental results under different scenarios are summarized and the conclusions are as follows:

[0109] Importance of each module: No matter which layer is replaced, the model performance decreases, which shows that all modules contribute significantly to the final performance, verifying their effectiveness in modeling hierarchical intra- and extra-layer relationships.

[0110] The key role of the SHR layer: When the SHR layer is replaced, the performance drops the most, indicating that in multi-level interaction constraint modeling, the rational use of student ability levels is crucial for accurately assessing students' knowledge status.

[0111] The above results fully demonstrate the advantages of the MHCD framework in multi-level modeling and the indispensability of its key modules.

[0112] The multi-level interactive constraint perception module in MHCD effectively captures hierarchical features while demonstrating strong versatility. To further validate its performance, the module was ported to two classic traditional cognitive diagnosis frameworks (IRT and DINA) and a neural network-based cognitive diagnosis framework (RCD), and experimental evaluations were conducted on both datasets. Figure 8 The experimental results of various classic cognitive diagnostic models combined with MHCD are summarized.

[0113] The experimental results provide the following key conclusions: First, models incorporating MHCD (e.g., MHCD-IRT) significantly outperform traditional models without it (e.g., IRT and DINA) in metrics such as AUC, ACC, and RMSE, demonstrating that incorporating multi-level constraint features into cognitive diagnosis can effectively improve model performance. Second, while the MHCD module still achieves top performance across all evaluation metrics when applied to the RCD model, the improvement compared to the traditional model is relatively small. This may be because the RCD framework already fully leverages graph neural networks for comprehensive modeling. In contrast, traditional non-neural network models benefit even more significantly from the introduction of MHCD, thanks to the deep capture of hierarchical features by multi-level interactive constraints. In summary, while the MHCD module demonstrates excellent versatility and cross-framework adaptability, the experimental results also suggest an important direction: redesigning a cognitive diagnosis framework based on the MHCD module can more systematically leverage its advantages and further improve the diagnosis of students' knowledge status.

[0114] The purpose of introducing multi-level interactive constraint perception is to enable the cognitive diagnosis model to more accurately reflect students' ability performance, thereby achieving reasonable inference of students' ability levels in various knowledge concepts. This constraint helps to ensure the rationality and effectiveness of the assessment results. Figure 2The performance of multi-level interaction constraints on two typical datasets (PISA-Read and PISA-Math) is demonstrated through kernel density plots. The experimental results reveal several important findings. First, both the traditional NCDM model and the MHCD model with multi-level interaction constraints can effectively distinguish the distribution of students' abilities on both datasets, and the performance is consistent with the statistical distribution trend of students' average scores, indicating that existing cognitive diagnostic techniques are highly effective in assessing students' knowledge status. Second, by comparing the performance of the two datasets, it is found that although the average level of knowledge status learned by the two models on the PISA-Read dataset is similar to the students' prior distribution, their overall distribution deviates significantly from the normal distribution. On the PISA-Math dataset, which has a larger data volume, the distribution of students' abilities is closer to the normal distribution, which is also consistent with the general law of ability development.

[0115] Third, experiments show that on the PISA-Math dataset, the student knowledge states learned by the NCDM model fluctuate significantly. However, after introducing multi-level interactive constraints, the student ability representation becomes smoother and more consistent with the prior statistical distribution of student evaluation scores. This phenomenon demonstrates that the MHCD model not only learns a more reasonable student ability state, but also avoids neglecting the rationality of the student ability representation in the pursuit of improved accuracy, thereby improving the model's interpretability. Finally, further analysis of the distribution of students at an ability level of 0.5 in both datasets, particularly those at an ability level of 0.5, reveals that the model incorporating multi-level interactive constraints achieves a distribution density closer to the prior distribution of students at this level. For example, in the PISA-Math dataset, the prior distribution of students at an ability level of 0.5 accounts for approximately 24.0%, while the NCDM model learns only 13.5%. In contrast, the MHCDNCDM model achieves a 24.5% representation of students at this level, which is more consistent with the prior distribution. This result further validates the positive role of introducing hierarchical interactive constraints in more reasonably reflecting student knowledge states.

[0116] In order to explore the impact of the number of hierarchical divisions on model performance, Figure 3The experimental results for the number of layers ranging from 4 to 30 are shown. It can be seen that when the number of layers increases from 4 to 20, the model performance gradually improves and reaches the optimal level when divided into 20 layers. However, when the number of layers is further increased to between 20 and 30, the model performance fluctuates and decreases slightly. This shows that the more detailed the layer division is, the better. The possible reason is that too detailed division will lead to increased data sparsity and noise amplification, which will have a negative impact on the learning effect and performance of the model. It is worth noting that, inspired by (Forehand et al., 2005; Sun et al., 2024c), dividing the layers into 6 capability levels has better interpretability. Therefore, although appropriately increasing the number of layer divisions can improve model performance to a certain extent, this embodiment still uses 6 layers for experiments to balance performance improvement and result interpretability.

[0117] A student's mastery of a specific knowledge point directly affects their problem-solving performance. For example, students with a higher mastery level are more likely to succeed on related problems (Chen et al., 2017). To evaluate the effectiveness of models in capturing this relationship, the Degree of Agreement (DOA) metric (Fouss et al., 2007) was introduced. Traditional models (such as IRT and MIRT) depict student ability through latent trait vectors, but they struggle to accurately reflect their mastery of specific knowledge points. This study proposes a multi-level interactive constraint perception model (MHCD) and compares it with models such as Random, DINA, NCDM, RCD, and DCD. The DOA metric is used to analyze the differences between these models in describing student knowledge mastery.

[0118] Analysis of the DOA results yielded the following key conclusions: First, the DINA model achieved the most outstanding DOA performance on all three datasets, demonstrating its significant advantage in explaining students' knowledge mastery. This result validates the DINA model's theoretical foundation, which relies on a clear connection between students and knowledge points. However, despite its excellent interpretability, the DINA model's predictive power remains limited in practical applications, failing to fully capture students' dynamic learning processes. Second, while neural network-based methods such as NCDM, RCD, and DCD achieved strong DOA performance, they still lagged behind the DINA model. The MHCD model's DOA performance was roughly on par with DINA, demonstrating that the introduction of multi-level interaction constraints significantly improved the neural network model's explanatory power. Furthermore, it is noteworthy that the DOA of the random assignment model remained stable at approximately 0.5, which was within expectations. The MHCD-Random model, which incorporates multi-level perception, achieved significant improvements in DOA, demonstrating that even in the absence of explicit knowledge mastery information, the introduction of perception mechanisms can effectively improve model performance. This phenomenon emphasizes the importance of multi-level perceptual learning in evaluating students' knowledge status, and provides a theoretical and practical basis for further improving the interpretability of the model.

[0119] The relationship between students’ cognitive diagnostic results, their abilities, and the difficulty of questions can be quantitatively analyzed using Eq. (23). A student and 20 questions answered by him were randomly selected from the PISA-Science dataset. Figure 5 The upper part shows the diagnosis results when the multi-level interactive perception mechanism is not introduced (the model degenerates into NCDM), and the lower part shows the diagnosis results of the MHCD framework.

[0120] The following key conclusions can be summarized from the charts: First, both frameworks can effectively learn student abilities and question difficulty and generate reasonable prediction results based on this, indicating that both methods have a certain degree of reliability in terms of modeling ability and difficulty.

[0121] Secondly, the MHCD framework achieved a prediction match rate of 75.0%, significantly higher than the NCDM framework’s 60.0%. This indicates that the MHCD framework is superior in terms of student performance prediction accuracy and can more accurately capture changes in students’ knowledge status.

[0122] Finally, the MHCD's interpretability score was 85.0%, significantly higher than the NCDM's 70.0%. In this study, interpretability was defined as: when a student's ability exceeds the difficulty of the question, the student should be able to correctly answer the question. If the prediction results are consistent with this rule, the diagnostic results have good interpretability; otherwise, they lack interpretability.

[0123] Specifically, for questions 122 and 38, although both frameworks predicted that the student would answer correctly, the actual results showed that the student did not answer correctly. In this case, MHCD provides a more reasonable explanation. For example, in question 122, MHCD learned that the student ability was θ = 0.52 and the question difficulty was hdiff = 0.46. Because the student ability was greater than the question difficulty, the predicted probability of a correct answer was 0.67, which is consistent with the laws of educational cognition and demonstrates good interpretability. However, the NCDM prediction failed to meet this law. Conversely, for questions 104 and 86, NCDM still predicted that the student would answer correctly even though the student ability was lower than the question difficulty, which violates the laws of educational cognition. In contrast, MHCD accurately identified the student's lack of ability and avoided incorrect predictions, further validating its advantage in interpretability.

[0124] In summary, the MHCD framework not only outperforms the NCDM in predictive accuracy but also demonstrates stronger interpretability through diagnostic results that conform to educational cognitive laws. This demonstrates the significant value of multi-level interactive perception mechanisms in cognitive diagnostic tasks.

[0125] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

[0126] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A cognitive diagnostic method for multi-level interactive constraint perception, characterized by: The following steps are involved: S1. First, identify and construct the hierarchical relationship of explicit records in real educational scenarios; S2. Secondly, we deeply explore the feature interactions between layers and achieve more reasonable knowledge state modeling through iterative constraint-aware strategies; S3. Finally, the integration of multi-level interactive perception features and personalized features significantly improves the expressive power of individual features and hierarchical features.

2. The multi-level interactive constraint perception cognitive diagnosis method according to claim 1, characterized in that: Define multiple core sets, including: Student set S={s1,s2,...,s n }, contains n students; Practice set E={e1,e2,...,e m }Includes m different exercises; Knowledge point set KC = {kc1, kc2, ..., kc k } contains k different knowledge points; in addition, it also includes multiple level sets: Knowledge point hierarchy set CH = {ch1,ch2,...,ch h }, describing the hierarchical structure of knowledge points; Exercise level set EH = {eh1,eh2,...,eh h }, indicating the hierarchical structure of the exercise questions; Student level set SH={sh1,sh2,...,sh h }, indicating students’ ability performance at different levels. Here, h represents the number of levels involved in the framework, which is set according to the knowledge points, exercises, and students' cognitive development level, and the number of levels h is the same for all level sets; Finally, the answer response set R = {0, 1} is used to represent the student's answer situation, where 1 indicates a completely correct answer and 0 indicates an incorrect answer.

3. The multi-level interactive constraint perception cognitive diagnosis method according to claim 2, characterized in that: The student's test record T is defined as a five-tuple set (s, e, ch, eh, sh), where s∈S represents a student in the student set, e∈E represents an exercise in the exercise set, ch∈CH represents the level to which the knowledge concept belongs, eh∈EH represents the level to which the exercise belongs, and sh∈SH represents the student's ability level. To describe the relationship between exercises and knowledge points, a predetermined Q matrix is introduced, denoted as Q=Q ij m×k , where m is the number of exercises and k is the number of knowledge points, if and only if exercise e i and knowledge kc j When associated, Q ij =1; otherwise, Q ij =0.

4. The multi-level interactive constraint perception cognitive diagnosis method according to claim 3, characterized in that: The student's personalized knowledge proficiency vector θ sp is a vector of length k, which is used to characterize the student's proficiency in each knowledge point. Its calculation process is composed of the student's unique hot encoding vector x s Multiply it with a trainable knowledge proficiency matrix S and map it through the sigmoid activation function, which is specifically expressed as follows: i sp =σ(x s ×S), θ sp ∈(0,1) 1×k represents the student’s personalized proficiency in each knowledge point, x s ∈{0,1} 1×n is the student’s unique hot encoding, used to identify the specific student, S∈R n×k Is a trainable parameter matrix that captures the proficiency relationship between students and knowledge points. Function σ(·) represents the sigmoid activation function, which is used to normalize the output value. The one-hot encoding of the exercise x e Used to characterize the characteristics of the exercise, the difficulty of the knowledge point test h diff p ∈(0,1) 1×K and practice discrimination h disc ∈(0,1) further characterizes the difficulty and discrimination ability of the exercise, and its calculation formulas are: h diff p =σ(x e ×A), h disc e =σ(x e ×B), where x e ∈{0,1} 1×m Represents the one-hot encoding of exercise e, A∈R M×K and B∈R M×1 is a trainable parameter matrix. At the same time, the Q matrix is used to describe the relationship between exercises and knowledge points. The calculation formula of its correlation vector is: Q e =x e ×Q, Among them, Q e ∈{0,1} 1×k It reflects the degree of relevance between the exercise and the knowledge point; The performance of students in completing the exercises can be described by the following model: r = M(θ,η e ), where r represents the student's response (such as the correct answer or score), θ represents the student's knowledge level, and η e is a parameter related to practice, and M is based on a neurocognitive diagnostic model that divides the student’s overall knowledge proficiency θ into two components: θ=S(θ sh ,i sp ), h diff =D(h diff p ,h diff h ), where θ sh represents the capability characteristics constrained by the hierarchical structure, and θ sp Represents the individual ability characteristics of students, h diff h ∈(0,1)1×K represents the difficulty of the questions constrained by the hierarchical structure, while functions S and D combine the two to reflect the multi-dimensional influence of students' knowledge mastery level and question difficulty.

5. The multi-level interactive constraint perception cognitive diagnosis method according to claim 4, characterized in that: Through the relationship matrix R(e i ,k j ) and R(s i ,e j ) represent the correlation between exercises and knowledge points, and between students and exercises respectively. The specific expressions are: ε ek ={(e i ,k j ),(k j ,e i )|R(e i ,k j )=1}, ε se ={(s i ,e j ),(e j ,s i )|R(s i ,e j )=1}, Among them, R(e i ,k j )=1 and R(s i ,e j )=1 respectively means constructing bidirectional edges to represent the relationship between exercises and knowledge points, and between students and exercises, and expressing the hierarchical relationship of all nodes through a unified hierarchical identification calculation function: H(v i )←h(v i ),v i ∈{e i ,c j ,s i }, where h(v i ) is a unified level identification calculation function, which calculates the level identification of different types of nodes respectively, specifically: By integrating the hierarchical identification of nodes and the relationship between edges, we construct two graph structures representing different hierarchical relationships: the hierarchical graph G of knowledge points and exercises. ec and exercises with students' hierarchical graph G se , which is expressed as: G ec =(H ec ,E ec ),G se =(H se ,E se ), Among them, H ec and H se Indicates the hierarchical identity of the corresponding node, E ec and E se The edge set of the graph depicts the complex hierarchical interactive relationships between knowledge points and exercises, and between exercises and students.

6. The multi-level interactive constraint perception cognitive diagnosis method according to claim 5, characterized in that: For the embedding of knowledge concept level (t) k ,First, calculate the attention correlation score e between node k and its neighbor node d (t) kd , the score is calculated using the following formula: Where ξ(·) represents the LeakyReLU activation function, eh (t) d is the embedding representation at the topic level, w a is a learnable weight vector, W is the weight matrix of the linear transformation, and the symbol ∥ represents the vector concatenation operation; Next, the calculated e (t) kd After normalization, the neighbor attention weight α is obtained (t) kd , and its calculation formula is: Among them, G^ ec Indicates the current node k in the knowledge point and exercise level graph G ec The set of neighbors in ; Then, by the attention weight α (t) kd The information of its related topic-level neighbor nodes is weightedly aggregated to obtain the aggregated knowledge concept hierarchical embedding ch^ (t) k : Where φ(·) is the Sigmoid activation function; Finally, the knowledge concept hierarchy is updated through multiple rounds of iterative embedding using the following formula: Among them, the fusion weight β k (t) Calculated by the following formula: Among them, f k is the score calculated by linear transformation, w b and b are learnable parameters.

7. The multi-level interactive constraint perception cognitive diagnosis method according to claim 6, characterized in that: Through the topic-knowledge point hierarchy diagram G ec and topic-student level graph G se It is iteratively updated, and the specific calculation process is as follows: First, the topic level embedding is updated through the knowledge concept level, and the calculation formula is: Among them, ch k (t) Represents the hierarchical embedding of the current knowledge concept, G ec is the set of neighborhood knowledge points of topic d, α dk (t) is the attention weight, representing the contribution of knowledge point k to question d, and φ(·) represents the activation function; Then, the question embedding is updated at the student level using the following formula: Among them, sh z (t) represents the ability embedding of student z, α dz (t) is the attention weight, G se is the neighborhood student set of question d; Next, the above two update results are weighted and fused to obtain the final title embedding update: Among them, the weight parameter β1 (t) and β2 (t) Adaptive calculation, as follows: Among them, q d (t) Indicates the current feature representation of topic d, and are the updated embedding representations of knowledge points and students respectively, [·∥·] is the feature concatenation operation, w c are trainable weight parameters.

8. The multi-level interactive constraint perception cognitive diagnosis method according to claim 7, characterized in that: The embedding of student ability levels is achieved through message aggregation of question-level information and dynamic self-adjustment. The specific calculation is as follows: First, based on the student's neighborhood question set G se , Among them, α zd (t) is the attention weight, which measures the contribution of question d to the embedding update of student z, and φ(·) is the activation function; Based on message aggregation, the update formula for embedding the student ability level is: Among them, sh z (t) is the embedding representation of student z at the current level, sh~ (t) is the update amount obtained by aggregating the topics.

9. The multi-level interactive constraint perception cognitive diagnosis method according to claim 8, characterized in that: Calculate the difficulty level of the problem h diff h and the hierarchical characteristics of students’ knowledge mastery θ sh , which is calculated as follows: h diff h =σ(W1·(eh (t+1) d ∥ch (t+1) k +b1), θsh=σ(W2·(eh (t+1) d ∥sh (t+1) z +b2), Among them, W1 and W2 are trainable weight matrices, b1 and b2 are bias terms, and the activation function σ(·) is the Sigmoid function.

10. The multi-level interactive constraint perception cognitive diagnosis method according to claim 9, characterized in that: By designing an adaptively optimized personalized weighting strategy, hierarchical features are effectively combined with personalized features to generate a comprehensive state representation of students: θ=a p ·i sh +(1-a p )·θ sp , h diff =b p ·h diff h +(1-β p )·h diff p Among them, α p and β p is an adjustable personalized weight coefficient, which is used to control the weight ratio of students' knowledge mastery characteristics and question difficulty characteristics in the fusion process. The basic paradigm of the cognitive diagnosis model is used to predict students' response to exercises. The specific formula is: The cross entropy loss function is used to measure the difference between the predicted value y and the student’s actual answer label r. The specific form of the loss function is as follows: Among them, r i represents the student’s actual answer label, y i represents the predicted probability of the model.

Images