A cold start cognitive diagnosis method based on a multi-granularity context enhanced large language model

Abstract

The application discloses a cold start cognitive diagnosis method based on a multi-granularity context enhanced large language model. The application comprises the following steps: firstly, a mixed skill graph is constructed by fusing the semantic association and the time sequence association between skills, historical skill records are extracted, and micro-skill historical information is classified and labeled; secondly, a perception state encoder is constructed, a cognitive state vector is obtained by fusing a question and environmental characteristics, and medium-scale answer performance information of a learning trajectory similar peer is retrieved; then, the objective correct rate of multi-dimensional educational attributes is extracted and mapped into a text difficulty label as macro reference correction information; finally, the multi-granularity context is converted into a structured prompt word to input a large language model, and the final prediction result and a diagnosis report are output through a four-step diagnosis reasoning chain. The application is suitable for an educational intelligent tutoring system, a new type of framework combining multi-granularity teaching context and large model reasoning is designed, and the prediction accuracy and interpretability of the model in a cold start data sparse environment are significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of educational data mining and natural language processing technology. Specifically, it relates to student cognitive modeling technology in intelligent tutoring systems, and more specifically, to a cold-start cognitive diagnosis method based on a multi-granularity context-enhanced large language model. Background Technology

[0002] Cognitive diagnosis and knowledge tracing, as fundamental and core tasks in educational data mining, aim to dynamically track the evolution of students' knowledge status by deeply modeling their historical interaction sequences, and thereby accurately predict their performance on specific questions or knowledge points in the future. In recent years, with the rapid development of educational informatization, various intelligent tutoring systems have been widely deployed and applied. To improve the accuracy of predictions, researchers have proposed many traditional Bayesian models and deep learning models, such as Bayesian Knowledge Trace (BKT), Deep Knowledge Trace (DKT), Attention-Aware Knowledge Trace (AKT), and Dynamic Key-Value Memory Network (DKVMN).

[0003] However, these existing models reveal significant limitations in practical applications: they heavily rely on long-term and intensive historical interaction data of students, inevitably leading to a severe "cold start" problem in real-world, high-risk educational scenarios. Specifically, when new students join the system or generate only a very small number of answer records, the extremely sparse or even missing observation samples prevent traditional models from establishing effective dependencies between these limited interaction sequences. This makes it difficult to accurately characterize students' cognitive trajectories and mastery levels, resulting in a sharp decline in predictive performance during the cold start phase.

[0004] In recent years, the rapid rise of large language models has demonstrated their superior natural language semantic understanding and complex logical reasoning capabilities, providing a novel research paradigm for addressing the cold-start challenge in recommender systems and educational assessment. Unlike traditional models that can only process ID sequences, large language models can deeply understand natural language descriptions containing rich pedagogical logic, such as specific skill names, question texts, and semantic relationships between knowledge points. This shift from understanding numerical identifiers to skill semantics can theoretically greatly compensate for the negative impact of insufficient historical interaction data.

[0005] However, directly applying general-purpose large language models to cognitive diagnostic tasks with highly specialized and fine-grained features still faces several severe technical challenges. First, traditional educational log data is mostly in the form of large-scale structured tables, and large language models inherently lack mechanisms for directly parsing and processing such structured data, making it difficult to effectively transform it into high-quality diagnostic information. Second, in the extremely sparse data environment of the cold start phase, large language models are prone to overfitting on the very few historical labels available. In the absence of group references and objective anchors, they are prone to generating logical "illusions" that deviate from objective teaching principles.

[0006] In summary, a new cognitive diagnostic framework needs to be designed to address the problems of data sparsity and large model illusion under cold start conditions. Summary of the Invention

[0007] In view of this, this invention provides a cold-start cognitive diagnosis method based on a multi-granularity augmented generation for knowledge tracing (MAG-KT). First, it constructs a semantic-temporal hybrid skill graph, integrating semantic and temporal associations between question skills. When the target predicted skill is missing or insufficient, it provides highly referential related skill performance, addressing the problem of historical sparsity of individual target skills at the micro level. Second, it proposes a perceptual state encoder, utilizing social learning theory to retrieve peers with similar cognitive states in the vector space, providing a reference for the answer performance of the meso-group. Third, it constructs global statistical data, transforming objective accuracy rates into specific text difficulty labels, reducing the reasoning illusion of the large model at the macro level. Fourth, through structured data transformation and a four-step diagnostic chain reasoning, it guides the large model to integrate multi-source evidence and output a highly interpretable diagnostic report.

[0008] The technical approach of this invention is as follows: First, a semantic-temporal hybrid skill graph is constructed to retrieve highly referential historical records of related skills; second, performance data of similar peers is obtained through a perceptual state encoder; next, objective accuracy and difficulty labels of multi-dimensional educational attributes are extracted from global statistical data; finally, the above multi-granular data is converted into structured natural language prompts and input into a large language model to output the final prediction through a four-step diagnostic reasoning chain.

[0009] The specific steps are as follows:

[0010] S1. Construct a semantic-temporal hybrid skill graph and perform associated skill retrieval to integrate semantic and temporal associations to solve the problem of sparsity in historical interaction data;

[0011] S2. Perform meso-level peer retrieval through a perception state encoder, extract real-time cognitive state and retrieve similar peers to introduce group reference information;

[0012] S3. Construct global statistical data to provide macro-level objective difficulty correction information, effectively alleviating the logical illusion of large models in zero-sample or sparse environments;

[0013] S4. Based on a large language model, perform a four-step thinking chain diagnostic reasoning process to enhance retrieval, and generate predictive results and interpretable diagnostic reports.

[0014] S11, Given a dataset of skills First, the semantic association between skills is calculated. A pre-trained model is then used to extract a dense vector representation of each skill name. Semantic association matrix between computational skills ;

[0015] S12. Calculate the frequency of direct temporal transitions between skills in students' historical answer records. Using a row maximum normalization strategy, divide the frequency of transitions between two specific skills by the maximum transition frequency originating from that source skill to obtain the logical matrix of the sequential transition relationships between skills. ;

[0016] S13. Weighted merge of the two types of matrices to generate an initial hybrid representation:

[0017] ,

[0018] Edges lacking transition records are penalized with a decay factor, and bidirectional dependency conflicts are resolved by retaining only edges with high dependencies, ultimately outputting a hybrid skill graph. ;

[0019] S14. When acquiring cold start history, perform graph retrieval based on the target skill, mark the associated skills that match the graph, and finally divide the skill records into three categories: target skill records, graph-related skill records, and other skill records, so as to provide skill history with different reference levels for subsequent large model inference.

[0020] S21, The model's input includes the first... Feature vector of the second answer question and set of environment variables Utilize the hidden state of the previous moment and current issues Calculate attention weights for different context types And then sum them by weight to obtain the fused context vector. ;

[0021] S22. The current time-to-date feature vector of the problem. With context vector Mapping to a unified feature space yields a comprehensive feature vector. Then, a binary answer result is introduced. The symbolic input is transformed into a directional signal that distinguishes between knowledge acquisition and cognitive punishment. Based on this symbolic input, the hidden state is updated using GRU. To record the learning trajectory and state transitions, and extract the hidden state after the Kth question. and attention vector The target state vector is constructed by concatenating and normalizing the vectors. ;

[0022] S23. Using a dual-index architecture, candidate peers are retrieved from the candidate pool. The average accuracy of the Top-K similar peers is obtained using the cosine similarity algorithm and used as a group reference performance for large model inference.

[0023] S31. Extract educational attribute entities from six dimensions: topic, template, skill, class, teacher, and question set, and set a minimum interaction threshold. The objective accuracy rate is calculated only for entities that meet the threshold conditions;

[0024] S32. Apply a preset threshold to map the objective pass rate to specific teaching difficulty labels, such as “Very Easy”, “Moderate”, and “Very Hard”. This label is injected into the prompt words as a calibration mechanism, thereby anchoring the model’s objective reasoning within the factual framework.

[0025] S41. Assemble the micro-granular skill records, meso-granular peer performance, and macro-granular global statistical information constructed above into text prompts containing individual history, peer references, and global attributes.

[0026] S42. The teacher model is used to process the training samples that integrate multi-granularity contexts, and a structured reasoning process is generated based on the four-step thinking chain to construct a distillation dataset with enhanced reasoning. Then, the distillation dataset is used to supervise and fine-tune the target small model to achieve high-performance cold-start cognitive diagnosis with strong interpretability.

[0027] S43. Guide the large model to execute a four-step reasoning process: First, analyze skill mastery; second, refer to peer performance information; third, analyze macro-level question difficulty indicators; fourth, synthesize evidence and make a final prediction. Through this chain-like reasoning process, the model outputs cold-start prediction results that combine educational logic and high interpretability. Attached Figure Description

[0028] Figure 1This is a schematic diagram of the overall framework of the cold-start cognitive diagnosis framework based on a multi-granularity context-enhanced large language model of the present invention. Detailed Implementation

[0029] To enable those skilled in the art to better understand the present invention, the specific embodiments of the present invention will be further described in detail below.

[0030] Please see Figure 1 , Figure 1 This is a schematic diagram of the overall framework of the present invention. The method systematically integrates three teaching contexts with different semantic granularities and injects them into a large language model using a retrieval-enhanced generation paradigm. Specifically, the present invention provides a multi-granularity context-enhanced cold-start cognitive diagnosis method based on a large language model, comprising the following steps:

[0031] S1. Constructing micro-granular context based on semantic-temporal hybrid skill graph: The micro-granularity revolves around the target student's personal historical answer records, integrates semantic and temporal associations between question skills to construct a hybrid skill graph, and classifies and filters historical answer records based on this graph to mark highly referential personal skill history;

[0032] S11. Use the pre-trained model BGE-M3 to extract the semantic vectors of each skill name in the dataset. And apply a preset nonlinear pruning threshold. If the cosine similarity is below the threshold, it is set to 0, thus filtering out links with low semantic relevance. The semantic relevance matrix between skills is then calculated. Its elements are represented as:

[0033] ;

[0034] in, and These represent the indices of the source skill and the target skill in the skill set, respectively. and These represent the semantic vectors of the source skill and the target skill, respectively.

[0035] S12. Calculate the direct temporal transition frequencies between skills in student answer records, and obtain the temporal correlation matrix after maximum value normalization. Its elements are represented as:

[0036] ;

[0037] in, This indicates that in the actual learning path of the system, from the first... The skill is directly transferred to the first skill. The statistical frequency of each skill and They represent the first The first skill and the first One skill, in the denominator An iterator index for all possible target skills in the dataset, used to find the maximum transition frequency from the source skill to any skill;

[0038] S13. Weighted fusion of semantic association matrix and temporal association matrix into a hybrid skill graph:

[0039] ,

[0040] Setting hyperparameters and For edges lacking skill transfer records, Multiply by the attenuation factor To suppress unsupported semantic links; if both bidirectional dependencies exceed the edge threshold Then only edges with higher dependencies are retained, thereby constructing a directed acyclic hybrid skill graph;

[0041] S14. Based on the hybrid skill graph, the cold start historical skill records are divided into three levels: target skill records, graph-related skill records, and other skill records, providing different levels of reference for subsequent large model predictions.

[0042] S2. Constructing a meso-granular context based on a perceptual state encoder: The meso-granular context revolves around the peer group's answer performance of the target student. By integrating the current question features and multi-source heterogeneous environment features, the current cognitive state vector of the student is dynamically tracked and extracted. Through a dual indexing architecture, peer performance data with similar cognitive states are retrieved in the vector space.

[0043] S21. Construct the current moment The problem's feature vector And a set of heterogeneous environmental feature vectors containing classes, teachers, question sets, and templates. ,in and Combined with the previous moment Hidden state of student answers Calculate the set using the following formula Attention weights for various environmental features :

[0044] ,

[0045] in, This represents a vector concatenation operation. For the transpose of the learnable projection matrix, For the learnable bias terms of the corresponding feature type, The unnormalized attention score is then calculated; subsequently, the attention weights are used. The environmental features are weighted and fused to obtain the context vector at the current time step. :

[0046] ;

[0047] S22. The current time is calculated using the following formula. The problem's feature vector With context vector Mapping to a unified feature space yields a comprehensive feature vector. :

[0048]

[0049] in, For activation function, and These are the learnable weight matrix and learnable bias term of the unified space mapping, respectively. This represents the vector concatenation operation; using a signed injection mechanism, students are monitored at time... binary answer response Convert to direction signal vector And compare it with the hidden state vector from the previous time step. The input is fed into a gated recurrent unit (GRU) for cognitive state evolution and update, resulting in the hidden state vector at the current time step. :

[0050] ,

[0051] Given a total length K of the historical interaction sequence of students during the cold start phase, what is the hidden state vector after extracting the Kth question? And the attention vector formed by concatenating the attention weights of each environmental feature in the Kth step. Then, the two are concatenated and normalized to form the target state vector used for retrieving companions. :

[0052] ;

[0053] S23, Based on the target state vector To construct a dual-index retrieval architecture for peer retrieval: first, based on the main index... Match students who have answered the current target question; if the number of candidate peers retrieved is insufficient, trigger the attribute combination fallback index. The candidate set is obtained by matching across a broader range of students using a composite key that includes skills, templates, and sequences; for the first candidate in the obtained candidate set... Extract the state vector of each candidate partner. The target student state vector is calculated using the following formula. With candidate partner state vector Cosine similarity between :

[0054] ,

[0055] in, This represents the vector transpose operation. The L2 norm of the vector is used to represent the vector. The vectors are sorted in descending order based on the calculated cosine similarity. The Top-N similar peers are extracted, and the average correct answer rate of these N students to the current question is calculated. This average correct answer rate is used as a reference for the peer group predicted by the large model.

[0056] S3. Constructing a macro-granular context based on global feature statistical information: The macro-granularity focuses on the global answer performance of the target student in the system, extracts the objective accuracy of multiple educational attributes, and maps them into discrete natural language difficulty labels as macro-calibration signals for the cold start answer prediction task.

[0057] S31. Obtain educational attribute data across six dimensions: questions, templates, skills, classes, teachers, and question sets. For each attribute category, count the historical interaction counts for each unique attribute identifier. Then, process data where the interaction count is below a preset threshold. The attribute identifiers were removed to obtain the retained attribute identifiers that met the statistical conditions for each dimension. ;

[0058] S32. Calculate the identifiers of each reserved attribute using the following formula. objective accuracy :

[0059]

[0060] in, Representative and attribute identifier Users who have interacted Indicates and The total number of users interacting, express right The binary classification response; subsequently, the accuracy rate is mapped to the corresponding text label according to a preset threshold range: when The time mapping is "Very Easy", when The time case is mapped to "Very Hard", and the rest are mapped to "Moderate". This is used to construct global statistics including accuracy and corresponding labels, which serve as objective correction information for the large model's predictions.

[0061] S4. Chain-based reasoning with multi-granularity context enhancement based on a large language model: Concatenate the multi-granularity context information obtained from S1 to S3 into structured natural language prompts; construct an enhanced dataset using the reasoning distillation paradigm to train the target small model, enabling it to have the reasoning ability of evidence integration; finally, guide the model to output prediction results and diagnostic reports through a four-step thinking chain.

[0062] S41. Assemble the micro-granular skill records, meso-granular peer performance, and macro-granular global statistical information constructed above into text prompts containing individual history, peer references, and global attributes.

[0063] S42. Construct a dataset based on the reasoning distillation paradigm and train the model: Use a higher-performance teacher model to batch process the training samples that integrate the multi-granularity context, and strictly guide the teacher model to generate a structured diagnostic reasoning process according to the preset four-step thinking chain, thereby constructing a reasoning-enhanced distillation dataset; Subsequently, use the distillation dataset to perform supervised fine-tuning of the target small model, so that it can learn the reasoning logic of the teacher model and acquire the complex logical reasoning ability from multi-source evidence to diagnostic conclusions, thereby enabling the target small model to achieve high-performance cold-start cognitive diagnosis with strong interpretability while maintaining low computational overhead;

[0064] S43. The large language model analyzes based on a four-step diagnostic reasoning chain, executing the following steps in sequence: Step 1, Skill Mastery Analysis: This step evaluates the individual's historical target skill records and highly relevant skill records matched to the graph, applying the recency principle to assign higher credibility to recent records; Step 2, Peer Performance Analysis: This step extracts group reference information based on the performance of peers with similar cognitive states on the current question; Step 3, Question Difficulty Analysis: This step obtains global reference information based on the six difficulty indicators related to the current question; Step 4, Synthesis and Prediction: This step synthesizes the above multi-source evidence, uses internal logical criteria to weigh the conflict between individual historical fluctuations and group / objective signals, and finally outputs a structured diagnostic report including the reasoning process, confidence score, and final correct / incorrect judgment.

[0065] To simulate the "cold start" environment of new user registration in a real system, this invention strictly divides the dataset into training and testing sets according to the ratio of 80 / 20 users. The system filters out users with fewer than 3 interactions and truncates the test trajectory to a maximum of 15 records to ensure that test students are completely invisible during the training phase, so as to rigorously evaluate the model's reasoning and generalization ability under the condition of sparse cold start data.

[0066] Based on the aforementioned data segmentation, to imbue the model with professional educational logic, this invention implements a reasoning distillation paradigm as a core step in the model training phase. The specific process is as follows: For the segmented training set, a pre-defined advanced teacher language model is invoked, requiring it to generate a step-by-step reasoning process. This process forces a strict alignment between complex reasoning trajectories and actual pedagogical logic (such as the historical transfer of skills and the reference value of peer performance), thereby constructing high-quality diagnostic reasoning data. Finally, using structured, assembled micro-level individual history, meso-level peer reference, and macro-level global label context as input features, and the reasoning process generated by the teacher model as output labels, a dedicated cognitive diagnostic fine-tuning dataset is constructed. This dataset is then used for supervised fine-tuning of the target small model, transforming it from a simple binary classification task performer into an expert system capable of outputting highly interpretable cognitive state assessments.

[0067] This invention conducted extensive experiments on two publicly available datasets. The data comes from the ASSISTments2009 and ASSISTments2012 datasets, which are widely used in the online education field. To verify the effectiveness of the cold-start cognitive diagnosis method based on a multi-granularity context-enhanced large language model, it was compared with several mainstream cognitive diagnosis baseline models such as BKT, DKT, and AKT. Evaluation metrics included Accuracy, F1 score, and AUC value, with AUC value being the core metric to address the imbalanced nature of educational data samples.

[0068] Table 1 shows the comparative experimental results, demonstrating that the method applied in this invention has significant advantages over traditional cognitive diagnostic baseline models in cold-start scenarios. Taking the supervised fine-tuned LLaMA3.1-8B-Instruct as an example, its AUC value reached 0.8400 on the ASSISTments2009 dataset, surpassing the best-performing traditional baseline model SAKT by 5.17%. On the ASSISTments2012 dataset, the method achieved even higher evaluation efficiency. Although the model was primarily fine-tuned on the ASSISTments2009 dataset, it still maintained extremely high prediction accuracy when directly applied to the unseen and more heterogeneous ASSISTments2012 dataset. For example, LLaMA-3.1-8B-Instruct achieved an AUC of 0.8184 on ASSISTments2012, representing a performance improvement of up to 7.86% compared to the best-performing traditional baseline AKT method on this dataset. Meanwhile, after integrating the multi-granularity context framework of this invention, the performance of the efficient 8B open-source model has extremely approached or even matched that of closed-source flagship models with a large number of parameters, such as Gemini-3-Flash-Preview and Qwen3.5-Plus, in some metrics. This proves that clear teaching logic guidance is more effective than simply increasing the parameter size.

[0069] Table 1 Comparative Experimental Results

Claims

1. A cold-start cognitive diagnosis method based on a multi-granularity context-enhanced large language model, characterized in that... Includes the following steps: S1. Constructing micro-granular context based on semantic-temporal hybrid skill graph: The micro-granularity revolves around the target student's personal historical answer records, integrates semantic and temporal associations between question skills to construct a hybrid skill graph, and classifies and filters historical answer records based on this graph to mark highly referential personal skill history; S2. Constructing a meso-granular context based on a perceptual state encoder: The meso-granular context revolves around the peer group's answer performance of the target student. By integrating the current question features and multi-source heterogeneous environment features, the current cognitive state vector of the student is dynamically tracked and extracted. Through a dual indexing architecture, peer performance data with similar cognitive states are retrieved in the vector space. S3. Constructing a macro-granular context based on global feature statistical information: The macro-granularity focuses on the global answer performance of the target student in the system, extracts the objective accuracy of multiple educational attributes, and maps them into discrete natural language difficulty labels as macro-calibration signals for the cold start answer prediction task. S4. Chain-based reasoning with multi-granularity context enhancement based on a large language model: Concatenate the multi-granularity context information obtained from S1 to S3 into structured natural language prompts; construct an enhanced dataset using the reasoning distillation paradigm to train the target small model, enabling it to have the reasoning ability of evidence integration; finally, guide the model to output prediction results and diagnostic reports through a four-step thinking chain.

2. The cold-start cognitive diagnosis method based on a multi-granularity context-enhanced large language model according to claim 1, characterized in that, Step S1 specifically includes: S11. Use the pre-trained model BGE-M3 to extract the semantic vectors of each skill name in the dataset. And apply a preset nonlinear pruning threshold. If the cosine similarity is below the threshold, it is set to 0, thus filtering out links with low semantic relevance. The semantic relevance matrix between skills is then calculated. Its elements are represented as: ； in, and These represent the indices of the source skill and the target skill in the skill set, respectively. and These represent the semantic vectors of the source skill and the target skill, respectively. S12. Calculate the direct temporal transition frequencies between skills in student answer records, and obtain the temporal correlation matrix after maximum value normalization. Its elements are represented as: ; in, This indicates that in the actual learning path of the system, from the first... The skill is directly transferred to the first skill. The statistical frequency of each skill and They represent the first The first skill and the first One skill, in the denominator An iterator index for all possible target skills in the dataset, used to find the maximum transition frequency from the source skill to any skill; S13. Weighted fusion of semantic association matrix and temporal association matrix into a hybrid skill graph: , Setting hyperparameters and For edges lacking skill transfer records, Multiply by the attenuation factor To suppress unsupported semantic links; if both bidirectional dependencies exceed the edge threshold Then only edges with higher dependencies are retained, thereby constructing a directed acyclic hybrid skill graph; S14. Based on the hybrid skill graph, the cold start historical skill records are divided into three levels: target skill records, graph-related skill records, and other skill records, providing different levels of reference for subsequent large model predictions.

3. The cold-start cognitive diagnosis method based on a multi-granularity context-enhanced large language model according to claim 1, characterized in that, Step S2 specifically includes: S21. Construct the current moment The problem's feature vector And a set of heterogeneous environmental feature vectors containing classes, teachers, question sets, and templates. ,in and Combined with the previous moment Hidden state of student answers Calculate the set using the following formula Attention weights for various environmental features : ， in, This represents a vector concatenation operation. For the transpose of the learnable projection matrix, For the learnable bias terms of the corresponding feature type, The unnormalized attention score is then calculated; subsequently, the attention weights are used. The environmental features are weighted and fused to obtain the context vector at the current time step. : ； S22. The current time is calculated using the following formula. The problem's feature vector With context vector Mapping to a unified feature space yields a comprehensive feature vector. : in, For activation function, and These are the learnable weight matrix and learnable bias term of the unified space mapping, respectively. This represents the vector concatenation operation; using a signed injection mechanism, students are monitored at time... binary answer response Convert to direction signal vector And compare it with the hidden state vector from the previous time step. The input is fed into a gated recurrent unit (GRU) for cognitive state evolution and update, resulting in the hidden state vector at the current time step. : ， Given a total length K of the historical interaction sequence of students during the cold start phase, what is the hidden state vector after extracting the Kth question? And the attention vector formed by concatenating the attention weights of each environmental feature in the Kth step. Then, the two are concatenated and normalized to form the target state vector used for retrieving companions. : ； S23, Based on the target state vector To construct a dual-index retrieval architecture for peer retrieval: first, based on the main index... Match students who have answered the current target question; if the number of candidate peers retrieved is insufficient, trigger the attribute combination fallback index. The candidate set is obtained by matching across a broader range of students using a composite key that includes skills, templates, and sequences; for the first candidate in the obtained candidate set... Extract the state vector of each candidate partner. The target student state vector is calculated using the following formula. With candidate partner state vector Cosine similarity between : ， in, This represents the vector transpose operation. The L2 norm of the vector is used to represent the L2 norm. The vectors are sorted in descending order based on the calculated cosine similarity. The Top-N similar peers are extracted, and the average correct answer rate of these N students to the current question is calculated. This average correct answer rate is used as a reference for the peer group predicted by the large model.

4. The cold-start cognitive diagnosis method based on a multi-granularity context-enhanced large language model according to claim 1, characterized in that, Step S3 specifically includes: S31. Obtain educational attribute data across six dimensions: questions, templates, skills, classes, teachers, and question sets. For each attribute category, count the historical interaction counts for each unique attribute identifier. Then, select classes with interaction counts below a preset threshold. The attribute identifiers were removed to obtain the retained attribute identifiers that met the statistical conditions for each dimension. ; S32. Calculate the identifiers of each reserved attribute using the following formula. objective accuracy : in, Representative and attribute identifier Users who have interacted, Indicates and The total number of users interacting, express right The binary classification response; subsequently, the accuracy rate is mapped to the corresponding text label according to a preset threshold range: when The time mapping is "Very Easy". The time case is mapped to "Very Hard", and the rest are mapped to "Moderate". This is used to construct global statistics including accuracy and corresponding labels, which serve as objective correction information for the large model's predictions.

5. The cold-start cognitive diagnosis method based on a multi-granularity context-enhanced large language model according to claim 1, characterized in that, Step S4 specifically includes: S41. Assemble the micro-granular skill records, meso-granular peer performance, and macro-granular global statistical information constructed above into text prompts containing individual history, peer references, and global attributes. S42. Construct a dataset based on the reasoning distillation paradigm and train the model: Use a higher-performance teacher model to batch process the training samples that integrate the multi-granularity context, and strictly guide the teacher model to generate a structured diagnostic reasoning process according to the preset four-step thinking chain, thereby constructing a reasoning-enhanced distillation dataset; Subsequently, use the distillation dataset to perform supervised fine-tuning of the target small model, so that it can learn the reasoning logic of the teacher model and acquire the complex logical reasoning ability from multi-source evidence to diagnostic conclusions, thereby enabling the target small model to achieve high-performance cold-start cognitive diagnosis with strong interpretability while maintaining low computational overhead; S43. The large language model analyzes based on a four-step diagnostic reasoning chain, executing the following steps in sequence: Step 1, Skill Mastery Analysis: This step evaluates the individual's historical target skill records and highly relevant skill records matched to the graph, applying the recency principle to assign higher credibility to recent records; Step 2, Peer Performance Analysis: This step extracts group reference information based on the performance of peers with similar cognitive states on the current question; Step 3, Question Difficulty Analysis: This step obtains global reference information based on the six difficulty indicators related to the current question; Step 4, Synthesis and Prediction: This step synthesizes the above multi-source evidence, uses internal logical criteria to weigh the conflict between individual historical fluctuations and group / objective signals, and finally outputs a structured diagnostic report including the reasoning process, confidence score, and final correct / incorrect judgment.

Images