A hierarchical forgetting method for protecting student privacy in cognitive portraits

Abstract

The application discloses a layered forgetting method for protecting privacy of students in cognitive portraits, and belongs to the field of data layered forgetting, and comprises the following steps: obtaining a student interaction dataset containing a student set, a question set and a knowledge point set, and receiving a forgetting request raised by a user; dividing the interaction dataset into a forgetting set and a reservation set according to the forgetting request; respectively calculating individual importance of each parameter of a cognitive portrait model on the forgetting set and the reservation set; grouping the parameters according to a function level of the cognitive portrait model, and calculating level importance of each function layer on the forgetting set; weighting and fusing the individual importance and the level importance to obtain smoothed parameter importance; and judging whether the parameters are highly related to the forgetting set according to the smoothed parameter importance, and selectively attenuating and updating the parameters highly related to the forgetting set. The application guarantees the accuracy and controllability of the forgetting operation, and simultaneously maximally maintains the diagnostic performance of the model on the reservation set.

Description

Technical Field

[0001] This invention belongs to the field of data hierarchical forgetting technology, and in particular relates to a hierarchical forgetting method for protecting student privacy in cognitive profiles. Background Technology

[0002] With the rapid development of artificial intelligence (AI) technology, education is accelerating its personalization. Cognitive profiling, as a core technology, aims to accurately infer students' mastery of potential knowledge components (KCs) by analyzing their answer data. Unlike traditional exams that only assess correctness, cognitive profiling can deeply reveal students' cognitive structures, accurately pinpoint knowledge weaknesses, and thus support personalized learning path planning and resource recommendation, promoting the realization of large-scale personalized education. Early cognitive profiling often employed statistical methods based on probabilistic graphical models, such as DINA and IRT models, using expert-constructed knowledge point-question matrices, i.e., Q matrices, to characterize the relationship between questions and knowledge points. In recent years, deep learning methods have significantly improved the expressive and predictive capabilities of models. For example, the DKT model uses recurrent neural networks to model the knowledge mastery process from time-series answer sequences; NeuralCDM, KaNCD, and KSCD, among other neural network-based cognitive profiling models, demonstrate excellent performance by capturing the inherent relationships within the data. While deep learning has driven significant progress in the diagnostic accuracy of cognitive profiling models, student privacy protection has been widely overlooked. Model training relies on large amounts of sensitive data, easily leading to compliance and security risks. In practical applications, on the one hand, regulations such as the General Data Protection Regulation (GDPR) grant users the "right to be forgotten," requiring models to completely erase the impact of specific user data; on the other hand, models may also be vulnerable to Membership Inference Attacks (MIA), which could leak the source of training data and trade secrets. Therefore, how to achieve an efficient and reliable "forgetting" mechanism for cognitive profiling models has become a key technical challenge in ensuring the privacy and performance of intelligent education systems.

[0003] To address these challenges, machine unlearning has gradually emerged as a promising solution. Current technical approaches mainly fall into two categories: Exact Unlearning, such as SISA, removes data by dividing the training data into multiple slices and training sub-models separately. However, this method suffers from drawbacks such as high storage overhead and poor scalability. Approximate Unlearning, on the other hand, seeks a balance between efficiency and effectiveness by directly modifying model parameters. Among these, methods based on the Fisher Information Matrix (FIM) have gradually become the core of research. FIM provides an efficient way to quantify the importance of model parameters, enabling researchers to identify and precisely adjust parameters related to forgotten data, avoiding complete model retraining.

[0004] While FIM-based approximate forgetting methods have shown promising potential, their direct application to modern cognitive profiling models, especially those based on deep neural networks, presents unique and novel challenges. Specifically, current mainstream deep learning cognitive profiling models typically contain multiple heterogeneous modules, such as a student embedding layer representing student abilities; an exercise embedding layer representing question attributes; and an interactive prediction network simulating the answering process. In this complex architecture, the importance signal of a single parameter calculated by FIM is often obscured by noise. For example, a parameter identified by FIM as "very important for the student to be forgotten" might only be one dimension of its embedding vector, while this dimension is also crucial for the knowledge representation of many other students. Indiscriminately suppressing this parameter can easily impair the representational ability of other normal samples, leading to a severe degradation in the model's diagnostic performance in retaining data. Therefore, effectively utilizing parameter importance information in such fine-grained models becomes a core bottleneck problem in achieving efficient data forgetting. Summary of the Invention

[0005] This invention proposes a hierarchical forgetting method for protecting student privacy in cognitive profiles, in order to solve the problems existing in the prior art.

[0006] To achieve the above objectives, this invention provides a hierarchical forgetting method for protecting student privacy in cognitive profiles, comprising the following steps:

[0007] Obtain a student interaction dataset containing a set of students, a set of questions, and a set of knowledge points, and receive forgetting requests from users;

[0008] The interactive dataset is divided into a forget set and a retention set based on the forgetting request;

[0009] Based on the Fisher information matrix, the individual importance of each parameter of the cognitive profiling model on the forget set and the retention set is calculated respectively.

[0010] The parameters are grouped according to the functional hierarchy of the cognitive profile model, and the hierarchical importance of each functional layer on the forgetting set is calculated.

[0011] The individual importance and the hierarchical importance are weighted and fused to obtain the smoothed parameter importance;

[0012] Based on the importance of the smoothed parameters, it is determined whether the parameters are highly correlated with the forgotten set, and highly correlated parameters are selectively decayed and updated.

[0013] Optionally, obtaining the student interaction dataset includes:

[0014] Collect student sets, question sets, and knowledge point sets;

[0015] Establish a Q-matrix mapping relationship between questions and knowledge points;

[0016] Interactive data is generated based on students' answer records, forming the student interaction dataset.

[0017] Optionally, the expression for calculating individual importance is:

[0018] ;

[0019] In the formula, The loss function of the model, and the expectation operation, are represented in the dataset. On average, This represents the partial derivative operation during gradient backpropagation. This represents the set of all trainable parameters for the entire cognitive profile generation model. These are the model parameters corresponding to student i.

[0020] Optionally, the expression for calculating hierarchical importance is:

[0021] ;

[0022] Where θ represents the set of all trainable parameters of the entire cognitive profile generation model; This represents a functional layer in the model architecture, and it is a subset of the model parameter set θ. This represents the dataset to be forgotten, which is the part of the data whose influence needs to be removed from the model. Indicates the importance of individual parameters, used to quantify a single parameter. For model memory of the dataset to be forgotten The degree of contribution or the strength of the correlation. A higher The value indicates that the parameter It has a stronger correlation with the data to be forgotten, D_f; Indicates the importance of the hierarchy. This value is used to quantify the importance of the entire functional layer. For model memory of the dataset to be forgotten The overall contribution level.

[0023] Optionally, the weighted fusion includes:

[0024] By introducing a smoothing hyperparameter, the individual importance and hierarchical importance are weighted and summed to obtain the adjusted parameter importance;

[0025] The importance of the adjusted parameters satisfies the following:

[0026] ;

[0027] In the formula, To determine the importance of parameters on the smoothed forgetting set, To smooth hyperparameters, Indicates the importance of individual parameters, Indicates the importance of the hierarchy.

[0028] Optionally, the cognitive profile model includes:

[0029] The representation module is used to generate embedding vectors for students, questions, and knowledge points;

[0030] The cognitive interaction module is used to calculate the student proficiency vector and the question difficulty vector based on the embedded vector;

[0031] The prediction module is used to input the proficiency vector and difficulty vector into a shared feedforward network and output the probability of answering correctly.

[0032] Optionally, the selective decay update includes:

[0033] Set a forgetting intensity hyperparameter, and perform selective decay update based on the forgetting intensity hyperparameter;

[0034] For forgotten target parameters that are determined to need updating, perform decay updates according to their importance ratio;

[0035] The original values ​​of the non-forgotten target parameters remain unchanged.

[0036] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0037] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method.

[0038] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method.

[0039] Compared with the prior art, the present invention has the following advantages and technical effects:

[0040] This invention discloses a hierarchical forgetting method for protecting student privacy in cognitive profiles, belonging to the field of hierarchical data forgetting. The method includes: acquiring a student interaction dataset containing a set of students, a set of questions, and a set of knowledge points, and receiving forgetting requests from users; dividing the interaction dataset into a forgetting set and a retention set based on the forgetting requests; calculating the individual importance of each parameter of the cognitive profile model in the forgetting set and the retention set respectively; grouping the parameters according to the functional hierarchy of the cognitive profile model and calculating the hierarchical importance of each functional layer in the forgetting set; weighting and fusing the individual importance and hierarchical importance to obtain smoothed parameter importance; determining whether a parameter is highly correlated with the forgetting set based on the smoothed parameter importance, and selectively decaying and updating highly correlated parameters. This invention, through hierarchical importance smoothing and selective parameter decay, accurately weakens parameters related to students to be deleted without retraining, balancing privacy-preserving forgetting with diagnostic performance, and offering advantages such as low computational overhead and plug-and-play functionality. Attached Figure Description

[0041] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0042] Figure 1 This is a schematic diagram of a method according to an embodiment of the present invention. Detailed Implementation

[0043] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0044] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0045] Example 1

[0046] like Figure 1 As shown, this embodiment provides a hierarchical forgetting method for protecting student privacy in cognitive profiles, including the following steps:

[0047] Obtain a student interaction dataset containing a set of students, a set of questions, and a set of knowledge points, and receive forgetting requests from users;

[0048] The interactive dataset is divided into a forget set and a retention set based on the forgetting request;

[0049] Based on the Fisher information matrix, the individual importance of each parameter of the cognitive profiling model on the forget set and the retention set is calculated respectively.

[0050] The parameters are grouped according to the functional hierarchy of the cognitive profile model, and the hierarchical importance of each functional layer on the forgetting set is calculated.

[0051] The individual importance and the hierarchical importance are weighted and fused to obtain the smoothed parameter importance;

[0052] Based on the importance of the smoothed parameters, it is determined whether the parameters are highly correlated with the forgotten set, and highly correlated parameters are selectively decayed and updated.

[0053] Optionally, obtaining the student interaction dataset includes:

[0054] Collect student sets, question sets, and knowledge point sets;

[0055] Establish a Q-matrix mapping relationship between questions and knowledge points;

[0056] Interactive data is generated based on students' answer records, forming the student interaction dataset.

[0057] Optionally, the expression for calculating individual importance is:

[0058] ;

[0059] In the formula, The loss function of the model, and the expectation operation, are represented in the dataset. On average, This represents the partial derivative operation in backpropagation of a neural network. The set of all trainable parameters for the entire cognitive profile generation model. Let be the neural network parameters corresponding to student i.

[0060] Optionally, the expression for calculating hierarchical importance is:

[0061] ;

[0062] Where θ represents the set of all trainable parameters of the entire cognitive profile generation model; This represents a functional layer in the model architecture, and it is a subset of the model parameters θ. This represents the dataset to be forgotten, which is the part of the data whose influence needs to be removed from the model. Indicates the importance of an individual parameter. This value is used to quantify a single parameter. For model memory of the dataset to be forgotten The degree of contribution or correlation strength. In embodiments of the present invention, this value can be calculated using a model in the dataset. loss function with respect to parameters The expected value of the squared gradient is used to obtain it. A higher The value indicates that the parameter It has a stronger correlation with the data to be forgotten, D_f; Indicates the importance of the hierarchy. This value is used to quantify the importance of the entire functional layer. For model memory of the dataset to be forgotten The overall contribution level. In embodiments of the present invention, this value is calculated as a functional layer. Individual parameter importance of all parameters The arithmetic mean. It provides a more macroscopic and statistically stable measure of importance, which can effectively smooth out the bias in the estimation of the importance of individual parameters caused by training noise.

[0063] Optionally, the weighted fusion includes:

[0064] By introducing a smoothing hyperparameter, the individual importance and hierarchical importance are weighted and summed to obtain the adjusted parameter importance;

[0065] The importance of the adjusted parameters satisfies the following:

[0066] ;

[0067] In the formula, To determine the importance of parameters on the smoothed forgetting set, To smooth hyperparameters, Indicates the importance of individual parameters, Indicates the importance of the hierarchy.

[0068] Optionally, the cognitive profile model includes:

[0069] The representation module is used to generate embedding vectors for students, questions, and knowledge points;

[0070] The cognitive interaction module is used to calculate the student proficiency vector and the question difficulty vector based on the embedded vector;

[0071] The prediction module is used to input the proficiency vector and difficulty vector into a shared feedforward network and output the probability of answering correctly.

[0072] Optionally, the selective decay update includes:

[0073] Set a forgetting intensity hyperparameter, and perform selective decay update based on the forgetting intensity hyperparameter;

[0074] For forgotten target parameters that are determined to need updating, perform decay updates according to their importance ratio;

[0075] The original values ​​of the non-forgotten target parameters remain unchanged.

[0076] This embodiment also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0077] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method.

[0078] This embodiment also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method.

[0079] Example 2

[0080] This embodiment provides a hierarchical forgetting method for protecting student privacy in cognitive profiles, comprising two parts: a forgetting-friendly cognitive profile framework and a hierarchical importance-guided forgetting mechanism (HIF). The first part includes three modules: Module 1 is a representation module, used to learn student embeddings, question embeddings, and knowledge point embeddings; Module 2 is a cognitive interaction module, used to calculate student proficiency vectors and question difficulty vectors based on interaction functions; Module 3 is a prediction module, used to input the vectors into a shared feedforward network to output the probability of correct answers. The second part comprises three modules: Module 4 is the individual parameter importance calculation module, which calculates the individual importance of each parameter on both the forgotten and retained sets based on the diagonal approximation of the Fisher information matrix; Module 5 is the hierarchical importance smoothing module, which first calculates the hierarchical importance of each functional layer in the model (such as the student embedding layer, interaction layer, and prediction layer) to obtain the hierarchical importance index, and then weights and fuses the individual importance and hierarchical importance to remove noise from single-parameter estimation and obtain a smoothed parameter importance index; Module 6 is the selective parameter decay module, which uses a threshold hyperparameter based on the adjusted importance index to determine whether the parameter is highly correlated with the forgotten set, determines the target parameter for forgetting, and performs selective decay updates on the target parameter, so that the model gradually eliminates its dependence on the forgotten data, thereby ensuring the accuracy and controllability of the forgetting operation, while maintaining the diagnostic performance of the model on the retained set to the greatest extent.

[0081] The student interaction dataset contains a collection of students. Problem Set and a collection of knowledge points The relationship between the test questions and the knowledge points is marked by experts. matrix express, Indicates a correct answer to the question. Knowledge points to master Upon receiving a user's forget request, the interactive dataset is divided into a forget set and a retain set, and the user's forget request is processed according to the forgetting strategy.

[0082] The six modules are described in detail below:

[0083] Module 1: Represents the module:

[0084] The representation module is used to learn the basic embedding representations in the cognitive profiling model, including student embeddings, question embeddings, and knowledge point embeddings. Specifically:

[0085] (1) Student embedding: for each student Learning a vector representation This is used to capture their potential level of mastery across all knowledge points;

[0086] (2) Question embedding: for each question Learning a vector representation This is used to characterize the attributes of the question, such as difficulty and discrimination.

[0087] (3) Knowledge point embedding: for each knowledge point Learning a vector representation It is used to describe the semantic features and interrelationships of knowledge points.

[0088] This module, through an independent embedded vector learning mechanism, achieves decoupled modeling of educational entities (students, questions, and knowledge points) at different levels, providing basic representation inputs for subsequent cognitive interaction modules. Simultaneously, this modular design distinguishes individual student characteristics from general cognitive patterns, providing clear parameter positioning for the hierarchical importance-guided forgetting algorithm of this invention.

[0089] Module Two: Cognitive Interaction Module

[0090] The role of the cognitive interaction module is to dynamically calculate the student's proficiency and the difficulty of the questions based on the embedded students, questions and knowledge points generated by the representation module, thereby characterizing the interaction between learners and questions.

[0091] (1) Student proficiency vector Module Two utilizes student embedding and knowledge point embedding through interactive functions (such as...). Dynamically generate a vector of students' proficiency in each knowledge point to reflect their mastery level in different knowledge points;

[0092] (2) Question Difficulty Vector Module 2 utilizes question embedding and knowledge point embedding to generate a difficulty vector for the questions, which is used to characterize the intensity of the questions' testing of different knowledge points;

[0093] This module implements a mapping from individual embedding to the interaction level, enabling it to capture the correspondence between students' knowledge states and question attributes. In this invention, the decoupled design of this module separates students' individual characteristics from general cognitive patterns, thereby providing a clearer hierarchical foundation for subsequent parameter importance estimation and forgetting operations.

[0094] Module 3: Prediction Module

[0095] The prediction module integrates the student proficiency vector and the test question difficulty vector generated by the cognitive interaction module, and combines them with... The matrix is ​​used to calculate the cognitive gap, thereby outputting the probability that a student answers a specific question correctly.

[0096] First, calculate the students In the title Cognitive gap:

[0097] ;

[0098] Here, ⊙ represents element-wise multiplication. This represents the vector of knowledge points involved in the question. This gap characterizes the discrepancy between the student's knowledge mastery and the cognitive requirements of the question.

[0099] Next, the cognitive gap vector is input into a shared feed-forward network (FFN), which outputs the predicted probability through nonlinear mapping. That is, students Answer the question correctly The probability of.

[0100] This module plays a role in the overall architecture, which not only completes the transformation from cognitive representation to prediction results, but also limits the influence of individual student information to its embedding vector, thus providing a localizable target parameter for the hierarchical importance-guided forgetting proposed in this invention.

[0101] Module Four: Individual Parameter Importance Calculation Module

[0102] The parameter importance assessment module quantifies the sensitivity of each parameter in the model to the forgetting and retention sets, providing a basis for subsequent hierarchical smoothing and selective forgetting. Specifically:

[0103] Using the diagonal approximation method of the Fisher Information Matrix (FIM), for each parameter The importance of the data on different datasets is calculated. The formula is as follows:

[0104] ;

[0105] in, The loss function of the model, and the expectation operation, are represented in the dataset. On average, For the parameters of the neural network model, Let be the neural network parameters corresponding to student i.

[0106] This indicator reflects the sensitivity and contribution of individual parameters. (For the forgetting set) With Reserved Set Individual parameter importance indices are calculated to characterize the degree to which parameters depend on the data to be forgotten and the data to be retained. This module provides input for the subsequent hierarchical importance smoothing module through micro-level parameter importance analysis, thereby ensuring the targetedness and effectiveness of the forgetting operation.

[0107] Module 5: Hierarchical Importance Smoothing Module

[0108] The role of the hierarchical importance smoothing module is to combine the importance of individual parameters with the overall importance of the hierarchy, and use hierarchical priors to correct individual estimates, thereby obtaining a more robust parameter importance index.

[0109] First, based on the structure of the cognitive profile model, the parameters are grouped according to their respective functional layers (such as student embedding layer, interaction layer, and prediction layer) for different functional layers in the model.

[0110] Then, the importance of all parameters at each level is averaged on the forgetting set to obtain the level importance index for that level:

[0111] ;

[0112] Where θ represents the set of all trainable parameters of the entire cognitive profile generation model; This represents a functional layer in the model architecture, and it is a subset of the model parameters θ. This represents the dataset to be forgotten, which is the part of the data whose influence needs to be removed from the model. Indicates the importance of an individual parameter. This value is used to quantify a single parameter. For model memory of the dataset to be forgotten The degree of contribution or correlation strength. In embodiments of the present invention, this value can be calculated using a model in the dataset. loss function with respect to parameters The expected value of the squared gradient is used to obtain it. A higher The value indicates that the parameter It has a stronger correlation with the data to be forgotten, D_f; Indicates the importance of the hierarchy. This value is used to quantify the importance of the entire functional layer. For model memory of the dataset to be forgotten The overall contribution level. In embodiments of the present invention, this value is calculated as a functional layer. Individual parameter importance of all parameters The arithmetic mean. It provides a more macroscopic and statistically stable measure of importance, which can effectively smooth out the bias in the estimation of the importance of individual parameters caused by training noise.

[0113] Hierarchical importance has stronger statistical stability and can reflect the overall sensitivity of the entire functional module (such as the student proficiency representation module) to forgotten data.

[0114] Next, a smoothing hyperparameter is introduced. The importance of individual parameters is weighted and integrated with the hierarchical importance of their respective layers to obtain the adjusted parameter importance:

[0115] ;

[0116] in, To determine the importance of parameters on the smoothed forgetting set, For smoothing hyperparameters.

[0117] This fusion mechanism effectively weakens the abnormal peaks in the importance of individual parameters caused by training noise, while retaining the stable forgetting signal reflected by the entire functional layer, thus obtaining more robust parameter importance assessment results.

[0118] Module Six: Selective Parameter Attenuation Module

[0119] The selective parameter decay module identifies parameters that are highly correlated with the forgotten set based on the adjusted parameter importance index and performs targeted decay updates on them, thereby achieving efficient removal of forgotten data while maximizing the diagnostic performance of the retained data.

[0120] First, the importance index of smoothed parameters obtained from Module 5. Introducing threshold hyperparameters To determine whether a parameter is overly sensitive to the forgetting set, it is considered highly correlated with the forgetting set when the following conditions are met:

[0121] ;

[0122] in, Indicates the importance of individual parameters.

[0123] Then, the parameters identified as targets for forgetting are used as decay targets, utilizing the forgetting intensity hyperparameter. These parameters are updated proportionally to gradually decrease their values, thereby weakening the model's memory of forgotten data.

[0124] ;

[0125] in This is a hyperparameter for the intensity of forgetting. Parameters not identified as targets for forgetting remain unchanged to ensure that knowledge related to the retained data is preserved.

[0126] Finally, the model after selective decay updates will perform at a near-random level on the forgotten set, indicating that it has effectively "forgotten" the target data. At the same time, its diagnostic performance on the retained set remains basically stable, thus achieving a balance between privacy protection and model practicality.

[0127] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A hierarchical forgetting method for protecting student privacy in cognitive profiles, characterized in that, Includes the following steps: Obtain a student interaction dataset containing a set of students, a set of questions, and a set of knowledge points, and receive forgetting requests from users; The interactive dataset is divided into a forget set and a retention set based on the forgetting request; Based on the Fisher information matrix, the individual importance of each parameter of the cognitive profiling model on the forget set and the retention set is calculated respectively. The expression for calculating individual importance is: ； In the formula, The loss function of the model, and the expectation operation, are represented in the dataset. On average, This represents the partial derivative operation in the backpropagation computation. This represents the set of all trainable parameters for the entire cognitive profile generation model. This represents the model parameters corresponding to student i; The parameters are grouped according to the functional hierarchy of the cognitive profile model, and the hierarchical importance of each functional layer on the forgetting set is calculated. The cognitive profiling model includes: The representation module is used to generate embedding vectors for students, questions, and knowledge points; The cognitive interaction module is used to calculate the student proficiency vector and the question difficulty vector based on the embedded vector; The prediction module is used to input the proficiency vector and the difficulty vector into a shared feedforward network and output the probability of answering correctly. The expression for calculating hierarchical importance is: ； Where θ represents the set of all trainable parameters of the entire cognitive profile generation model; It represents a functional layer in the model architecture and is a subset of the model parameter set θ; Represents the dataset to be forgotten; Indicates the importance of individual parameters; Indicates the importance of hierarchy; The individual importance and the hierarchical importance are weighted and fused to obtain the smoothed parameter importance; Based on the importance of the smoothed parameters, it is determined whether the parameters are highly correlated with the forgotten set, and highly correlated parameters are selectively decayed and updated.

2. The method according to claim 1, characterized in that, Obtaining the student interaction dataset includes: Collect student sets, question sets, and knowledge point sets; Establish a Q-matrix mapping relationship between questions and knowledge points; Interactive data is generated based on students' answer records, forming the student interaction dataset.

3. The method according to claim 1, characterized in that, The weighted fusion includes: By introducing a smoothing hyperparameter, the individual importance and hierarchical importance are weighted and summed to obtain the adjusted parameter importance; The importance of the adjusted parameters satisfies the following: ； In the formula, To determine the importance of parameters on the smoothed forgetting set, To smooth hyperparameters, Indicates the importance of individual parameters, Indicates the importance of hierarchy. This represents the model parameters corresponding to student i. It represents a functional layer in the model architecture and is a subset of the model parameter set θ; This represents the dataset to be forgotten.

4. The method according to claim 1, characterized in that, The selective decay update includes: Set a forgetting intensity hyperparameter, and perform selective decay update based on the forgetting intensity hyperparameter; For forgotten target parameters that are determined to need updating, perform decay updates according to their importance ratio; The original values ​​of the non-forgotten target parameters are kept unchanged.

5. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-4.

7. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-4.

Images