Dual noise-aware knowledge tracing system, method, apparatus, and storage medium

Abstract

The application provides a double-noise-aware knowledge tracking system, method, device and storage medium, and relates to the technical field of knowledge tracking, and comprises an embedding representation module, which is used for constructing a question embedding sequence and an interaction embedding sequence based on a historical learning interaction sequence; a knowledge state modeling module, which is used for performing self-attention coding on the question embedding sequence and the interaction embedding sequence, performing comparative modeling on the question coding embedding sequence through a differential attention mechanism, obtaining a differential attention score, and determining a knowledge state representation based on the differential attention score and the interaction coding embedding sequence; a knowledge state correction module, which is used for denoising and correcting the knowledge state representation through a conditional diffusion model; and a prediction module, which is used for splicing and fusing the corrected knowledge state representation and the question coding embedding sequence, and predicting a correct probability of a student's answer to each question based on fused features. The scheme can suppress attention noise and improve the robustness of the prediction result.

Description

Technical Field

[0001] This application relates to the field of knowledge tracking technology, and more particularly to a knowledge tracking system, method, device, and storage medium with dual noise perception. Background Technology

[0002] Knowledge tracking (KT) models the evolution of students' knowledge status based on their historical answer records and predicts their future performance to support personalized learning. Accurately modeling students' knowledge status and reliably predicting their performance are the core objectives of knowledge tracking.

[0003] However, student interaction data in real-world educational settings is often influenced by various factors, such as emotions and uneven data distribution errors, introducing unavoidable noise. Existing KT models may ignore the impact of this noise on the learning process, limiting their ability to reliably reflect students' true knowledge state. Furthermore, while the attention mechanism widely used in KT models in recent years can achieve weighted modeling of historical interactions, it is susceptible to attention noise caused by irrelevant context, thus affecting the modeling of key historical information. Summary of the Invention

[0004] In view of this, this application proposes a dual noise-aware knowledge tracking system, method, apparatus, and storage medium.

[0005] In a first aspect, this application provides a dual noise-aware knowledge tracking system, comprising: an embedded representation module, a knowledge state modeling module, a knowledge state correction module, and a prediction module; The embedding representation module is used to construct a question embedding sequence that integrates trainable difficulty embedding vectors based on historical learning interaction sequences, and to construct an interaction embedding sequence that reflects the answer to the question and the correctness of the answer. The knowledge state modeling module is used to perform self-attention encoding on the question embedding sequence and the interaction embedding sequence through a multi-head self-attention mechanism to obtain a question encoding embedding sequence and an interaction encoding embedding sequence. The module then performs comparative modeling on the question encoding embedding sequence through a differential attention mechanism to obtain a differential attention score. Finally, the module determines the knowledge state representation based on the differential attention score and the interaction encoding embedding sequence. The knowledge state correction module is used to denoise and correct the knowledge state representation through a conditional diffusion model to obtain a corrected knowledge state representation. The prediction module is used to concatenate and fuse the corrected knowledge state representation with the question encoding embedding sequence to obtain joint features, and predict the probability of a student answering each question correctly based on the joint features.

[0006] In one embodiment, the embedded representation module includes: a question embedding unit and an interaction embedding unit; The question embedding unit is used to extract the questions and bound knowledge points corresponding to each time step in the historical learning interaction sequence, assign a trainable difficulty embedding vector to each question, and determine the knowledge point variant embedding and the knowledge point basic embedding based on the bound knowledge points. The question embedding sequence is constructed based on the difficulty embedding vector, the knowledge point variant embedding and the knowledge point basic embedding. The interaction embedding unit is used to form knowledge point-response pairs based on the questions, knowledge points and binary answer results at each time step in the same interaction sequence, construct matching interaction variant embeddings and joint variant embeddings containing questions and responses based on the knowledge point-response pairs, and construct the interaction embedding sequence based on the corresponding interaction variant embeddings, joint variant embeddings and difficulty embedding vectors.

[0007] In one embodiment, the knowledge state modeling module includes: a self-attention encoding submodule and a state evolution submodule; The self-attention encoding submodule is used to take the question embedding sequence and the interaction embedding sequence themselves as the query, key and value, respectively, and output the question encoding embedding sequence and the interaction encoding embedding sequence through projection transformation of each attention head, scaling dot product attention weight calculation and multi-head output splicing linear transformation, respectively. The state evolution submodule is used to divide the feature dimension of the question encoding embedding sequence into two parallel independent subspaces to construct a dual-path attention channel. In each channel, a decay term based on time interval is introduced to calculate attention weights with temporal constraints. Based on the dual-path attention weights and learnable differential weights, a differential attention score is determined. The differential attention score is applied to the interactive encoding embedding sequence and processed by normalization and residual scaling to obtain the knowledge state representation.

[0008] In one embodiment, the state evolution submodule includes: a differential modeling unit and a knowledge state generation unit; The differential modeling unit is used to divide the feature dimension of the question encoding embedding sequence into two parallel independent subspaces to construct a dual-path attention channel. In each channel, a scaling dot product operation is performed with the subspace feature of the current time step as the query and the subspace feature of the previous time step as the key. At the same time, a decay term based on the time interval is introduced to apply temporal weight constraints to the attention score. After normalization, two single-path attention weights with temporal constraints are obtained. The average distribution of the dual-path attention weights is calculated and the local sequence entropy of the average distribution is solved. The differential attention score is generated based on the differential weights, the local sequence entropy and the dual-path attention weights. The knowledge state generation unit is used to apply the differential attention score as a weighting weight to the interactive coding embedding sequence, perform root mean square normalization on the result, and use the structural prior stability coefficient to perform residual scaling adjustment on the normalization result, and output the knowledge state representation.

[0009] In one embodiment, the knowledge state correction module includes: a forward diffusion submodule and a reverse denoising submodule; The forward diffusion submodule is used to determine the noise intensity of each diffusion time step using a linear noise scheduling strategy, and gradually injects Gaussian noise into the knowledge state representation during the multi-step diffusion process to generate a series of noisy knowledge state representations with noise levels ranging from low to high. The reverse denoising submodule is used to generate diffusion time step embeddings based on the noisy knowledge state representations of each diffusion time step. It introduces student knowledge mastery level condition constraints and question embedding condition information, respectively calls the conditional denoising network to obtain the denoised knowledge state estimates under the corresponding condition constraints, and adaptively weights and fuses the two denoising results through a conditional gating mechanism to obtain the corrected knowledge state representation.

[0010] In one embodiment, the conditional denoising network includes: a conditional fusion modulation unit, an adaptive causal attention unit, a feedforward feature modulation unit, and a temporal convolutional output unit; The conditional fusion modulation unit is used to generate a sine and cosine form of diffusion time step embedding, and fuses the diffusion time condition and the knowledge tracking observation condition into a unified conditional representation through linear projection and nonlinear transformation. Based on the unified conditional representation, a multilayer perceptron generates three types of modulation parameters, namely scaling, translation and gating, which are respectively adapted to the attention sublayer and the feedforward sublayer. The adaptive causal attention unit is used to fuse the noisy knowledge state representation with the learnable position code to obtain the fused code feature. Based on the modulation parameters corresponding to the attention sub-layer, the position code fused feature is subjected to adaptive layer normalization modulation. A multi-head self-attention mechanism is used to model the internal dependency of the modulation result sequence. During the modeling process, the attention correction intensity is dynamically controlled through the gated residual mechanism, and the first feature sequence after attention enhancement is output. The feedforward feature modulation unit is used to perform adaptive layer normalization modulation on the first feature sequence based on the modulation parameters corresponding to the feedforward sub-layer, use a feedforward neural network to perform nonlinear feature transformation on the modulation result, and connect it with the residual of the first feature sequence to output the second feature sequence after global semantic transformation. The temporal convolution output unit is used to apply a one-dimensional causal convolution to the second feature sequence, and after being processed by layer normalization, nonlinear activation and random deactivation in sequence, it outputs the corrected knowledge state representation by connecting it with the residual of the second feature sequence.

[0011] In one embodiment, the prediction module includes: a feature fusion unit and a mapping unit; The feature fusion unit is used to perform a concatenation operation along the feature dimension on the corrected knowledge state representation and the question encoding embedding sequence to obtain the joint feature; The mapping unit is used to input the joint features into a fully connected neural network with two hidden layers and output the probability of the student answering the corresponding question correctly.

[0012] Secondly, this application also provides a dual noise-aware knowledge tracking method, including: Based on historical learning interaction sequences, construct question embedding sequences that fuse trainable difficulty embedding vectors, and construct interactive embedding sequences that reflect the answers to questions and the correctness of the answers. The question embedding sequence and the interaction embedding sequence are self-attention encoded by a multi-head self-attention mechanism to obtain a question encoding embedding sequence and an interaction encoding embedding sequence. The question encoding embedding sequence is compared and modeled by a differential attention mechanism to obtain a differential attention score. The knowledge state representation is determined based on the differential attention score and the interaction encoding embedding sequence. The knowledge state representation is denoised and corrected using a conditional diffusion model to obtain a corrected knowledge state representation. The corrected knowledge state representation is concatenated and fused with the question encoding embedding sequence to obtain joint features. Based on the joint features, the probability of a student answering each question correctly is predicted.

[0013] Thirdly, this application also provides an electronic device, including a processor and a memory; the memory has a computer program stored thereon, wherein the computer program, when executed by the processor, implements the dual noise perception knowledge tracking method as described in the second aspect.

[0014] Fourthly, this application also provides a non-transitory computer storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the dual noise perception knowledge tracking method as described in the second aspect.

[0015] The dual-noise-sensing knowledge tracking system proposed in this application has the following advantages over related technologies: 1. The dual-noise-aware knowledge tracking system of this application constructs a question embedding sequence that integrates trainable difficulty embedding vectors and an interaction embedding sequence of corresponding answer results through an embedding representation module, providing complete feature inputs covering question attributes and answer performance for temporal modeling; then, the sequence context-dependent encoding is completed through multi-head self-attention of the knowledge state modeling module, and further, the question encoding embedding is modeled by dual-path comparison through differential attention mechanism. The differential constraint of dual-path attention filters out the weight interference of irrelevant context, suppresses attention noise, and achieves accurate weighting of key historical learning interactions, resulting in an initial knowledge state representation that is more in line with the real learning evolution law, effectively improving the anti-interference ability of temporal attention modeling.

[0016] 2. By introducing a conditional diffusion model through the knowledge state correction module, multi-scale noise simulation and conditional reverse denoising correction are performed on the initial knowledge state. Combined with knowledge tracking-related constraint information, noise disturbances in the data are removed, restoring a more stable and reliable corrected knowledge state. Finally, the prediction module embeds and merges the corrected knowledge state with the question code, and completes the prediction of the correct answer probability based on joint features. From feature input and state modeling to noise correction, a full-link noise resistance system is formed, which significantly improves the model's ability to characterize the student's true knowledge state and the robustness of the prediction results. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A schematic diagram illustrating the principle of introducing attention noise into the existing KT model; Figure 2 This is a schematic diagram of the structure of a dual-noise-sensing knowledge tracking system in one embodiment of this application; Figure 3 This is a schematic diagram of the framework flow of a dual-noise-sensing knowledge tracking system in one embodiment of this application; Figure 4 This is a flowchart illustrating a knowledge tracking method with dual noise perception in one embodiment of this application.

[0019] Explanation of reference numerals in the attached figures: 10- Dual noise perception knowledge tracking system, 11- Embedded representation module, 12- Knowledge state modeling module, 13- Knowledge state correction module, 14- Prediction module. Detailed Implementation

[0020] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0021] As described in the background section, Knowledge Tracking (KT) models the evolution of a student's knowledge state based on their historical answer records and predicts their future performance to support personalized learning. Accurately modeling a student's knowledge state and reliably predicting their performance is the core objective of Knowledge Tracking.

[0022] However, student interaction data in real-world educational settings is often influenced by various factors, such as emotions and uneven data distribution errors, introducing unavoidable noise. Existing KT models may ignore the impact of this noise on the learning process, limiting their ability to reliably reflect students' true knowledge state. Furthermore, while the attention mechanism widely used in KT models in recent years can achieve weighted modeling of historical interactions, it is susceptible to attention noise caused by irrelevant context, thus affecting the modeling of key historical information.

[0023] Specifically, attention-based KT models essentially assign different weights to previous learning interactions to identify the most influential historical interactions on a student's current response. This aims to achieve high-precision prediction while maintaining model interpretability and effectively addressing complex challenges in educational scenarios such as data sparsity, forgetting behavior, anomalous interactions, and cognitive fluctuations. However, attention mechanisms tend to overemphasize irrelevant context when modeling sequences, potentially allocating only a small amount of attention to the correct position, thus introducing attention noise and bias that affects the accuracy of knowledge state modeling. Furthermore, most existing attention-based KT models rely on weighted aggregation of historical interactions to represent the current knowledge state. This modeling approach focuses on depicting the importance distribution of historical information but lacks explicit modeling of the relative changes in knowledge state relative to the current learning behavior, thus limiting the model's ability to express the dynamic evolution of knowledge state during the learning process. In addition, in real-world learning scenarios, students' responses are influenced by both subjective and objective factors, making it impossible for observed interaction data to fully and accurately reflect their true knowledge level. For example, subjective factors include guessing, mistakes, emotions, and stress, while objective factors include question quality and recording errors. Perfect response records do not exist; learning records are essentially observations of latent knowledge states under noise interference. Therefore, it is necessary to denoise both behavioral noise and attention noise observed in the KT task.

[0024] According to learning transfer theory, effective positive transfer of learning depends on the extraction of structurally relevant cognitive features. Attention weights should be allocated preferentially to historical interaction information that has a substantial cognitive connection to the current task, rather than cognitively irrelevant redundant information. For example... Figure 1 As shown in (a) in the prediction At that time, the model is Different attention weights were assigned to them. , , It received the highest top three weights. The knowledge points tested are and The model contains of Assigning higher weights is reasonable, but for those without... Related and Assigning high weights leads to unavoidable errors. This indicates that attention mechanisms are susceptible to interference from irrelevant context, thus introducing attentional noise. Figure 1 The radar chart in the image illustrates the changes in students' knowledge mastery; blue represents the model's knowledge mastery level at the previous time step. (The initial knowledge mastery level is shown in pink, representing the student's predicted knowledge mastery level, while orange represents the student's actual knowledge mastery level.) arrive At what time, the model predicts , , Their level of mastery has increased, and , The level of mastery has decreased; arrive , , reduce, , , Therefore, students' knowledge acquisition is not simply a continuous increase, but rather an evolutionary process in which reinforcement and inhibition alternate through practice and feedback. The model predicts during the process The students' knowledge mastery fluctuated dramatically: first decreasing rapidly and then increasing rapidly; however, their actual knowledge mastery remained almost stable. This highlights the problem of a certain error between the model's predicted knowledge state and the actual knowledge state, which we call observation noise. Specifically, such as... Figure 1As shown in (b), students' subjective factors, such as guessing and emotions, directly affect their answers. Unavoidable objective factors, such as recording errors and uneven distribution of answer data, can also lead to deviations between observed records and students' actual abilities. Based on the latent variable modeling concept in educational measurement, the observed answers are essentially random observations of students' latent knowledge state under noise interference. In other words, the answer records we obtain may inherently contain some bias, and the student's knowledge state predicted by the model is itself a latent variable affected by noise. Figure 1 (c) further illustrates each knowledge point. Explanation of the specific knowledge points.

[0025] Based on this, in some embodiments, such as Figure 2 and Figure 3 As shown, the knowledge tracking system 10 with dual noise perception provided in this application includes: an embedded representation module 11, a knowledge state modeling module 12, a knowledge state correction module 13, and a prediction module 14.

[0026] The embedding representation module 11 is used to construct a question embedding sequence that integrates trainable difficulty embedding vectors based on historical learning interaction sequences, and to construct an interaction embedding sequence that reflects the answer to the question and the correctness of the answer.

[0027] Knowledge tracking aims to infer a student's latent knowledge state over time based on their historical interactions with the learning system, and to predict their mastery of future learning content. In intelligent tutoring systems, there exists a set of students. Practice set and knowledge point collection Each exercise is linked to one or more knowledge points, and each student answers based on what they have learned. Therefore, by sorting each student's answer records chronologically, their interaction sequence during the learning process (i.e., the historical learning interaction sequence) can be represented as:

[0028] in The title item indicating the time step. This indicates the knowledge points related to the question. This represents a student's binary response to the question, with 0 indicating a correct answer and 1 indicating an incorrect answer. The goal of Knowledge Tracking (KT) is to predict a student's performance on a new question at the next time step, based on historical information and the next question, given a sequence of student interactions. The prediction results of the KT task can further support personalized tutoring and optimize learning paths.

[0029] Based on students' historical learning interaction sequences as input, on the one hand, for the questions and their associated knowledge points corresponding to each time step in the sequence, a trainable difficulty embedding vector is assigned to each question. This is combined with basic knowledge point embeddings and knowledge point variant embeddings to vectorize the question features, constructing a question embedding sequence covering the multidimensional attributes of the questions. On the other hand, based on the questions, knowledge points, and binary answer results at each time step in the same interaction sequence, knowledge point-response pairs are formed. These are further used to generate matching interaction variant embeddings and question-response joint variant embeddings. Combined with the difficulty embedding vectors, these are used to construct an interaction embedding sequence that integrates question information and correct / incorrect answer results. This fully encodes students' learning interaction behaviors and question attribute features, providing a standardized feature input foundation for subsequent knowledge state deduction.

[0030] The knowledge state modeling module 12 is used to perform self-attention encoding on the question embedding sequence and the interaction embedding sequence through a multi-head self-attention mechanism to obtain the question encoding embedding sequence and the interaction encoding embedding sequence. The question encoding embedding sequence is compared and modeled through a differential attention mechanism to obtain the differential attention score. The knowledge state representation is determined based on the differential attention score and the interaction encoding embedding sequence.

[0031] The knowledge state modeling module 12 performs self-attention encoding on the question embedding sequence and the interaction embedding sequence respectively through a multi-head self-attention mechanism. It models the contextual associations and temporal dependencies between features at each time step, enhances the global semantic expression capability of features, and outputs the corresponding question encoding embedding sequence and interaction encoding embedding sequence. On this basis, a differential attention mechanism is used to perform dual-path comparison modeling on the question encoding embedding sequence. It filters attention noise caused by irrelevant context through feature subspace splitting and differential constraints, accurately measures the influence weight of each historical learning node, generates a differential attention score, and finally uses this score as a weighted weight to apply to the interaction encoding embedding sequence. It aggregates the answer feedback information in historical interactions to obtain a knowledge state representation that can reflect the dynamic evolution of students' knowledge mastery level.

[0032] The knowledge state correction module 13 is used to denoise and correct the knowledge state representation through the conditional diffusion model to obtain the corrected knowledge state representation.

[0033] The diffusion model is a deep learning model based on a probabilistic generation process. Its core idea is to progressively add noise to the data, transforming it into Gaussian noise through a Markov chain, and then learning how to recover the original data from the noise. The entire process can be divided into two stages: forward diffusion and backward diffusion. Assume... This represents a sample from high-dimensional data, where the real data distribution is... The goal of the diffusion model is to build a generative model. , so that: .

[0034] In the forward propagation process, the diffusion model gradually adds noise to the real data, eventually transforming it into noise samples that approximate an isotropic Gaussian distribution. From the data... Departure, at each diffusion time step Add Gaussian noise to the data to generate a series Based on the superposition property of Gaussian distribution and the reparameterization technique, noise samples at any time step can be obtained without explicitly sampling each intermediate step. You can directly from generate:

[0035] in Indicates from arrive The cumulative signal scaling factor, The signal preservation coefficient, For time steps noise intensity, It is the identity matrix. When the total number of diffusion steps... hour, Converging to a standard Gaussian distribution .

[0036] The reverse process of a diffusion model starts from the endpoint of the forward diffusion process (a known pure noise distribution, usually a standard Gaussian distribution) and gradually removes noise to recover the original data. This process is the core of the diffusion model's generation of new data, involving learning how to recover useful information from noise. From Gaussian noise... Initially, a clear sample was eventually recovered. Specifically, the reverse process is the inverse operation of the forward process, achieved by learning a denoising function. From noise China Resumption .

[0037] in These are neural network parameters. and These represent the mean and variance, respectively. Furthermore, substituting these values ​​into the relevant formulas for the forward process yields the inverse solution formula:

[0038] In actual training, a simplified loss can be obtained using a fixed variance approach to minimize prediction noise. With real noise Mean square error: It is noise sampled from a standard Gaussian distribution, while It is a parameterized model that needs to be trained to predict noise.

[0039] The conditional diffusion model introduces conditional information to guide the generation process based on the standard diffusion model, making the output more consistent with expected constraints. By performing noise simulation and denoising correction under conditional constraints on the knowledge state representation through the conditional diffusion model, the bias effects of various disturbance factors are removed while retaining the core laws of knowledge state evolution. The initial knowledge state is corrected and smoothed, and finally, a corrected knowledge state representation that is more in line with students' actual knowledge mastery level and has stronger stability can be obtained.

[0040] The prediction module 14 is used to concatenate and fuse the corrected knowledge state representation with the question encoding embedding sequence to obtain joint features, and predict the probability of the student answering each question correctly based on the joint features.

[0041] The prediction module 14 deeply integrates the student's true knowledge mastery status with the attribute features of the corresponding questions through feature splicing and fusion, forming a joint feature that simultaneously covers the learner's cognitive level and question feature information. Based on this joint feature, the module then conducts reasoning on the student's answer performance and finally outputs the probability of the student's correct answer on each corresponding question, thus completing the quantitative assessment of the student's learning mastery and the prediction of answer performance by the knowledge tracking model.

[0042] The aforementioned dual-noise-aware knowledge tracking system 10 constructs a question embedding sequence that integrates trainable difficulty embedding vectors and an interaction embedding sequence of corresponding answer results through the embedding representation module 11, providing complete feature inputs covering question attributes and answer performance for temporal modeling. Then, the knowledge state modeling module 12 performs multi-head self-attention to complete the sequence context-dependent encoding. Furthermore, a differential attention mechanism is used to perform dual-path comparison modeling on the question encoding embedding, filtering out weight interference from irrelevant context through differential constraints of dual-path attention, suppressing attention noise, and achieving accurate weighting of key historical learning interactions. This results in an initial knowledge state representation that more closely reflects the actual learning evolution, effectively improving the anti-interference capability of temporal attention modeling. The knowledge state correction module 13 introduces a conditional diffusion model to perform multi-scale noise simulation and conditional reverse denoising correction on the initial knowledge state. Combined with knowledge tracking-related constraint information, noise disturbances in the data are removed, restoring a more stable and reliable corrected knowledge state. The final prediction module 14 integrates the corrected knowledge state with the question encoding, and completes the prediction of the correct answer probability based on joint features. It forms a full-link noise resistance system from feature input, state modeling to noise correction, which significantly improves the model's ability to characterize the student's real knowledge state and the robustness of the prediction results.

[0043] In some embodiments, the embedded representation module 11 includes a question embedding unit and an interaction embedding unit.

[0044] The question embedding unit is used to extract the questions and associated knowledge points corresponding to each time step in the historical learning interaction sequence, assign a trainable difficulty embedding vector to each question, and determine the knowledge point variant embedding and the knowledge point basic embedding based on the associated knowledge points. The question embedding sequence is constructed based on the difficulty embedding vector, the knowledge point variant embedding, and the knowledge point basic embedding.

[0045] The interactive embedding unit is used to form knowledge point-response pairs based on the questions, knowledge points and binary answer results at each time step in the same interactive sequence. Based on the knowledge point-response pairs, it constructs matching interactive variant embeddings and joint variant embeddings containing questions and responses. Based on the corresponding interactive variant embeddings, joint variant embeddings and difficulty embedding vectors, it constructs an interactive embedding sequence.

[0046] In application, to capture richer and more personalized input representations, we introduce the Rasch model structure, widely adopted in psychometrics and highly interpretable, into the model. We treat the difficulty of each item as a learnable latent parameter, enabling the model to characterize the inherent differences in difficulty among different items in a uniform way. Based on this, we assign a trainable difficulty embedding vector to each item and incorporate it as a conditional factor into the item embedding, allowing the model to explicitly model fine-grained differences between items in the representation space. Specifically, the item embedding is ultimately represented as:

[0047] in Indicates the topic and the knowledge points it contains At time step Embedded, Knowledge point variant embeddings represent the related knowledge points contained in the questions, used to capture the micro-differences between different questions that cover the same knowledge points. This represents a learnable scalar difficulty parameter related to the question, i.e., a difficulty embedding vector. This indicates the basic embedding of knowledge points contained in the question. and The operations represent element-wise multiplication and addition, respectively.

[0048] Building upon the question embedding, we further constructed a knowledge point-response pair interactive embedding using a similar approach, reflecting what questions students answered and whether their answers were correct or incorrect. The specific construction method is as follows:

[0049] in, Indicates the student's time step The comprehensive interactive representation includes the problem reaction . Representative issue Key concepts - interaction variant embeddings of response pairs, vectors This represents a joint variant embedding containing both items and responses, reflecting the joint effect of the variants.

[0050] In some embodiments, the knowledge state modeling module 12 includes a self-attention encoding submodule and a state evolution submodule.

[0051] The self-attention encoding submodule is used to take the question embedding sequence and the interaction embedding sequence themselves as the query, key and value, respectively. After projection transformation of each attention head, scaling dot product attention weight calculation and multi-head output splicing linear transformation, the question encoding embedding sequence and the interaction encoding embedding sequence are output respectively.

[0052] The state evolution submodule is used to divide the feature dimension of the question encoding embedding sequence into two parallel independent subspaces to construct a dual-path attention channel. In each channel, a time-interval-based decay term is introduced to calculate the attention weight with temporal constraints. The differential attention score is determined based on the dual-path attention weight and the learnable differential weight. The differential attention score is applied to the interactive encoding embedding sequence and processed by normalization and residual scaling to obtain the knowledge state representation.

[0053] To characterize the latent knowledge states within students' learning sequences, a self-attention encoding submodule is used to perform contextual modeling of question and response interactions. Question encoding captures the relatively stable cognitive demand structure inherent in the learning process, while response interaction encoding reflects dynamically changing learning behavioral feedback. Based on this, a multi-head self-attention mechanism is employed to embed the question sequence. and interactive embedded sequences Encode, and Each element is treated as a Query, Key, and Value, respectively. A self-attention network is implemented using dot product and vector multiplication.

[0054] in It is the dimension of each attention head, and , , It is the projection transformation function for each head.

[0055] In real-world learning scenarios, a student's ability to answer questions correctly depends not only on the questions themselves but also on the combined influence of their historical learning experiences. A student's knowledge state changes through continuous practice and feedback, and the key lies in the relative impact of different historical experiences on the current state update. Therefore, the state evolution submodule uses an improved differential attention mechanism to comparatively model historical interactions, characterizing their relative influence on the current knowledge state update.

[0056] The state evolution submodule first divides the feature dimension of the sequence into two parallel independent feature subspaces to construct a dual-path attention calculation channel. During the attention calculation process of each channel, a decay term based on the historical interaction time interval is introduced to apply temporal constraints, giving higher attention weights to historical interactions closer to the current learning node. This results in two single-path attention weights with temporal dependencies. Based on this, learnable differential weights are used to compare and model the dual-path attention weights. The differential constraints of the dual-path features filter attention noise caused by irrelevant context, highlighting the weight contribution of key historical interactions and generating differential attention scores. Subsequently, the differential attention scores are used as weights in the interaction encoding embedding sequence to complete the weighted aggregation of historical response feedback information. The aggregation results are then subjected to normalization and residual scaling to stabilize the numerical distribution and improve model robustness, ultimately yielding a knowledge state representation that accurately depicts the dynamic evolution of students' knowledge mastery levels.

[0057] In some embodiments, the state evolution submodule includes: a differential modeling unit and a knowledge state generation unit.

[0058] The differential modeling unit is used to divide the feature dimension of the question encoding embedding sequence into two parallel independent subspaces to construct a dual-path attention channel. In each channel, a scaling dot product operation is performed with the subspace feature of the current time step as the query and the subspace feature of the previous time step as the key. At the same time, a decay term based on the time interval is introduced to impose temporal weight constraints on the attention score. After normalization, two single-path attention weights with temporal constraints are obtained. The average distribution of the dual-path attention weights is calculated and the local sequence entropy of the average distribution is solved. The differential attention score is generated based on the differential weights, local sequence entropy and dual-path attention weights.

[0059] The knowledge state generation unit applies differential attention scores as weights to the interactive encoding embedding sequence, performs root mean square normalization on the result, uses structural prior stability coefficients to perform residual scaling on the normalization result, and outputs a knowledge state representation.

[0060] First, within each attention head, the differential modeling unit divides the feature dimension into two parallel subspaces, each corresponding to a set of independent attention computation channels, and models in parallel across multiple head dimensions: The superscripts (1) and (2) represent two parallel subspaces.

[0061] In each attention path, the model employs scaled dot product attention with temporal distance decay. For the For each attention path, we use the representation of the current step. As a query, the previous step represents... As the key, the attention score is calculated by multiplying the scaled dot product by a time-interval-based decay term, assigning more weight to those historical time steps that are more relevant to the current time step. The attention weights are calculated as follows:

[0062] in, , , It is the dimension of each head. This represents the relative distance between different time steps. This is a learnable decay parameter used to characterize memory decay over time. The average of the outputs of the two attention paths is used as the comprehensive attention distribution.

[0063] in, and These represent the attention weight matrices corresponding to the two differential attention paths. Then, the local sequence entropy is calculated based on this distribution:

[0064] in, To measure the dispersion of attention distribution, It is a very small positive number (e.g.) ), used to avoid numerical instability problems. This indicates the current query position, i.e., the time step. This represents the location index in the historical interaction sequence, i.e., all historical locations that the current query location may be interested in. This means that when the query position is At that time, for position The average attention weights are assigned. Furthermore, to achieve contrast modeling between the two attention paths, the model introduces a learnable differential weight:

[0065] in, The a priori stability coefficient of the structure is adaptively adjusted with network depth, progressively strengthening the stability constraint on the knowledge state. , , , It is a learnable latent vector. The differential attention score is thus constructed as:

[0066] in, These are learnable parameters. Differential attention score. This reflects the current model's selection and focus on key information from past learning trajectories. The knowledge state generation unit applies the adjusted differential attention score to the value vector, then uses RMSnorm and residual scaling to ensure numerical stability, and combines this with historical student interactions to obtain the time step. Knowledge status:

[0067] in, It includes students' historical answer records. Through a differential attention mechanism, the model increases the magnitude of knowledge updates when cognitive uncertainty is high, while suppressing over-updates when the knowledge state is stable.

[0068] In some embodiments, the knowledge state correction module 13 includes a forward diffusion submodule and a reverse denoising submodule.

[0069] The forward diffusion submodule is used to determine the noise intensity of each diffusion time step using a linear noise scheduling strategy. During the multi-step diffusion process, Gaussian noise is gradually injected into the knowledge state representation to generate a series of noisy knowledge state representations with varying noise levels from low to high.

[0070] The reverse denoising submodule is used to generate diffusion time step embeddings based on the noisy knowledge state representations of each diffusion time step. It introduces student knowledge mastery level condition constraints and question embedding condition information, respectively calls the conditional denoising network to obtain the denoised knowledge state estimates under the corresponding condition constraints, and adaptively weights and fuses the two denoising results through a conditional gating mechanism to obtain the corrected knowledge state representation.

[0071] The forward diffusion submodule receives the original knowledge state representation. That is, in the student's answer sequence The knowledge state at step 0 of the diffusion process at time step n. To simulate a series of interfering factors during the learning process, Gaussian noise is gradually injected into the forward diffusion process across multiple diffusion time steps to construct a series of intermediate states ranging from low noise to high noise. This also provides multi-scale training samples for subsequent reverse denoising learning. Specifically, the knowledge state at step 0 of the forward diffusion process at time step n. The noisy knowledge state of each diffusion step can be defined as:

[0072] in It is noise sampled from a standard normal distribution with a mean of 0 and a variance of 1, used to simulate cognitive perturbations. , This represents the maximum number of steps in the diffusion process. As the number of diffusion steps increases, the weight of the original knowledge state in the current representation gradually decreases, while the impact of noise perturbations continuously increases. Furthermore, we use a linear scheduling strategy to obtain the noise intensity. This ensures the smoothness and stability of the diffusion process and avoids abrupt changes in the knowledge state.

[0073] The reverse process of the diffusion model can be viewed as a conditional representation recovery process. Specifically, we train a conditional denoising network. Directly estimate the knowledge state after denoising In the first One diffusion step, the reverse denoising submodule receives the noisy knowledge state. diffusion step embedding and condition information As input, the output is the denoised knowledge state estimate under the corresponding conditions:

[0074] Choosing appropriate conditional information is crucial for effectively reconstructing the student's knowledge state representation. Therefore, two key conditional information sets from KT were primarily used: the student's knowledge mastery level and the question embedding. The denoised knowledge state was obtained through two conditional paths:

[0075] in It represents the degree of students' knowledge mastery, and describes their potential abilities in different knowledge points, reflecting their knowledge accumulation and individual learning status. The question embedding contains the semantics, difficulty, and corresponding knowledge points of the question. By combining these two types of conditions, the reverse denoising process can perform conditional corrections based on individual circumstances and question characteristics, achieving a refined model of student performance. To integrate these two types of information, we designed a conditional gating mechanism that adaptively weights the denoising results of the two paths. Specifically, we first construct a gating vector. This vector is calculated by concatenating two denoised knowledge states and passing them through a linear transformation layer with a sigmoid activation function. To further control the information retention ratio of the two paths, we use a gating vector. and its calculation The denoised knowledge states obtained under different conditions are multiplied separately to obtain the fused and denoised knowledge states. The specific calculation process is expressed by the following formula:

[0076] in for Activation function. It is a gated weight matrix used to perform linear projection on two input vectors. This represents element-wise multiplication. To prevent the gating weights from becoming too extreme, a slight regularization term is introduced into the gating loss function:

[0077] in For batch size, For sequence length, This indicates the gating weight. This represents the regularization loss generated by the gating network, used to constrain the distribution of the gating weights. During training, the model optimizes the deviation between the prediction and the target denoised state using weighted mean square error, and adjusts the loss weights at each step using the signal-to-noise ratio (SNR).

[0078] in, It is the loss function in the diffusion process, which uses the signal-to-noise ratio term to weight the errors of different diffusion steps. This indicates the added noise. For diffusion step The signal-to-noise ratio is usually determined by the noise scheduling function. This represents the denoised knowledge state predicted by the conditional denoising network. This represents the initial knowledge state.

[0079] In some embodiments, the conditional denoising network includes: a conditional fusion modulation unit, an adaptive causal attention unit, a feedforward feature modulation unit, and a temporal convolutional output unit.

[0080] The conditional fusion modulation unit is used to generate the diffusion time step embedding in sine and cosine form. The diffusion time condition and the knowledge tracking observation condition are fused into a unified conditional representation through linear projection and nonlinear transformation. Based on the unified conditional representation, three types of modulation parameters, namely scaling, translation and gating, are generated by the multilayer perceptron to adapt to the attention sublayer and the feedforward sublayer, respectively.

[0081] The adaptive causal attention unit is used to fuse noisy knowledge state representation with learnable position encoding to obtain encoded fusion features. Based on the modulation parameters corresponding to the attention sub-layer, adaptive layer normalization modulation is performed on the position encoding fusion features. A multi-head self-attention mechanism is used to model the internal dependencies of the modulation result sequence. During the modeling process, the attention correction intensity is dynamically controlled through a gated residual mechanism, and the first feature sequence after attention enhancement is output.

[0082] The feedforward feature modulation unit is used to perform adaptive layer normalization modulation on the first feature sequence based on the modulation parameters corresponding to the feedforward sub-layer. The feedforward neural network is used to perform nonlinear feature transformation on the modulation result, and the result is connected with the residual of the first feature sequence to output the second feature sequence after global semantic transformation.

[0083] The temporal convolution output unit is used to apply a one-dimensional causal convolution to the second feature sequence, and after passing through layer normalization, nonlinear activation and random deactivation in sequence, it outputs the corrected knowledge state representation by connecting it with the residual of the second feature sequence.

[0084] To ensure the rationality of the diffusion process, we explicitly introduce two types of complementary conditional information during denoising, constraining the reverse update from different levels: diffusion time conditions and knowledge tracking observation conditions. The former characterizes the noise intensity and denoising stage of the current state, while the latter provides task-related external information to guide the denoising direction. In the diffusion model, different diffusion steps correspond to different signal-to-noise ratios. States at higher diffusion steps typically contain stronger noise and require more significant structural corrections, while those at lower diffusion steps focus more on fine-tuning details. To enable the denoising network to perceive this difference, the conditional fusion modulation unit uses sine and cosine time embeddings to map discrete diffusion time steps into continuous time embedding vectors:

[0085] in The exponentially decaying frequency term is constructed as follows: in, This indicates the dimension of the diffusion time embedding. This indicates the number of frequency components used to construct the sinusoidal time-coded array. Hyperparameter This is a preset maximum time scale used to control the range of variation corresponding to different frequency dimensions. Through the above construction, different frequency components can cover different time scales, enabling the temporal embedding to have periodic representation capabilities, thereby helping the denoising network distinguish approximate diffusion steps. Then, we align the diffusion time step conditions and knowledge tracking conditions to jointly guide the diffusion process:

[0086] The above formula indicates that the diffusion time condition and the knowledge tracking observation condition are combined through linear projection and nonlinear transformation to form a unified conditional representation. Wherein... It is a multilayer perceptron. It is a weight matrix. It is a bias term. It is conditional information. The embedding dimension. Next, to inject temporal and knowledge tracking information into the network's internal features, a lightweight multilayer perceptron generates six sets of modulation parameters for translation, scaling, and gating control of the attention and feedforward sublayers, respectively. The corresponding formulas are as follows:

[0087] Each parameter has one dimension. , It is the sigmoid linear unit smooth activation function. Used for scaling and translation modulation of multi-head self-attention sublayers. This is used for scaling and translation modulation of the feedforward sublayer. Based on the above modulation parameters, the adaptive causal attention unit employs an adaptive layer normalization modulation operation to dynamically inject conditional information into the feature representations of each layer of the network. The modulation operation of the attention layer can be formally represented as:

[0088] in, This represents the trainable linear weight matrix corresponding to the attention sublayer of the denoising network. This represents the layer normalization operation LayerNorm, which does not contain learnable affine parameters. It uses learnable positional encoding to preserve sequence structure information. Given the modulated feature representation, the adaptive causal attention unit models the intra-sequence dependencies using a multi-head causal self-attention structure. The temporal unit vector is computed in parallel through multi-head attention and masked to block future information. The outputs of each head are then concatenated and subjected to a linear transformation.

[0089] in, It is an upper triangular mask matrix that guarantees causality. Represents the number of attention heads, It is the dimension of each head. , These are trainable parameters. It's a softmax operation. This indicates that the results from multiple head outputs are concatenated. To further control the influence of attention corrections at different diffusion stages on the final denoising result, the adaptive causal attention unit introduces a gated residual mechanism to obtain the first feature sequence:

[0090] Among them, the gating coefficient Modulation parameters adaptively generated from conditional information. This is the knowledge state obtained after adding noise in the forward stage. After attention modeling, the feedforward feature modulation unit further performs position-by-position feature transformation through the feedforward network. Then, using the previously obtained modulation parameters, a linear transformation is performed on the feedforward sublayer to further obtain the knowledge state vector for each time unit, i.e., the second feature sequence:

[0091] These are the modulation parameters obtained earlier, used for scaling and translation modulation of the feedforward sublayer. Layer Normalization (LayerNorm) It is a learnable location encoding. Then, residual connections and layer normalization are performed again to output the knowledge state vector of each time unit.

[0092] A student's knowledge state may change in consecutive responses. Therefore, to explicitly model the evolution pattern of knowledge state over a short period, the temporal convolutional output unit uses one-dimensional causal convolution to locally model the sequence dimension to extract information, and employs residual connections to obtain the final denoised knowledge state vector, i.e., the corrected knowledge state representation:

[0093] in This represents a one-dimensional causal convolution applied along the time dimension, with a kernel size of 3. We use left-hand zero-padding for the convolution operation to ensure causality. Presentation layer normalization operation, GELU represents the element-wise nonlinear activation function. It is a random deactivation operator.

[0094] In some embodiments, the prediction module 14 includes a feature fusion unit and a mapping unit.

[0095] The feature fusion unit is used to perform a concatenation operation on the corrected knowledge state representation and the question encoding embedding sequence along the feature dimension to obtain joint features.

[0096] The mapping unit is used to input joint features into a fully connected neural network with two hidden layers and output the probability of a student answering the corresponding question correctly.

[0097] To achieve accurate prediction of students' answer performance, a feature fusion unit is used to process the final denoised knowledge state. And student exercises embedded sequence The concatenated features are fused to obtain joint features, which are then sequentially input into a fully connected neural network with two hidden layers by the mapping unit. This network employs the following computational structure: the first and second hidden layers undergo nonlinear mapping processing via the ReLU activation function, and the output is then compressed to the (0,1) interval using the sigmoid function to obtain the student's correct answer probability for each question. The specific calculation process is represented by the following formula:

[0098] To measure the difference between the predicted probability and the actual answer, we use the binary cross-entropy loss function to evaluate the predicted probability. With real labels Supervised learning is performed, and missing answer locations are masked. Ultimately, the model's total training loss is a weighted combination of prediction loss, diffusion loss, and gating loss.

[0099] ReLU is the activation function. These are learnable parameters. It is the denoised knowledge state after gating fusion. It is a question embedding after self-attention encoding. It is the total loss of the model. and These are the losses generated by the KT prediction task, the diffusion process, and the gating operation, respectively. As the main loss It is a hyperparameter used to balance the weight of each part of the loss in the total loss.

[0100] In some embodiments, please refer to Figure 4 This application provides a dual noise perception knowledge tracking method, including the following steps S401 to S404.

[0101] S401: Construct a question embedding sequence that integrates trainable difficulty embedding vectors based on historical learning interaction sequences, and construct an interactive embedding sequence that reflects the answer to the question and the correctness of the answer.

[0102] S402: Self-attention encoding of the question embedding sequence and the interaction embedding sequence is performed through a multi-head self-attention mechanism to obtain the question encoding embedding sequence and the interaction encoding embedding sequence. The question encoding embedding sequence is compared and modeled through a differential attention mechanism to obtain the differential attention score. The knowledge state representation is determined based on the differential attention score and the interaction encoding embedding sequence.

[0103] S403: Denoise and correct the knowledge state representation by using a conditional diffusion model to obtain a corrected knowledge state representation.

[0104] S404: The corrected knowledge state representation is concatenated and fused with the question encoding embedding sequence to obtain joint features. Based on the joint features, the probability of a student answering each question correctly is predicted.

[0105] It should be noted that the dual noise perception knowledge tracking system and the dual noise perception knowledge tracking method provided in this application are based on the same inventive concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned dual noise perception knowledge tracking system, and the repeated parts will not be described again.

[0106] In some embodiments, an electronic device provided in this application includes a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the aforementioned dual noise perception knowledge tracking method.

[0107] Specifically, the processor may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor may also include onboard memory for caching purposes. The processor may be a single processing unit or multiple processing units for performing different actions of the method flow according to embodiments of this application.

[0108] Memory can be any medium capable of containing, storing, transmitting, propagating, or transmitting instructions. For example, memory can include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, instruments, or propagation media. Specific examples of memory include: magnetic storage devices such as magnetic tape or hard disk drives (HDDs); optical storage devices such as optical discs (CD-ROMs); and also random access memory (RAM) or flash memory; and / or wired / wireless communication links.

[0109] This application also provides a non-transitory computer storage medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned dual-noise-aware knowledge tracking method. This computer-readable medium may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into that device / apparatus / system. The aforementioned computer-readable medium carries one or more programs, which, when executed, implement the method as described in the embodiments of this application.

[0110] According to embodiments of this application, a computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wired, optical fiber, radio frequency signals, etc., or any suitable combination thereof.

[0111] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0112] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application, even if such combinations or combinations are not explicitly described in this application. All such combinations and / or combinations fall within the scope of this application. Therefore, the scope of this application should not be limited to the above embodiments. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A knowledge tracking system with dual noise perception, characterized in that, include: The module includes an embedded representation module, a knowledge state modeling module, a knowledge state correction module, and a prediction module. The embedding representation module is used to construct a question embedding sequence that integrates trainable difficulty embedding vectors based on historical learning interaction sequences, and to construct an interaction embedding sequence that reflects the answer to the question and the correctness of the answer. The knowledge state modeling module is used to perform self-attention encoding on the question embedding sequence and the interaction embedding sequence through a multi-head self-attention mechanism to obtain a question encoding embedding sequence and an interaction encoding embedding sequence. The module then performs comparative modeling on the question encoding embedding sequence through a differential attention mechanism to obtain a differential attention score. Finally, the module determines the knowledge state representation based on the differential attention score and the interaction encoding embedding sequence. The knowledge state correction module is used to denoise and correct the knowledge state representation through a conditional diffusion model to obtain a corrected knowledge state representation. The prediction module is used to concatenate and fuse the corrected knowledge state representation with the question encoding embedding sequence to obtain joint features, and predict the probability of a student answering each question correctly based on the joint features.

2. The knowledge tracking system with dual noise perception as described in claim 1, characterized in that, The embedded representation module includes: a question embedding unit and an interaction embedding unit; The question embedding unit is used to extract the questions and bound knowledge points corresponding to each time step in the historical learning interaction sequence, assign a trainable difficulty embedding vector to each question, and determine the knowledge point variant embedding and the knowledge point basic embedding based on the bound knowledge points. The question embedding sequence is constructed based on the difficulty embedding vector, the knowledge point variant embedding and the knowledge point basic embedding. The interaction embedding unit is used to form knowledge point-response pairs based on the questions, knowledge points and binary answer results at each time step in the same interaction sequence, construct matching interaction variant embeddings and joint variant embeddings containing questions and responses based on the knowledge point-response pairs, and construct the interaction embedding sequence based on the corresponding interaction variant embeddings, joint variant embeddings and difficulty embedding vectors.

3. The knowledge tracking system with dual noise perception as described in claim 1, characterized in that, The knowledge state modeling module includes: a self-attention encoding submodule and a state evolution submodule; The self-attention encoding submodule is used to take the question embedding sequence and the interaction embedding sequence themselves as the query, key and value, respectively, and output the question encoding embedding sequence and the interaction encoding embedding sequence through projection transformation of each attention head, scaling dot product attention weight calculation and multi-head output splicing linear transformation, respectively. The state evolution submodule is used to divide the feature dimension of the question encoding embedding sequence into two parallel independent subspaces to construct a dual-path attention channel. In each channel, a decay term based on time interval is introduced to calculate attention weights with temporal constraints. Based on the dual-path attention weights and learnable differential weights, a differential attention score is determined. The differential attention score is applied to the interactive encoding embedding sequence and processed by normalization and residual scaling to obtain the knowledge state representation.

4. The knowledge tracking system with dual noise perception as described in claim 3, characterized in that, The state evolution submodule includes: a differential modeling unit and a knowledge state generation unit; The differential modeling unit is used to divide the feature dimension of the question encoding embedding sequence into two parallel independent subspaces to construct a dual-path attention channel. In each channel, a scaling dot product operation is performed with the subspace feature of the current time step as the query and the subspace feature of the previous time step as the key. At the same time, a decay term based on the time interval is introduced to apply temporal weight constraints to the attention score. After normalization, two single-path attention weights with temporal constraints are obtained. The average distribution of the dual-path attention weights is calculated and the local sequence entropy of the average distribution is solved. The differential attention score is generated based on the differential weights, the local sequence entropy and the dual-path attention weights. The knowledge state generation unit is used to apply the differential attention score as a weighting weight to the interactive coding embedding sequence, perform root mean square normalization on the result, and use the structural prior stability coefficient to perform residual scaling adjustment on the normalization result, and output the knowledge state representation.

5. The knowledge tracking system with dual noise perception as described in claim 1, characterized in that, The knowledge state correction module includes: a forward diffusion submodule and a reverse denoising submodule; The forward diffusion submodule is used to determine the noise intensity of each diffusion time step using a linear noise scheduling strategy, and gradually injects Gaussian noise into the knowledge state representation during the multi-step diffusion process to generate a series of noisy knowledge state representations with noise levels ranging from low to high. The reverse denoising submodule is used to generate diffusion time step embeddings based on the noisy knowledge state representations of each diffusion time step. It introduces student knowledge mastery level condition constraints and question embedding condition information, respectively calls the conditional denoising network to obtain the denoised knowledge state estimates under the corresponding condition constraints, and adaptively weights and fuses the two denoising results through a conditional gating mechanism to obtain the corrected knowledge state representation.

6. The knowledge tracking system with dual noise perception as described in claim 5, characterized in that, The conditional denoising network includes: a conditional fusion modulation unit, an adaptive causal attention unit, a feedforward feature modulation unit, and a temporal convolutional output unit; The conditional fusion modulation unit is used to generate a sine and cosine form of diffusion time step embedding, and fuses the diffusion time condition and the knowledge tracking observation condition into a unified conditional representation through linear projection and nonlinear transformation. Based on the unified conditional representation, a multilayer perceptron generates three types of modulation parameters, namely scaling, translation and gating, which are respectively adapted to the attention sublayer and the feedforward sublayer. The adaptive causal attention unit is used to fuse the noisy knowledge state representation with the learnable position code to obtain the fused code feature. Based on the modulation parameters corresponding to the attention sub-layer, the position code fused feature is subjected to adaptive layer normalization modulation. A multi-head self-attention mechanism is used to model the internal dependency of the modulation result sequence. During the modeling process, the attention correction intensity is dynamically controlled through the gated residual mechanism, and the first feature sequence after attention enhancement is output. The feedforward feature modulation unit is used to perform adaptive layer normalization modulation on the first feature sequence based on the modulation parameters corresponding to the feedforward sub-layer, use a feedforward neural network to perform nonlinear feature transformation on the modulation result, and connect it with the residual of the first feature sequence to output the second feature sequence after global semantic transformation. The temporal convolution output unit is used to apply a one-dimensional causal convolution to the second feature sequence, and after being processed by layer normalization, nonlinear activation and random deactivation in sequence, it outputs the corrected knowledge state representation by connecting it with the residual of the second feature sequence.

7. The knowledge tracking system with dual noise perception as described in claim 1, characterized in that, The prediction module includes: a feature fusion unit and a mapping unit; The feature fusion unit is used to perform a concatenation operation along the feature dimension on the corrected knowledge state representation and the question encoding embedding sequence to obtain the joint feature; The mapping unit is used to input the joint features into a fully connected neural network with two hidden layers and output the probability of the student answering the corresponding question correctly.

8. A knowledge tracking method with dual noise perception, characterized in that, include: Based on historical learning interaction sequences, construct question embedding sequences that fuse trainable difficulty embedding vectors, and construct interactive embedding sequences that reflect the answers to questions and the correctness of the answers. The question embedding sequence and the interaction embedding sequence are self-attention encoded by a multi-head self-attention mechanism to obtain a question encoding embedding sequence and an interaction encoding embedding sequence. The question encoding embedding sequence is compared and modeled by a differential attention mechanism to obtain a differential attention score. The knowledge state representation is determined based on the differential attention score and the interaction encoding embedding sequence. The knowledge state representation is denoised and corrected using a conditional diffusion model to obtain a corrected knowledge state representation. The corrected knowledge state representation is concatenated and fused with the question encoding embedding sequence to obtain joint features. Based on the joint features, the probability of a student answering each question correctly is predicted.

9. An electronic device, characterized in that, It includes a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the dual noise perception knowledge tracing method as described in claim 8.

10. A non-transitory computer storage medium, characterized in that, It stores a computer program, wherein the computer program, when executed by a processor, implements the knowledge tracking method with dual noise perception as described in claim 8.

Images