Dual-flow attention knowledge tracing method based on knowledge tracing neural network

By employing a dual-stream attention method based on a knowledge-tracking neural network, learners' knowledge factors and emotional and behavioral factors are decoupled, solving the problem of low purity in knowledge representation in existing technologies. This enables the quantitative modulation of emotional and behavioral factors, improving the model's interpretability and support for teaching interventions.

CN123070221APending Publication Date: 2026-11-10GUANGDONG UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611029039.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-10
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing knowledge tracing methods struggle to effectively decouple learners' knowledge factors from their emotional and behavioral factors, resulting in low purity of knowledge representation and a lack of quantitative modulation of emotional and behavioral factors, which impacts their ability to support teaching interventions and learning diagnoses.

Method used

We employ a dual-stream attention method based on a knowledge-tracking neural network. By using an input feature layer, an embedding and feature fusion layer, a temporal coding layer, an internally constrained dual-stream attention layer, and dual prediction heads, we model the knowledge stream and the emotion stream respectively. We also introduce an emotion modulation effect analysis module to achieve decoupled representation and quantized modulation of knowledge and emotion.

Benefits of technology

It improves the model's ability to characterize the dynamic fluctuations of the learning process, enhances the modeling accuracy of changes in learning behavior, and improves the model's interpretability, providing a better basis for decision-making in teaching intervention and learning support.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070221A_ABST
    Figure CN123070221A_ABST
Patent Text Reader

Abstract

The application relates to a double-flow attention knowledge tracking method based on a knowledge tracking neural network. The method comprises the following steps: acquiring original interaction data and dividing the data set; setting network construction parameters and network training parameters; generating response time residuals, long-term and short-term historical accuracy, sample sequences and weakly supervised emotion prior features according to the data set; constructing a knowledge tracking neural network according to the network construction parameters; training the knowledge tracking neural network according to the response time residuals, long-term and short-term historical accuracy and sample sequences to obtain a trained knowledge tracking neural network; and completing knowledge tracking analysis according to the trained knowledge tracking neural network and the weakly supervised emotion prior features. The application enhances the modeling precision of learning behavior changes by constructing a multi-dimensional behavior feature system containing response time residuals, long-term and short-term historical accuracy and weakly supervised emotion priors, thereby improving the ability to depict dynamic fluctuations in the learning process.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the fields of learning analytics and intelligent teaching technology, and in particular to a dual-stream attention knowledge tracking method based on a knowledge tracking neural network. Background Technology

[0002] Knowledge tracing (KT) is a core task in educational data mining, learning analytics, and intelligent tutoring systems. It aims to dynamically assess learners' knowledge mastery based on their historical interaction sequences and predict their responses to subsequent questions. Accurately characterizing learners' cognitive states allows educational systems to plan personalized learning paths, recommend questions, and provide precise teaching interventions, thereby effectively improving learning efficiency and the learning experience. For example, in large-scale online education platforms or intelligent question bank systems, the performance of the knowledge tracing model directly impacts the quality of learning recommendations and learners' long-term performance. Given the highly dynamic and complex nature of the learning process, and the multidimensional and multimodal characteristics of interaction data, traditional rule-based or shallow statistical methods are insufficient to meet practical needs. Therefore, deep learning-based knowledge tracing methods have gradually become a research focus.

[0003] Existing knowledge tracing methods primarily rely on deep learning models to model learners' historical interaction sequences. For example, early methods introduced recurrent neural networks to model these sequences; subsequently, dynamic key-value memory networks and models based on self-attention mechanisms further improved the ability to model long-range dependencies. These methods typically focus on the association between questions and knowledge points, predicting future performance by learning from past responses. However, these methods generally assume that learners' answers are primarily determined by their knowledge mastery, concentrating the model's focus on modeling the knowledge state.

[0004] However, in real-world learning scenarios, learners' responses are influenced not only by their knowledge acquisition but also by a combination of emotional and behavioral factors, including fluctuations in response time, short-term error accumulation, changes in behavioral rhythm, and changes in emotional state. Current technologies typically encode both knowledge and emotional / behavioral factors into a unified representation, leading to a high degree of coupling between information from different sources in the representation space. On one hand, behavioral features such as response time and historical accuracy are often introduced in a simplistic manner, lacking systematic modeling based on context or statistical structures, making it difficult to stably characterize fine-grained behavioral changes during the learning process. On the other hand, the lack of explicit constraints on different factors during representation learning causes knowledge and emotional / behavioral information to become intertwined, affecting the purity and stability of the knowledge representation. Furthermore, since model outputs typically provide only a single prediction, it's impossible to distinguish the specific sources of influence of knowledge and emotional / behavioral factors within the prediction, thus hindering effective support for instructional intervention and learning diagnosis.

[0005] To address the aforementioned issues, existing technologies have proposed several knowledge tracing methods that integrate multiple features or employ a dual-stream structure. These methods typically extract learner response sequences and auxiliary behavioral features, constructing different branches for modeling, and then aggregating them later through feature concatenation or result-level fusion. However, these methods remain relatively crude in their behavioral feature modeling, lacking a systematic construction based on residuals or contextual information, and failing to reflect the dynamic changes during the learning process. Furthermore, the constraints between knowledge representation and emotion representation are often applied after attention calculation, failing to suppress mutual interference during the representation formation stage, resulting in limited decoupling effects. In addition, these methods often lack mechanisms for simultaneously modeling single-branch knowledge results and joint knowledge-emotion results, making it difficult to further quantify the modulating effect of emotional and behavioral factors on the prediction results, thus limiting the model's application in interpretable analysis and instructional intervention support.

[0006] Therefore, simply using emotional and behavioral features as auxiliary inputs or fusing them at the prediction layer makes it difficult to guarantee the purity of the knowledge representation itself, and also makes it difficult to distinguish the relative roles of knowledge factors and emotional and behavioral factors in the prediction results. There is an urgent need for a knowledge tracking method that can effectively decouple knowledge factors and emotional and behavioral factors at the representation learning stage, while taking into account both prediction performance and interpretability, so as to improve the expressive power of the model and enhance its support for teaching decisions. Summary of the Invention

[0007] Therefore, it is necessary to provide a dual-stream attention knowledge tracking method based on a knowledge tracking neural network that can predict the relative effects of knowledge factors and emotional and behavioral factors in the results, in order to address the above-mentioned technical problems.

[0008] One type of knowledge tracking neural network includes: Input feature layer: used to obtain multiple features; Embedding and Feature Fusion Layer: Embeds and fuses features to obtain fused features; Temporal coding layer: Temporal coding is performed based on the fused features to obtain time series data; Internally constrained dual-stream attention layer: includes knowledge flow attention branch and emotion flow attention branch; time series data are processed separately into knowledge flow attention branch and emotion flow attention branch to obtain knowledge single-branch representation and emotion single-branch representation; Dual prediction heads: Obtain knowledge single-branch results and sentiment single-branch results based on knowledge single-branch representation and sentiment single-branch representation; Emotion Modulation Effect Analysis Module: Keeping the knowledge single-branch representation unchanged, the emotion single-branch representation is replaced with different emotion anchors to analyze the emotion modulation effect.

[0009] Furthermore, specifically: Multiple features are obtained from the input feature layer, including: question identifier encoding, knowledge point identifier encoding, current answer result, response time residual, long-term historical accuracy, short-term historical accuracy, and weakly supervised sentiment prior label; Embedding and Feature Fusion Layer: The question identifier encoding, knowledge point identifier encoding, current answer result, response time residual, long-term historical accuracy, short-term historical accuracy, and weakly supervised sentiment prior label are embedded and fused to obtain the fused features; the discrete variables such as knowledge point identifier, question identifier, and answer result are converted into low-dimensional dense vector representations and fused with continuous features such as response time residual, long-term historical accuracy, and short-term historical accuracy. The temporal coding layer uses a GRU network to temporally encode the input vector output by the embedding layer. This is used to model the dynamic process of learners' response behavior evolving over time and to provide a shared temporal context representation for subsequent knowledge flow and emotion flow branches. The internally constrained dual-stream attention layer includes a knowledge-stream attention branch and an emotion-stream attention branch. The knowledge-stream focuses on extracting knowledge representations related to answering ability, while the emotion-stream focuses on extracting emotional representations related to behavioral fluctuations. Specifically, the knowledge-stream is represented by question conditions. For queries, the basic behavior state is represented. For keys and values, generate knowledge pre-representation and knowledge post-representation; emotion stream is represented by basic behavioral states. Generate pre-representation and post-representation of sentiment for the query, key, and value; Knowledge pre-representation and emotion pre-representation are represented as follows: ,

[0010] The computational process of knowledge flow attention is represented as follows:

[0011]

[0012]

[0013]

[0014] in, Causality masking is used to ensure that the model only uses information from the current time step and earlier. The computational process for emotion-flow attention is represented as follows:

[0015]

[0016]

[0017]

[0018] in, and The following indicates the mapping function; The prediction layer includes knowledge-based single-branch prediction and knowledge-sentiment joint prediction heads.

[0019] This invention also provides a two-stream attention knowledge tracking method based on a knowledge tracking neural network, including: S1: Obtain the raw interaction data and divide the dataset; S2: Set the parameters for building and training the knowledge tracing neural network; S3: Generate response time residuals, long and short-term historical accuracy, sample sequences, and weakly supervised sentiment prior features based on the dataset; S4: The knowledge tracking neural network is trained based on the response time residual, long and short-term historical accuracy, and sample sequence to obtain the trained knowledge tracking neural network; S5: Complete knowledge tracking analysis based on the trained knowledge tracking neural network and weakly supervised emotion prior features.

[0020] Furthermore, in step S2, the network construction parameters are specifically: embedding dimension, hidden layer dimension, number of attention heads, and number of recurrent network layers; Network training parameters: activation function, number of training epochs, batch size, learning rate, and weight decay coefficient.

[0021] Furthermore, in step S3, specifically: S3.1: Discrete variable encoding; establishing a knowledge point identifier mapping table and a question identifier mapping table based on the training dataset; A knowledge point identifier mapping table and a question identifier mapping table are established based on the training dataset to convert the discrete variables in the original learning interaction data into an index form that the model can process; for knowledge points or questions that do not appear in the training set, they are uniformly mapped to unknown labels. Set learners In the The knowledge points in this interaction are marked as follows: The question is marked as Then the knowledge point encoding function is expressed as:

[0022] The coding function for the problem is represented as follows:

[0023] in, and The mapping function fitted based on the training set is used to map knowledge point identifiers and question identifiers to their corresponding integer indices, respectively. S3.2: Construction of response time residuals: The response time features are standardized to construct the response time residual features; S3.3: Construction of short- and long-term historical accuracy; S3.4: Construction of weakly supervised emotion prior labels. For learners who lack explicit emotion labels, weakly supervised emotion prior labels are constructed to characterize their behavioral and emotional states. S3.5: Construct sequence samples to form a sequence sample suitable for causal time series modeling, and align it with the target to be predicted at the next time step; For any learner sequence Using a length of By dividing the sliding window, we can obtain window fragments:

[0024] in, Indicate learner The A window of samples; In each window, the front The current input sequence is formed by taking one time point as the current input sequence and the last time point is used to align with the previous position to form the prediction target; thus, the current input sequence, the next question condition sequence, and the prediction target sequence can be constructed; among them, the current input sequence is used to represent the historical information that the model can see at the current time point, the next question condition sequence is used to provide the question information corresponding to the position to be predicted, and the prediction target sequence is used to supervise the model output.

[0025] Further, in step S3.2, specifically: To mitigate the impact of extreme values ​​on the modeling structure, a logarithmic transformation of the response time is performed. The expression for this transformation is:

[0026] in, Indicate learner In the The original response time in this interaction. Let represent the response time after logarithmic transformation; then, based on the training dataset, calculate the mean and standard deviation of the response time for each knowledge point, denoted as . and Therefore, the first The response time residual corresponding to this interaction can be expressed as:

[0027] in, Indicate learner The response time residual in the t-th interaction, A very small constant is introduced to prevent the denominator from being zero; when the number of samples for a certain knowledge point in the training dataset is insufficient, the global response time mean can be used. and global standard deviation As an alternative statistic.

[0028] Further, in step S3.3, specifically: Long-term historical accuracy features describe the learner's overall accuracy level up to the current moment, while short-term historical accuracy features describe local performance fluctuations over a recent period. Long-term historical accuracy is defined as the mean of all answers given up to the current moment, and its expression is:

[0029] in, Indicate learner The response result in the i-th interaction; when t=1, since there is no historical response record, let ; Short-term historical accuracy is defined as the length of the historical data. The average of the answers within the sliding window is expressed as:

[0030] in, Indicates the short-term window length.

[0031] Further, in step S3.4, specifically: Weakly supervised emotion prior labels are jointly constructed based on the answer results, response time residuals, and long-term historical accuracy. Weak supervisory emotional prior label ,in, Indicates a positive state. Indicates a neutral state. Indicates a negative state; the weakly supervised emotion prior label is generated according to the following rules: when season This indicates that the learner is currently in a relatively positive, flow-like behavioral state; when and season This indicates that learners are in a negative state when they answer incorrectly and respond significantly slowly; when , and season This indicates that while the learner has consistently performed well, they are currently making mistakes and exhibiting abnormal fluctuations in their behavior; in other cases, let... This indicates that the learner is in a neutral state.

[0032] Further, in step S4, the specific steps are as follows: the knowledge tracking neural network is trained on the dataset, and the target model with the best performance is selected through the validation dataset; during the training process, the training dataset is input into the model in batches, and joint optimization is performed based on the real answer results and various constraint signals; the AdamW optimizer is used during training, and the learning rate is adjusted using a cosine annealing strategy; after each round of training, the model index is calculated using the validation set, and the optimal model parameters are selected based on the joint prediction head AUC value on the validation set.

[0033] Furthermore, in step S5, specifically: S5.1: Based on samples with weakly supervised emotion labels in the dataset, statistically analyze the emotion representation centers corresponding to different emotion categories to construct emotion anchors; S5.2: For any sample, after obtaining its knowledge, represent... Emotional expression And the conditions for the next question; the original joint prediction result can be written as:

[0034] In maintaining Under the condition that the question embedding and knowledge point embedding remain unchanged, the original sentiment representation is replaced with a preset anchor point representation; when positive sentiment anchor points and negative sentiment anchor points are injected, the corresponding prediction results are as follows:

[0035]

[0036] By using the above substitutions, under the same knowledge state and the same question conditions, we can examine only the predictive differences caused by changes in emotional expression, thus providing a clear benchmark for the effect analysis. S5.3: Define the emotion modulation effect on a single sample as: , The global sentiment modulation effect can be obtained by averaging over all samples:

[0037] S5.4: For samples with a negative emotion label under weak supervision, the average improvement of the original prediction result after injecting a positive emotion anchor can be statistically analyzed to characterize the potential improvement space under the negative behavior state; for samples with the original prediction probability in the uncertainty interval, the average prediction difference after injection under positive and negative emotion anchors can be compared to identify the influence strength of emotion factors in the model's hesitation interval. S5.5 Based on the above effect analysis, it is possible to distinguish whether a certain prediction result is mainly affected by the knowledge state or by the emotional and behavioral state; and to identify learning segments that are highly sensitive to emotional factors.

[0038] This invention enhances the model's ability to characterize the dynamic fluctuations of the learning process by constructing a multi-dimensional behavioral feature system that includes response time residuals, long- and short-term historical accuracy, and weakly supervised emotional priors, thereby improving the modeling accuracy of changes in learning behavior. By introducing an effect analysis mechanism based on latent representation substitution, the modulating effect of emotional and behavioral factors on prediction results is analyzed while keeping the knowledge state unchanged. This allows for a quantitative assessment of prediction differences under different emotional states, improving the model's interpretability and providing a basis for subsequent teaching intervention strategies and learning support decisions. Attached Figure Description

[0039] Figure 1 A flowchart of a dual-stream attention knowledge tracking method based on a knowledge tracking neural network, provided in an embodiment of the present invention.

[0040] Figure 2 This is a schematic diagram of a knowledge tracking model structure based on knowledge-emotion decoupled representation and internally constrained two-stream attention, provided for an embodiment of the present invention. It is used to illustrate the connection relationship and information flow mode between the input feature layer, embedding and feature fusion layer, temporal coding layer, internally constrained two-stream attention layer and dual prediction heads in the model.

[0041] Figure 3 This is a schematic diagram illustrating the emotion modulation effect analysis based on emotion anchor replacement provided in an embodiment of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0043] Example 1: This embodiment provides, as follows: Figure 1 The dual-stream attention knowledge tracking method based on a knowledge tracking neural network, as shown, includes: S1: Obtain the raw interaction data and divide the dataset; S2: Set network construction parameters and network training parameters; S3: Generate response time residuals, long and short-term historical accuracy, sample sequences, and weakly supervised sentiment prior features based on the dataset; S4: Construct a knowledge tracking neural network based on the network construction parameters; S5: The knowledge tracking neural network is trained based on the response time residual, long and short-term historical accuracy, and sample sequence to obtain the trained knowledge tracking neural network; S6: Complete knowledge tracking analysis based on the trained knowledge tracking neural network and weakly supervised emotion prior features.

[0044] This embodiment enhances the model's ability to characterize the dynamic fluctuations of the learning process by constructing a multi-dimensional behavioral feature system that includes response time residuals, long-term and short-term historical accuracy, and weakly supervised emotional priors, thereby improving the modeling accuracy of changes in learning behavior. By introducing an effect analysis mechanism based on latent representation substitution, the modulating effect of emotional and behavioral factors on prediction results is analyzed while keeping the knowledge state unchanged. This allows for a quantitative assessment of prediction differences under different emotional states, improving the model's interpretability and providing a basis for subsequent teaching intervention strategies and learning support decisions.

[0045] Example 2: This embodiment further discloses information based on Embodiment 1: Furthermore, in step S2, the network construction parameters are specifically: embedding dimension, hidden layer dimension, number of attention heads, and number of recurrent network layers; Network training parameters: activation function, number of training epochs, batch size, learning rate, and weight decay coefficient.

[0046] Furthermore, in step S3, specifically: S3.1: Discrete variable encoding; establishing a knowledge point identifier mapping table and a question identifier mapping table based on the training dataset; A knowledge point identifier mapping table and a question identifier mapping table are established based on the training dataset to convert the discrete variables in the original learning interaction data into an index form that the model can process; for knowledge points or questions that do not appear in the training set, they are uniformly mapped to unknown labels. Set learners In the The knowledge points in this interaction are marked as follows: The question is marked as Then the knowledge point encoding function is expressed as:

[0047] The coding function for the problem is represented as follows:

[0048] in, and The mapping function fitted based on the training set is used to map knowledge point identifiers and question identifiers to their corresponding integer indices, respectively. S3.2: Construction of response time residuals: The response time features are standardized to construct the response time residual features; S3.3: Construction of short- and long-term historical accuracy; S3.4: Construction of weakly supervised emotion prior labels. For learners who lack explicit emotion labels, weakly supervised emotion prior labels are constructed to characterize their behavioral and emotional states. S3.5: Construct sequence samples to form a sequence sample suitable for causal time series modeling, and align it with the target to be predicted at the next time step; For any learner sequence Using a length of By dividing the sliding window, we can obtain window fragments:

[0049] in, Indicate learner The A window of samples; In each window, the front The current input sequence is formed by taking one time point as the current input sequence and the last time point is used to align with the previous position to form the prediction target; thus, the current input sequence, the next question condition sequence, and the prediction target sequence can be constructed; among them, the current input sequence is used to represent the historical information that the model can see at the current time point, the next question condition sequence is used to provide the question information corresponding to the position to be predicted, and the prediction target sequence is used to supervise the model output.

[0050] Further, in step S3.2, specifically: To mitigate the impact of extreme values ​​on the modeling structure, a logarithmic transformation of the response time is performed. The expression for this transformation is:

[0051] in, Indicate learner In the The original response time in this interaction. Let represent the response time after logarithmic transformation; then, based on the training dataset, calculate the mean and standard deviation of the response time for each knowledge point, denoted as . and Therefore, the first The response time residual corresponding to this interaction can be expressed as:

[0052] in, Indicate learner The response time residual in the t-th interaction, A very small constant is introduced to prevent the denominator from being zero; when the number of samples for a certain knowledge point in the training dataset is insufficient, the global response time mean can be used. and global standard deviation As an alternative statistic.

[0053] Further, in step S3.3, specifically: Long-term historical accuracy features describe the learner's overall accuracy level up to the current moment, while short-term historical accuracy features describe local performance fluctuations over a recent period. Long-term historical accuracy is defined as the mean of all answers given up to the current moment, and its expression is:

[0054] in, Indicate learner The response result in the i-th interaction; when t=1, since there is no historical response record, let ; Short-term historical accuracy is defined as the length of the historical data. The average of the answers within the sliding window is expressed as:

[0055] in, Indicates the short-term window length.

[0056] Further, in step S3.4, specifically: Weakly supervised emotion prior labels are jointly constructed based on the answer results, response time residuals, and long-term historical accuracy. Weak supervisory emotional prior label ,in, Indicates a positive state. Indicates a neutral state. Indicates a negative state; the weakly supervised emotion prior label is generated according to the following rules: when season This indicates that the learner is currently in a relatively positive, flow-like behavioral state; when and season This indicates that learners are in a negative state when they answer incorrectly and respond significantly slowly; when , and season This indicates that while the learner has consistently performed well, they are currently making mistakes and exhibiting abnormal fluctuations in their behavior; in other cases, let... This indicates that the learner is in a neutral state.

[0057] Further, in step S4, the specific steps are as follows: the knowledge tracking neural network is trained on the dataset, and the target model with the best performance is selected through the validation dataset; during the training process, the training dataset is input into the model in batches, and joint optimization is performed based on the real answer results and various constraint signals; the AdamW optimizer is used during training, and the learning rate is adjusted using a cosine annealing strategy; after each round of training, the model index is calculated using the validation set, and the optimal model parameters are selected based on the joint prediction head AUC value on the validation set.

[0058] Furthermore, in step S5, specifically: S5.1: Based on samples with weakly supervised emotion labels in the dataset, statistically analyze the emotion representation centers corresponding to different emotion categories to construct emotion anchors; S5.2: For any sample, after obtaining its knowledge, represent... Emotional expression And the conditions for the next question; the original joint prediction result can be written as:

[0059] In maintaining Under the condition that the question embedding and knowledge point embedding remain unchanged, the original sentiment representation is replaced with a preset anchor point representation; when positive sentiment anchor points and negative sentiment anchor points are injected, the corresponding prediction results are as follows:

[0060]

[0061] By using the above substitutions, under the same knowledge state and the same question conditions, we can examine only the predictive differences caused by changes in emotional expression, thus providing a clear benchmark for the effect analysis. S5.3: Define the emotion modulation effect on a single sample as: , The global sentiment modulation effect can be obtained by averaging over all samples:

[0062] S5.4: For samples with a negative emotion label under weak supervision, the average improvement of the original prediction result after injecting a positive emotion anchor can be statistically analyzed to characterize the potential improvement space under the negative behavior state; for samples with the original prediction probability in the uncertainty interval, the average prediction difference after injection under positive and negative emotion anchors can be compared to identify the influence strength of emotion factors in the model's hesitation interval. S5.5 Based on the above effect analysis, it is possible to distinguish whether a certain prediction result is mainly affected by the knowledge state or by the emotional and behavioral state; and to identify learning segments that are highly sensitive to emotional factors.

[0063] This embodiment enhances the model's ability to characterize the dynamic fluctuations of the learning process by constructing a multi-dimensional behavioral feature system that includes response time residuals, long-term and short-term historical accuracy, and weakly supervised emotional priors, thereby improving the modeling accuracy of changes in learning behavior. By introducing an effect analysis mechanism based on latent representation substitution, the modulating effect of emotional and behavioral factors on prediction results is analyzed while keeping the knowledge state unchanged. This allows for a quantitative assessment of prediction differences under different emotional states, improving the model's interpretability and providing a basis for subsequent teaching intervention strategies and learning support decisions.

[0064] Example 3: This embodiment further discloses information based on Embodiment 2: like Figure 1 The diagram shows the implementation flowchart of the dual-stream attention knowledge tracking method based on a knowledge tracking neural network provided in this embodiment of the invention. The specific steps are as follows: S1: Data acquisition, cleaning and initialization configuration: acquire raw learning interaction data exported from the learning platform, standardize the fields, and complete the dataset partitioning; The raw data includes at least the fields of user identifier, question identifier, knowledge point identifier, answer result, and answer time. During data preprocessing, samples with empty user identifiers, answer order, knowledge point identifiers, answer results, and answer times are first deleted. Then, outlier samples with response times less than or equal to zero are removed. Finally, the data is sorted according to user identifier and answer order. Furthermore, to avoid data leakage, the dataset is divided into training, validation, and test sets at the user level, with a ratio of 8:1:1, ensuring that a complete interaction sequence for the same user appears in only one dataset.

[0065] S2: Initialization parameters include embedding dimension, hidden layer dimension, number of attention heads, number of recurrent network layers, activation function, number of training epochs, batch size, learning rate, and weight decay coefficient. In a preferred embodiment, the embedding dimension can be set to 32, the hidden layer dimension to 64, the number of attention heads to 4, the temporal coding layer to a 2-layer GRU structure, the activation function to ReLU, the number of training epochs to 20, the batch size to 64, and the learning rate to [missing information]. The weight decay coefficient can be set to .

[0066] S3: Build a feature processor based on the training dataset and generate model inputs in a unified format for the training set, validation set, and test set; The specific process is as follows: S3.1: Discrete variable encoding; establishing a knowledge point identifier mapping table and a question identifier mapping table based on the training dataset; In one specific embodiment, a knowledge point identifier mapping table and a question identifier mapping table are established based on the training dataset to convert the discrete variables in the original learning interaction data into an index form that the model can process. For knowledge points or questions that do not appear in the training set, they are uniformly mapped to unknown identifiers, thereby ensuring that unseen categories in the validation set and test set can still be processed stably.

[0067] Set learners In the The knowledge points in this interaction are marked as follows: The question is marked as Then the knowledge point encoding function is expressed as:

[0068] The coding function for the problem is represented as follows:

[0069] in, and The mapping function, fitted based on the training set, is used to map knowledge point identifiers and question identifiers to their corresponding integer indices. After encoding, the knowledge point information and question information in the original interaction sequence can be converted into a unified discrete index representation, which is then input into the subsequent embedding layer for vectorization processing.

[0070] S3.2 Response Time Residual Construction: To characterize the differences in learners' behavioral rhythm during the answering process and reduce the direct impact of differences in the original response time scale between questions of different knowledge points, response time features are standardized to construct response time residual features. This feature reflects the degree of deviation of the learner's current answering speed from the general answering level for similar knowledge points.

[0071] In a specific implementation, the response time is first logarithmically transformed to reduce the impact of extreme large values ​​on the modeling structure. The expression is as follows:

[0072] in, Indicate learner In the The original response time in this interaction. This represents the response time after logarithmic transformation. Then, based on the training dataset, the mean and standard deviation of the response time are calculated for each knowledge point, denoted as... and Therefore, the first The response time residual corresponding to this interaction can be expressed as:

[0073] in, Indicate learner The response time residual in the t-th interaction, A very small constant is introduced to prevent the denominator from being zero. When the number of samples for a particular knowledge point in the training dataset is insufficient, the global response time average can be used. and global standard deviation As an alternative statistic, the above processing transforms response time information into comparable behavioral bias signals, providing a basis for subsequent emotion prior construction and model input representation.

[0074] S3.3 Construction of long-term and short-term historical accuracy: In order to reflect learners’ overall mastery level over a longer period of time as well as recent changes in their answering status, long-term historical accuracy features and short-term historical accuracy features are constructed respectively.

[0075] In a specific implementation, the long-term historical accuracy feature describes the learner's overall accuracy level up to the current moment, while the short-term historical accuracy feature describes local performance fluctuations over a recent period. The long-term historical accuracy can be defined as the average of all answers given up to the current moment, and its expression is:

[0076] in, Indicate learner The response result in the i-th interaction. When t=1, since there are no historical response records, we can let .

[0077] Furthermore, short-term historical accuracy can be defined as the accuracy over a period of time. The average of the answers within the sliding window is expressed as:

[0078] in, This represents the short-term window length, which can be 5 in the example. By simultaneously introducing the two features mentioned above, learners' response behavior can be characterized from both long-term trends and short-term fluctuations, providing richer behavioral contextual information for subsequent emotion prior generation and temporal state modeling.

[0079] S3.4 Weakly supervised emotion prior label construction: For learners who lack explicit emotion labels, weakly supervised emotion prior labels are constructed to characterize their behavioral and emotional states.

[0080] In a specific embodiment, the emotion is used to represent the learning behavior state. This label is not directly equivalent to a real emotion in a psychological sense. For cases where there is no explicit emotion label, a weakly supervised emotion prior label is jointly constructed based on the answer result, response time residual, and long-term historical accuracy. This label is used to represent positive, stable, or negative behavioral states during the learning process, thereby providing a supervisory signal for subsequent emotion branch learning.

[0081] Weak supervisory emotional prior label ,in, Indicates a positive state. Indicates a neutral state. This represents a negative state. The weakly supervised emotion prior label is generated according to the following rules: when... season This indicates that the learner is currently in a relatively positive, flow-like behavioral state; when and season This indicates that learners are in a negative state when they answer incorrectly and respond significantly slowly; when , and season This indicates that while the learner has consistently performed well, they are currently making mistakes and exhibiting abnormal fluctuations in their behavior; in other cases, let... This indicates that the learner is in a neutral state.

[0082] Based on the above rules, weakly supervised emotion priors can be constructed from observable learning behaviors, providing auxiliary supervisory information for the model to explicitly separate knowledge factors from emotional and behavioral factors.

[0083] S3.5 Construct sequence samples to form a sequence sample suitable for causal time series modeling, and align it with the target to be predicted at the next time step.

[0084] Specifically, to enable the model to predict the next moment's response performance based on the learner's historical interaction sequence, the learner's interaction records are windowed and segmented to construct sequence samples suitable for causal time series modeling. The purpose of this process is to preserve historical dependencies.

[0085] In a specific implementation, for any learner sequence Using a length of By dividing the sliding window, we can obtain window fragments:

[0086] in, Indicate learner The A window of samples.

[0087] In each window, the front The current input sequence is constructed from the first time step, and the last time step is used to align with the previous position to form the prediction target. This results in the current input sequence, the next question condition sequence, and the prediction target sequence. The current input sequence represents the historical information visible to the model at the current time step, the next question condition sequence provides the question information corresponding to the position to be predicted, and the prediction target sequence is used to supervise the model output. Through this sequence construction method, the original interactive data can be transformed into a sample format suitable for training and inference in knowledge tracking tasks.

[0088] S4: Model Construction. A knowledge tracking neural network capable of simultaneously characterizing both knowledge state and emotional / behavioral state is constructed. The model as a whole comprises an embedding layer, a temporal encoding layer, an internally constrained two-stream attention layer, and dual prediction heads. These components are sequentially connected to complete the mapping from raw input features to predicted output results.

[0089] S4.1 To unify the mapping of discrete identifiers and continuous behavioral features to an input space that the neural network can process, an embedding layer is constructed. This layer converts discrete variables such as knowledge point identifiers, question identifiers, and answer results into low-dimensional dense vector representations, and fuses them with continuous features such as response time residuals, long-term historical accuracy, and short-term historical accuracy. The discrete variables are represented using embedding layers as follows:

[0090] in, , and These represent the knowledge point embedding function, the question embedding function, and the answer result embedding function, respectively.

[0091] Furthermore, by concatenating the above discrete embedding representation with continuous behavioral features such as response time residuals, long-term historical accuracy, and short-term historical accuracy, the input vector at the current moment can be obtained:

[0092] Through the above processing, multi-source learning behavior information can be integrated into a unified vector representation, which can then be used as input for subsequent temporal coding layers.

[0093] S4.2 performs temporal encoding on the input vector output by the embedding layer to extract the learner's basic behavioral state in the historical interaction sequence. This layer is used to model the dynamic process of the learner's response behavior evolving over time and to provide a shared temporal context representation for subsequent knowledge flow and emotion flow branches.

[0094] In this embodiment, the timing coding layer may employ a GRU network, and its basic behavioral state representation can be written as:

[0095] in, Indicates the first The basic behavioral state at any given moment.

[0096] Furthermore, in order to incorporate question condition information related to the current answering task into subsequent knowledge flows, question condition representations can be further constructed:

[0097] in, The projection function is composed of a linear mapping and an activation function. Through the above design, the model can both preserve the dynamic behavioral states in the learning history and explicitly introduce the current question context, laying the foundation for subsequent two-stream attention modeling.

[0098] S4.3 models the knowledge state and emotional / behavioral state separately, and reduces mutual contamination between them in the representation space. An internally constrained dual-stream attention layer is constructed after the temporal encoding layer. This layer includes a knowledge stream attention branch and an emotion stream attention branch. The knowledge stream focuses on extracting knowledge representations related to answering ability, while the emotion stream focuses on extracting emotional representations related to behavioral fluctuations. Specifically, the knowledge stream is represented by question conditions. For queries, the basic behavior state is represented. For keys and values, generate knowledge pre-representation and knowledge post-representation; emotion stream is represented by basic behavioral states. Generate a pre-representation and a post-representation of sentiment for the query, key, and value.

[0099] In a specific embodiment, knowledge pre-representation and sentiment pre-representation can be represented as follows: ,

[0100] Furthermore, the computational process of knowledge flow attention can be represented as follows:

[0101]

[0102]

[0103]

[0104] in, Causality masking is used to ensure that the model only uses information from the current moment and earlier.

[0105] Correspondingly, the computational process for emotion stream attention can be expressed as:

[0106]

[0107]

[0108]

[0109] in, and This serves as the post-representation mapping function. Through the aforementioned two-stream structure, knowledge post-representations and sentiment post-representations can be obtained separately, thus providing a foundation for subsequent constraint learning and output prediction.

[0110] S4.4 To ensure that knowledge representations better represent learners' knowledge states, and emotion representations better represent behavioral fluctuations, emotion supervision constraints, adversarial decoupling constraints, and orthogonal constraints are introduced within the dual-stream attention layer. To enable emotion pre-representations to learn weakly supervised prior emotion information, emotion pre-representations are classified to obtain emotion prediction results: ; To suppress sentiment information mixed in with the knowledge pre-representation, a gradient inversion layer is used to input the knowledge pre-representation into the adversarial classifier, in the form of: ; To reduce the correlation between knowledge pre-representation and sentiment pre-representation, an orthogonal constraint is constructed by limiting the similarity of their included angles:

[0111] S4.5 To simultaneously obtain prediction results based solely on knowledge state and joint prediction results integrating knowledge state and emotional state, a dual prediction head is set at the model output. This design can not only improve prediction performance but also facilitate subsequent comparisons of the modulating effects of emotional and behavioral factors on the prediction results.

[0112] In this embodiment, the prediction layer includes a knowledge-single-branch prediction head and a knowledge-sentiment joint prediction head.

[0113] The single-branch prediction result can be expressed as:

[0114] The joint prediction result of knowledge and sentiment can be expressed as:

[0115] S5: Model Training. Based on the training dataset, the parameters of the knowledge tracking neural network described above are learned, and the optimal target model is selected using the validation dataset. During training, the training dataset is input into the model in batches, and joint optimization is performed based on the actual responses and various constraint signals. The AdamW optimizer is used during training, and a cosine annealing strategy is employed to adjust the learning rate. After each round of training, the model metrics are calculated using the validation set, and the optimal model parameters are selected based on the joint predictor head AUC value on the validation set.

[0116] In this embodiment, after model training is completed, the test dataset is input into the trained target model to obtain predictions of the learner's performance in the next moment. This step simultaneously outputs joint results for actual system predictions and single-branch knowledge results for interpretive analysis. The single-branch knowledge prediction probability can be used as an interpretive analysis result to help identify the relative influence of knowledge factors and emotional / behavioral factors in a given prediction result.

[0117] S6: After obtaining the trained target model, to further analyze the impact of emotional and behavioral factors on the knowledge tracking prediction results, this embodiment introduces an effect analysis mechanism based on latent representation replacement during the testing phase. This mechanism does not directly change the knowledge state, but observes the degree to which changes in emotional representation modulate the prediction results while keeping the post-knowledge representation and the conditional information of the next question unchanged. Therefore, it can maintain the integrity of the technical logic and enhance the interpretability of the model output.

[0118] S6.1 Based on samples with weakly supervised emotion labels in the training or validation dataset, statistically analyze the emotion representation centers corresponding to different emotion categories to construct emotion anchors. S6.2 For any sample in the test set, during the normal reasoning process of the model, its knowledge representation can be obtained. Emotional expression And the conditions for the next question. The original joint prediction result can be written as:

[0119] To identify the modulating effect of emotional factors on prediction results, this embodiment further maintains... Under the condition that the question embedding and knowledge point embedding remain unchanged, the original sentiment representation is replaced with a preset anchor point representation. When positive sentiment anchor points and negative sentiment anchor points are injected, the corresponding prediction results are as follows:

[0120]

[0121] By using the above substitutions, under the same knowledge state and the same question conditions, we can examine only the predictive differences caused by changes in emotional expression, thus providing a clear benchmark for effect analysis.

[0122] S6.3 defines the emotion modulation effect on a single sample as: , The global sentiment modulation effect can be obtained by averaging over all samples:

[0123] S6.4 In this embodiment, to enhance model interpretability, further analysis is performed on specific subgroups. For samples with a negative emotion label under weak supervision, the average improvement relative to the original prediction result after injecting a positive emotion anchor can be statistically analyzed to characterize the potential improvement space under negative behavioral states. For samples whose original predicted probabilities are in the uncertainty interval (e.g., 0.4 to 0.6), the average prediction difference after injection under positive and negative emotion anchors can be compared to identify the influence strength of emotion factors in the model's hesitation interval.

[0124] S6.5 Based on the above effect analysis, this embodiment can at least distinguish whether a prediction result is mainly affected by knowledge state or by emotional and behavioral state; and identify learning segments that are highly sensitive to emotional factors. This provides more granular intervention criteria for the teaching system, such as prioritizing encouragement, prompts, or rhythm adjustments for emotionally sensitive samples when their knowledge levels are similar; it also improves the interpretability of the knowledge tracking model, enabling the system to output not only predicted values ​​but also the degree to which the predicted values ​​are modulated by emotional states.

[0125] This analysis can characterize the interpretable effects and potential intervention sensitivity within the model, but it is not equivalent to the direct identification of the causal effects of external teaching interventions in real educational scenarios.

[0126] This embodiment introduces an internal constraint mechanism before attention convergence to effectively separate knowledge representation from emotional behavior representation during the representation formation stage, thereby reducing mutual interference between different factors and improving the purity and stability of knowledge representation. By constructing a multi-dimensional behavioral feature system including response time residuals, long and short-term historical accuracy, and weakly supervised emotional priors, it enhances the model's ability to characterize dynamic fluctuations in the learning process, thus improving the modeling accuracy of changes in learning behavior. By designing a dual-flow structure of knowledge flow and emotional flow, it explicitly distinguishes between knowledge state and emotional behavior state within the model, enabling clearer semantic division of different factors in the representation space, thereby improving the overall interpretability of the model. By setting knowledge... The single-branch prediction head and the knowledge-emotion joint prediction head enable collaborative modeling of different information sources, allowing the model to maintain high predictive performance while also possessing the ability to analyze and diagnose the sources of prediction results. Furthermore, the method of this invention has good versatility and can be applied to various scenarios such as online education platforms, adaptive practice systems, intelligent tutoring systems, and learning analytics platforms. Further, by introducing an effect analysis mechanism based on latent representation substitution, the modulating effect of emotional and behavioral factors on prediction results is analyzed while maintaining the knowledge state. This allows for a quantitative assessment of prediction differences under different emotional states, improving the model's interpretability and providing a basis for subsequent teaching intervention strategies and learning support decisions.

Claims

1. A knowledge-tracking neural network, characterized in that, include: Input feature layer: used to obtain multiple features; Embedding and Feature Fusion Layer: Embeds and fuses features to obtain fused features; Temporal coding layer: Temporal coding is performed based on the fused features to obtain time series data; Internally constrained dual-stream attention layer: includes knowledge flow attention branch and emotion flow attention branch; time series data are processed separately into knowledge flow attention branch and emotion flow attention branch to obtain knowledge single-branch representation and emotion single-branch representation; Dual prediction heads: Obtain knowledge single-branch results and sentiment single-branch results based on knowledge single-branch representation and sentiment single-branch representation; Emotion Modulation Effect Analysis Module: Keeping the knowledge single-branch representation unchanged, the emotion single-branch representation is replaced with different emotion anchors to analyze the emotion modulation effect.

2. The knowledge tracking neural network according to claim 1, characterized in that, Specifically: Multiple features are obtained from the input feature layer, including: question identifier encoding, knowledge point identifier encoding, current answer result, response time residual, long-term historical accuracy, short-term historical accuracy, and weakly supervised sentiment prior label; Embedding and Feature Fusion Layer: The question identifier encoding, knowledge point identifier encoding, current answer result, response time residual, long-term historical accuracy, short-term historical accuracy, and weakly supervised sentiment prior label are embedded and fused to obtain the fused features; the discrete variables such as knowledge point identifier, question identifier, and answer result are converted into low-dimensional dense vector representations and fused with continuous features such as response time residual, long-term historical accuracy, and short-term historical accuracy. The temporal coding layer uses a GRU network to temporally encode the input vector output by the embedding layer. This is used to model the dynamic process of learners' response behavior evolving over time and to provide a shared temporal context representation for subsequent knowledge flow and emotion flow branches. The internally constrained dual-stream attention layer includes a knowledge-stream attention branch and an emotion-stream attention branch. The knowledge-stream focuses on extracting knowledge representations related to answering ability, while the emotion-stream focuses on extracting emotional representations related to behavioral fluctuations. Specifically, the knowledge-stream is represented by question conditions. For queries, the basic behavior state is represented. For keys and values, generate knowledge pre-representation and knowledge post-representation; emotion stream is represented by basic behavioral states. Generate pre-representation and post-representation of sentiment for the query, key, and value; Knowledge pre-representation and emotion pre-representation are represented as follows: , The computational process of knowledge flow attention is represented as follows: in, Causality masking is used to ensure that the model only uses information from the current time step and earlier. The computational process for emotion-flow attention is represented as follows: in, and The following indicates the mapping function; The prediction layer includes knowledge-based single-branch prediction and knowledge-sentiment joint prediction heads.

3. A two-stream attention knowledge tracking method based on a knowledge tracking neural network, characterized in that, include: S1: Obtain the raw interaction data and divide the dataset; S2: Set the parameters for building and training the knowledge tracing neural network; S3: Generate response time residuals, long and short-term historical accuracy, sample sequences, and weakly supervised sentiment prior features based on the dataset; S4: The knowledge tracking neural network is trained based on the response time residual, long and short-term historical accuracy, and sample sequence to obtain the trained knowledge tracking neural network; S5: Complete knowledge tracking analysis based on the trained knowledge tracking neural network and weakly supervised emotion prior features.

4. The dual-stream attention knowledge tracking method based on a knowledge tracking neural network according to claim 3, characterized in that, In step S2, the network construction parameters are specifically: embedding dimension, hidden layer dimension, number of attention heads, and number of recurrent network layers; Network training parameters: activation function, number of training epochs, batch size, learning rate, and weight decay coefficient.

5. The dual-stream attention knowledge tracking method based on a knowledge tracking neural network according to claim 3, characterized in that, In step S3, specifically: S3.1: Discrete variable encoding; establishing a knowledge point identifier mapping table and a question identifier mapping table based on the training dataset; A knowledge point identifier mapping table and a question identifier mapping table are established based on the training dataset to convert the discrete variables in the original learning interaction data into an index form that the model can process. For knowledge points or questions that do not appear in the training set, they are uniformly mapped to unknown labels; Assume learners In the The knowledge points in this interaction are marked as follows: The question is marked as Then the knowledge point encoding function is expressed as: The problem's encoding function is represented as follows: in, and The mapping function fitted based on the training set is used to map knowledge point identifiers and question identifiers to their corresponding integer indices, respectively. S3.2: Construction of response time residuals: The response time features are standardized to construct the response time residual features; S3.3: Construction of short- and long-term historical accuracy; S3.4: Construction of weakly supervised emotion prior labels. For learners who lack explicit emotion labels, weakly supervised emotion prior labels are constructed to characterize their behavioral and emotional states. S3.5: Construct sequence samples to form a sequence sample suitable for causal time series modeling, and align it with the target to be predicted at the next time step; For any learner sequence Using a length of By dividing the sliding window, we can obtain window fragments: in, Indicate learner The A window of samples; In each window, the front The current input sequence is formed by taking one time point as the current input sequence and the last time point is used to align with the previous position to form the prediction target; thus, the current input sequence, the next question condition sequence, and the prediction target sequence can be constructed; among them, the current input sequence is used to represent the historical information that the model can see at the current time point, the next question condition sequence is used to provide the question information corresponding to the position to be predicted, and the prediction target sequence is used to supervise the model output.

6. The dual-stream attention knowledge tracking method based on a knowledge tracking neural network according to claim 3, characterized in that, In step S3.2, specifically: To mitigate the impact of extreme values ​​on the modeling structure, a logarithmic transformation of the response time is performed. The expression for this transformation is: in, Indicate learner In the The original response time in this interaction. Let represent the response time after logarithmic transformation; then, based on the training dataset, calculate the mean and standard deviation of the response time for each knowledge point, denoted as . and Therefore, the first The response time residual corresponding to this interaction can be expressed as: in, Indicate learner The response time residual in the t-th interaction, A very small constant is introduced to prevent the denominator from being zero; when the number of samples for a certain knowledge point in the training dataset is insufficient, the global response time mean can be used. and global standard deviation As an alternative statistic.

7. The dual-stream attention knowledge tracking method based on a knowledge tracking neural network according to claim 3, characterized in that, In step S3.3, specifically: Long-term historical accuracy features describe the learner's overall accuracy level up to the current moment, while short-term historical accuracy features describe local performance fluctuations over a recent period. Long-term historical accuracy is defined as the mean of all answers given up to the current moment, and its expression is: in, Indicate learner The response result in the i-th interaction; when t=1, since there is no historical response record, let ; Short-term historical accuracy is defined as the length of the historical data. The average of the answers within the sliding window is expressed as: in, Indicates the short-term window length.

8. The dual-stream attention knowledge tracking method based on a knowledge tracking neural network according to claim 3, characterized in that, In step S3.4, specifically: Weakly supervised emotion prior labels are jointly constructed based on the answer results, response time residuals, and long-term historical accuracy. Weak supervisory emotional prior label ,in, Indicates a positive state. Indicates a neutral state. Indicates a negative state; the weakly supervised emotion prior label is generated according to the following rules: when season This indicates that the learner is currently in a relatively positive, flow-like behavioral state; when and season This indicates that learners are in a negative state when they answer incorrectly and respond significantly slowly; when , and season This indicates that while the learner has consistently performed well, they are currently making mistakes and exhibiting abnormal fluctuations in their behavior; in other cases, let... This indicates that the learner is in a neutral state.

9. The dual-stream attention knowledge tracking method based on a knowledge tracking neural network according to claim 3, characterized in that, In step S4, specifically: the knowledge tracking neural network is trained on the dataset, and the target model with the best performance is selected by verifying the dataset; during the training process, the training dataset is input into the model in batches, and joint optimization is performed based on the actual answer results and various constraint signals; The AdamW optimizer was used during training, and a cosine annealing strategy was employed to adjust the learning rate. After each round of training, the model metrics are calculated using the validation set, and the optimal model parameters are selected based on the joint predictor head AUC value on the validation set.

10. The dual-stream attention knowledge tracking method based on a knowledge tracking neural network according to claim 3, characterized in that, In step S5, specifically: S5.1: Based on samples with weakly supervised emotion labels in the dataset, statistically analyze the emotion representation centers corresponding to different emotion categories to construct emotion anchors; S5.2: For any sample, after obtaining its knowledge, represent... Emotional expression And the conditions for the next question; the original joint prediction result can be written as: In maintaining Under the condition that the question embedding and knowledge point embedding remain unchanged, the original sentiment representation is replaced with a preset anchor point representation; when positive sentiment anchor points and negative sentiment anchor points are injected, the corresponding prediction results are as follows: By using the above substitutions, under the same knowledge state and the same question conditions, we can examine only the predictive differences caused by changes in emotional expression, thus providing a clear benchmark for the effect analysis. S5.3: Define the emotion modulation effect on a single sample as: , The global sentiment modulation effect can be obtained by averaging over all samples: S5.4: For samples with a negative emotion label under weak supervision, the average improvement of the original prediction result after injecting a positive emotion anchor can be statistically analyzed to characterize the potential improvement space under the negative behavior state; for samples with the original prediction probability in the uncertainty interval, the average prediction difference after injection under positive and negative emotion anchors can be compared to identify the influence strength of emotion factors in the model's hesitation interval. S5.5 Based on the above effect analysis, it is possible to distinguish whether a certain prediction result is mainly affected by the knowledge state or by the emotional and behavioral state; and to identify learning segments that are highly sensitive to emotional factors.