Student knowledge state evaluation method and device fusing kan and evidence theory, equipment and medium

Abstract

This application discloses a method, apparatus, device, and medium for assessing student knowledge status by integrating KAN and evidence theory, relating to the field of artificial intelligence technology. The method includes: constructing an enhanced feature representation that integrates question difficulty and discrimination; performing deep temporal modeling using a transformer encoder with an embedded memory decay operator; and employing KAN to map hidden state to beta-distributed evidence parameters, thereby achieving decoupled output of prediction probability and cognitive uncertainty. The beneficial effects include improving model parameter efficiency and interpretability, quantifying cognitive uncertainty to distinguish between true mastery and accidental guessing, enhancing the reliability of prediction results from the knowledge tracking model, and providing a credible assessment basis for smart education.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, and medium for assessing student knowledge status that integrates KAN and evidence theory. Background Technology

[0002] Current mainstream knowledge tracing methods are mainly divided into modeling methods based on learning behavior sequences and modeling methods based on graph structures. Among them, deep knowledge tracing, knowledge tracing based on self-attention mechanisms, and Transformer architecture-related models are the most widely used. These methods typically use recurrent neural networks or Transformer encoders to capture the dependencies in the answer sequence, use multilayer perceptrons as feature transformation and output components, and output single-point probability values ​​through Sigmoid or Softmax layers to predict students' knowledge mastery status.

[0003] Existing deep learning-based knowledge tracking models generally use multilayer perceptrons as the core feature mapping component. Their fully connected structure and fixed activation function result in a lack of mathematical interpretability in the internal feature transformation process, as well as high parameter redundancy and low computational efficiency. At the same time, the model only outputs single-point probability estimates, which cannot quantify the cognitive uncertainty of the prediction results and makes it difficult to distinguish between students' "true mastery of knowledge points" and "accidental correct guesses." In scenarios with sparse data and cold starts, unreliable predictions based on blind confidence are prone to occur. In addition, conventional attention mechanisms do not incorporate the forgetting curve of human memory and cannot reflect the actual cognitive process of students.

[0004] Therefore, improving the interpretability and parameter efficiency of knowledge tracing models, and achieving the quantification and reliable evaluation of the uncertainty in predicting students' knowledge status, has become an urgent problem to be solved. Summary of the Invention

[0005] The main objective of this application is to provide a method, apparatus, device, and medium for assessing student knowledge status that integrates KAN and evidence theory, aiming to solve the technical problem of how to improve the reliability of the prediction results of knowledge tracing models.

[0006] To achieve the above objectives, this application proposes a method for assessing student knowledge status that integrates KAN and evidence theory, including: Obtain students' historical answer time sequence data, wherein the historical answer time sequence data includes question identifiers, concept identifiers, answer results, and answer timestamps; An enhanced feature representation that integrates question difficulty and discrimination is constructed based on the historical answer time series data; The enhanced feature representation is input into a preset student knowledge state assessment model to obtain the student knowledge state assessment result, wherein the preset student knowledge state assessment model includes a Transformer encoder with an implanted memory decay operator and a KAN; The step of inputting the enhanced feature representation into a preset student knowledge status assessment model to obtain the student knowledge status assessment result includes: The enhanced feature representation is modeled in deep temporal sequence using a Transformer encoder to obtain a hidden state vector containing context dependencies. The hidden state vector is mapped to evidence parameters of a beta distribution using KAN, and the evidence parameters include cumulative evidence for correct answers and cumulative evidence for incorrect answers. Based on the cumulative evidence of correct answers and the cumulative evidence of incorrect answers, the expected prediction probability and cognitive uncertainty are calculated, and the student's knowledge status assessment results are output.

[0007] In one embodiment, the step of performing deep temporal modeling on the enhanced feature representation using a Transformer encoder to obtain a hidden state vector containing context dependencies includes: The relative position decay factor is calculated based on the time distance between the current moment and the historical moment. The relative position decay factor decays exponentially as the time distance increases. The enhanced feature representation is input into the multi-head self-attention sub-layer to calculate the attention score between the current query vector and the historical key-value vector; The attention score is multiplied element-wise by the relative position decay factor to obtain the weighted attention score; The attention output features are obtained by weighted summing of the value vectors at historical moments based on the weighted attention scores. The attention output features and enhanced feature representations are fused by residual connections and layer normalization to obtain the intermediate hidden layer state; The intermediate hidden layer state is input into a position-based feedforward neural network and subjected to nonlinear transformation. The hidden layer state vector is then obtained through residual connections and layer normalization.

[0008] In one embodiment, the step of calculating the relative position attenuation factor based on the time distance between the current time and historical times includes: Obtain the current time index and the historical time index, calculate the absolute value of the difference between the current time index and the historical time index, and obtain the time distance value; Obtain learnable decay rate parameters; Multiply the attenuation rate parameter by the time distance value to obtain the exponent; The relative position attenuation factor is obtained by calculating the value of an exponential function with the natural constant as the base and the opposite of the exponent.

[0009] In one embodiment, before the step of inputting the enhanced feature representation into a preset student knowledge state assessment model to obtain the student knowledge state assessment result, the following steps are included: Obtain a training sample set and construct an initial student knowledge status assessment model, wherein the training sample set includes multiple students' historical answer time series data and corresponding real answer result labels; The historical answer time series data in the training sample set is input into the initial student knowledge state assessment model, and the expected prediction probability, cumulative evidence of correct answers, and cumulative evidence of incorrect answers are obtained through forward calculation. The evidence loss value is calculated based on the expected predicted probability, cumulative evidence for correct answers, cumulative evidence for incorrect answers, and the actual answer label. The evidence loss value is composed of a weighted sum of the prediction error term and the variance regularization term. Data augmentation is performed on the historical time-series answer data in the training sample set to generate a first augmented view and a second augmented view; The first enhanced view and the second enhanced view are respectively input into the initial student knowledge state evaluation model to obtain the first hidden state vector and the second hidden state vector; The contrastive learning loss value is calculated based on the first hidden layer state vector and the second hidden layer state vector. The evidence loss value and the contrastive learning loss value are weighted and summed to obtain the joint loss value; The network parameters of the initial student knowledge state evaluation model are updated by backpropagation based on the joint loss value to obtain the optimized evaluation model. The current optimized evaluation model is output as the preset student knowledge state evaluation model when the maximum number of iterations is reached or the joint loss value is less than the preset loss threshold.

[0010] In one embodiment, the step of mapping the hidden state vector to a beta distribution evidence parameter using KAN includes: The hidden layer state vector is input into the linear layer of KAN, and a learnable spline basis function is applied to each feature dimension of the hidden layer state vector for nonlinear activation to obtain the spline basis function output. The spline activation output is obtained by weighted summation of the spline basis function outputs for each feature dimension. Aggregate the spline activation outputs of all feature dimensions to obtain the aggregated feature representation; The aggregated feature representation is subjected to non-negative constraint processing by a soft activation function to obtain non-negative evidence values; Add a Bayesian prior offset to the non-negative evidence value to obtain a two-dimensional evidence vector; The first dimension of the two-dimensional evidence vector is used as the cumulative evidence for correct answers, and the second dimension is used as the cumulative evidence for incorrect answers.

[0011] In one embodiment, the step of inputting the hidden layer state vector into a linear layer of a KAN, applying a learnable spline basis function to each feature dimension of the hidden layer state vector for nonlinear activation, and obtaining the spline basis function output includes: Determine the number of feature dimensions of the hidden layer state vector; For each feature dimension, determine the input value for each feature dimension; Initialize a preset number of B-spline basis functions, wherein the B-spline basis functions are piecewise polynomial functions, and each B-spline basis function is assigned learnable control coefficients; The input values ​​are substituted into each B-spline basis function for calculation to obtain the response values ​​of each basis function; The response values ​​of each basis function are weighted and summed according to the control coefficients to obtain the spline basis function output.

[0012] In one embodiment, the step of calculating the expected prediction probability and cognitive uncertainty based on the cumulative evidence of correct answers and the cumulative evidence of incorrect answers, and outputting the student's knowledge status assessment result includes: The total amount of evidence is obtained by summing the cumulative evidence of correct answers and the cumulative evidence of incorrect answers. Divide the cumulative evidence of correct answers by the total amount of evidence to obtain the expected prediction probability; The cognitive uncertainty is calculated based on the preset number of categories and the total amount of evidence, wherein the cognitive uncertainty is inversely proportional to the total amount of evidence. The predicted probability expectation is compared with a preset probability threshold. When the predicted probability expectation is greater than or equal to the preset probability threshold, the mastery result is determined to be that the student has mastered the corresponding knowledge point. When the predicted probability expectation is less than the preset probability threshold, the mastery result is determined to be that the student has not mastered the corresponding knowledge point. The predicted probability expectation, cognitive uncertainty, and mastery results are used as the assessment results of students' knowledge status.

[0013] Furthermore, to achieve the above objectives, this application also proposes a student knowledge status assessment device that integrates KAN and evidence theory, the student knowledge status assessment device integrating KAN and evidence theory comprising: The data acquisition module is used to acquire students' historical answer time series data, wherein the historical answer time series data includes question identifiers, concept identifiers, answer results, and answer timestamps; The feature construction module is used to construct an enhanced feature representation that integrates question difficulty and discrimination based on the historical answer time series data; The model evaluation module is used to input the enhanced feature representation into a preset student knowledge state evaluation model to obtain the student knowledge state evaluation result. The preset student knowledge state evaluation model includes a Transformer encoder with an embedded memory decay operator and a KAN. It is also used to perform deep temporal modeling on the enhanced feature representation using the Transformer encoder to obtain a hidden state vector containing context dependencies. The KAN is used to map the hidden state vector to beta-distributed evidence parameters, including cumulative evidence for correct answers and cumulative evidence for incorrect answers. Based on the cumulative evidence for correct answers and the cumulative evidence for incorrect answers, the expected prediction probability and cognitive uncertainty are calculated, and the student knowledge state evaluation result is output.

[0014] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the student knowledge state assessment method that integrates KAN and evidence theory as described above.

[0015] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the student knowledge state assessment method that integrates KAN and evidence theory as described above.

[0016] This application constructs an enhanced feature representation that integrates question difficulty and discrimination, utilizes a transformer encoder with an embedded memory decay operator for deep temporal modeling, and employs KAN to map hidden state to beta-distributed evidence parameters, thereby decoupling the output of prediction probability and cognitive uncertainty. The beneficial effects include improving model parameter efficiency and interpretability, quantifying cognitive uncertainty to distinguish between true mastery and accidental guessing, enhancing the reliability of the prediction results of the knowledge tracing model, and providing a credible evaluation basis for smart education. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the first embodiment of the student knowledge status assessment method that integrates KAN and evidence theory in this application. Figure 2This is a flowchart illustrating the second embodiment of the student knowledge status assessment method that integrates KAN and evidence theory in this application. Figure 3 This is a schematic diagram of the modular structure of the student knowledge status assessment device that integrates KAN and evidence theory in this application. Figure 4 This is a schematic diagram of the device structure of the hardware operating environment involved in the student knowledge status assessment method that integrates KAN and evidence theory in the embodiments of this application.

[0019] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0021] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0022] Current mainstream knowledge tracing methods are mainly divided into modeling methods based on learning behavior sequences and modeling methods based on graph structures. Among them, deep knowledge tracing, knowledge tracing based on self-attention mechanisms, and Transformer architecture-related models are the most widely used. These methods typically use recurrent neural networks or Transformer encoders to capture the dependencies in the answer sequence, and use multilayer perceptrons as feature transformation and output components. They output single-point probability values ​​through Sigmoid or Softmax layers to predict students' knowledge mastery status. Existing deep learning-based knowledge tracing models generally use multilayer perceptrons as the core feature mapping component. Its fully connected structure and fixed activation function lead to a lack of mathematical interpretability in the internal feature transformation process, as well as high parameter redundancy and low computational efficiency. At the same time, the model only outputs single-point probability estimates, which cannot quantify the cognitive uncertainty of the prediction results and makes it difficult to distinguish between students' "true mastery of knowledge points" and "accidental correct guesses." In scenarios with sparse data and cold starts, it is prone to unreliable predictions based on blind confidence. In addition, conventional attention mechanisms do not incorporate the forgetting curve of human memory and cannot reflect students' actual cognitive processes. Therefore, improving the interpretability and parameter efficiency of knowledge tracing models, and achieving the quantification and reliable evaluation of the uncertainty in predicting students' knowledge status, has become an urgent problem to be solved.

[0023] Based on the above, this application also provides a method for assessing student knowledge status that integrates KAN and evidence theory, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the student knowledge status assessment method that integrates KAN and evidence theory in this application.

[0024] In this embodiment, the student knowledge status assessment method integrating KAN and evidence theory includes steps S10~S30: Step S10: Obtain the student's historical answer sequence data.

[0025] It should be noted that historical answer sequence data is a collection of student answer-related information recorded in chronological order. This data fully preserves the sequential and temporal characteristics of student answering actions. Historical answer sequence data includes question identifiers, concept identifiers, answer results, and answer timestamps. The question identifier is a unique code used to distinguish each practice question, accurately locating the corresponding question content and avoiding confusion between different questions. The question identifier is a core index linking attributes such as question difficulty and discrimination. The concept identifier is a unique code used to identify the knowledge point corresponding to the question, clarifying the knowledge category to which the question belongs and reflecting the correspondence between the question and the knowledge point. The concept identifier helps distinguish answers to different knowledge points. The answer result is the student's feedback on the corresponding question, directly reflecting the student's mastery of the relevant knowledge points. Answer results are usually categorized as correct or incorrect, serving as a direct basis for judging the student's knowledge status. The answer timestamp records the time information when the student completed the corresponding question, clearly indicating the specific moment the answering action occurred. The timestamp for each response is used to calculate the time interval between responses.

[0026] Specifically, the system first retrieves historical answer sequence data from the database of the student online learning platform. Then, it cleans the historical answer sequence data, removing invalid records with missing timestamps or abnormal results. The cleaned records are then sorted according to the chronological order of their timestamps, resulting in a chronologically ordered sequence of student answers. Finally, the student answer sequence is truncated or padded according to a preset length to obtain a fixed-length historical answer sequence. This is done to ensure the reliability of the data input to the model and the correct temporal relationships, avoiding biases in knowledge state estimation due to missing data or disordered temporal sequences. Furthermore, the fixed sequence length facilitates batch parallel computation, improving the efficiency of model training and inference.

[0027] Step S20: Construct an enhanced feature representation that integrates question difficulty and discrimination based on historical answer time series data.

[0028] It should be noted that question identifiers, concept identifiers, and answer results are extracted from historical answer time series data and mapped to basic semantic embedding vectors, question difficulty parameter embedding vectors, and answer result interaction embedding vectors, respectively; the question discrimination parameter embedding vector corresponding to the question identifier is obtained; the question discrimination parameter embedding vector and the question difficulty parameter embedding vector are multiplied element-wise to obtain a weighted difficulty feature vector; the basic semantic embedding vector, the weighted difficulty feature vector, and the answer result interaction embedding vector are superimposed to obtain an enhanced feature representation.

[0029] Specifically, firstly, question identifiers, concept identifiers, and answer results are extracted from historical answer sequence data. The question identifiers are input into a pre-defined question embedding matrix for lookup mapping to obtain a basic semantic embedding vector. The concept identifiers are input into a pre-defined concept embedding matrix for lookup mapping to obtain a question difficulty parameter embedding vector. The answer results are converted into numerical labels and input into a pre-defined interaction embedding matrix for lookup mapping to obtain an answer result interaction embedding vector. Then, based on the question identifier, the corresponding question discrimination parameter embedding vector is retrieved from a pre-defined discrimination parameter library. The question discrimination parameter embedding vector is then multiplied element-wise with the question difficulty parameter embedding vector to obtain a weighted difficulty feature vector. This operation uses discrimination to weight and adjust the difficulty, essentially using the attribute of "whether the question can differentiate between students of different levels" to adjust the numerical strength of "how difficult the question is." Finally, the basic semantic embedding vector, the weighted difficulty feature vector, and the answer result interaction embedding vector are summed element-wise to obtain an enhanced feature representation. The reason for this is that simply using the embedding vector of the question identifier would lose the educational semantic information of the question itself. By explicitly introducing difficulty and discrimination parameters, the model has psychometric prior knowledge from the beginning. For example, a question with high discrimination and difficulty has a higher weighted difficulty feature vector value, which can better distinguish the knowledge mastery boundaries of high-achieving students and low-achieving students, so that subsequent time series modeling is based on more discriminative features.

[0030] Step S30: Input the enhanced feature representation into the preset student knowledge status assessment model to obtain the student knowledge status assessment result.

[0031] It should be noted that the preset student knowledge state assessment model is a deep learning model used to track and assess student knowledge states. After training, this model can stably perform feature processing and result output. The preset student knowledge state assessment model includes a Transformer encoder with embedded memory decay operators and KAN. The Transformer encoder with embedded memory decay operators is a Transformer coding structure with an added memory decay mechanism, used for deep modeling of temporal features. This encoder retains the advantages of parallel computing while conforming to the forgetting curve of human memory, improving the rationality of temporal feature extraction. KAN, or Kolmogorov-Arnold network, is a neural network structure that places learnable activation functions on the network connection edges. KAN has good mathematical interpretability and can efficiently complete complex nonlinear feature mapping.

[0032] Specifically, firstly, the enhanced feature representation is input into a pre-defined student knowledge state assessment model. A Transformer encoder is used to perform deep temporal modeling on the enhanced feature representation. An exponential decay operator based on temporal distance is introduced into the multi-head self-attention mechanism, giving higher attention weight to recent learning behaviors and automatically decaying the weight of long-term behaviors, thus simulating the forgetting curve of human memory and obtaining a hidden state vector containing contextual dependencies. Next, KAN is used to map the hidden state vector to beta-distributed evidence parameters. A refined nonlinear mapping of the features is performed using edge-learnable spline basis functions, replacing the black-box transformation of the traditional multilayer perceptron, to obtain cumulative evidence for correct answers and cumulative evidence for incorrect answers. Finally, the expected prediction probability and cognitive uncertainty are calculated based on the cumulative evidence for correct and incorrect answers. The expected prediction probability is compared with a pre-defined probability threshold to obtain the mastery judgment result. The expected prediction probability, cognitive uncertainty, and mastery judgment result are then output as the student knowledge state assessment result. The reason for this is to deeply integrate interpretable function networks with uncertainty quantification mechanisms, so that while maintaining high prediction accuracy, the model can explicitly distinguish between "true mastery" and "accidental guessing", providing a reliable risk reference for adaptive learning systems and avoiding blindly pushing difficult questions to students whose knowledge status is unclear.

[0033] Further, before step S30, the method includes: acquiring a training sample set and constructing an initial student knowledge state assessment model, wherein the training sample set includes multiple historical answer time series data of students and corresponding real answer result labels; inputting the historical answer time series data in the training sample set into the initial student knowledge state assessment model, and obtaining the expected prediction probability, cumulative evidence for correct answers, and cumulative evidence for incorrect answers through forward computation; calculating the evidence loss value based on the expected prediction probability, cumulative evidence for correct answers, cumulative evidence for incorrect answers, and real answer result labels, wherein the evidence loss value is composed of a weighted sum of the prediction error term and the variance regularization term; and performing data augmentation on the historical answer time series data in the training sample set. The system generates a first augmented view and a second augmented view. These are then input into the initial student knowledge state evaluation model to obtain a first hidden state vector and a second hidden state vector. The contrastive learning loss value is calculated based on the first and second hidden state vectors. The evidence loss value and the contrastive learning loss value are weighted and summed to obtain a joint loss value. The network parameters of the initial student knowledge state evaluation model are updated via backpropagation based on the joint loss value to obtain an optimized evaluation model. The optimized evaluation model is output as the preset student knowledge state evaluation model when the maximum number of iterations is reached or the joint loss value is less than a preset loss threshold.

[0034] Specifically, firstly, historical answer records are retrieved from the database of the student online learning platform. Valid records with complete answer timestamps and results are selected and grouped according to student identifiers. Each student's answer records are then sorted from earliest to latest by answer timestamp, resulting in multiple historical answer time series data points for each student. The answer result corresponding to the next moment in each historical answer time series data point is used as the label for the actual answer result, constructing a training sample set. Simultaneously, the network parameters of the transformer encoder, the learnable control coefficients of the Kolmogorov-Arnold network, and the decay rate parameter are initialized to construct an initial student knowledge state evaluation model. This is done to ensure the correct temporal relationship of the training data and that the model parameters are in a state of optimization, providing a foundation for subsequent iterative training.

[0035] Then, a batch of historical answer time-series data is extracted from the training sample set and input into the initial student knowledge state assessment model. The model then sequentially performs enhanced feature representation construction, time-series modeling of the transformer encoder with embedded memory decay operators, Kolmogorov-Arnold network evidence parameter mapping, and prediction probability expectation calculation. Forward propagation yields the prediction probability expectation, cumulative evidence for correct answers, and cumulative evidence for incorrect answers. This is done to obtain the prediction output under the current model parameters, providing a basis for subsequent loss calculation and parameter updates.

[0036] Next, the evidence loss is calculated based on the expected predicted probability, cumulative evidence for correct answers, cumulative evidence for incorrect answers, and the labels of the actual answers. Specifically, the mean squared error between the expected predicted probability and the labels of the actual answers is calculated to obtain the prediction error term. The product of the expected predicted probability and one minus the expected predicted probability is divided by the sum of the cumulative evidence for correct answers and the cumulative evidence for incorrect answers, plus one, to obtain the variance regularization term. The prediction error term and the variance regularization term are then added to obtain the evidence loss value. The reason for this is that the prediction error term ensures the accuracy of the model's predictions, while the variance regularization term suppresses the model's overconfidence when evidence is insufficient, i.e., when the sum of the cumulative evidence for correct answers and the cumulative evidence for incorrect answers is minimized, thus avoiding unreliable high-confidence predictions in sparse data scenarios.

[0037] Subsequently, two different data augmentation operations were performed on the same batch of historical answer time-series data. The first data augmentation used a random timestamp perturbation and question masking strategy to generate a first augmented view. The second data augmentation used a different random seed and executed the same strategy to generate a second augmented view. The first and second augmented views were then input into the initial student knowledge state assessment model, respectively. The first and second hidden state vectors were obtained through forward computation. The cosine similarity loss between the first and second hidden state vectors was calculated to obtain the contrastive learning loss value. The reason for this approach is to enhance the feature consistency under different data augmented views, enabling the model to learn a stable representation robust to input perturbations and improving feature generalization ability.

[0038] Next, the evidence loss value and the contrastive learning loss value are weighted and summed according to preset weight coefficients to obtain the joint loss value. The gradient of the joint loss value with respect to each network parameter is calculated using the backpropagation algorithm. A preset optimizer is then used to update the network parameters of the initial student knowledge state evaluation model according to a preset learning rate, resulting in an optimized evaluation model. The reason for this is to jointly optimize prediction accuracy and feature consistency, enabling the model to have stable feature extraction capabilities while making accurate predictions.

[0039] Finally, it is determined whether the current iteration count has reached the preset maximum iteration count, or whether the joint loss value is less than the preset loss threshold. If the current iteration count has reached the preset maximum iteration count or the joint loss value is less than the preset loss threshold, the iteration stops and the current optimized evaluation model is output as the preset student knowledge state evaluation model; otherwise, the next batch of data is drawn from the training sample set and the above training process is repeated. The reason for this is to ensure that the model fully converges to the optimal parameter state, while avoiding overfitting caused by overtraining, so that the final model has good generalization performance.

[0040] This embodiment constructs an enhanced feature representation that integrates question difficulty and discrimination, utilizes a transformer encoder with an embedded memory decay operator for deep temporal modeling, and employs KAN to map hidden state to beta-distributed evidence parameters, achieving decoupled output of prediction probability and cognitive uncertainty. The beneficial effects include improved model parameter efficiency and interpretability, quantification of cognitive uncertainty to distinguish between true mastery and accidental guessing, and enhanced reliability of the knowledge tracing model's prediction results, providing a credible evaluation basis for smart education.

[0041] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 The student knowledge status assessment method integrating KAN and evidence theory, step S30, further includes steps S201-S203: Step S201: Deep temporal modeling of the enhanced feature representation is performed using a Transformer encoder to obtain a hidden state vector containing context dependencies.

[0042] It should be noted that step S201 includes: First, calculating the relative position decay factor based on the time distance between the current time and the historical time. Specifically, this involves obtaining the current time index and the historical time index, calculating the absolute value of the difference between the current time index and the historical time index to obtain the time distance value; obtaining a learnable decay rate parameter; multiplying the decay rate parameter by the time distance value to obtain the exponent; and calculating the value of an exponential function with the natural constant as the base and the negative of the exponent as the exponent to obtain the relative position decay factor, which decays exponentially with increasing time distance. Specifically, first, obtaining the current time index and the historical time index, calculating the absolute value of the difference between the current time index and the historical time index to obtain the time distance value, which represents the number of time steps between the current time and the historical time. For example, if the current time is the 10th time step and the historical time is the 3rd time step, then the time distance value is 7. Then, obtaining the learnable decay rate parameter from the model parameters, which is a scalar value automatically optimized during model training to control the rate of memory decay. Next, the attenuation rate parameter is multiplied by the time distance value to obtain an exponent, which is positive and proportional to the time distance; the larger the time distance, the larger the exponent. Finally, the relative position attenuation factor is obtained by calculating the exponential function value with the natural constant as the base and the negative of the exponent as the exponent. The specific formula is as follows: in The elements of the attenuation matrix represent the relative position attenuation factors. γ For learnable decay rate, Represents the current moment i With historical moments j The time distance is considered as the relative position decay factor. The greater the distance, the larger the value. The value of this decay factor ranges from 0 to 1. When the time distance is 0, the decay factor is 1, indicating complete retention. As the time distance increases, the decay factor decreases exponentially, tending towards 0, indicating gradual forgetting. This is done so that the forgetting rate can be adaptively adjusted through learnable parameters, allowing the model to flexibly adjust the degree of memory decay according to the characteristics of different knowledge points and individual student differences, avoiding the rigidity of modeling caused by a fixed decay rate.

[0043] Secondly, the enhanced feature representation is input into the multi-head self-attention sub-layer to calculate the attention score between the current query vector and the key-value vectors from previous time steps. Specifically, the enhanced feature representation is input into the multi-head self-attention sub-layer, and a linear transformation is used to map the enhanced feature representation into a query vector matrix, a key vector matrix, and a value vector matrix. The transposes of the query vector matrix and the key vector matrix are then multiplied, and the result is scaled by dividing by the square root of the feature dimension to obtain the attention score between the current query vector and the key-value vectors from previous time steps. This is done to calculate the association strength between the current and previous time steps, providing a foundation for subsequent weighted aggregation.

[0044] The attention score is obtained by element-wise multiplying the attention score with the relative position decay factor. Then, the value vectors of historical moments are weighted and summed based on the weighted attention score to obtain the attention output feature. Specifically, the weighted attention score is normalized to obtain the attention weight distribution, and the value vectors of historical moments are weighted and summed based on the attention weight distribution to obtain the attention output feature. The reason for this is to incorporate the memory decay mechanism into the attention calculation, so that the model considers both relevance and time proximity. For example, a question answered correctly a month ago has a higher weight than a question answered incorrectly yesterday, thus more accurately reflecting the current state of knowledge.

[0045] Finally, the attention output features and enhanced feature representations are fused through residual connections and layer normalization to obtain intermediate hidden state. The intermediate hidden state is then input into a location-based feedforward neural network for nonlinear transformation, and the hidden state vector is obtained through residual connections and layer normalization. Specifically, the attention output features and enhanced feature representations are added element-wise through residual connections, and the result is then distributed and stabilized through layer normalization to obtain the intermediate hidden state. This intermediate hidden state is then input into the location-based feedforward neural network, where linear transformation, nonlinear activation, and a second linear transformation are performed sequentially to obtain feedforward output features. These feedforward output features are then added element-wise to the intermediate hidden state through residual connections, and the result is distributed and stabilized through layer normalization to obtain the hidden state vector. The specific formula is as follows: in This represents the hidden state vector. Representation layer normalization, This represents a feedforward neural network. Indicates the intermediate hidden state; This indicates the self-attention mechanism. This represents the attention output features. This represents an enhanced feature representation; Represents a non-linear activation function. This represents the weight matrix of the first layer of the feedforward neural network. This represents the bias vector of the first layer of the feedforward neural network. This represents the weight matrix of the second layer of the feedforward neural network. This represents the bias vector of the second layer of the feedforward neural network. The reason for this is that residual connections alleviate the vanishing gradient problem in deep networks, ensuring effective information transfer; layer normalization stabilizes the input distribution of each layer; and the feedforward neural network independently performs nonlinear feature transformations at each position. These three elements work together to enable the model to extract deep semantic features while maintaining training stability.

[0046] Step S202: Use KAN to map the hidden state vector to evidence parameters of a beta distribution. The evidence parameters include cumulative evidence for correct answers and cumulative evidence for incorrect answers.

[0047] It should be noted that step S202 includes: firstly, inputting the hidden layer state vector into the linear layer of the KAN; applying learnable spline basis functions for nonlinear activation to each feature dimension of the hidden layer state vector to obtain the spline basis function output. It is important to understand that: the number of feature dimensions of the hidden layer state vector is determined; for each feature dimension, the input value for each feature dimension is determined; a preset number of B-spline basis functions are initialized, where the B-spline basis functions are piecewise polynomial functions, and each B-spline basis function is assigned learnable control coefficients; the input values ​​are substituted into each B-spline basis function for calculation to obtain the response value of each basis function; and the response values ​​of each basis function are weighted and summed according to the control coefficients to obtain the spline basis function output. Specifically, firstly, the dimension configuration information of the hidden layer state vector is obtained, and the number of feature dimensions of the hidden layer state vector is determined. This number of feature dimensions is determined by the output dimension setting of the transformer encoder, for example, a preset dimension value of 256 or 512 dimensions. Then, for each feature dimension, the corresponding value of that dimension is extracted from the hidden layer state vector as the input value, which is a scalar floating-point number. Next, a preset number (5 to 20) of B-spline basis functions are initialized. These B-spline basis functions are piecewise polynomial functions, each composed of multiple low-order polynomials concatenated within its domain, maintaining continuity and smoothness between adjacent segments. Learnable control coefficients are assigned to each B-spline basis function; these control coefficients are floating-point parameters automatically optimized through backpropagation during model training. Subsequently, the input value is substituted into each B-spline basis function for calculation, obtaining the response value of each basis function. Specifically, based on the position of the input value within the domain of the B-spline basis function, the segmented interval into which it falls is determined, and then the polynomial coefficients corresponding to that interval are used to calculate the output value. Finally, the response values ​​of each basis function are weighted and summed according to the control coefficients to obtain the spline basis function output, which is the activation value of each feature dimension after independent nonlinear transformation. The specific formula is expressed as follows: in This represents the total number of preset quantities. This indicates the number of B-spline basis functions currently in use. Indicates the first The first dimension The learnable control coefficients corresponding to each basis function Indicates the first Each B-spline basis function is a set of piecewise polynomials responsible for capturing local nonlinear features. This is because B-spline basis functions possess local support properties; each basis function produces a significant response only within a small range. This ensures that small changes in the input value only affect the output in a local area, thus achieving refined nonlinear fitting. Simultaneously, the learnability of the control coefficients allows the model to adaptively adjust the weight combination of each basis function, accurately characterizing the complex nonlinear boundary between knowledge mastery and test-taking performance.

[0048] Secondly, the spline basis function outputs for each feature dimension are weighted and summed to obtain the spline activation output. Then, the spline activation outputs for all feature dimensions are aggregated to obtain the aggregated feature representation, expressed by the following formula: in Indicates the input dimension. This represents the dimension of the hidden state vector. Indicates the first j The spline basis functions output results for each input dimension. The reason for this is to integrate the independent nonlinear mapping results of each dimension into a unified feature representation, providing a comprehensive feature foundation for subsequent evidence parameter prediction.

[0049] Finally, a non-negativity constraint is applied to the aggregated feature representation using a soft activation function to obtain non-negative evidence values. A Bayesian prior offset is then added to these non-negative evidence values ​​to obtain a two-dimensional evidence vector. Specifically, a non-linear transformation is performed on the aggregated feature representation using a soft activation function to ensure that the output value is always non-negative, resulting in non-negative evidence values. Then, a Bayesian prior offset is added to these non-negative evidence values ​​to obtain a two-dimensional evidence vector, expressed by the following formula: in This represents a soft activation function. This represents the hidden state vector. The accumulated evidence representing students' correct answers The cumulative evidence representing incorrect answers is represented by a Bayesian prior bias of 1, ensuring that each component of the two-dimensional evidence vector satisfies a validity constraint greater than or equal to 1. This is done to prevent the beta distribution definition from becoming invalid due to zero or negative values ​​in the evidence parameters. The prior bias is equivalent to introducing a uniform prior when observation data is insufficient, ensuring that the model can still output reasonable probability estimates in abnormal scenarios. The first dimension of the two-dimensional evidence vector is used as the cumulative evidence for correct answers, and the second dimension as the cumulative evidence for incorrect answers. This is done to explicitly decouple the aggregated features into two opposing evidence components: the cumulative evidence for correct answers represents the total amount of historical evidence observed by the model supporting the student's correct answer, and the cumulative evidence for incorrect answers represents the total amount of historical evidence supporting the student's incorrect answer. Together, they determine the shape of the beta distribution, thereby achieving a quantitative separation between prediction probability and uncertainty.

[0050] Step S203: Calculate the expected prediction probability and cognitive uncertainty based on the cumulative evidence of correct answers and the cumulative evidence of incorrect answers, and output the student's knowledge status assessment results.

[0051] It should be noted that step S203 includes: summing the cumulative evidence of correct answers and the cumulative evidence of incorrect answers to obtain the total amount of evidence; dividing the cumulative evidence of correct answers by the total amount of evidence to obtain the expected predicted probability; calculating cognitive uncertainty based on the preset number of categories and the total amount of evidence, wherein cognitive uncertainty is inversely proportional to the total amount of evidence; comparing the expected predicted probability with a preset probability threshold, and determining that the student has mastered the corresponding knowledge point when the expected predicted probability is greater than or equal to the preset probability threshold, and determining that the student has not mastered the corresponding knowledge point when the expected predicted probability is less than the preset probability threshold; and using the expected predicted probability, cognitive uncertainty, and mastery result as the student's knowledge status assessment result.

[0052] Specifically, the cumulative evidence of correct answers and the cumulative evidence of incorrect answers are first added element by element to obtain the total amount of evidence. The total amount of evidence represents the total amount of comprehensive observational evidence the model uses to predict current student answers. A larger sum of cumulative evidence for correct answers and cumulative evidence for incorrect answers indicates more comprehensive information the model possesses. Then, dividing the cumulative evidence for correct answers by the total amount of evidence yields the expected prediction probability, as shown in the formula: in Represents the expected probability of prediction. This indicates that the cumulative evidence is correct. This represents the total amount of evidence. The expected value of this predicted probability ranges from 0 to 1. The closer the value is to 1, the more the model tends to predict that the student will answer correctly; the closer the value is to 0, the more the model tends to predict that the student will answer incorrectly. The reason for this is to use the expectation formula of the beta distribution to transform the two opposing evidence components into an intuitive probability estimate, which is consistent with the understanding of "probability of mastery" in the educational context.

[0053] Next, cognitive uncertainty is calculated based on the preset number of classification categories (2 in this embodiment) and the total amount of evidence. The specific formula is as follows: in This indicates cognitive uncertainty. Indicates the total amount of evidence. The number of preset classification categories is represented by the formula. Cognitive uncertainty is inversely proportional to the total amount of evidence; the larger the total amount of evidence, the smaller the cognitive uncertainty, and vice versa. This is because when the model observes more consistent historical data, the sum of accumulated evidence for correct and incorrect answers increases, reducing cognitive uncertainty and indicating higher confidence in the prediction results. Conversely, in scenarios with sparse data or cold starts, the total amount of evidence is small, and cognitive uncertainty remains high, thus identifying samples the model is unsure about and avoiding overconfidence. Subsequently, the expected predicted probability is compared with a preset probability threshold (e.g., 0.5). If the expected predicted probability is greater than or equal to the preset probability threshold, the result is determined as the student mastering the corresponding knowledge point; if the expected predicted probability is less than the preset probability threshold, the result is determined as the student not mastering the corresponding knowledge point. This is done by discretizing the continuous probability value into a binary classification decision, making it easier for the education system to directly determine whether a student needs reinforcement learning for that knowledge point. Finally, the expected predicted probability, cognitive uncertainty, and mastery results are combined and encapsulated as the student's knowledge status assessment result. This assessment result simultaneously includes the predicted value, confidence level, and decision conclusion. This is done because traditional models only output a single probability value, failing to distinguish between "stable predictions after full mastery" and "random guesses due to a lack of information." By decoupling probability and uncertainty in the output, the education system can identify students in a state of confusion, thus recommending more basic exercises or intervening with human tutoring, rather than blindly recommending highly difficult problems. This provides a reliable risk reference for adaptive learning.

[0054] This embodiment uses enhanced feature representations as input and employs a Transformer encoder with an embedded memory decay operator to deeply model temporal dependencies, obtaining hidden state vectors. Then, it utilizes KAN mapping to accumulate evidence of correct and incorrect answers in a beta distribution. Finally, based on the evidence parameters, it calculates the expected prediction probability and cognitive uncertainty, outputting the student's knowledge state assessment result. This approach improves model interpretability and parameter efficiency, quantifies the uncertainty of knowledge state prediction, and makes the assessment results more reliable.

[0055] Based on the first embodiment of this application, this application also provides a student knowledge status assessment device that integrates KAN and evidence theory. Please refer to... Figure 3 The device includes: The data acquisition module 10 is used to acquire students' historical answer time sequence data, which includes question identifiers, concept identifiers, answer results, and answer timestamps.

[0056] Feature construction module 20 is used to construct an enhanced feature representation that integrates question difficulty and discrimination based on historical answer time series data.

[0057] The model evaluation module 30 is used to input the enhanced feature representation into a preset student knowledge state evaluation model to obtain the student knowledge state evaluation result. The preset student knowledge state evaluation model includes a Transformer encoder with an embedded memory decay operator and a KAN. It is also used to perform deep temporal modeling on the enhanced feature representation through the Transformer encoder to obtain a hidden state vector containing context dependencies. The hidden state vector is mapped to evidence parameters of a beta distribution using KAN. The evidence parameters include cumulative evidence for correct answers and cumulative evidence for incorrect answers. The expected prediction probability and cognitive uncertainty are calculated based on the cumulative evidence for correct answers and cumulative evidence for incorrect answers, and the student knowledge state evaluation result is output.

[0058] The student knowledge status assessment device integrating KAN and evidence theory provided in this application, employing the student knowledge status assessment method integrating KAN and evidence theory in the above embodiments, can solve the technical problem of how to improve the reliability of the prediction results of the knowledge tracking model. Compared with the prior art, the beneficial effects of the student knowledge status assessment device integrating KAN and evidence theory provided in this application are the same as the beneficial effects of the student knowledge status assessment method integrating KAN and evidence theory provided in the above embodiments, and other technical features in the student knowledge status assessment device integrating KAN and evidence theory are the same as the features disclosed in the methods of the above embodiments, and will not be repeated here.

[0059] This application provides a student knowledge status assessment device that integrates KAN and evidence theory. The student knowledge status assessment device integrating KAN and evidence theory includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the student knowledge status assessment method integrating KAN and evidence theory in the above embodiment 1.

[0060] The following is for reference. Figure 4This document illustrates a structural schematic diagram of a student knowledge state assessment device suitable for implementing embodiments of this application, which integrates KAN and evidence theory. The student knowledge state assessment device integrating KAN and evidence theory in embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 4 The student knowledge status assessment device shown, which integrates KAN and evidence theory, is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0061] like Figure 4 As shown, the student knowledge status assessment device integrating KAN and evidence theory may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 1002 or a program loaded from storage device 1003 into random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the student knowledge status assessment device integrating KAN and evidence theory. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the student knowledge state assessment device incorporating KAN and evidence theory to wirelessly or wiredly communicate with other devices to exchange data. Although various student knowledge state assessment devices incorporating KAN and evidence theory are shown in the figures, it should be understood that it is not required to implement or possess all of them. More or fewer may be implemented alternatively.

[0062] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0063] The student knowledge status assessment device integrating KAN and evidence theory provided in this application, employing the student knowledge status assessment method integrating KAN and evidence theory in the above embodiments, can solve the technical problem of how to improve the reliability of the prediction results of the knowledge tracking model. Compared with the prior art, the beneficial effects of the student knowledge status assessment device integrating KAN and evidence theory provided in this application are the same as the beneficial effects of the student knowledge status assessment method integrating KAN and evidence theory provided in the above embodiments, and other technical features in this student knowledge status assessment device integrating KAN and evidence theory are the same as those disclosed in the method of the previous embodiment, and will not be repeated here.

[0064] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0065] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0066] This application provides a computer-readable medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the student knowledge state assessment method that integrates KAN and evidence theory in the above embodiments.

[0067] The computer-readable medium provided in this application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor devices, or any combination thereof. More specific examples of computer-readable media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable medium may be any tangible medium containing or storing a program that can be executed by instructions, used by a device, or used in conjunction with it. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0068] The aforementioned computer-readable medium may be included in a student knowledge status assessment device that integrates KAN and evidence theory; or it may exist independently and not be incorporated into a student knowledge status assessment device that integrates KAN and evidence theory.

[0069] The aforementioned computer-readable medium carries one or more programs that, when executed by a student knowledge state assessment device incorporating KAN and evidence theory, enable the device to write computer program code for performing the operations of this application in one or more programming languages ​​or combinations thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0070] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, all blocks in the flowcharts or block diagrams may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that all blocks in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using dedicated hardware-based implementations that perform the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0071] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0072] The readable medium provided in this application is a computer-readable medium that stores computer-readable program instructions (i.e., a computer program) for executing the aforementioned student knowledge state assessment method that integrates KAN and evidence theory. This addresses the technical problem of improving the reliability of prediction results from knowledge tracing models. Compared to existing technologies, the beneficial effects of the computer-readable medium provided in this application are the same as those of the student knowledge state assessment method integrating KAN and evidence theory provided in the above embodiments, and will not be elaborated upon here.

[0073] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the student knowledge status assessment method that integrates KAN and evidence theory as described above.

[0074] The computer program product provided in this application solves the technical problem of how to improve the reliability of the prediction results of knowledge tracing models. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the student knowledge state assessment method that integrates KAN and evidence theory provided in the above embodiments, and will not be repeated here.

[0075] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for assessing students' knowledge status by integrating KAN and evidence theory, characterized in that, The method includes: Obtain students' historical answer time sequence data, wherein the historical answer time sequence data includes question identifiers, concept identifiers, answer results, and answer timestamps; An enhanced feature representation that integrates question difficulty and discrimination is constructed based on the historical answer time series data; The enhanced feature representation is input into a preset student knowledge state evaluation model to obtain the student knowledge state evaluation result, wherein the preset student knowledge state evaluation model includes a Transformer encoder with an implanted memory decay operator and a KAN. The step of inputting the enhanced feature representation into a preset student knowledge status assessment model to obtain the student knowledge status assessment result includes: The enhanced feature representation is modeled in deep temporal sequence using a Transformer encoder to obtain a hidden state vector containing context dependencies. The hidden state vector is mapped to evidence parameters of a beta distribution using KAN, and the evidence parameters include cumulative evidence for correct answers and cumulative evidence for incorrect answers. Based on the cumulative evidence of correct answers and the cumulative evidence of incorrect answers, the expected prediction probability and cognitive uncertainty are calculated, and the student's knowledge status assessment results are output.

2. The method as described in claim 1, characterized in that, The step of performing deep temporal modeling on the enhanced feature representation using a Transformer encoder to obtain a hidden state vector containing context dependencies includes: The relative position decay factor is calculated based on the time distance between the current moment and the historical moment. The relative position decay factor decays exponentially as the time distance increases. The enhanced feature representation is input into the multi-head self-attention sub-layer to calculate the attention score between the current query vector and the historical key-value vector; The attention score is multiplied element-wise by the relative position decay factor to obtain the weighted attention score; The attention output features are obtained by weighted summing of the value vectors at historical moments based on the weighted attention scores. The attention output features and enhanced feature representations are fused by residual connections and layer normalization to obtain the intermediate hidden layer state; The intermediate hidden layer state is input into a position-based feedforward neural network and subjected to nonlinear transformation. The hidden layer state vector is then obtained through residual connections and layer normalization.

3. The method as described in claim 2, characterized in that, The step of calculating the relative position attenuation factor based on the time distance between the current moment and historical moments includes: Obtain the current time index and the historical time index, calculate the absolute value of the difference between the current time index and the historical time index, and obtain the time distance value; Obtain learnable decay rate parameters; Multiply the attenuation rate parameter by the time distance value to obtain the exponent; The relative position attenuation factor is obtained by calculating the value of an exponential function with the natural constant as the base and the opposite of the exponent.

4. The method as described in claim 1, characterized in that, Before the step of inputting the enhanced feature representation into the preset student knowledge status assessment model to obtain the student knowledge status assessment result, the following steps are included: Obtain a training sample set and construct an initial student knowledge status assessment model, wherein the training sample set includes multiple students' historical answer time series data and corresponding real answer result labels; The historical answer time series data in the training sample set is input into the initial student knowledge state assessment model, and the expected prediction probability, cumulative evidence of correct answers, and cumulative evidence of incorrect answers are obtained through forward calculation. The evidence loss value is calculated based on the expected predicted probability, cumulative evidence for correct answers, cumulative evidence for incorrect answers, and the actual answer label. The evidence loss value is composed of a weighted sum of the prediction error term and the variance regularization term. Data augmentation is performed on the historical time-series answer data in the training sample set to generate a first augmented view and a second augmented view; The first enhanced view and the second enhanced view are respectively input into the initial student knowledge state evaluation model to obtain the first hidden state vector and the second hidden state vector; The contrastive learning loss value is calculated based on the first hidden layer state vector and the second hidden layer state vector. The evidence loss value and the contrastive learning loss value are weighted and summed to obtain the joint loss value; The network parameters of the initial student knowledge state evaluation model are updated by backpropagation based on the joint loss value to obtain the optimized evaluation model. The current optimized evaluation model is output as the preset student knowledge state evaluation model when the maximum number of iterations is reached or the joint loss value is less than the preset loss threshold.

5. The method as described in claim 1, characterized in that, The step of mapping the hidden state vector to a beta distribution evidence parameter using KAN includes: The hidden layer state vector is input into the linear layer of KAN, and a learnable spline basis function is applied to each feature dimension of the hidden layer state vector for nonlinear activation to obtain the spline basis function output. The spline activation output is obtained by weighted summation of the spline basis function outputs for each feature dimension. Aggregate the spline activation outputs of all feature dimensions to obtain the aggregated feature representation; The aggregated feature representation is subjected to non-negative constraint processing by a soft activation function to obtain non-negative evidence values; Add a Bayesian prior offset to the non-negative evidence value to obtain a two-dimensional evidence vector; The first dimension of the two-dimensional evidence vector is used as the cumulative evidence for correct answers, and the second dimension is used as the cumulative evidence for incorrect answers.

6. The method as described in claim 5, characterized in that, The steps of inputting the hidden layer state vector into the linear layer of the KAN, applying a learnable spline basis function to each feature dimension of the hidden layer state vector for nonlinear activation, and obtaining the spline basis function output include: Determine the number of feature dimensions of the hidden layer state vector; For each feature dimension, determine the input value for each feature dimension; Initialize a preset number of B-spline basis functions, wherein the B-spline basis functions are piecewise polynomial functions, and each B-spline basis function is assigned learnable control coefficients; The input values ​​are substituted into each B-spline basis function for calculation to obtain the response values ​​of each basis function; The response values ​​of each basis function are weighted and summed according to the control coefficients to obtain the spline basis function output.

7. The method as described in claim 1, characterized in that, The step of calculating the expected prediction probability and cognitive uncertainty based on the cumulative evidence of correct answers and the cumulative evidence of incorrect answers, and outputting the student's knowledge status assessment results, includes: The total amount of evidence is obtained by summing the cumulative evidence of correct answers and the cumulative evidence of incorrect answers. Divide the cumulative evidence of correct answers by the total amount of evidence to obtain the expected prediction probability; The cognitive uncertainty is calculated based on the preset number of categories and the total amount of evidence, wherein the cognitive uncertainty is inversely proportional to the total amount of evidence. The predicted probability expectation is compared with a preset probability threshold. When the predicted probability expectation is greater than or equal to the preset probability threshold, the mastery result is determined to be that the student has mastered the corresponding knowledge point. When the predicted probability expectation is less than the preset probability threshold, the mastery result is determined to be that the student has not mastered the corresponding knowledge point. The predicted probability expectation, cognitive uncertainty, and mastery results are used as the assessment results of students' knowledge status.

8. A student knowledge status assessment device integrating KAN and evidence theory, characterized in that, The device includes: The data acquisition module is used to acquire students' historical answer time series data, wherein the historical answer time series data includes question identifiers, concept identifiers, answer results, and answer timestamps; The feature construction module is used to construct an enhanced feature representation that integrates question difficulty and discrimination based on the historical answer time series data; The model evaluation module is used to input the enhanced feature representation into a preset student knowledge state evaluation model to obtain the student knowledge state evaluation result. The preset student knowledge state evaluation model includes a Transformer encoder with an embedded memory decay operator and a KAN. It is also used to perform deep temporal modeling on the enhanced feature representation using the Transformer encoder to obtain a hidden state vector containing context dependencies. The KAN is used to map the hidden state vector to beta-distributed evidence parameters, including cumulative evidence for correct answers and cumulative evidence for incorrect answers. Based on the cumulative evidence for correct answers and the cumulative evidence for incorrect answers, the expected prediction probability and cognitive uncertainty are calculated, and the student knowledge state evaluation result is output.

9. A student knowledge status assessment device integrating KAN and evidence theory, characterized in that, The device includes: a memory, a processor, and a student knowledge status assessment program that integrates KAN and evidence theory, stored on the memory and running on the processor, the student knowledge status assessment program that integrates KAN and evidence theory is configured to implement the steps of the student knowledge status assessment method that integrates KAN and evidence theory as described in any one of claims 1-7.

10. A storage medium, characterized in that, The storage medium stores a student knowledge status assessment program that integrates KAN and evidence theory. When the student knowledge status assessment program that integrates KAN and evidence theory is executed by the processor, it implements the steps of the student knowledge status assessment method that integrates KAN and evidence theory as described in any one of claims 1-7.

Images