Explanatable knowledge tracking method, online learning platform and electronic equipment

Abstract

The invention discloses an interpretable knowledge tracking method, an online learning platform and electronic equipment, and belongs to the technical field of data mining and interpretable artificial intelligence, the method comprises the steps of question definition and data preprocessing, question representation, knowledge acquisition, knowledge application, knowledge updating and model training, vectorization processing is carried out on preprocessed data, and the vectorization processing is carried out on the preprocessed data; the knowledge states of the learner at different moments are fused through a multi-layer perceptron, an increment index is constructed, a loss value is minimized, the knowledge states of the learner are dynamically updated, and a personalized teaching strategy is generated. According to the method, the process of solving the specific problem by the learner is deconstructed, the knowledge acquisition degree of the learner is modeled before the problem is solved, the knowledge application degree of the learner is calculated when the problem is solved, and the knowledge updating degree of the learner is measured after the problem is solved, so that the interpretability of the problem solving process is realized; and an educator and a learner can master own teaching and learning states at the first time.

Description

Technical Field

[0001] The present invention relates to a knowledge tracking method, an online learning platform and an electronic device, and in particular to an explainable knowledge tracking method, an online learning platform and an electronic device, belonging to the technical field of data mining and explainable artificial intelligence. Background Art

[0002] There is a knowledge reversal phenomenon in existing educational technologies: teaching procedures that are effective for beginners are ineffective for experts. Therefore, educators must be able to continuously evaluate and monitor learners' knowledge mastery during the learning process to dynamically determine appropriate teaching resources, especially in online learning scenarios. Knowledge tracking is an influential research field in distance education. It can automatically track students' knowledge levels at different stages and has been widely used in adaptive learning systems.

[0003] In recent years, some researchers have begun to pay attention to the interpretability of knowledge tracking methods. They have introduced item response theory, adopted attention mechanisms, or designed educationally meaningful learnable parameters to enhance the interpretability of the model, thereby better explaining students' learning performance. However, these methods mainly focus on performance prediction and ignore the learning process, resulting in unreasonable evaluation of knowledge mastery patterns. From a cognitive perspective, learners' knowledge mastery curves are relatively stable and will not fluctuate dramatically in a short period of time.

[0004] Practice is a typical problem-solving process. Students enhance their understanding of problems through deep thinking, constantly explore solutions, and ultimately achieve knowledge consolidation. The entire problem-solving process involves the acquisition, application, internalization, and forgetting of memorized knowledge in the brain. However, existing knowledge tracking methods ignore the process of learners solving problems, equating memorized knowledge with acquired or applied knowledge, resulting in drastic fluctuations in the state of knowledge. According to problem-solving theory, knowledge acquisition involves learners organizing implicit and explicit knowledge related to the problem, while knowledge application involves applying this knowledge to seek solutions. Remembering does not mean applying, and being familiar with discrete knowledge concepts does not mean that students can effectively organize them or convert abstract knowledge into concrete knowledge or skills when answering questions - as the saying goes, knowing the way and doing the way are two different things, which reflects this phenomenon.

[0005] In order to improve the irrationality of knowledge status assessment, enhance the interpretability of the model and improve the accuracy of prediction, it is a new idea to reconstruct the students' learning process from the perspective of human problem solving. It explores an explainable knowledge tracking method for the problem-solving process. By using deep learning technology, it deeply analyzes the various stages of learners' problem solving, and analyzes in detail how learners understand problems, acquire knowledge, apply knowledge, internalize knowledge, and forget knowledge, thereby reproducing their learning process. It can more accurately diagnose and attribute learners' knowledge status, which has important research significance and broad application value. Summary of the invention

[0006] The purpose of the present invention is to provide an explainable knowledge tracing method, an online learning platform and an electronic device, which not only provides a knowledge tracing method, but also enhances the explainability of the knowledge tracing method in view of the phenomenon that the learner's problem-solving process is ignored in the prior art, thereby solving the shortcomings of the prior art.

[0007] The present invention provides the following scheme:

[0008] A problem-solving process-oriented interpretable knowledge tracing method is applied to a problem-solving process-oriented interpretable knowledge tracing system, including:

[0009] Before solving the problem: obtaining the learner's practice records, performing data preprocessing on the practice records, and obtaining corresponding preprocessed data;

[0010] The preprocessed data is vectorized to obtain the problem embedding representation vector P n , combined with the knowledge concepts, difficulty and discrimination in the practice records, to obtain the enhanced question embedding representation vector

[0011] When solving problems: The learner’s previous knowledge state H is integrated through a multi-layer perceptron n-1 and the enhanced problem embedding representation vector Model the learner's personalized problem space, which includes the initial state SP of the problem n , target state EP n and the adaptation operator OP n , and use activation functions and gating networks to calculate the knowledge acquisition degree vector

[0012] Use a multi-layer perceptron and activation function to generate a corresponding solution from the adaptation operator of the personalized problem space to obtain a knowledge application degree vector The probability of learners answering questions correctly is calculated through the four-parameter item response theory, and the learners' responses are attributed.

[0013] Based on the knowledge acquisition degree vector and knowledge application degree vector Construct incremental indicators to quantify the learner's knowledge internalization degree KI n , and the forgetting effect of learners over time Modeling and internalizing knowledge KI n and the forgetting effect Acting on the knowledge state H at the previous moment n-1 , get the learner's updated knowledge state H at the current moment n ;

[0014] After solving the problem: By using the cross entropy loss function, the loss value between the probability of the learner answering the question correctly and his actual response is calculated, and the loss value is minimized by optimizing the model;

[0015] The above process is iterated repeatedly to dynamically update the learner's knowledge status and generate corresponding learner personalized teaching strategies.

[0016] Furthermore, the step of obtaining the learner's practice record and performing data preprocessing on the practice record to obtain corresponding preprocessed data further includes:

[0017] The exercise record includes the answer time ts n , answer question p n , the knowledge concept k tested by the question n and the answer result r n , the data preprocessing includes cleaning, deduplication, and deletion of missing values.

[0018] Furthermore, the preprocessed data is vectorized to obtain a problem embedding representation vector P n , combined with the knowledge concepts, difficulty and discrimination in the practice records, to obtain the enhanced question embedding representation vector Further including:

[0019] The enhanced question embedding representation vector Embedding vector K with knowledge concept n As the main body, the difficulty of the problem b n and discrimination α n is the coefficient of variation, the enhanced problem embedding representation vector Satisfies the following formula:

[0020]

[0021] Where: K n is the knowledge concept embedding vector, is the coefficient of variation, φ and is the trainable weight, P n is the question embedding representation vector.

[0022] Furthermore, the multi-layer perceptron is used to fuse the learner's previous knowledge state H n-1 and the enhanced problem embedding representation vector Model the learner's personalized problem space, which includes the initial state SP of the problem n , target state EP n and the adaptation operator OP n , and use activation functions and gating networks to calculate the degree of knowledge acquisition Further including:

[0023] The learner’s previous knowledge state H is integrated through a multi-layer perceptron n-1 and the enhanced question embedding representation vector Model the learner's personalized problem space and obtain the initial state SP of the problem n , target state EP n and the adaptation operator OP n ;

[0024] Using tanh activation function from the initial state SP n and the previous knowledge state H n-1 The knowledge state candidate value KA is obtained from n , specifically:

[0025] By integrating the learner’s prior knowledge n-1 and enhanced question embedding representation To model the initial state SP in the personalized problem space n and the target state EP n , using the target state EP n and the initial state SP n The difference between n , the adaptation operator OP n Satisfies the following formula:

[0026]

[0027] OP n =EP n -SP n

[0028] in: represents vector concatenation, is a learnable matrix, |K| is the total number of knowledge concepts;

[0029] The knowledge state candidate value KA nEnter the initial state SP n , target state EP n , Adaptation operator OP n After the gating network composed of sigmoid activation function, the final knowledge acquisition degree is obtained Specifically:

[0030] Use the tanh activation function to generate the knowledge state candidate value KA for knowledge acquisition n , and retain the knowledge state candidate value KA through a gating network guided by question information n A necessary part of .

[0031] Further, the adaptation operator OP from the personalized problem space using a multilayer perceptron and an activation function n Generate the corresponding solution and get the knowledge application degree vector The probability of learners answering questions correctly is calculated through the four-parameter item response theory, and the learners' responses are attributed, further including:

[0032] Using a multilayer perceptron and tanh activation function, we can adapt the operator OP from the personalized problem space. n Generate possible solutions according to the problem target vector EP in the personalized problem space n and the knowledge acquisition degree vector to determine the learner’s final solution;

[0033] The learners' responses were simulated through the four-parameter item response theory, and the probability of learners answering questions correctly was calculated.

[0034] Furthermore, the knowledge acquisition degree vector and knowledge application degree vector Construct incremental indicators to quantify the learner's knowledge internalization degree KI n , and the forgetting effect of learners over time Modeling and internalizing knowledge KI n and the forgetting effect Acting on the knowledge state H at the previous moment n-1 , get the learner's updated knowledge state H at the current moment n , further comprising:

[0035] In quantifying the learner's knowledge internalization degree KI n To ensure that the incremental index is non-negative, (tanh(x)+1) / 2 is used as the non-negative constraint operator. The knowledge internalization degree KI n Satisfies the following formula:

[0036]

[0037] in: Represents vector concatenation, SP n is the initial state, EP n is the target state, OP n is the adaptation operator, b KI is the bias term;

[0038] The following formula is used to calculate the learner's personalized knowledge increment

[0039]

[0040] in: is the degree of knowledge acquisition, is the knowledge application degree vector, R n is the answer result, b ΓKI is the bias term, KI n is the degree of knowledge internalization, H n-1 It is the state of knowledge at the previous moment.

[0041] Furthermore, it also includes:

[0042] The degree of knowledge reduction of learners in the answer interval is calculated according to the following formula:

[0043]

[0044] Where: The time interval is it n =ts n+1 -ts n , which indicates the time difference between two responses, in minutes, IT n It n The embedding representation vector of is the next enhancement problem embedding representation vector, is the current state of knowledge, b ΓKF is the bias term, ts n+1 It is the next answer time, ts n It is time to answer this question.

[0045] Furthermore, the method further comprises: calculating the loss value between the probability of the learner correctly answering the question and his actual response by using the cross entropy loss function, and minimizing the loss value by optimizing the model, further comprising:

[0046] The cross entropy loss function satisfies the following formula:

[0047]

[0048] in: is the cross entropy loss function, r n is the result of the answer. is the predicted probability.

[0049] An online learning platform for implementing the problem-solving process-oriented explainable knowledge tracking method, comprising:

[0050] The raw data preprocessing module is used to obtain the learner's practice records, perform data preprocessing on the practice records, and obtain corresponding preprocessed data;

[0051] The problem representation module vectorizes the preprocessed data to obtain the problem embedding representation vector P n , combined with the knowledge concepts, difficulty and discrimination in the practice records, to obtain the enhanced question embedding representation vector

[0052] The knowledge acquisition module integrates the learner’s previous knowledge state H through a multi-layer perceptron n-1 and the enhanced problem embedding representation vector Model the learner's personalized problem space, which includes the initial state SP of the problem n , target state EP n and the adaptation operator OP n , and use activation functions and gating networks to calculate the knowledge acquisition degree vector

[0053] The knowledge application module uses a multi-layer perceptron and an activation function to generate a corresponding solution from the adaptation operator of the personalized problem space and obtain a knowledge application degree vector The probability of learners answering questions correctly is calculated through the four-parameter item response theory, and the learners' responses are attributed.

[0054] A knowledge updating module based on the knowledge acquisition degree vector and knowledge application degree vector Construct incremental indicators to quantify the learner's knowledge internalization degree KI n , and the forgetting effect of learners over time Modeling and internalizing knowledge KI n and the forgetting effect Acting on the knowledge state H at the previous moment n-1 , get the learner's updated knowledge state H at the current moment n ;

[0055] The model training module calculates the loss between the probability of a learner correctly answering a question and their actual response by using the cross entropy loss function, and minimizes the loss by optimizing the model;

[0056] The above process is iterated repeatedly to dynamically update the learner's knowledge status and generate corresponding learner personalized teaching strategies.

[0057] An electronic device comprises: a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the steps of the method.

[0058] Compared with the prior art, the present invention has the following advantages:

[0059] The present invention takes problem solving as the core in the learning process, and deconstructs the process of learners solving specific problems through artificial intelligence means such as problem representation, knowledge acquisition, knowledge application, knowledge updating and model training, and quantifies and optimizes this process. It models the learner's knowledge acquisition degree before solving the problem, calculates the learner's knowledge application degree when solving the problem, and measures the learner's knowledge updating degree after solving the problem. In the process of knowledge tracking, it realizes the interpretability of the problem-solving process, which helps educators and learners to grasp their own teaching and learning status at the first time.

[0060] The present invention adopts a multi-layer perceptron to learn the difficulty and discrimination of problems from problem embedding, takes knowledge concept embedding as the main body, and uses difficulty and discrimination as coefficient of variation to construct problem representation, thereby achieving accurate and comprehensive representation of the problem.

[0061] The present invention combines the four-parameter item response theory of educational psychology with probability statistics, and can calculate the probability of learners correctly answering test questions, attribute the learners' performance to subjective problem factors (difficulty, discrimination) and objective learners' knowledge factors (knowledge application) and behavioral factors (guessing and slipping), extract cognitive features from learners' practice records in the process of teaching and learning, measure the degree of knowledge internalization, model the degree of forgetting according to the answer interval, so as to achieve the purpose of dynamically updating the knowledge state, generate corresponding learners' personalized teaching strategies, and make the learners' learning process more targeted and personalized.

[0062] The present invention can scientifically and comprehensively diagnose the learner's knowledge mastery and quantify the learner's performance in both knowledge acquisition and knowledge application. In personalized education, the learner's knowledge status can be quantified to dynamically identify the type of learning activities that the learner needs to focus on at different time points. If it is detected that the learner has a high degree of knowledge acquisition but weak knowledge application ability, the present invention can use precise analysis to determine that the learner has an understanding of a certain knowledge point, but has difficulties in actual application - based on this quantitative result, more application exercises can be recommended to the learner, rather than simply reviewing the knowledge points.

[0063] The data-based personalized guidance strategy of the present invention can effectively help students improve their problem-solving ability, ensure that learning activities are more targeted, improve teaching efficiency, accurately predict future performance, and explain the reasons behind the prediction, helping teachers and students understand the key factors that affect learning outcomes, thereby improving the pertinence and effectiveness of personalized education. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0065] Figure 1 It is a flowchart of an interpretable knowledge tracing method for the problem-solving process.

[0066] Figure 1A It is a flow chart of the optimization technical solution of step S3.

[0067] Figure 2 It is an architectural diagram of an explainable knowledge tracking system for the problem-solving process.

[0068] Figure 3 It is an example diagram of the three stages of learners’ problem solving and knowledge transformation.

[0069] Figure 4 It is a model diagram of the embodiment of the present application that can explain the knowledge tracking method in a specific application scenario.

[0070] Figure 5 It is a structural diagram of an electronic device. DETAILED DESCRIPTION

[0071] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0072] The purpose of the multiple embodiments of the present application listed below is to address the shortcomings of existing knowledge tracking methods and provide an explainable knowledge tracking method for the problem-solving process. It comprehensively utilizes data mining technology, deep learning and natural language processing technology to model the learner's knowledge state transformation from the three stages of problem solving - before problem solving, during problem solving and after problem solving, and deeply explores how students transform theoretical knowledge into practical skills.

[0073] In the process of the explainable knowledge tracking method provided in the embodiment of the present application, the learner's performance is decomposed into three major factors: objective problem factors (test difficulty and discrimination), learner knowledge factors (knowledge acquisition and knowledge application), and learner behavior factors (guessing and slipping). These factors combined help the personalized learning system to comprehensively and scientifically diagnose the learner's knowledge mastery, and through quantitative generation of the two core indicators of "knowledge acquisition value" and "knowledge application value", accurately reflect the learner's learning status.

[0074] In the personalized learning system, the method and system provided in the embodiment of the present application will evaluate the learning needs of learners at different time points in real time according to these quantitative indicators and other behavioral data (such as guessing and slipping), and judge whether they need to focus on strengthening knowledge acquisition or focusing on knowledge application training. For example, suppose that a student's knowledge acquisition value in the mathematics 'calculus' module is 80%, while the knowledge application value is only 50%, and the system also detects that the student has more guessing behaviors in answering, indicating that he has a solid grasp of theoretical knowledge, but encounters difficulties in practical application, and there is an insufficient understanding of the problem or blind guessing. Through these analyses, the system determines that it currently needs more targeted application training. Based on this analysis result, the method and system provided in the embodiment of the present application will automatically push more challenging exercises to the student from the preset exercise question bank. These exercises include not only more difficult calculus problems, but also more differentiated questions to help students identify problems at different levels. At the same time, the system will reduce the simple review of basic knowledge, but guide students to further improve their practical application ability of knowledge through application and high differentiation exercises.

[0075] In addition, the method and system provided in the embodiments of the present application can also combine students' learning habits, answering time, error distribution and slip situations to further optimize the pushed content. For example, if the system finds that a student often "slips" in the problem-solving process (that is, he clearly has the ability to solve the problem but makes mistakes in certain steps), the system will recommend some exercises that help consolidate the problem-solving steps and prevent common mistakes. The method and system provided in the embodiments of the present application are based on the dynamic adjustment of quantitative data and behavioral analysis to ensure that personalized learning is more targeted, help teachers and students understand the key obstacles in the learning process, and provide teachers with specific strategies to guide students and improve teaching efficiency.

[0076] like Figure 1 The problem-solving process-oriented interpretable knowledge tracing method shown in the figure is applied to the problem-solving process-oriented interpretable knowledge tracing system, including:

[0077] Before resolving the issue:

[0078] Step S1, problem definition and data preprocessing: obtaining the learner's practice records, performing data preprocessing on the practice records, and obtaining corresponding preprocessed data;

[0079] Step S2, problem representation: vectorize the preprocessed data to obtain a problem embedding representation vector P n , combined with the knowledge concepts, difficulty and discrimination in the practice records, to obtain the enhanced question embedding representation vector

[0080] When troubleshooting:

[0081] Step S3, knowledge acquisition: The learner’s previous knowledge state H is integrated through a multi-layer perceptron n-1 and the enhanced problem embedding representation vector Model the learner's personalized problem space, which includes the initial state SP of the problem n , target state EP n and the adaptation operator OP n , and use activation functions and gating networks to calculate the knowledge acquisition degree vector

[0082] Step S4, knowledge application: Use a multi-layer perceptron and an activation function to generate a corresponding solution from the adaptation operator of the personalized problem space to obtain a knowledge application degree vector The probability of learners answering questions correctly is calculated through the four-parameter item response theory, and the learners' responses are attributed.

[0083] Step S5, knowledge update: based on the knowledge acquisition degree vector and knowledge application degree vector Construct incremental indicators to quantify the learner's knowledge internalization degree KI n , and the forgetting effect of learners over time Modeling and internalizing knowledge KI n and the forgetting effect Acting on the knowledge state H at the previous moment n-1 , get the learner's updated knowledge state H at the current moment n ; In step S5, the forgetting effect Describes the phenomenon that learners gradually forget the knowledge they have mastered over time, which will have an impact on the learner's knowledge state. In order to characterize the forgetting effect The impact on learners can be modeled in different ways, for example, by reducing the learner's probability of mastering a particular knowledge point, or by introducing a decay function to simulate the gradual weakening of the learner's memory.

[0084] In step S5, the knowledge internalization level KI n and the forgetting effect Acting on the knowledge state H at the previous moment n-1 , we can update and get the learner's knowledge state H at the current moment n , that is: the degree of knowledge internalization KI n and the forgetting effect Combined, it acts on the knowledge state H at the previous moment n-1 , get the current knowledge state H n .

[0085] The degree of knowledge internalization KI n and the forgetting effect The combined process includes analyzing the learner's performance in answering questions, fitting the forgetting curve, and evaluating the effect of intervention measures, and finally obtaining the updated current knowledge state H n , the current knowledge state H n It can dynamically track learners' knowledge status and provide real-time feedback and adjustment suggestions for teaching.

[0086] Exemplarily, in other embodiments, the current knowledge state H n Applied to the knowledge tracking model (CMKT), the educational concept map is used to model the learner, which solves the problem of sparse learner data. The CMKT model combines the educational relationship information and topological information in the concept map and uses a recursive neural network to perform the knowledge tracking task. The above current knowledge state H n The application scenario provides a way to quantify the learner's knowledge internalization degree and model the forgetting effect, thereby updating the learner's knowledge status.

[0087] After resolving the issue:

[0088] Step S6, model training: by using the cross entropy loss function, the loss value between the probability of the learner correctly answering the question and his actual response is calculated, and the loss value is minimized by optimizing the model;

[0089] Step S7, repeating the process from step S1 to step S6, dynamically updating the learner's knowledge status, and generating a corresponding learner personalized teaching strategy.

[0090] The perceptron (also known as a multilayer perceptron, MLP) used in the technical solution of steps S1 to S7 is a type of feedforward artificial neural network consisting of multiple layers, including an input layer, one or more hidden layers, and an output layer. MLP is trained using a supervised learning technique called backpropagation, usually in combination with a gradient descent algorithm. The technical solution of steps S1 to S7 can be summarized as follows:

[0091] (1) Problem definition and data preprocessing, including the definition of the knowledge tracking problem, data collection and data preprocessing of the online learning platform;

[0092] (2) Problem representation: Propose a knowledge-centric problem representation that adjusts the variability of problems by learning the difficulty and discrimination of problems;

[0093] (3) Knowledge acquisition: Before solving a problem, the learner’s personalized problem space (including the problem’s initial state, target state, and adaptation operator) is modeled, and their acquisition of problem-related knowledge is simulated through a gating mechanism;

[0094] (4) Knowledge application: capturing learners’ goal-oriented knowledge application during problem solving and attributing their responses to problems using the four-parameter item response theory;

[0095] (5) Knowledge updating: After solving a problem, incremental indicators are used to quantify the learner’s knowledge internalization and model the forgetting effect over time;

[0096] (6) Model training: By using the cross-entropy loss function, the loss value between the predicted probability of the learner answering the question correctly and his actual response is calculated, and the loss value is minimized by optimizing the model.

[0097] (7) Iterative update: Repeat the above process from (1) to (6) to dynamically update the learner's knowledge status and generate corresponding personalized teaching strategies for the learner.

[0098] In the technical solution provided in steps S1 to S7, the three stages of learners' problem solving are deconstructed to model the learners' knowledge transformation, thereby predicting the learners' future performance. In the specific technical solution of steps S1 to S7, the learners' historical practice records are first collected and preprocessed from the online learning platform, and then the rich attributes (difficulty and discrimination) of the learning problems are learned according to the practice records, the knowledge-centered problem representation is enhanced, and then the knowledge transformation in the learners' problem solving process is simulated. Before solving the problem (i.e., the knowledge acquisition module), the learners' personalized problem perception and knowledge acquisition degree are modeled; when solving the problem (i.e., the knowledge application module), the knowledge application level is measured according to the knowledge acquisition, and the problem response is calculated by the four-parameter item response theory; after the problem is solved (i.e., the knowledge update module), the degree of learners' knowledge internalization and forgetting is quantified. Finally, the model is trained to ensure the rationality of the knowledge status diagnosis and the accuracy of the performance prediction.

[0099] Preferably, in step S1, the acquiring of the learner's practice record, and performing data preprocessing on the practice record to obtain corresponding preprocessed data further comprises:

[0100] The exercise record includes the answer time ts n , answer question p n , the knowledge concept k tested by the question n and the answer result r n , the data preprocessing includes cleaning, deduplication, and deletion of missing values.

[0101] Exemplarily, in step S1, the problem definition and data preprocessing are specifically as follows:

[0102] Step S11, problem definition: Given a learner’s N practice records, each practice record contains the answer time ts n , answer question p n , the knowledge concept k tested by the question n and the answer result r n The task of knowledge tracking is to evaluate the learner's knowledge status based on his / her practice records and predict the probability of correctly answering the next question, for example:

[0103] Given a learner's historical learning sequence X N ={X1,X2,...,X N}, N represents the total number of student exercises. n =(ts n ,p n ,k n ,r n ), n∈N is the most basic exercise unit, indicating that students n time, to test knowledge concept kn The problem n The reaction is r n , r n =1 means the answer is correct, r n =0 indicates an incorrect answer. The task of knowledge tracking is to N Evaluate the student's knowledge status and predict the probability that the learner will correctly answer the next question, i.e., p(r n+1 =1|p n+1 ,X n ).

[0104] Step S12, data preprocessing: collect learners' answer behavior data from the online learning platform, cleanse, remove duplicates, delete missing values, etc. to ensure data quality, for example:

[0105] Collect learners’ answering behavior data on the online learning platform, including the time when each student answered each time, the specific questions answered, the knowledge concepts tested by the questions, and the answer results. Then, clean, remove duplicates, delete missing values, and other operations on these data to ensure data quality. Next, encode the features so that they can be input into the model for subsequent analysis and modeling.

[0106] Preferably, in step S2, the preprocessed data is vectorized to obtain a question embedding representation vector P n , combined with the knowledge concepts, difficulty and discrimination in the practice records, to obtain the enhanced question embedding representation vector Further including:

[0107] The enhanced question embedding representation vector Embedding vector K with knowledge concept n As the main body, the difficulty of the problem is d n and discrimination α n is the coefficient of variation, the enhanced problem embedding representation vector Satisfies the following formula:

[0108]

[0109] Where: K n is the knowledge concept embedding vector, is the coefficient of variation, φ and is the trainable weight, P n is the question embedding representation vector.

[0110] In step S2, the problem representation is enhanced by learning rich properties of the problem, so the problem representation includes learning the difficulty and discrimination properties of the problem and modeling the knowledge-centric problem representation:

[0111] Step S21, learning the difficulty and discrimination of the problem: learning the difficulty and discrimination of the problem from the problem embedding vector through a multi-layer perceptron:

[0112]

[0113] Note: σ represents the sigmoid activation function. is a dimension d h The learnable vector, P n is the question embedding vector, b pa is the bias. n represents the difficulty of the problem, α n represents the problem discrimination, both are scalars.

[0114] Step S22, question representation: The goal of the exercise is to learn knowledge, so the knowledge concepts tested by the questions are crucial. In addition, questions testing the same knowledge concepts may have different effects on learning and performance due to differences in difficulty and discrimination. This method considers both the primary and secondary relationships and question attributes to enhance question representation:

[0115]

[0116] Note: K n is the knowledge concept embedding, is the coefficient of variation, φ and are trainable weights.

[0117] like Figure 1A As shown, preferably, in step S3, the previous moment knowledge state H of the learner is fused by the multilayer perceptron n-1 and the enhanced problem embedding representation vector Model the learner's personalized problem space, which includes the initial state SP of the problem n , target state EP n and the adaptation operator OP n , and use activation functions (tanh activation function or sigmoid activation function) and gating networks to calculate the degree of knowledge acquisition Further including:

[0118] Step S31: fuse the learner’s previous knowledge state H through a multi-layer perceptron n-1 and the enhanced question embedding representation vector Model the learner's personalized problem space and obtain the initial state SP of the problem n , target state EP n and the adaptation operator OP n ;

[0119] Step S32, using the tanh activation function from the initial state SP n and the previous knowledge state H n-1 The knowledge state candidate value KA is obtained from n , specifically:

[0120] By integrating the learner’s prior knowledge n-1 and problem representation To model the initial state SP in the personalized problem space n and the target state EP n , using the target state EP n and the initial state SP n The difference between n , the adaptation operator OP n Satisfies the following formula:

[0121]

[0122] OP n =EP n -SP n ……(5)

[0123] in: represents vector concatenation, is a learnable matrix, |K| is the total number of knowledge concepts;

[0124] Step S33: The knowledge state candidate value KA n Enter the initial state SP n , target state EP n , Adaptation operator OP n After the gating network composed of sigmoid activation function, the final knowledge acquisition degree is obtained Specifically:

[0125] Use the tanh activation function to generate the knowledge state candidate value KA for knowledge acquisition n , and retain the knowledge state candidate value KA through a gating network guided by question information n A necessary part of .

[0126] The specific content of knowledge acquisition in step S3 can be summarized as follows:

[0127] Problem perception: The initial state and target state in the personalized problem space are modeled by integrating the learner’s prior knowledge and problem representation, and the difference between the target state and the initial state is used to represent the adaptation operator in the problem space.

[0128] The learner’s solution to the problem begins with constructing an internal representation of the external problem statement, namely the “problem space” (including the initial state, the target state, and the adaptation operator). This method integrates the learner’s prior knowledge H n-1 and enhanced question embedding representation To model the initial state SP in the personalized problem space n and the target state EP n , using the target state EP n and the initial state SP n The difference between n :

[0129]

[0130] OP n =EP n -SP n ……(7)

[0131] Remark: represents vector concatenation, is the learnable matrix, |K| is the total number of knowledge concepts.

[0132] Knowledge acquisition: A tanh activation function is used to generate candidate values ​​for knowledge acquisition from the initial state of the problem space and the learner’s knowledge state, and then a gating network guided by problem information is used to select and retain the necessary parts of the candidate values.

[0133] Knowledge acquisition includes acquiring knowledge from the external environment and constructing new knowledge through internal thinking processes. In the knowledge tracking scenario, external knowledge comes from the initial state of the problem (known conditions) SP n , internal knowledge is the learner’s knowledge state at the previous moment H n-1 , the embodiment of the present application uses the tanh activation function to generate the knowledge state candidate value KA for knowledge acquisition n :

[0134]

[0135] Since learners have limited memory capacity, they will selectively focus on the most relevant and core aspects of the problem during the knowledge acquisition process in order to make more effective planning. To address this problem, the present application embodiment further designs a gating network to select and retain the necessary parts through question information:

[0136]

[0137] Preferably, in step S4, the adaptation operator OP of the personalized problem space is selected using a multilayer perceptron and an activation function. nGenerate the corresponding solution and get the knowledge application degree vector The probability of learners answering questions correctly is calculated through the four-parameter item response theory, and the learners' responses are attributed, further including:

[0138] Using a multilayer perceptron and tanh activation function, we can adapt the operator OP from the personalized problem space. n Generate possible solutions according to the problem target vector EP in the personalized problem space n and the knowledge acquisition degree vector to determine the learner’s final solution;

[0139] The learners' responses were simulated through the four-parameter item response theory, and the probability of learners answering questions correctly was calculated.

[0140] In step S4, the specific content of knowledge application can be summarized as follows:

[0141] Step S41, knowledge application: using a multilayer perceptron and a tanh activation function to generate possible solutions from the adaptation operator of the problem space, and further determining the learner's final solution based on the problem goal and the degree of knowledge acquisition in the problem space;

[0142] Knowledge application involves converting abstract knowledge into concrete action plans and adjusting and improving these plans in practice. The solution to the problem is contained in the adaptation operator in the problem space. This method uses the tanh activation function to extract the solution from the adaptation operator OP. n possible solutions and designed an EP with the problem objective n and knowledge acquisition Leading door To determine the learner's final solution:

[0143]

[0144]

[0145] Step S42, question answering: The four-parameter item response theory is used to calculate the probability of students correctly answering questions, and the learner's response is attributed to the subjective question factor (difficulty d n , discrimination α n ), objective learner knowledge factors (knowledge application ) and learner behavior factors (guessing G n and slip S n ):

[0146]

[0147] In step S42, the four-parameter item response theory is integrated to simulate the student's response by fully considering the subjective problem factors and objective learner factors. First, the learner's knowledge application is integrated with the predicted problem to determine the learner's knowledge application degree that is most relevant to the predicted problem:

[0148]

[0149] Then model the learner’s guess G on the prediction problem n and slip S n Probability:

[0150]

[0151] Finally, the student response is simulated by fully considering the problem factors (difficulty, discrimination), learner knowledge factors (knowledge application) and learner behavior factors (guessing and slipping):

[0152]

[0153] Note: D is a constant whose value is 4×1.7.

[0154] The 4-Parameter Item Response Theory (4PL) in step S4 is a model in psychometrics and educational measurement, which is used to evaluate the difficulty of test questions (items) and the ability of test subjects. The 4-Parameter Item Response Theory is an extension of the Item Response Theory (IRT), which can describe the test subject's response to the questions in more detail.

[0155] In IRT, the most commonly used model is the three-parameter model (3PL), which includes three parameters:

[0156] Difficulty parameter (b): indicates the difficulty of the question.

[0157] Discrimination parameter (a): indicates the ability of the item to distinguish between test takers of different ability levels.

[0158] Guessing parameter (c): It represents the probability that the subject can answer the question correctly even if his ability is very low.

[0159] The four-parameter model adds one more parameter to the three-parameter model: 4. Pseudo-chance parameter (d): It represents the probability that the subject will answer the question incorrectly even if he or she has high ability.

[0160] The probability curve of the four-parameter model can be expressed as:

[0161] P(X=1∣θ,a,b,c,d)=c+(1-cd)ea(θ-b)1+ea(θ-b)P(X=1∣

[0162] θ,a,b,c,d)=c+(1-cd)1+ea(θ-b)ea(θ-b)……(17)

[0163] Where: P(X=1) is the probability that the subject answers the question correctly, θ is the subject's ability level, a is the discrimination parameter, b is the difficulty parameter, c is the guessing parameter, and d is the slip parameter (pseudo-chance parameter).

[0164] The four-parameter item response theory used in the present application embodiment is particularly suitable for situations where the ability levels are widely distributed or the difficulty of the questions varies greatly. For example, the addition of the slip parameter (pseudo-chance parameter) enables the model to better handle the following situations: the test subject may answer simple questions incorrectly due to carelessness or other non-ability factors, that is, the learner clearly has the ability to solve the problem but makes mistakes in some steps.

[0165] Preferably, in step S5, the knowledge acquisition degree vector and knowledge application degree vector Construct incremental indicators to quantify the learner's knowledge internalization degree KI n , and the forgetting effect of learners over time Modeling and internalizing knowledge KI n and the forgetting effect Acting on the knowledge state H at the previous moment n-1 , get the learner's updated knowledge state H at the current moment n , further comprising:

[0166] In quantifying the learner's knowledge internalization degree KI n To ensure that the incremental index is non-negative, (tanh(x)+1) / 2 is used as the non-negative constraint operator. The knowledge internalization degree KI n Satisfies the following formula:

[0167]

[0168] in: Represents vector concatenation, SP n is the initial state, SP n is the target state, OP n is the adaptation operator, b KI is the bias term;

[0169] The following formula is used to calculate the learner's personalized knowledge increment

[0170]

[0171] in: is the degree of knowledge acquisition, is the knowledge application degree vector, R n is the answer result, b ΓKI is the bias term, KI n is the degree of knowledge internalization, H n-1 is the knowledge state at the previous moment,

[0172] The degree of knowledge reduction of learners in the answer interval is calculated according to the following formula:

[0173]

[0174] Where: The time interval is it n =ts n+1 -ts n , which indicates the time difference between two responses, in minutes, IT n It n The embedding representation vector of is the next enhancement problem embedding representation vector, is the current state of knowledge, b ΓKF is the bias term;

[0175] In step S5, the learner's knowledge state will change after practice. The increase in knowledge mastery reflects the internalization of knowledge, while forgetting will lead to a decrease in knowledge. Therefore, knowledge updating includes two modules: knowledge internalization and forgetting.

[0176] Step S51, knowledge internalization: The learner's new knowledge comes from problem practice, so all the information of the problem is integrated to obtain the candidate value of knowledge increment. Constructivist learning theory shows that learning is an active construction process, and the whole practice process has a positive impact on learners, so use As a non-negativity constraint operator to ensure that the increment is non-negative:

[0177]

[0178] Through internalization, learners not only accept information, but also need to integrate it into the cognitive framework. To this end, the present embodiment proposes a growth index that considers knowledge acquisition, knowledge application, and answer response to determine the learner's personalized knowledge increment:

[0179]

[0180] Step S52, knowledge forgetting: Due to the limitations of the human memory system, current knowledge may be forgotten over time. In order to simulate the complex forgetting effect, the embodiment of the present application provides a forget gate to model the degree of knowledge reduction of the learner in the answer interval. The specific formula is as follows:

[0181]

[0182]

[0183] Note: The time interval is it n =ts n+1 -ts n , which is the time difference between the two responses, in minutes. n It n The embedding representation of the embodiment of the present application sets all time intervals greater than 1 month to 1 month.

[0184] Preferably, in step S6, the calculation of the loss value between the probability of the learner correctly answering the question and his actual response by using the cross entropy loss function and minimizing the loss value by optimizing the model further includes:

[0185] The cross entropy loss function satisfies the following formula:

[0186]

[0187] in: is the cross entropy loss function, r n is the answer result, r n is the answer result (actual response), is the predicted probability (predicted response).

[0188] In step S6, to train all parameters θ, the predicted responses are used and the actual response r n The cross entropy logarithmic loss between is used as the objective function, and the parameters in the model are updated using the Adam optimization algorithm. The specific steps when using the Adam optimization algorithm to update the model parameters are:

[0189] Step S61, defining a model and a loss function: defining a model and a loss function, wherein the architecture of the model includes the type of layers and the number of neurons, cross entropy is selected as the loss function, and the model is specifically a deep learning model;

[0190] Step S62, initializing parameters: before starting training, initializing the parameters of the model (such as the Xavier initialization method), the parameters including weights and bias values;

[0191] Step S63, forward propagation: performing the prediction output of forward propagation for each training sample through the sigmoid function, converting the linear output of the model into a probability distribution;

[0192] Step S64, calculating the loss, using a cross entropy loss function to calculate the difference between the model prediction and the true label;

[0193] Step S65, back propagation: calculating the gradient of the loss function with respect to the model parameters based on the chain rule;

[0194] Step S66, parameter update: use an adaptive learning rate optimization algorithm (such as the Adam optimization algorithm) to update the parameters of the model.

[0195] The optimization technology provided in step S61 to step S66 minimizes the difference between model prediction and actual data by constructing and training a deep learning model and using the cross entropy loss function as the optimization target. In the process of model construction, the model architecture is first designed, including selecting the appropriate network layer type, the number of neurons in each layer, etc., giving initial values ​​to the weights and biases of the model, initializing the parameters before training, calculating the predicted output of the model for each training sample, converting the linear output into a probability distribution using the sigmoid function, and then using the cross entropy loss function to measure the difference between the probability distribution predicted by the model and the true label, performing loss calculation, and then calculating the gradient of the loss function with respect to the model parameters by the chain rule, providing a basis for parameter updating, and finally applying the Adam optimization algorithm (adaptive learning rate optimization algorithm) to update the model parameters according to the gradient, thereby minimizing the loss function.

[0196] The optimization technology solution provided in steps S61 to S66 enables the model to learn complex patterns and features in the data through iterative training, and with the continuous updating of parameters, the prediction performance of the model on classification tasks is gradually improved, and the generalization ability is enhanced. Then, through automated gradient calculation and parameter updating, human intervention is reduced and the efficiency of the training process is improved. Finally, based on the adaptive learning rate optimization algorithm (Adam optimization algorithm), it adapts to different training data and tasks, thereby improving the applicability of the model.

[0197] The optimization technology solution provided in steps S61 to S66 realizes automated learning and high-precision prediction of classification problems by constructing, training and optimizing a neural network model, while ensuring that the (deep learning) model has good generalization and adaptability, and provides a good optimization solution for calculating the loss value between the probability of a learner correctly answering a question and his or her actual response, thereby minimizing the loss value.

[0198] For the method steps disclosed in the above embodiments, the method steps are described as a series of action combinations for the purpose of simple description, but those skilled in the art should know that the embodiments of the present invention are not limited by the order of the actions described, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present invention.

[0199] like Figure 2 The problem-solving process-oriented interpretable knowledge tracking system shown is used to implement the problem-solving process-oriented interpretable knowledge tracking method described in the embodiment of the present application, including:

[0200] The raw data preprocessing module is used to obtain the learner's practice records, perform data preprocessing on the practice records, and obtain corresponding preprocessed data;

[0201] The problem representation module vectorizes the preprocessed data to obtain the problem embedding representation vector P n , combined with the knowledge concepts, difficulty and discrimination in the practice records, to obtain the enhanced question embedding representation vector

[0202] The knowledge acquisition module integrates the learner’s previous knowledge state H through a multi-layer perceptron n-1 and the enhanced problem embedding representation vector Model the learner's personalized problem space, which includes the initial state SP of the problem n , target state EP n and the adaptation operator OP n , and use activation functions and gating networks to calculate the knowledge acquisition degree vector

[0203] The knowledge application module uses a multi-layer perceptron and an activation function to generate a corresponding solution from the adaptation operator of the personalized problem space and obtain a knowledge application degree vector The probability of learners answering questions correctly is calculated through the four-parameter item response theory, and the learners' responses are attributed.

[0204] A knowledge updating module based on the knowledge acquisition degree vector and knowledge application degree vector Construct incremental indicators to quantify the learner's knowledge internalization degree KI n , and the forgetting effect of learners over time Modeling and internalizing knowledge KI n and the forgetting effect Acting on the knowledge state H at the previous moment n-1 , get the learner's updated knowledge state H at the current moment n ;

[0205] The model training module calculates the loss between the probability of a learner correctly answering a question and their actual response by using the cross entropy loss function, and minimizes the loss by optimizing the model;

[0206] The above process is iterated repeatedly to dynamically update the learner's knowledge status and generate corresponding learner personalized teaching strategies.

[0207] The implementation methods of the system described above are merely illustrative, for example, the various functional modules, units or subsystems in the system may or may not be physically separated, or may or may not be physical units, that is, they may be located in the same place, or they may be distributed to multiple different systems and their subsystems or modules. Those skilled in the art may select some or all of the functional modules, units or subsystems according to actual needs to achieve the purpose of the embodiments of the present invention, and those of ordinary skill in the art may understand and implement the above-mentioned situations without creative work.

[0208] like Figure 3 and Figure 4 As shown, the present invention provides a corresponding online learning platform based on the interpretable knowledge tracing method and system for problem-solving process, in which the interpretable knowledge tracing system for problem-solving process is provided to execute the interpretable knowledge tracing method for problem-solving process.

[0209] like Figure 5 As shown, the present invention not only provides an explainable knowledge tracking method and system, but also provides corresponding electronic equipment and storage media:

[0210] An electronic device, characterized in that it includes: a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the steps of an explainable knowledge tracking method for a problem-solving process.

[0211] A computer-readable storage medium stores a computer program executable by an electronic device. When the computer program runs on the electronic device, the electronic device executes the steps of an interpretable knowledge tracing method for a problem-solving process.

[0212] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM has many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0213] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0214] In addition, those skilled in the art will appreciate that, although some embodiments described herein include certain features included in other embodiments but not other features, the combination of features of different embodiments is meant to be within the scope of the present invention and form different embodiments. For example, any one of the embodiments claimed in the claims can be used in any combination in the embodiments of the present invention.

[0215] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0216] In addition, the technical solutions between the various embodiments of the present invention can be combined with each other, but it must be based on the fact that ordinary technicians in the field can implement it. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0217] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An interpretable knowledge tracing method for problem-solving process, applied to an interpretable knowledge tracing system for problem-solving process, characterized in that: include: Before solving the problem: obtaining the learner's practice records, performing data preprocessing on the practice records, and obtaining corresponding preprocessed data; The preprocessed data is vectorized to obtain the problem embedding representation vector P n , combined with the knowledge concepts, difficulty and discrimination in the practice records, to obtain the enhanced question embedding representation vector When solving problems: The learner’s previous knowledge state H is integrated through a multi-layer perceptron n-1 and the enhanced problem embedding representation vector Model the learner's personalized problem space, which includes the initial state SP of the problem n , target state EP n and the adaptation operator OP n , and use activation functions and gating networks to calculate the knowledge acquisition degree vector Use a multi-layer perceptron and activation function to generate a corresponding solution from the adaptation operator of the personalized problem space to obtain a knowledge application degree vector The probability of learners answering questions correctly is calculated through the four-parameter item response theory, and the learners' responses are attributed. Based on the knowledge acquisition degree vector and knowledge application degree vector Construct incremental indicators to quantify the learner's knowledge internalization degree KI n , and the forgetting effect of learners over time Modeling and internalizing knowledge KI n and the forgetting effect Acting on the knowledge state H at the previous moment n-1 , get the learner's updated knowledge state H at the current moment n ; After solving the problem: By using the cross entropy loss function, the loss between the probability of the learner answering the question correctly and their actual response is calculated, and the loss value is minimized by optimizing the model; The above process is iterated repeatedly to dynamically update the learner's knowledge status and generate corresponding learner personalized teaching strategies.

2. The interpretable knowledge tracing method for problem solving process according to claim 1 is characterized in that: The step of obtaining the learner's practice record and performing data preprocessing on the practice record to obtain corresponding preprocessed data further includes: The exercise record includes the answer time ts n , answer question p n , the knowledge concept k tested by the question n and the answer result r n , the data preprocessing includes cleaning, deduplication, and deletion of missing values.

3. The interpretable knowledge tracing method for problem solving process according to claim 1 is characterized in that: The preprocessed data is vectorized to obtain a problem embedding representation vector P n , combined with the knowledge concepts, difficulty and discrimination in the practice records, to obtain the enhanced question embedding representation vector Further including: The enhanced question embedding representation vector Embedding vector K with knowledge concept n As the main body, the difficulty of the problem is d n and discrimination α n is the coefficient of variation, the enhanced problem embedding representation vector Satisfies the following formula: Where: K n is the knowledge concept embedding vector, is the coefficient of variation, φ and is the trainable weight, P n is the question embedding representation vector.

4. The interpretable knowledge tracing method for problem solving process according to claim 1, characterized in that: The multi-layer perceptron is used to fuse the learner’s previous knowledge state H n-1 and the enhanced problem embedding representation vector Model the learner's personalized problem space, which includes the initial state SP of the problem n , target state EP n and the adaptation operator OP n , and use activation functions and gating networks to calculate the degree of knowledge acquisition Further including: The learner’s previous knowledge state H is integrated through a multi-layer perceptron n-1 and the enhanced question embedding representation vector Model the learner's personalized problem space and obtain the initial state SP of the problem n , target state EP n and the adaptation operator OP n ; Using tanh activation function from the initial state SP n and the previous knowledge state H n-1 The knowledge state candidate value KA is obtained from n , specifically: By integrating the learner’s prior knowledge n-1 and problem representation To model the initial state SP in the personalized problem space n and the target state EP n , using the target state EP n and the initial state SP n The difference between n , the adaptation operator OP n Satisfies the following formula: OP n =EP n -SP n in: represents vector concatenation, is a learnable matrix, |K| is the total number of knowledge concepts; The knowledge state candidate value KA n Enter the initial state SP n , target state EP n , Adaptation operator OP n After the gating network composed of sigmoid activation function, the final knowledge acquisition degree is obtained Specifically: Use the tanh activation function to generate the knowledge state candidate value KA for knowledge acquisition n , and retain the knowledge state candidate value KA through a gating network guided by question information n A necessary part of .

5. The interpretable knowledge tracing method for problem solving process according to claim 1, characterized in that: The adaptation operator OP from the personalized problem space using a multilayer perceptron and an activation function n Generate the corresponding solution and get the knowledge application degree vector The probability of learners answering questions correctly is calculated through the four-parameter item response theory, and the learners' responses are attributed, further including: Using a multilayer perceptron and tanh activation function, we can adapt the operator OP from the personalized problem space. n Generate possible solutions according to the problem target vector EP in the personalized problem space n and the knowledge acquisition degree vector to determine the learner’s final solution; The learners' responses were simulated through the four-parameter item response theory, and the probability of learners answering questions correctly was calculated.

6. The interpretable knowledge tracing method for problem solving process according to claim 1, characterized in that: The knowledge acquisition degree vector and knowledge application degree vector Construct incremental indicators to quantify the learner's knowledge internalization degree KI n , and the forgetting effect of learners over time Modeling and internalizing knowledge KI n and the forgetting effect Acting on the knowledge state H at the previous moment n-1 , get the learner's updated knowledge state H at the current moment n , further comprising: In quantifying the learner's knowledge internalization degree KI n To ensure that the incremental index is non-negative, (tanh(x)+1) / 2 is used as the non-negative constraint operator. The knowledge internalization degree KI n Satisfies the following formula: in: Represents vector concatenation, SP n is the initial state, EP n is the target state, OP n is the adaptation operator, b KI is the bias term; The following formula is used to calculate the learner's personalized knowledge increment in: is the degree of knowledge acquisition, is the knowledge application degree vector, R n is the answer result, b ΓKI is the bias term, KI n is the degree of knowledge internalization, H n-1 It is the state of knowledge at the previous moment.

7. The interpretable knowledge tracing method for problem solving process according to claim 6 is characterized in that: Also includes: The degree of knowledge reduction of learners in the answer interval is calculated according to the following formula: Where: The time interval is it n =ts n+1 -ts n , which indicates the time difference between two responses, in minutes, IT n It n The embedding representation vector of is the next enhancement problem embedding representation vector, is the current state of knowledge, b ΓKF is the bias term, ts n+1 It is the next answer time, ts n It is time to answer this question.

8. The interpretable knowledge tracing method for problem solving process according to claim 1, characterized in that: The method further includes: calculating the loss value between the probability of the learner correctly answering the question and the actual response thereof by using the cross entropy loss function, and minimizing the loss value by optimizing the model. The cross entropy loss function satisfies the following formula: in: is the cross entropy loss function, r n is the answer result, is the predicted probability.

9. An online learning platform for implementing the interpretable knowledge tracking method for problem solving process as claimed in any one of claims 1 to 8, characterized in that: include: The raw data preprocessing module is used to obtain the learner's practice records, perform data preprocessing on the practice records, and obtain corresponding preprocessed data; The problem representation module vectorizes the preprocessed data to obtain the problem embedding representation vector P n , combined with the knowledge concepts, difficulty and discrimination in the practice records, to obtain the enhanced question embedding representation vector The knowledge acquisition module integrates the learner’s previous knowledge state H through a multi-layer perceptron n-1 and the enhanced problem embedding representation vector Model the learner's personalized problem space, which includes the initial state SP of the problem n , target state EP n and the adaptation operator OP n , and use activation functions and gating networks to calculate the knowledge acquisition degree vector The knowledge application module uses a multi-layer perceptron and an activation function to generate a corresponding solution from the adaptation operator of the personalized problem space and obtain a knowledge application degree vector The probability of learners answering questions correctly is calculated through the four-parameter item response theory, and the learners' responses are attributed. A knowledge updating module based on the knowledge acquisition degree vector and knowledge application degree vector Construct incremental indicators to quantify the learner's knowledge internalization degree KI n , and the forgetting effect of learners over time Modeling and internalizing knowledge KI n and the forgetting effect Acting on the knowledge state H at the previous moment n-1 , get the learner's updated knowledge state H at the current moment n ; The model training module calculates the loss between the probability of a learner correctly answering a question and their actual response by using the cross entropy loss function, and minimizes the loss by optimizing the model; The above process is iterated repeatedly to dynamically update the learner's knowledge status and generate corresponding learner personalized teaching strategies.

10. An electronic device, characterized in that: include: A processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the steps of the method described in any one of claims 1 to 8.

Images