Knowledge tracking method and device based on counterfactual personalization augmentation

Abstract

The application discloses a kind of knowledge tracking method and device based on counterfactual personalization enhancement, belong to knowledge tracking technical field.The method includes: obtaining student information, question and answer result, define knowledge tracking, construct knowledge tracking causal diagram based on knowledge tracking, construct answer result prediction model based on knowledge tracking causal diagram, answer result prediction model includes learning tendency driving path submodel, knowledge state mapping path submodel, question representation influence path submodel and fusion output submodel;Based on knowledge tracking causal diagram, calculate path total effect, path direct effect and path indirect effect;Based on path total effect, path direct effect, path indirect effect and answer result prediction model, obtain the counterfactual prediction debiasing formula of student's question answer result;Based on counterfactual prediction debiasing formula, obtain the counterfactual prediction value of student's question answer result, the method reduces the global bias of question in knowledge tracking task.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of knowledge tracking, and particularly relates to a knowledge tracking method and device based on counterfactual personalized enhancement. BACKGROUND

[0002] Under the background of deep integration of artificial intelligence and education, knowledge tracking, as a core technology for evaluating learning effect, has significance for the development of personalized education in terms of modeling accuracy. The dynamic evolution characteristics of knowledge state directly affect the learning path planning effect. If there is systematic bias in the modeling process, it will lead to inaccurate teaching decisions, and thus affect the quality of talent training.

[0003] Existing knowledge tracking methods are mainly based on deep learning frameworks, and still face significant technical bottlenecks. First, the binary evaluation mechanism is difficult to depict the multi-dimensional characteristics of learning behavior, resulting in limited representation ability of the model for complex cognitive state. Second, the global bias problem causes false correlation feature capture, causing the knowledge state evaluation result to deviate from the true cognitive level. In addition, traditional bias correction methods lack effective mechanisms in maintaining the authenticity of data distribution, making it difficult to balance model robustness and practical application adaptability.

[0004] The related technology simply uses binary results to infer the knowledge state of students, which is prone to over-reliance on the distribution of answers to these questions, and incorrectly interprets high accuracy or low accuracy as the student's knowledge mastery level, thereby ignoring other factors (such as actual understanding ability). It fails to fully capture the dynamic changes in students' knowledge state over time, and the prediction accuracy of students' knowledge state mastery is low. SUMMARY

[0005] The present application aims to at least solve one of the technical problems existing in the prior art. To this end, the present application proposes a knowledge tracking method and device based on counterfactual personalized enhancement, which reduces the global bias of the problem in the knowledge tracking task.

[0006] In a first aspect, the present application provides a knowledge tracking method based on counterfactual personalized enhancement, the method comprising:

[0007] Obtaining student information, questions and answer results, defining knowledge tracking, the knowledge tracking being based on the answer results of students to historical questions and student information to predict the probability of students answering questions correctly;

[0008] Constructing a knowledge tracking causal diagram based on the knowledge tracking, the knowledge tracking causal diagram comprising a node set and a path set, the node set comprising students, questions, knowledge states and answer results, and the path set comprising a learning tendency driving path, a question representation influencing path and a knowledge state mapping path;

[0009] constructing an answer result prediction model based on the knowledge tracking causal diagram, the answer result prediction model comprising a learning tendency driven path sub-model, a knowledge state mapping path sub-model, a question representation influence path sub-model and a fusion output sub-model;

[0010] calculating a path total effect, a path direct effect and a path indirect effect based on the knowledge tracking causal diagram;

[0011] obtaining a counterfactual prediction debiasing formula of a student's answer result to a question based on the path total effect, the path direct effect, the path indirect effect and the answer result prediction model;

[0012] obtaining a counterfactual prediction value of the student's answer result to the question based on the counterfactual prediction debiasing formula.

[0013] According to an embodiment of the present application, the learning tendency driven path is student-answer result, used for describing the direct influence of student cognition on the answer result, the question representation influence path is question-answer result, used for describing the direct influence of the question's own attribute on the answer result, and the knowledge state mapping path is student and question-knowledge state-answer result.

[0014] According to an embodiment of the present application, the knowledge state mapping path comprises a capability question matching path and an answer result prediction path, the capability question matching path being student and question-knowledge state, used for describing the decision effect of the matching degree of student capability and question on the knowledge state, and the answer result prediction path being knowledge state-answer result, used for describing the prediction effect of the knowledge state on the answer result.

[0015] According to an embodiment of the present application, the learning tendency driven path sub-model comprises a first full connection network and a first multi-layer perception, used for obtaining a learning tendency driven path prediction result, the knowledge state mapping path sub-model is a knowledge tracking model, the question representation influence path sub-model comprises a second full connection network and a second multi-layer perception, used for obtaining a question representation influence path prediction result, and the fusion output sub-model comprises a third multi-layer perception, used for obtaining a prediction value of the student's answer result to the question.

[0016] According to an embodiment of the present application, the calculation formula of the fusion output sub-model is as follows:

[0017]

[0018] wherein, is a fusion state, is a fusion function, is an intermediate output state of the learning tendency driven path sub-model, is a knowledge mastery state of the knowledge state mapping path sub-model, an intermediate output state of a path submodel of a problem representation, denotes concatenation, is a trainable matrix, is a trainable bias vector, is a predicted value of a problem answer result, is a calculated value of a real-world problem answer result, and MLP is a third multilayer perceptron.

[0019] According to an embodiment of the present application, the calculation formula of the counterfactual prediction debiasing formula is as follows:

[0020]

[0021] wherein, is a predicted value of a problem answer result, is a learning tendency driven path prediction result, is a problem representation influenced path prediction result, TIE is an indirect effect of a path, and NDE is a direct effect of a path, is a weight of a learning tendency driven path prediction result, is a weight of a problem representation influenced path prediction result, is a hyperparameter, is a counterfactual prediction value of a student for a problem answer result, S is a student, Q is a problem, and Y is a student answer result, is a local learning tendency, and t is a time step, is a global problem representation of a problem answered at time step t+1.

[0022] According to an embodiment of the present application, the training process of the answer result prediction model comprises:

[0023] constructing a preset answer result prediction model;

[0024] obtaining different student information, different problems, and different answer results as a data set;

[0025] training the preset answer result prediction model according to the data set based on a loss function to obtain the answer result prediction model, wherein the loss function is constructed based on a binary cross-entropy loss function.

[0026] In a second aspect, the present application provides a knowledge tracking device based on counterfactual personalized enhancement, the device comprising:

[0027] an acquisition module configured to acquire student information, problems, and answer results, and define a knowledge tracking, wherein the knowledge tracking is a probability of a student answering a problem correctly based on the answer results of the student to historical problems and the student information;

[0028] a first constructing module configured to construct a knowledge tracing causal graph based on the knowledge tracing, the knowledge tracing causal graph comprising a node set and a path set, the node set comprising a student, a question, a knowledge state and an answer result, the path set comprising a learning tendency driven path, a question representation influenced path and a knowledge state mapped path;

[0029] a second constructing module configured to construct an answer result prediction model based on the knowledge tracing causal graph, the answer result prediction model comprising a learning tendency driven path submodel, a knowledge state mapped path submodel, a question representation influenced path submodel and a fusion output submodel;

[0030] a first calculating module configured to calculate a path total effect, a path direct effect and a path indirect effect based on the knowledge tracing causal graph;

[0031] a second calculating module configured to obtain an counterfactual prediction debiasing formula of an answer result of a student to a question based on the path total effect, the path direct effect, the path indirect effect and the answer result prediction model;

[0032] a predicting module configured to obtain a counterfactual prediction value of an answer result of a student to a question based on the counterfactual prediction debiasing formula.

[0033] In a third aspect, the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the knowledge tracing method based on counterfactual personalized enhancement according to the first aspect.

[0034] In a fourth aspect, the present application provides a non-transitory computer readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to implement the knowledge tracing method based on counterfactual personalized enhancement according to the first aspect.

[0035] In a fifth aspect, the present application provides a chip, comprising a processor and a communication interface, wherein the communication interface is coupled to the processor, and the processor is configured to execute a program or an instruction to implement the knowledge tracing method based on counterfactual personalized enhancement according to the first aspect.

[0036] In a sixth aspect, the present application provides a computer program product comprising a computer program, wherein the computer program is executable by a processor to implement the knowledge tracing method based on counterfactual personalized enhancement according to the first aspect.

[0037] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter.

[0038] The application provides a knowledge tracking method based on counterfactual personalized enhancement.

[0039] (1) The application defines a knowledge tracking problem based on student information, questions and answer results, combines a counterfactual prediction debiasing formula, calculates path total effects, path direct effects and path indirect effects of a knowledge tracking causal diagram, comprehensively utilizes deep learning and causal inference technology to explore student local learning tendencies and global problem representations, designs a method capable of supporting different knowledge tracking models, captures student local learning tendencies and global problem representations, and effectively alleviates global bias in the knowledge tracking task by using counterfactual inference, and enhances personalized performance of students.

[0040] (2) The application calculates question answer results by using a counterfactual prediction debiasing formula, combines prediction results of learning tendency driven paths and problem representation influence paths, adjusts direct effects of learning tendency driven paths and problem representation influence paths, combines indirect effects of knowledge state mapping paths, realizes effect control counterfactual prediction, can alleviate global bias in the knowledge tracking task, improves prediction accuracy of students on question answer results, and helps more accurately analyze knowledge mastery of students, and provides a more reliable basis for formulating personalized learning strategies.

[0041] (3) The application fuses an output submodel to fuse intermediate output states of a learning tendency driven path submodel, a knowledge state mapping path submodel and a problem representation influence path submodel, combines a trainable matrix and a bias vector, and finally predicts by using a multilayer perception machine, fuses output information of different paths, and improves accuracy of the answer result prediction model. BRIEF DESCRIPTION OF DRAWINGS

[0042] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which:

[0043] Figure 1 is a flowchart of the knowledge tracking method based on counterfactual personalized enhancement provided by the embodiments of the application;

[0044] Figure 2 is a structural schematic diagram of the knowledge tracking causal diagram provided by the embodiments of the application;

[0045] Figure 3 is a knowledge tracking framework structure diagram provided by the embodiments of the application;

[0046] Figure 4 is a structural schematic diagram of the knowledge tracking device based on counterfactual personalized enhancement provided by the embodiments of the application;

[0047] Figure 5 is a structural schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present application will be clearly described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all of them. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art belong to the scope of protection of the present application.

[0049] The terms "first", "second", and the like in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are usually a category, and do not limit the number of objects, for example, the first object can be one or more. In addition, "and / or" in the specification and claims means at least one of the connected objects, and the character " / ", generally represents that the front and rear associated objects are in a "or" relationship.

[0050] The knowledge tracking method based on counterfactual personalization enhancement, the knowledge tracking device based on counterfactual personalization enhancement, the electronic device and the readable storage medium provided by the embodiments of the present application will be described in detail below in combination with the drawings and specific embodiments and their application scenarios.

[0051] The knowledge tracking method based on counterfactual personalization enhancement can be applied to a terminal, and can be specifically executed by hardware or software in the terminal.

[0052] The terminal includes, but is not limited to, a portable communication device such as a mobile phone or a tablet computer having a touch-sensitive surface (e.g., a touchscreen display and / or a touchpad). It should also be understood that in some embodiments, the terminal can not be a portable communication device, but a desktop computer having a touch-sensitive surface (e.g., a touchscreen display and / or a touchpad).

[0053] In each of the following embodiments, a terminal including a display and a touch-sensitive surface is described. However, it should be understood that the terminal can include one or more other physical user interface devices such as a physical keyboard, a mouse, and a joystick.

[0054] The execution subject of the knowledge tracking method based on counterfactual personalized enhancement provided in the embodiments of the present application can be an electronic device or a functional module or functional entity capable of implementing the knowledge tracking method based on counterfactual personalized enhancement in the electronic device. The electronic device mentioned in the embodiments of the present application includes but is not limited to a mobile phone, a tablet computer, a computer, a camera, a wearable device and the like. The knowledge tracking method based on counterfactual personalized enhancement provided in the embodiments of the present application is described below by taking an electronic device as an execution subject.

[0055] In the current rapid development of artificial intelligence, combining advanced technology with education has become an important trend in recent years. Students are no longer satisfied with traditional face-to-face learning, but are increasingly inclined to use online platforms and intelligent education tools for learning. The continuous accumulation of online learning activity data provides valuable opportunities for us to understand students' learning process and better meet the needs of personalized education.

[0056] There is an internal knowledge state in the learning process of students, which reflects their mastery of the content learned and is constantly updated as the learning deepens. To understand the learning process behind students, the core task is to track the evolution of students' knowledge state, which is the key to knowledge tracking technology. Knowledge tracking aims to model students' knowledge mastery based on learning records and predict future performance.

[0057] The knowledge tracking task usually assumes that the answer result is a binary value, that is, the answer to each question has only two cases of "correct" or "incorrect". However, this simplified evaluation method may not reflect the students' mastery level, especially when students show some bias on certain questions. For example, some overly simple questions will make most students answer correctly, while some more challenging questions will produce more incorrect answers due to excessive difficulty. If the model simply uses such binary results to infer the students' knowledge state, it is easy to over-rely on the answer distribution of these questions. Specifically, the model can find a shortcut to incorrectly interpret high accuracy as the students' knowledge mastery level, thereby ignoring other factors (such as actual understanding ability) and incorrectly judging that the students have mastered a certain knowledge point, although they do not perform well on these complex questions. Therefore, this global bias caused by the imbalance of correct and incorrect answers to questions not only makes the model overfit to these features, but also fails to fully capture the dynamic changes in students' knowledge state over time, leading to incorrect modeling of the knowledge state.

[0058] To mitigate global bias, a simple solution is to resample the dataset to balance the proportion of correct and incorrect answers for each question. This is a root cause of bias, but in fact this method has certain limitations. First, it may change the actual distribution of data, affecting the applicability of the model in the actual scene. Second, in the actual learning scene, some questions themselves have significant bias due to their content characteristics. Therefore, simple resampling cannot effectively identify and isolate the source of global bias.

[0059] Causal inference provides a systematic method to explicitly determine the causal relationship between factors by constructing a causal graph. From the perspective of causal inference, global bias can be regarded as interference of confounding variables on model prediction. Therefore, by introducing counterfactual inference, the student's answer performance under the influence of bias can be estimated by controlling the influence of confounding variables. On the other hand, if the global bias in knowledge tracking is carefully analyzed, it can be found that global bias affects model prediction from a global perspective, that is, it affects the model's inference of the student's knowledge state according to the global question representation. Therefore, another key point to solve this problem is to introduce a more refined personalized modeling strategy, taking into account the local learning tendency, and avoiding the model's simple reliance on global data distribution. By defining the student's local learning tendency, the student's strengths and weaknesses are revealed, helping the model to identify learning needs that are not fully reflected by global bias and make inferences that are more consistent with the student's learning state.

[0060] Although causal inference is widely used in many fields, the related research of integrating it into the knowledge tracking task is still in the initial exploration stage. Therefore, exploring a knowledge tracking method based on counterfactual personalized enhancement using counterfactual inference, personalized modeling and deep learning technology to model the evolution of student knowledge has important research significance and application value for mitigating bias and improving the inference of knowledge state.

[0061] Figure 1 is a flowchart of the knowledge tracking method based on counterfactual personalized enhancement provided by the embodiments of the present application, as Figure 1 shown, the knowledge tracking method based on counterfactual personalized enhancement includes steps 110, 120, 130 and 140.

[0062] Step 110, obtaining student information, questions and answer results, defining knowledge tracking, the knowledge tracking being a probability of predicting a student's correct answer to a question based on the student's answer results to historical questions and the student's information;

[0063] It is easy to understand that the knowledge tracking task dynamically models the change process of the knowledge state of the learner by analyzing the historical learning sequence of the learner, and predicts the mastery degree of the future to a specific knowledge point.

[0064] In the knowledge tracing task, first, student information, questions, and answer results are obtained, and knowledge tracing, global question representation, and local learning tendency are defined. The student information includes student ID.

[0065] Specifically, in an online learning platform, students will leave a large amount of historical data through learning exercises. Each student will answer a series of questions associated with a specific knowledge concept and make a corresponding answer result, thereby promoting knowledge acquisition and mastery.

[0066] For example, assume that there are I different students in a student set, represented as follows:

[0067]

[0068] where, is the ith student.

[0069] There are N different questions in the question set, represented as follows:

[0070]

[0071] where, is the nth question.

[0072] The knowledge concept set is the knowledge concept involved in all questions, represented as follows:

[0073] where,

[0074] is the mth question. A student will answer a question containing a knowledge concept at time step t and get an answer result, which is represented as:

[0075]

[0076] where,

[0077] is the answer result of the question at time step t, 1 for correct answer and 0 for incorrect answer. Therefore, the learning sequence of each student can be regarded as the interaction record with the exercise in multiple time steps, represented as follows:

[0078]

[0079] where,

[0080] is a single question, is the knowledge concept involved in the question, is the answer result of the single question. ​

[0081] It is easy to understand that the evolution of knowledge state is contained in the learning sequence of students, and the knowledge tracking task aims to evaluate the knowledge state of students through the historical learning sequence, and then predict whether they can correctly answer the next question, and the calculation formula is as follows:

[0082]

[0083] It should be noted that knowledge tracking mainly relies on the interaction records between students and problems to obtain knowledge state, but these interaction records lack student information and ignore the differences in knowledge absorption ability of individual students. Because even if there are the same interaction records, due to the differences in cognitive ability, learning strategy, learning motivation and other factors of each student, they will show different knowledge state growth for the same learning content. Therefore, simply using interaction records to model knowledge state may not reflect the true situation, and including student information and interaction records in model construction together can more accurately depict the knowledge state of students. Therefore, the final calculation formula of the knowledge tracking task is as follows:

[0084]

[0085] Global problem representation uses problem difficulty to reflect the degree of global bias. Global bias usually comes from the attributes of the problem itself and the answer mode of high exposure problems in the data, which makes the right and wrong distribution of the problem uneven. If this phenomenon is expressed in statistics, it can usually be expressed by problem difficulty, so here the global problem representation, that is, the problem difficulty, is used to reflect the degree of global bias. Since the problem is related to the knowledge concept, the same problem involves a knowledge point, but the same knowledge point involves different problems, so the difficulty brought by the knowledge point will be different from the difficulty of the problem itself. In addition, the exposure rate also affects the reliability of the evaluation difficulty, if the exposure rate is low, then the difficulty measured by the correct rate of the problem will be unreliable. Therefore, the global problem difficulty is jointly regulated by the problem difficulty and the knowledge point difficulty. According to the Bayesian estimation, first calculate the difficulty of the knowledge concept, and the calculation formula is as follows:

[0086]

[0087] Where, is the number of times the knowledge concept is answered, is the correct rate of the knowledge concept, is the average number of times all knowledge concepts are answered, is the average correct rate of all knowledge concepts, is the knowledge concept, is the difficulty of the knowledge concept.

[0088] Then, the difficulty of the problem is adjusted by the difficulty of the knowledge concept, and the calculation formula of the difficulty of the problem is as follows:

[0089]

[0090] wherein, is the number of times the question is answered, is the accuracy of the answer to the question, is the average number of times all questions are answered, is the difficulty of the knowledge concept corresponding to the question, is the difficulty of the jth question.

[0091] The local learning tendency of a student is dynamically reflected by analyzing the student's answering performance at a time step, thereby reflecting the individualized learning state and progress trend. The calculation of the local learning tendency is based on the student's question answering result and the global question representation. At a time step , the difference between the student's answer result and the global question representation for the question is calculated, and then the cumulative average of the difference at the historical time steps is calculated, which is the local learning tendency of the student at different time steps. The local learning tendency of the student can help to alleviate the global bias, although a certain question shows high difficulty in the dataset, it does not mean that the question has the same challenge for all students; for some students, the question can still be answered correctly, which needs to start from the learning tendency of the student and consider individual differences.

[0092] The local learning tendency can be obtained by combining the student's question answering result and the global question representation. The student's answer result for each question can reflect his learning performance, but due to the different difficulty of the questions, the value reflected by the correct and wrong answers is not completely the same. If a question with high difficulty is answered correctly, it may better indicate that the student performs well, and vice versa, if a question with low difficulty is answered incorrectly, it may indicate that the student still has weaknesses. The local learning tendency needs to reflect the individualized performance of the student, and the answer result itself contains the difficulty information of the question, so the influence of the question itself on the answer result needs to be removed. In addition, the local learning tendency needs to dynamically depict the student's performance at different time steps, so the cumulative performance of the student from the beginning of learning to the current time step needs to be calculated. Specifically, for the question at time step , first, the difference between the answer result and the global question representation is calculated to correct the answering performance, so that it is more in line with the individualized learning situation. Then, the cumulative average of the difference at the historical time steps is calculated, so that it can reflect the dynamic learning trend and avoid the accidental influence of a single answer. The specific formula is as follows:

[0093]

[0094] wherein, is the answer result at time step , is the global question representation, is the answer result at time step a global problem representation of the question answered at the time, t is a local learning tendency, and t is a time step.

[0095] The local learning tendency has certain interpretability. In the case of answering correctly, the local learning tendency gives higher positive feedback if the question is difficult. In the case of answering incorrectly, the negative impact of the error is greater if the question is less difficult. The local learning tendency can more fairly evaluate the individual performance of students' learning state and progress trend.

[0096] In step 120, a knowledge tracking causal diagram is constructed based on the knowledge tracking, the knowledge tracking causal diagram includes a node set and a path set, the node set includes a student, a question, a knowledge state, and an answer result, and the path set includes a learning tendency driving path, a question representation influencing path, and a knowledge state mapping path.

[0097] Figure 2 A structural schematic diagram of the knowledge tracking causal diagram provided by the embodiments of the present application is shown in Figure 2 The knowledge tracking causal diagram includes a node set and a path set, the node set includes a student, a question, a knowledge state, and an answer result, wherein the node S represents a student, the node Q represents a question, the node H represents a student knowledge state, and the node Y represents a student answer result.

[0098] It should be noted that most knowledge tracking models are based on the causal diagram in (a) in Figure 2 , but the causal diagram in (b) in Figure 2 can more accurately express the model prediction process. Specifically, the student and the question themselves can also directly affect the prediction result. The former covers the individual differences of the student, and its own characteristics can directly affect the answer result. For example, a student with strong learning ability may answer correctly even if he or she faces a difficult problem. It should be noted that the difference between the learning ability of the student and the knowledge state is that the knowledge state is local and only reflects the student's mastery of certain knowledge points. The learning ability is diverse, such as cognitive ability, which has exceeded the scope of the knowledge state. The latter contains the attribute characteristics of the question, because a question that is too simple or too difficult may directly lead to the student answering correctly or incorrectly without relying too much on the student's knowledge state.

[0099] In some embodiments, the causal diagram in (b) in Figure 2 is adopted, the path set includes a learning tendency driving path S-Y, a question representation influencing path Q-Y, and a knowledge state mapping path S and Q-H-Y, and the knowledge state mapping path can be decomposed into a capability question matching path S and Q-H and an answer result prediction path H-Y.

[0100] Step 130, constructing an answer result prediction model based on the knowledge tracking causal diagram, the answer result prediction model including a learning tendency driven path submodel, a knowledge state mapping path submodel, a question representation influence path submodel, and a fusion output submodel;

[0101] Further, constructing an answer result prediction model according to the knowledge tracking causal diagram, specifically including the following steps:

[0102] (1) Embedding representation: the input elements of the answer result prediction model are student information, questions, answer results, global question representation, and local learning tendency, wherein the global question representation and the local learning tendency use continuous embedding representation, and other elements use embedding matrix for embedding representation, and the learning sequence is regarded as the splicing of question / knowledge concept and answer response embedding.

[0103] (2) Constructing a learning tendency driven path submodel: used for describing the learning tendency driven path in the causal diagram, the local learning tendency embedding at time step t is input into the module, and two fully connected networks are used to obtain the local learning tendency state, and finally a multilayer perceptron is used to obtain the prediction result.

[0104] (3) Constructing a knowledge state mapping path submodel: used for describing the knowledge state mapping path in the causal diagram, the existing knowledge tracking model generally follows this path to predict the answer result, and the knowledge state mapping path submodel is constructed based on the existing knowledge tracking model, and the student embedding and the interaction sequence embedding are input, the knowledge state is obtained through the state updating layer, and the output result is obtained through the mapping layer.

[0105] (4) Constructing a question representation influence path submodel: used for describing the question representation influence path in the causal diagram. The global question representation embedding at time step t+1 is input into the module, two fully connected networks are used to obtain the question representation state, and finally a multilayer perceptron is used to obtain the prediction result.

[0106] (5) Constructing a fusion output submodel: the answer result is jointly affected by the learning tendency driven path, the knowledge state mapping path, and the question representation influence path, so firstly the state outputs of the learning tendency driven path submodel, the knowledge state mapping path submodel, and the question representation influence path submodel are fused through splicing to obtain a new knowledge mastery state, and then a multilayer perceptron is used to obtain the answer result prediction value.

[0107] Step 140, calculating the path total effect, the path direct effect, and the path indirect effect based on the knowledge tracking causal diagram;

[0108] The path total effect is the total effect of the student and the question on the answer result. The path total effect is the comparison of the answer result in the real world with the result of the variable in the counterfactual world in the reference state (the state not affected by a specific intervention or influence), which is used to measure the change in the answer result in the two states.

[0109] The direct effect in the causal diagram includes the effect of the student and the question on the answer result through the direct path. In the case where the intermediate variable of the student's knowledge state is blocked, the natural direct effect caused by the change of the independent variable can be calculated, and the comprehensive direct effect of the learning tendency driving path and the question representation influencing path on the answer result can be obtained through joint analysis.

[0110] The knowledge state mapping path includes the intermediate variable of the knowledge state, and the student and the question indirectly affect the answer result through the intermediate variable of the knowledge state, that is, the indirect effect of the student and the question on the answer result. The total effect can be decomposed into the natural direct effect and the total indirect effect.

[0111] Step 150, based on the path total effect, the path direct effect, the path indirect effect, and the answer result prediction model, obtaining a counterfactual prediction debiasing formula of the student's answer result for the question;

[0112] It is worth noting that the student's local learning tendency helps to reveal the influence of individual differences on the learning process, thereby alleviating the global bias, so on the basis of the indirect effect, by adjusting the direct effect of the path, a debiasing formula integrating the student's local learning tendency is obtained, and according to the formula, the counterfactual prediction debiasing formula of the student's answer result for the question is obtained in combination with the implementation mode of each path.

[0113] Step 160, based on the counterfactual prediction debiasing formula, obtaining a counterfactual prediction value of the student's answer result for the question.

[0114] According to the knowledge tracking method based on counterfactual personalized enhancement provided in the embodiments of the present application, by obtaining student information, questions and answer results, and defining a knowledge tracking question based thereon, combining the counterfactual prediction debiasing formula, the path total effect, the path direct effect and the path indirect effect of the knowledge tracking causal diagram are calculated, the deep learning and causal inference technology are comprehensively utilized to explore the student's local learning tendency and global question representation, a method capable of supporting different knowledge tracking models is designed, the student's local learning tendency and global question representation are captured, and the global bias in the knowledge tracking task is effectively alleviated by using counterfactual inference, and the student's personalized performance is enhanced.

[0115] In some embodiments, the learning tendency driven path is a student-answer result, used to describe the direct influence of student cognition on the answer result, the question representation influence path is a question-answer result, used to describe the direct influence of the question itself on the answer result, and the knowledge state mapping path is a student and question-knowledge state-answer result.

[0116] It is easy to understand that the learning tendency driven path is the direct influence of the student on the answer result, and a high-ability student can flexibly use the learned knowledge points to improve the accuracy of the problem answer, and a low-ability student may lack strategies to affect the answer performance.

[0117] The question representation influence path is the direct influence of the question itself on the answer result, and a question that is too simple or too difficult may directly lead to the student's correct or incorrect answer, resulting in global bias, so that the model can better predict the answer result by only memorizing the characteristics of the question.

[0118] In this embodiment, by inputting the student information, the question, the answer result, the global question representation, and the local learning tendency into the learning tendency driven path, the question representation influence path, and the knowledge state mapping path, the prediction value of the student's answer result for the next question is obtained. The comprehensive analysis of the student's individual cognitive level, the question characteristics, and the knowledge state is realized, the accuracy of the student's answer result prediction is improved, the personalized education and the question design are provided with strong support, and the teaching effect and the learning experience are optimized.

[0119] In some embodiments, the knowledge state mapping path includes a capability question matching path and an answer result prediction path, the capability question matching path is a student and question-knowledge state, used to describe the determining effect of the matching degree of the student's ability and the question on the knowledge state, and the answer result prediction path is a knowledge state-answer result, used to describe the prediction effect of the knowledge state on the answer result.

[0120] In some embodiments, the capability question matching path is that the student and the question jointly determine the knowledge state, and the higher the matching degree of the student's ability and the question, the higher the degree of the student's mastery of the question. The answer result prediction path makes a prediction of the answer result through the knowledge state.

[0121] In this embodiment, the knowledge state mapping path is composed of the capability question matching path and the answer result prediction path, which can more comprehensively and accurately evaluate the student's knowledge mastery, and enhance the student's individual performance.

[0122] In some embodiments, the learning tendency driven path sub-model includes a first fully connected network and a first multi-layer perceptron, which are used to obtain the learning tendency driven path prediction results, the knowledge state mapping path sub-model is a knowledge tracking model, the problem representation influence path sub-model includes a second fully connected network and a second multi-layer perceptron, which are used to obtain the problem representation influence path prediction results, and the fusion output sub-model includes a third multi-layer perceptron, which is used to obtain the predicted value of the student's answer to the question.

[0123] Figure 3 This is a diagram of the knowledge tracking framework provided by the embodiment of the present application, such as Figure 3 As shown in the figure, the elements are first embedded and represented. Then, the answer result prediction model is constructed according to the path in the causal graph, including the learning tendency driving path sub-model, the knowledge state mapping path sub-model, the question representation influence path sub-model and the fusion output sub-model. Finally, the intermediate output states of the learning tendency driving path sub-model, the knowledge state mapping path sub-model and the question representation influence path sub-model are used to obtain the predicted value of the question answer result through the fusion output sub-model. The specific process is as follows:

[0124] (1) Embedded representation

[0125] The input elements of the framework are student information, questions, answer results, global problem representation and local learning tendency, and each input element is embedded. Specifically, the student embedding is achieved through the embedding matrix (in is the embedding dimension), the problem is embedded through the embedding matrix (in is the embedding dimension), knowledge concept embedding is represented by the embedding matrix (in is the embedding dimension), the answer response is embedded through the embedding matrix (in is the embedding dimension). The global problem representation embedding and the local learning tendency embedding are represented by continuous embedding, specifically and ,in and is a trainable vector.

[0126] Given a learning sequence, each interaction can be viewed as a concatenation of questions / knowledge concepts and answer responses, i.e. The choice of questions and knowledge concepts depends on the requirements of the knowledge tracking model.

[0127] (2) Learning tendency-driven path sub-model

[0128] It is easy to understand that the model is constructed based on the learning tendency driving path and is used to model students' personalized local learning tendencies. The prediction result of the learning tendency driving path is defined as ,in Structural equations for modeling the pathways driving learning dispositions.

[0129] To more accurately model the personalized dynamic learning process, the time step Locally learned tendency embedding on express Specifically, we first use a simple network layer to capture the student's local learning tendency. This process uses a linear layer, uses the ReLU activation function, and then outputs it through another linear layer to obtain the local learning tendency state. The calculation formula for the local learning tendency state is as follows:

[0130] ,

[0131] in, is the local learning tendency state, and is a trainable matrix, and is the trainable bias vector.

[0132] Then, a multi-layer perceptron (MLP) is used to perform nonlinear mapping on the local learning tendency state, and the prediction result output is: .

[0133] (3) Knowledge state mapping path sub-model

[0134] It is easy to understand that this model corresponds to a knowledge state mapping path, and the output of this path is ,in It is a structural equation about the knowledge state update, It is a structural equation about the output. Existing knowledge tracking models generally follow this path for prediction, so this path is built based on existing knowledge tracking models. For example, for the deep knowledge tracking model DKT, using student embedding express , using the learning interaction sequence express , and use both as input to the state update module (RNN, LSTM) to obtain the updated knowledge state , and then get the output result through the mapping layer.

[0135] (4) Problem representation impact path sub-model

[0136] The problem characterization is defined as the impact path prediction result: ,in To model the problem representation influence path, a structural equation is used to characterize the path of influence. Similar to the learning propensity driven path, this module will embed the global problem representation at time step is computed as follows:

[0137]

[0138] where is the problem representation state, and are trainable matrices, and are trainable bias vectors.

[0139] Then, a multi-layer perceptron (MLP) is used to obtain the predicted output, computed as follows:

[0140]

[0141] In this embodiment, by fusing the intermediate output states of the learning propensity driven path submodel, the knowledge state mapping path submodel, and the problem representation influence path submodel, and using a multi-layer perceptron for the final prediction, the output information of different path models is combined, the understanding of the student learning state is improved, thus achieving more accurate personalized learning prediction, optimizing education decision support, and further improving the accuracy of learning effect evaluation.

[0142] In some embodiments, the calculation formula of the fusion output submodel is as follows:

[0143]

[0144] where is the fusion state, is the fusion function, is the intermediate output state of the learning propensity driven path submodel, is the knowledge mastery state of the knowledge state mapping path submodel, is the intermediate output state of the problem representation influence path submodel, denotes concatenation, is a trainable matrix, is a trainable bias vector, is the predicted value of the problem answer result, is the calculated value of the real-world problem answer result, and MLP is a third multi-layer perceptron.

[0145] It is easy to understand that, from the perspective of causality, Figure 2 ​​​The answer result in (b) is jointly affected by the learning tendency driving path, the knowledge state mapping path, and the problem representation influencing path. The state outputs of the learning tendency driving path module, the knowledge state mapping path module, and the problem representation influencing path module are 、 、 The goal of the three-path fusion is to predict the probability of answering the question correctly at the next time step, so it is necessary to use the time step on the three paths Time , used to predict the time step When answering the result and time steps The above question , and fuse them. According to the ancestor node, define the time step superior The calculation formula is as follows:

[0146]

[0147] in, is the structural equation, is the fusion function.

[0148] It is worth noting that in order to ensure the clarity of the symbol expression, the uniform use of and Represents related variables in different paths, although they are represented differently in different paths. There are many ways to construct the fusion function. Here, the three states are concatenated and output: the predicted value is output through a multi-layer perceptron.

[0149] In some embodiments, the first multi-layer perceptron, the second multi-layer perceptron, and the third multi-layer perceptron are the same multi-layer perceptron.

[0150] In this embodiment, the fusion output sub-model integrates the intermediate output states of the learning tendency driving path sub-model, the knowledge state mapping path sub-model and the problem representation influencing path sub-model, combines the trainable matrix and the bias vector, and uses a multi-layer perceptron to make the final prediction. By fusing the output information of different paths, the accuracy of the answer result prediction model is improved.

[0151] In some embodiments, the calculation formula of the counterfactual prediction debiasing formula is as follows:

[0152]

[0153] in, is the predicted value of the answer to the question, Predicting outcomes for learning propensity-driven pathways, The problem characterization affects the path prediction results, TIE is the path indirect effect, NDE is the path direct effect, The weight of the predicted results of the learning tendency driving path, The weight that affects the path prediction result for the problem representation, is a hyperparameter, is the counterfactual prediction value of the student's answer to the question, S is the student, Q is the question, and Y is the student's answer result. is the local learning tendency, t is the time step, is the global question representation of the question answered at time step t+1.

[0154] In some embodiments, expressions for the total path effect, direct effect, and indirect effect are obtained based on causal graph analysis. Then, by controlling the direct effect and combining the fusion output sub-model, a counterfactual prediction debiasing formula for answering the question is obtained. The specific process is as follows:

[0155] (1) Calculate the total path effect

[0156] The total effect is the change in the dependent variable when the independent variable moves from a reference state to a target state in causal inference. In this causal diagram, the total effect of the path is and right The total effect is the real world Results The variables in the counterfactual world are in the reference state Results The comprehensive reflection of the differences, the calculation formula of the total effect is as follows,

[0157]

[0158] Where TE represents the total path effect, Computational values ​​for answering real-world questions, The calculated value of the answer to the counterfactual world question.

[0159] It should be noted that , and It can be understood as the reference value of a variable, used to represent a hypothetical state. The total effect TE can be understood as the difference between the expected value and the reference value, that is, the total impact of the independent variable on the dependent variable.

[0160] (2) Calculation of direct effects

[0161] exist Figure 2In the causal diagram shown in (b) of FIG. 1, the direct effect, i.e., the effect of the student and the problem on the answer result through the learning tendency driving path and the problem representation influencing path, respectively, is referred to as the natural direct effect. The natural direct effect represents the influence of the independent variable on the dependent variable through the direct path. Specifically, this effect reflects the amount of change in the dependent variable directly caused by the change in the independent variable when the mediating variable is blocked. For example, in the causal diagram shown in (b) of FIG. 1, the natural direct effect of the student on the answer result is the effect of the student on the answer result through the student-answer result path, i.e., the effect of the student on the answer result when the knowledge state is blocked. Figure 2 In the causal diagram shown in (c) of FIG. 1, when the mediating variable knowledge state is fixed to and , the calculation formula of the direct effect is as follows:

[0162]

[0163] where NDE is the direct effect, is the calculated value of the counterfactual world problem answer result without considering the knowledge state.

[0164] The direct effect estimates the effect produced in the counterfactual world when only the student learning tendency and the problem representation directly act on the answer result.

[0165] (3) Calculation of indirect effect

[0166] According to the causal diagram structure, the total effect can be decomposed into the natural direct effect and the total indirect effect, and the total indirect effect represents the influence on the dependent variable through the mediating variable. Therefore, for (b) of FIG. 1, the total effect of the student and the problem on the answer result can be decomposed into the direct effect of the student and the problem on the answer result under the action of the direct path student-answer result and problem-answer result, and the indirect effect of the student and the problem on the answer result through the mediating variable knowledge state under the action of the knowledge state mapping path. The calculation formula between the total effect, the direct effect, and the indirect effect is as follows: Figure 2

[0167]

[0168] where TIE is the indirect effect.

[0169] To alleviate the global bias in knowledge tracing, the ultimate goal is to achieve bias alleviation by changing the causal effect of the two direct paths, i.e., the student-answer result path and the problem-answer result path. First, the causal effect other than the two paths, i.e., the indirect effect, needs to be obtained. The calculation formula of the indirect effect of the knowledge state mapping path is as follows:

[0170]

[0171] (4) Calculation of counterfactual prediction debiasing formula for answering problems

[0172] ​The global bias of knowledge tracking can be considered as the influence of the problem-answer outcome path, while the student-answer outcome path can be considered as the effect of the student's local learning tendency. The global bias ignores the individual behavior characteristics of students due to the over-reliance on the general behavior characteristics of all students. The local learning tendency focuses on the individualized understanding of students on the question, rather than relying on the overall data distribution. Using the local learning tendency of students helps to reveal the influence of individual differences on the learning process, thereby reducing the impact of global bias. Therefore, it is necessary to control the direct effect and retain the positive signal belonging to the local learning tendency of students. The counterfactual prediction based on the initial calculation formula of the debiasing formula is as follows:

[0173]

[0174] wherein, is a hyperparameter for controlling the influence of NDE on the prediction of future question performance.

[0175] Further, in combination with the fusion output formula, the on the time step can be represented as:

[0176]

[0177] It is worth noting that, since the knowledge state variable is in the reference state, it can be considered as a constant term, and its influence on the prediction is consistent, which can be directly ignored. The above formula can be simplified as:

[0178]

[0179] For , wherein , , all three variables are in the reference state, i.e. fixed and not affected by other changes. Therefore, the causal effects of , and are assumed to be constant, representing a baseline case. Under this setting, is considered as a constant term, which can be directly ignored.

[0180] Finally, the calculation formula of the counterfactual prediction based on the debiasing formula is as follows:

[0181]

[0182] It should be noted that and are adjustable hyperparameters, which adjust the influence of the student's local learning tendency and the global problem representation to better achieve bias mitigation.

[0183] In this embodiment, by using the counterfactual prediction debiasing formula to calculate the problem answer result, combining the prediction results of the learning tendency driving path and the problem representation influencing path, adjusting the direct effects of the learning tendency driving path and the problem representation influencing path, and combining the indirect effects of the knowledge state mapping path, the effect control counterfactual prediction can alleviate the global bias in the knowledge tracking task, improve the prediction accuracy of the student's problem answer result, and help more accurately analyze the student's knowledge mastery, providing a more reliable basis for the development of personalized learning strategies.

[0184] In some embodiments, the training process of the answer result prediction model comprises:

[0185] constructing a preset answer result prediction model;

[0186] obtaining different student information, different problems and different answer results as a data set;

[0187] training the preset answer result prediction model based on the data set based on a loss function, wherein the loss function is constructed based on a binary cross-entropy loss function.

[0188] In some embodiments, the training of the answer result prediction model is a multi-task learning training. To evaluate the prediction effect of the model, a suitable knowledge tracking public data set is selected for training and prediction. These public data sets, such as ASSISTments2009, ASSISTments2017, etc., contain basic information such as student ID, knowledge concept ID, problem ID, and answer result. The data set is divided into a training verification set and a test set in the ratio of 8:2, and then the training verification set is divided into a training set and a verification set in the ratio of 8:2. Traditional evaluation strategies usually predict based on randomly selected test data, but since these data still exhibit global bias characteristics, they cannot accurately reflect the prediction ability of the model after causal inference. To test the debiasing ability of the model, the test set needs to be stratified sampled according to the problem category. Specifically, the correct and incorrect interactions in each problem category in the test set are sampled so that the number of correct and incorrect problems is equal. To ensure that the size of the entire test set does not change, some problem interactions may be repeated or lost. Note that the training set is not debiased.

[0189] Further, the prediction process of the causal graph shown in (b) of the Figure 2 The loss function of this task is calculated as follows:

[0190]

[0191] wherein, is the first loss function.

[0192] Then, for learning tendency driving path prediction task and problem representation influence path prediction task, both use binary cross entropy loss function to measure prediction accuracy, and the calculation formula is as follows:

[0193]

[0194] wherein, is the second loss function, is the third loss function.

[0195] Based on the first loss function, the second loss function and the third loss function, the loss function of the answer result prediction model is obtained, and the calculation formula is as follows:

[0196]

[0197] wherein, is the first loss function, is the second loss function, is the third loss function, is the first hyperparameter, is the second hyperparameter.

[0198] In this embodiment, by constructing a preset answer result prediction model and training based on different student information, different problems and different answer results as a data set, an answer result prediction model is obtained. The binary cross entropy loss function is used for optimization to improve the prediction accuracy, thereby providing support for the development of individualized learning scheme, improving the learning effect of students and reducing the waste of teaching resources.

[0199] The knowledge tracking method based on counterfactual individualized enhancement provided in the embodiments of the application can be executed by a knowledge tracking device based on counterfactual individualized enhancement. In the embodiments of the application, the knowledge tracking device based on counterfactual individualized enhancement is taken as an example to illustrate the knowledge tracking device based on counterfactual individualized enhancement provided in the embodiments of the application.

[0200] The embodiments of the application also provide a knowledge tracking device based on counterfactual individualized enhancement, as shown in Figure 4 The knowledge tracking device based on counterfactual individualized enhancement includes an acquisition module 410, a first construction module 420, a second construction module 430, a first calculation module 440, a second calculation module 450 and a prediction module 460.

[0201] The acquisition module 410 is configured to acquire student information, questions and answer results, and define a knowledge tracking based on the answer results of the students to historical questions and the student information, to predict a probability of the students to answer a question correctly

[0202] The first construction module 420 is configured to construct a knowledge tracking causal diagram based on the knowledge tracking, the knowledge tracking causal diagram including a node set and a path set, the node set including students, questions, knowledge states and answer results, and the path set including a learning tendency driven path, a question representation influenced path and a knowledge state mapped path.

[0203] The second construction module 430 is configured to construct an answer result prediction model based on the knowledge tracking causal diagram, the answer result prediction model including a learning tendency driven path submodel, a knowledge state mapped path submodel, a question representation influenced path submodel and a fusion output submodel.

[0204] The first calculation module 440 is configured to calculate a total path effect, a direct path effect and an indirect path effect based on the knowledge tracking causal diagram.

[0205] The second calculation module 450 is configured to obtain an anti-factual prediction debiasing formula of the students to answer results of the questions based on the total path effect, the direct path effect, the indirect path effect and the answer result prediction model.

[0206] The prediction module 460 is configured to obtain an anti-factual prediction value of the students to answer results of the questions based on the anti-factual prediction debiasing formula.

[0207] The knowledge tracking method based on anti-factual personalized enhancement provided by the embodiments of the present application can acquire student information, questions and answer results, define a knowledge tracking question based thereon, combine an anti-factual prediction debiasing formula, calculate a total path effect, a direct path effect and an indirect path effect of a knowledge tracking causal diagram, comprehensively utilize deep learning and causal inference technology to explore local learning tendencies of the students and global question representations, design a method capable of supporting different knowledge tracking models, capture the local learning tendencies of the students and the global question representations, and effectively alleviate global bias in a knowledge tracking task by using anti-factual inference, thereby enhancing personalized performance of the students.

[0208] The knowledge tracking apparatus based on anti-factual personalized enhancement provided by the embodiments of the present application can realize each process of the knowledge tracking method based on anti-factual personalized enhancement Figures 1 to 3 embodied by the embodiments of the present application, and thus details are not repeated here.

[0209] In some embodiments, as Figure 5As shown, the electronic device 500 includes a processor 501, a memory 502, and a computer program stored in the memory 502 and executable on the processor 501. The computer program is executed by the processor 501 to implement each process of the knowledge tracking method based on counterfactual personalization enhancement, and achieve the same technical effects. To avoid repetition, details are not described herein.

[0210] It should be noted that the electronic device in the embodiments of the present application includes the mobile electronic device and the non-mobile electronic device described above.

[0211] The embodiments of the present application also provide a non-transitory computer readable storage medium having a computer program stored thereon. The computer program is executed by the processor to implement each process of the knowledge tracking method based on counterfactual personalization enhancement, and achieve the same technical effects. To avoid repetition, details are not described herein.

[0212] The processor is the processor in the electronic device in the above embodiments. The readable storage medium includes a computer readable storage medium, such as a computer readable memory (Read-Only Memory, ROM), a random access memory (Random Access Memory, RAM), a magnetic disk or an optical disk, etc.

[0213] The embodiments of the present application also provide a computer program product including a computer program. The computer program is executed by the processor to implement the knowledge tracking method based on counterfactual personalization enhancement.

[0214] The processor is the processor in the electronic device in the above embodiments. The readable storage medium includes a computer readable storage medium, such as a computer readable memory (Read-Only Memory, ROM), a random access memory (Random Access Memory, RAM), a magnetic disk or an optical disk, etc.

[0215] The embodiments of the present application also provide a chip including a processor and a communication interface. The communication interface is coupled to the processor. The processor is configured to execute a program or an instruction to implement each process of the knowledge tracking method based on counterfactual personalization enhancement, and achieve the same technical effects. To avoid repetition, details are not described herein.

[0216] It should be understood that the chip mentioned in the embodiments of the present application can also be referred to as a device-level chip, a device chip, a chip device, or a system-on-chip device, etc.

[0217] It should be noted that, in the present document, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises a", "comprising", or "includes a", does not, without more constraints, preclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the element. Additionally, it should be noted that the scope of the methods and apparatus of the present embodiments are not limited by the order of the steps or the sequence for performing the steps, as some steps can occur simultaneously, in other steps can occur sequentially, or in between other steps can occur, as can be readily appreciated by those skilled in the art. Further, the features of the various examples described can be combined in other examples.

[0218] From the above description of the embodiments, it is clear that the above-described method of the embodiments can be realized by means of software and the necessary universal hardware platform, of course, it can also be realized by hardware, but in many cases, the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a computer software product in essence or in the form of a part of the prior art that makes a contribution. The computer software product is stored in a storage medium (such as ROM / RAM, magnetic disc, optical disc), and includes a plurality of instructions for causing a terminal (which can be a mobile phone, computer, server, or network device, etc.) to execute the knowledge tracking method based on counterfactual personalized augmentation of each embodiment of the present application.

[0219] In the description of the present application, "first feature" and "second feature" can include one or more of the features.

[0220] In the description of the present application, "a plurality of" means two or more.

[0221] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-described specific embodiments, and the above-described specific embodiments are merely illustrative and not limiting. Those skilled in the art can make many forms under the inspiration of the present application without departing from the scope of the present application and the scope of protection of the claims.

[0222] In the description of the application, reference has been made to descriptive terms such as "one embodiment", "some embodiments", "an embodiment", "example", "specific example" or "some examples" etc. It is emphasized that each of these terms refers to a specific feature, structure, material or characteristic described in connection with a particular embodiment or example. The descriptive terms are not necessarily referring to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0223] While the embodiments of the application have been shown and described, it is to be understood that the embodiments described are only by way of example and that various changes, modifications, substitutions and alterations can be made thereto without departing from the spirit and scope of the application as set forth in the claims and equivalents thereof.

Claims

1. A knowledge tracking method based on counterfactual personalized enhancement, characterized in that: The method comprises: Obtain student information, questions, and answers, and define knowledge tracking, which is to predict the probability of a student's answer to a question being correct based on the student's answers to historical questions and the student information; Constructing a knowledge tracking causal graph based on knowledge tracking, wherein the knowledge tracking causal graph includes a node set and a path set, wherein the node set includes students, questions, knowledge states, and answer results, and the path set includes a learning tendency driving path, a question representation influencing path, and a knowledge state mapping path; Constructing an answer result prediction model based on the knowledge tracing causal graph, the answer result prediction model includes a learning tendency driving path sub-model, a knowledge state mapping path sub-model, a question representation impact path sub-model and a fusion output sub-model; Calculate the total path effect, the direct path effect, and the indirect path effect based on the knowledge tracing causal graph; Based on the total path effect, the direct path effect, the indirect path effect and the answer result prediction model, a counterfactual prediction debiasing formula for the student's answer result to the question is obtained; Based on the counterfactual prediction debiasing formula, obtaining the counterfactual prediction value of the student's answer to the question; The calculation formula of the counterfactual prediction debiasing formula for the student's answer to the question is as follows: ； in, is the predicted value of the answer to the question, Predicting outcomes for learning propensity-driven pathways, The problem characterization affects the path prediction results, TIE is the path indirect effect, NDE is the path direct effect, The weight of the predicted results of the learning tendency driving path, The weight that affects the path prediction result for the problem representation, is a hyperparameter, is the counterfactual prediction value of the student's answer to the question; MLP is the third multi-layer perceptron, is the local learning tendency state, and is a trainable matrix, and is the trainable bias vector; To learn local tendency embedding representation, is a trainable vector; is the time step The answer result when is the time step The global problem representation of the question answered when is the local learning tendency, t is the time step; The learning tendency driving path is student-answer result, which is used to describe the direct impact of student cognition on the answer result. The problem representation influence path is question-answer result, which is used to describe the direct impact of the question's own attributes on the answer result. The knowledge state mapping path is student and question-knowledge state-answer result.

2. The knowledge tracking method based on counterfactual personalized enhancement according to claim 1 is characterized in that: The knowledge state mapping path includes an ability-problem matching path and an answer result prediction path. The ability-problem matching path is student and problem-knowledge state, which is used to describe the decisive effect of the matching degree between student ability and problem on the knowledge state. The answer result prediction path is knowledge state-answer result, which is used to describe the predictive effect of knowledge state on answer result.

3. The knowledge tracking method based on counterfactual personalized enhancement according to claim 1 is characterized in that: The learning tendency driven path sub-model includes a first fully connected network and a first multi-layer perceptron, which are used to obtain the learning tendency driven path prediction results. The knowledge state mapping path sub-model is a knowledge tracking model. The problem representation influence path sub-model includes a second fully connected network and a second multi-layer perceptron, which are used to obtain the problem representation influence path prediction results. The fusion output sub-model includes a third multi-layer perceptron, which is used to obtain the predicted value of the student's answer to the question.

4. The knowledge tracking method based on counterfactual personalized enhancement according to claim 3 is characterized in that: The calculation formula of the fusion output sub-model is as follows: ； in, In fusion state, is the fusion function, is the intermediate output state of the learning tendency driving pathway sub-model, The knowledge mastery state of the knowledge state mapping path sub-model, The intermediate output state of the problem representation impacts the path sub-model, Indicates splicing, is a trainable matrix, is the trainable bias vector, is the predicted value of the answer to the question, It is the calculation value of the answer to the real-world question, and MLP is the third multi-layer perceptron.

5. The knowledge tracking method based on counterfactual personalized enhancement according to claim 1 is characterized in that: The training process of the answer result prediction model includes: Construct a prediction model for preset answer results; Obtain different student information, different questions, and different answer results as a data set; Based on the loss function, a preset answer result prediction model is trained according to the data set to obtain the answer result prediction model, wherein the loss function is constructed based on the binary cross entropy loss function.

6. A knowledge tracking device based on counterfactual personalized enhancement, implemented using the knowledge tracking method based on counterfactual personalized enhancement according to any one of claims 1 to 5, characterized in that: The device comprises: An acquisition module is used to obtain student information, questions, and answer results, and define knowledge tracking, which is to predict the probability of a student's answer to a question being correct based on the student's answer results to historical questions and student information; A first construction module is configured to construct a knowledge tracing causal graph based on knowledge tracing, wherein the knowledge tracing causal graph includes a node set and a path set, wherein the node set includes students, questions, knowledge states, and answer results, and the path set includes a learning tendency driving path, a question representation influencing path, and a knowledge state mapping path; A second construction module is configured to construct an answer result prediction model based on the knowledge tracing causal graph, wherein the answer result prediction model includes a learning tendency driving path sub-model, a knowledge state mapping path sub-model, a question representation impact path sub-model, and a fusion output sub-model; A first calculation module is used to calculate the total effect of the path, the direct effect of the path and the indirect effect of the path based on the knowledge tracing causal graph; A second calculation module is used to obtain a counterfactual prediction debiasing formula for the student's answer to the question based on the total path effect, the direct path effect, the indirect path effect and the answer result prediction model; The prediction module is used to obtain the counterfactual prediction value of the student's answer to the question based on the counterfactual prediction debiasing formula.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the knowledge tracking method based on counterfactual personalized enhancement is implemented as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the knowledge tracking method based on counterfactual personalized enhancement is implemented as described in any one of claims 1 to 5.

Images