A deep knowledge tracking and interpretation method and device

Abstract

The present application provides a deep knowledge tracking and interpretation method and device, which relates to the field of intelligent guidance analysis. The method includes: obtaining the historical answer records and questions to be predicted of the target students, and constructing the learning interaction sequence corresponding to the target students according to the answer time sequence based on the historical answer records; inputting the learning interaction sequence into the knowledge tracking model to output the answer probability prediction value of the target students who correctly answer the questions to be predicted; based on the preset strategy, pre-screening operations are performed on the learning interaction sequence to obtain candidate subsequences; combining the educational constraint method to calculate the causal contribution of the candidate subsequence to the answer probability prediction value, the educational constraint method includes the cognitive load reduction method and the time decay weighted method; according to the causal contribution, a multi-objective particle swarm optimization algorithm is used to generate the target explanation subsequence, and the answer probability prediction value is tracked and interpreted through the target explanation subsequence. The present application solves the problem that the DLKT model cannot intuitively present which key interaction pairs mainly drive the probability prediction of students correctly answering a certain question.

Description

Technical Field

[0001] The present application relates to the field of intelligent learning guidance analysis, and in particular to a method and device for deep knowledge tracking and interpretation. Background Art

[0002] Knowledge tracing technology is widely used in education to analyze students' mastery of specific knowledge points during the learning process. Advances in deep learning technology have enabled KT models to fit complex nonlinear relationships. Deep knowledge tracing models (DLKT), based on deep learning technology, utilize neural networks to effectively model students' historical learning behaviors and predict future responses. However, the internal state and prediction processes of these models are often black boxes, making it difficult to explain why the model makes a certain prediction. Users cannot verify the logical rationality of the model's output, and if the model exhibits implicit bias (such as underestimating the abilities of certain student types), it is difficult to trace the cause. Therefore, interpretability is a prerequisite for the implementation of deep knowledge tracing models in educational scenarios.

[0003] Existing interpretable approaches to DLKT models fall into two general categories: one calculates the contribution of individual interaction pairs within a learning interaction sequence, thereby ignoring the joint effects between interaction pairs; the other seeks to maximize the overall causal effect of the learning interaction sequence, thereby ignoring internal factor redundancy and including a large number of irrelevant, spurious, or redundant factors. In reality, complex interactions exist between interaction pairs. For example, two interaction pairs containing similar information can lead to redundancy. Redundant factors cannot provide an explanatory effect and may instead provide misleading information. In other words, the DLKT model fails to intuitively identify key interaction pairs and primarily drives predictions about the probability of a student correctly answering a question.

[0004] Therefore, there is an urgent need for a deep knowledge tracking and interpretation method and device. Summary of the Invention

[0005] This application provides a deep knowledge tracing interpretation method and device, which solves the problem that the DLKT model cannot intuitively present which key interaction pairs, and mainly drives the probability prediction of students correctly answering a certain question.

[0006] In a first aspect of the present application, a deep knowledge tracking and interpretation method and device are provided, the method comprising: obtaining the historical answer records and questions to be predicted of a target student, and constructing a learning interaction sequence corresponding to the target student according to the answer time sequence based on the historical answer records; inputting the learning interaction sequence into a knowledge tracking model to output a predicted value of the answer probability of the target student correctly answering the question to be predicted; based on a preset strategy, performing a pre-screening operation on the learning interaction sequence to obtain a candidate subsequence; calculating the causal contribution of the candidate subsequence to the answer probability prediction value in combination with an educational constraint method, the educational constraint method including a cognitive load reduction method and a time decay weighted method; based on the causal contribution, using a multi-objective particle swarm optimization algorithm to generate a target explanation subsequence, and tracking and interpreting the answer probability prediction value through the target explanation subsequence.

[0007] Optionally, based on knowledge in the field of pedagogy, a pre-screening operation is performed on the learning interaction sequence to obtain candidate subsequences, specifically including: using particle swarm encoding to assign corresponding particles to each subsequence in the learning interaction sequence; obtaining a target candidate subsequence, where the target candidate subsequence is any subsequence in multiple learning interaction sequences; obtaining the position vector of the particle corresponding to the target candidate subsequence, and confirming whether the values ​​of each dimension of the position vector meet the preset screening conditions; if the values ​​of each dimension of the position vector meet the preset screening conditions, then taking the target candidate subsequence as the candidate subsequence.

[0008] Optionally, the preset strategy includes a hidden state observation-based generation strategy and a gradient prior-based generation strategy.

[0009] Optionally, the causal contribution of the candidate subsequence to the answer probability prediction value is calculated in combination with the educational constraint method, specifically including: calculating the predictive ability score corresponding to the explanation subsequence, the predictive ability score is used to evaluate the degree of predictive influence of the interactive information contained in the explanation subsequence on the answer probability prediction value; calculating the sufficiency score and necessity score corresponding to the explanation subsequence, the sufficiency score and necessity score are used to evaluate the internal information redundancy in the explanation subsequence; based on the predictive ability score, sufficiency score and necessity score, and based on the time exponential decay function, the causal contribution of the explanation subsequence to the answer probability prediction value is calculated.

[0010] Optionally, calculating the prediction ability score corresponding to the interpretation subsequence includes: calculating the prediction ability score using the following formula:

[0011] ;

[0012] in, For the The predictive ability score corresponding to the explanatory subsequence, For the Explanation subsequence, Indicates an intervention operation. Indicates that the knowledge tracking model is based only on To perform prediction operations, Indicates that the knowledge tracking model is based only on The output result obtained by making predictions is is the original predicted value of the answer probability, Used for judgment and Whether they are in the same decision direction, the decision direction includes correct prediction and wrong prediction.

[0013] Optionally, calculating the sufficiency score and the necessity score corresponding to the interpretation subsequence includes: calculating the sufficiency score using the following formula:

[0014] ;

[0015] in, is the sufficiency score, Indicates the The predictive ability score corresponding to the explanatory subsequence, Indicates removal The prediction ability score corresponding to the remaining explanation subsequence is calculated by the following formula:

[0016] ;

[0017] in is the necessity score, Represents any interaction data point in the explanation subsequence, express The minimum necessary score, Indicates the summation of the necessity scores of all interaction data points whose necessity scores are less than 0.

[0018] Optionally, based on the causal contribution, a multi-objective particle swarm optimization algorithm is used to generate a target explanation subsequence, and the answer probability prediction value is tracked and explained through the target explanation subsequence, specifically including: dynamically stratifying the candidate subsequences according to the causal contribution to generate a hierarchical explanation subsequence set, the hierarchical explanation subsequence set including a first set layer, a second set layer, a third set layer and a fourth set layer; updating the first set layer, the second set layer and the third set layer through elite learning to obtain a first target explanation subsequence; updating the fourth set layer through reverse learning to obtain a second target explanation subsequence; using the first target explanation subsequence and the second target explanation subsequence as the target explanation subsequence, and tracking and explaining the answer probability prediction value through the target explanation subsequence.

[0019] In a second aspect of the present application, a deep knowledge tracking and interpretation device is provided, which includes an acquisition module and a processing module, wherein:

[0020] The acquisition module is used to obtain the historical answer records and questions to be predicted of the target students, and to construct the learning interaction sequence corresponding to the target students according to the historical answer records in the order of answer time.

[0021] The processing module is used to input the learning interaction sequence into the knowledge tracking model to output the predicted value of the target student's correct answer probability to the predicted question; based on the preset strategy, the learning interaction sequence is pre-screened to obtain candidate subsequences; the causal contribution of the candidate subsequences to the predicted answer probability is calculated in combination with the educational constraint method, which includes the cognitive load reduction method and the time decay weighted method; based on the causal contribution, the multi-objective particle swarm optimization algorithm is used to generate the target explanation subsequence, and the answer probability prediction value is tracked and explained through the target explanation subsequence.

[0022] In the third aspect of the present application, an electronic device is provided, including a processor, a memory, a user interface and a network interface, the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device performs any of the methods described above.

[0023] In a fourth aspect of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a computer program, and the computer program is executed by a processor to perform any of the above methods.

[0024] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0025] 1. Obtain the historical answer records and questions to be predicted of the target student, and construct the learning interaction sequence corresponding to the target student based on multiple historical answer records; input the learning interaction sequence into the knowledge tracking model to output the target student's answer probability prediction value for the predicted question; generate multiple explanation subsequences based on the learning interaction sequence and particle swarm coding; calculate multiple causal contributions of multiple explanation subsequences to the answer probability prediction value; obtain the target causal contribution that meets the preset ranking conditions among the multiple causal contributions, and track and explain the answer probability prediction value based on particle swarm iteration and the explanation subsequence corresponding to the target causal contribution, thereby generating explanation subsequences through particle swarm coding, and combining causal inference to calculate the causal contribution of each subsequence to the answer probability prediction value, which can effectively solve the problem that the DLKT model cannot intuitively present which interaction pairs mainly drive the probability prediction of students correctly answering a certain question, thereby making the answer probability prediction value more transparent and providing a clear basis for personalized teaching.

[0026] 2. Initialize the particle swarm on the candidate subsequences by using multiple preset strategies, including but not limited to designing question-related strategies, prior-related strategies, answer-related strategies, and distance-first strategies, thereby improving the quality of the initial population, ensuring that the initial particle swarm can cover more representative candidate sub-interaction sequences, and improving the effectiveness of the search space and optimization efficiency.

[0027] 3. A predictive ability score is calculated to evaluate the degree of predictive influence of the interactive information contained in the explanation subsequence on the answer probability prediction value. At the same time, the sufficiency score and necessity score corresponding to the explanation subsequence are calculated to evaluate the internal information redundancy in the explanation subsequence, and the causal contribution of the explanation subsequence to the answer probability prediction value is calculated based on the predictive ability score, the sufficiency score and the necessity score, so as to accurately identify and screen the key interactive data points that have the greatest influence on the answer probability prediction value of the knowledge tracking model, thereby improving the interpretability of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a flowchart of a deep knowledge tracking and interpretation method and device provided by an embodiment of the present application;

[0029] Figure 2 This is a schematic diagram of a method for constructing an attribution quality function provided in an embodiment of the present application;

[0030] Figure 3 Schematic diagram of an iterative update method for a learning interaction sequence provided in an embodiment of the present application;

[0031] Figure 4 This is a module diagram of a deep knowledge tracking and interpretation device provided by an embodiment of the present application;

[0032] Figure 5 This is a structural diagram of an electronic device provided in an embodiment of the present application.

[0033] Explanation of the reference numerals: 41, acquisition module; 42, processing module; 501, processor; 502, communication bus; 503, user interface; 504, network interface; 505, memory. DETAILED DESCRIPTION

[0034] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments.

[0035] The terms used in the following examples of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the specification of the present application, the singular expressions "a", "an", "said", "above", "the", and "this" are intended to include plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in the present application refers to and includes any or all possible combinations of one or more of the listed items.

[0036] In the following, the terms "first" and "second" are used for descriptive purposes only and should not be understood to imply or suggest relative importance or implicitly indicate the number of the technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of this application, unless otherwise specified, "plurality" means two or more.

[0037] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0038] Please refer to Figure 1 , which shows a flow chart of a deep knowledge tracking and interpretation method and device provided in an embodiment of the present application, the flow chart mainly includes the following steps: S101 to S105.

[0039] Step S101: Obtain the historical answer records and the questions to be predicted of the target student, and construct a learning interaction sequence corresponding to the target student according to the historical answer records in the order of answer time.

[0040] Specifically, the historical answer records of multiple students are collected. Taking the target student as an example, the target student can be any one of the multiple students. The historical answer record includes the collection of questions answered by the target student in the historical learning process, as well as the correct or incorrect information of the target student's answer to each question. The question collection can include many different types of questions, such as mathematics type, physics type, chemistry type, and each question can involve different knowledge points. Construct the learning interaction sequence of the target student based on the historical answer record, that is, the questions he answered in the past. Question to , and the corresponding answer results to , then the target student's interaction sequence can be: ,in, , Indicates whether the target student answered correctly .

[0041] Step S102: input the learning interaction sequence into the knowledge tracking model to output a predicted value of the probability of the target student correctly answering the predicted question.

[0042] Specifically, assuming the question to be predicted is , then the interaction sequence will be learned Input the knowledge tracking model to output the target student's answer probability prediction value for the predicted question. The answer probability prediction value is the target student's next moment in The probability prediction of answering correctly is , model about The interpretation subsequence is denoted as , specifically:

[0043] ;

[0044] in, for The Explanation subsequence, Represents the attribution quality function. Therefore, in the knowledge tracking model, the predicted value of the target student's answer probability to the predicted question is obtained by maximizing the attribution quality function.

[0045] Step S103: Based on a preset strategy, a pre-screening operation is performed on the learning interaction sequence to obtain a candidate subsequence.

[0046] Specifically, a particle swarm is constructed and the particles in the particle swarm are mapped to subsequences in the learning interaction sequence, with one particle corresponding to one subsequence. The particles in the particle swarm are encoded. During the encoding process, binary or continuous numerical encoding is used to represent the selection of data points in the interaction sequence. For example, each data point in the interaction sequence is mapped to a dimension of the particle position vector. 1 indicates that the data point is selected into the interpretation subsequence, and 0 indicates that it is not selected. The subsequence corresponding to the selected particle is used as the interpretation subsequence.

[0047] In a possible implementation, step S103 further includes: the preset strategy includes a hidden state observation-based generation strategy and a gradient prior-based generation strategy.

[0048] Specifically, the pre-set strategies include, but are not limited to, generation strategies based on hidden state observation, generation strategies based on gradient priors, and generation strategies based on experience. The generation strategy based on hidden state observation: Independent of the model, for each time step, the student's implicit knowledge state / correct answer probability for the problem being predicted is output, and the probability is observed as the probability changes over time. For time steps where the upward or downward trend corresponds to the positive or negative prediction value to be explained, the corresponding interaction pairs are retained. The generation strategy based on gradient priors: Using a gradient-based method, the K interaction pairs with the highest correlation with the predicted answer probability are selected to form the initial candidate subsequence. The generation strategy based on experience: Interaction pairs whose correct or incorrect answers in the input interaction sequence are consistent with the predicted answer output by the model to be explained are retained, while those with inconsistent answers are ignored. Furthermore, interaction pairs with distant time steps or irrelevant concepts are ignored to form the initial candidate subsequences. The stochastic strategy maintains diversity and can, to a certain extent, prevent regression into local optima.

[0049] In a possible implementation, step S103 further includes: sorting the learning interaction sequence according to its length, and obtaining multiple candidate subsequences based on the sorting result; using particle swarm encoding to assign corresponding particles to the multiple candidate subsequences; obtaining a target candidate subsequence, where the target candidate subsequence is any subsequence among the multiple candidate subsequences; obtaining a position vector of the particle corresponding to the target candidate subsequence, and confirming whether each dimension value of the position vector meets a preset screening condition; if each dimension value of the position vector meets the preset screening condition, then using the target candidate subsequence as the interpretation subsequence.

[0050] Specifically, if the original length of the learning interaction sequence is t, there are 2^t possible combinations of candidate subsequences. At this time, take any candidate subsequence in 2^t as an example and assign it a corresponding particle. The dimension of the particle is the same as the length of the learning interaction sequence. When generating interpretation subsequences based on particle swarm coding, the interpretation problem can be transformed into an optimization problem in a binary solution space. Obtain the position vector of the particle corresponding to the target candidate subsequence, and record the various dimension values ​​of the position vector as ,like , then the preset filtering conditions are met, indicating that The candidate subsequences of time steps are selected as factors; if , then the preset filtering conditions are not met, indicating that The candidate subsequences of time steps are not selected as factors. The above process is repeated until all the explanation subsequences that meet the preset screening conditions are selected.

[0051] Step S104 , calculating the causal contribution of the candidate subsequence to the answer probability prediction value in combination with the pedagogical constraint method, where the pedagogical constraint method includes a cognitive load reduction method and a time decay weighting method.

[0052] Specifically, to construct the attribution quality function, please refer to Figure 2 , which presents a schematic diagram of an attribution quality function construction method provided in an embodiment of the present application, and measures the causal contribution of each explanation subsequence to the answer probability prediction value through the attribution quality function to discover the causal relationship between the input and output of the knowledge tracking model, and one explanation subsequence corresponds to one causal contribution.

[0053] Regarding the cognitive load reduction method: According to educational theory, it is necessary to ensure the credibility of the explanation and avoid generating misleading explanations, such as over-reliance on certain irrelevant redundant features, and introduce sufficiency and necessity to verify the independent explanatory power of the candidate explanation subsequence; according to cognitive load theory, human memory capacity is limited, and it is necessary to reduce redundant information, optimize the presentation of explanations, and avoid information overload. Therefore, a necessity score is designed to suppress redundant information (such as 5 wrong reasons are presented at the same time, the two most critical ones are retained, and the ones that provide less information are considered redundant) to avoid repeated explanations of similar errors. A suitable attribution quality function can help the optimization algorithm select the globally optimal explanation subsequence. The present invention mainly considers the explanation attribution quality of the selected subsequence from two aspects, namely the overall explanation sufficiency and internal information redundancy of the explanation subsequence. Among them, for the overall explanation sufficiency: by calculating the predictive ability score corresponding to the explanation subsequence, the degree of predictive influence of the interactive information contained in the explanation subsequence on the predicted value of the answer probability can be evaluated based on the predictive ability score. Intervention means in causal inference can be used to utilize Arithmetic operations directly manipulate the values ​​of input variables to study how the model predictions will change. The calculation formula is as follows:

[0054] ;

[0055] in, For the The predictive ability score corresponding to the explanatory subsequence, For the Explanation subsequence, Indicates an intervention operation. Indicates that the knowledge tracking model is based only on To perform prediction operations, Indicates that the knowledge tracking model is based only on The output result obtained by making predictions is is the original predicted value of the answer probability, Used for judgment and Whether they are in the same decision direction, the decision direction includes correct prediction and wrong prediction; for internal information redundancy: by calculating the sufficiency score and necessity score corresponding to the explanation subsequence, the internal information redundancy in the explanation subsequence is evaluated based on the sufficiency score and necessity score. Among them, for the sufficiency score, based on the principle of causal sufficiency, ensure that the explanation part contains factors that affect the decision-making of the knowledge tracking model, while the non-explanation part does not contain factors that affect the decision-making of the knowledge tracking model. In order to avoid the over-justification effect, that is, excessive pursuit of prediction score consistency, the incorrect introduction of relevant but non-causal information and the neglect of other explainable information. Therefore, based on the principle of causal sufficiency, the calculation formula of the sufficiency score is as follows:

[0056] ;

[0057] in, is the sufficiency score, Indicates the The predictive ability score corresponding to the explanatory subsequence, Indicates removal The predictive ability score corresponding to the remaining explanatory subsequences; when an explanatory subsequence with the best overall explanatory effect is found through causal sufficiency, the redundancy problem follows. Because there are complex interactions, mutual dependence, and mutual collisions between the internal factors of a subsequence, there are also synergistic effects and redundant effects. The synergistic effect means that the overall explanatory power will be improved due to the addition of factors, while the redundant effect means that the addition of certain factors does not bring about an improvement in explanatory power. Therefore, it is necessary to introduce a necessity score to evaluate the contribution of each factor in the explanatory subsequence. Specifically, based on causal screening, when the existence of certain factors in the explanatory subsequence does not provide more information for the target variable, or only provides less information, it is called redundant. Therefore, based on the redundancy effect, the necessity of each factor in the explanatory subsequence is measured:

[0058] ;

[0059] Regarding the time decay weighting method: According to the forgetting curve theory, as time goes by, the influence of past learning behavior on the current behavior tends to decay exponentially, that is, students' more distant historical experience is unlikely to have an impact on the current situation. By introducing the exponential decay factor of time distance and using the time domain decay function to reduce the weight of distant time steps, the necessity measurement function is improved. The formula is:

[0060] ;

[0061] in, represents the time step of the target prediction, represent At the time step of the original interaction sequence, is the interval between two time steps, is the attenuation factor.

[0062] Improved The necessity score is measured as:

[0063] ;

[0064] Afterwards, the necessity score is calculated using the following formula:

[0065] ;

[0066] in is the necessity score, Represents any interaction data point in the explanation subsequence, express The minimum necessary score, Indicates the sum of the necessity scores of all interaction data points whose necessity scores are less than 0. Indicates that when Remove back, Finally, the causal contribution of the explanation subsequence to the predicted value of the answer probability is calculated based on the predictive ability score, sufficiency score, and necessity score. The calculation formula is as follows:

[0067] ;

[0068] in, For the The predicted value of the probability of answering the subsequence The quality of causal explanation, Indicates the The predicted value of the probability of answering the subsequence The causal sufficiency of Indicates the The predicted value of the probability of answering the subsequence causal necessity.

[0069] In step S105 , a multi-objective particle swarm optimization algorithm is used to generate a target explanation subsequence based on the causal contribution, and the answer probability prediction value is tracked and explained through the target explanation subsequence.

[0070] Specifically, a multi-objective particle swarm optimization algorithm is used to optimize the subsequence combination, ensuring both explanation sufficiency and explanation necessity, and generating the final explanation subsequence. Using the particle swarm optimization algorithm to optimize the search space, given a knowledge tracking model, the optimal explanation subsequence is explored. The explanation subsequence corresponding to the target causal contribution in the learning interaction sequence that meets the preset ranking conditions is determined and taken as the optimal explanation subsequence. The learning interaction sequence is then updated based on the particle swarm iteration and the explanation subsequence corresponding to the target causal contribution.

[0071] In a possible embodiment, step S105 also includes: dynamically stratifying the candidate subsequences according to the causal contribution to generate a hierarchical explanation subsequence set, the hierarchical explanation subsequence set including a first set layer, a second set layer, a third set layer and a fourth set layer; updating the first set layer, the second set layer and the third set layer through elite learning to obtain a first target explanation subsequence; updating the fourth set layer through reverse learning to obtain a second target explanation subsequence; using the first target explanation subsequence and the second target explanation subsequence as the target explanation subsequence, and tracking and interpreting the answer probability prediction value through the target explanation subsequence.

[0072] Specifically, a dynamic stratification strategy and constraint processing technology are introduced: First, all particles in the population are sorted by fitness value and initially divided into two layers. Constraint processing technology is used on the N / 2 particles in the upper layer to differentiate the first and second set layers, namely the elite layer G and the better layer. Specifically, the particle with the longest length in the upper layer As the bound, find the fitness value better than of particles form the elite layer G, representing short and good particles, and the rest The particles form a better layer; the N / 2 particles in the lower layer are also constrained, that is, the particle with the shortest length As the boundary, find the fitness value worse than of particles, representing long and poor particles, constitute the redundant layer, that is, the fourth set layer, and the rest The particles form a poor layer, namely the third set layer. The first set layer, the second set layer, the third set layer and the fourth set layer are collectively referred to as the hierarchical interpretation subsequence set. The hierarchical relationship of the hierarchical interpretation subsequence set is as follows: Figure 2 As shown in .

[0073] A divide-and-conquer approach uses an elite retention strategy to maintain the elite layer, the better layer, and the worse layer G, and introduces an Opposition-Based Learning (OBL) strategy to handle redundancy. Specifically, unlike traditional OBL (Opposition-Based Learning) methods, this approach does not generate an inverse solution for all particles, but only generates an inverse solution for redundant particles in the population that are both long and poor:

[0074]

[0075] Where r is a random number between 0 and 1. The fitness values ​​of the current solution and the reverse solution are compared, and the best one is selected to enter the next generation, thereby enhancing the population diversity in the search space.

[0076] The ordinary particle swarm iterative algorithm is used to update both the better and worse layers. The learning objects are the elite layer particles and the individual historical optimality. The update formula is as follows:

[0077]

[0078] in, For the Particles in The speed of the wheel update determines the direction and amplitude of the particle's movement in the next step. For the Particles in The speed of the wheel update, Inertia weight controls the influence of the particle's current velocity. Larger values ​​make the particle more likely to move in the direction of its original velocity, exploring unknown areas. Smaller values ​​make it more susceptible to new solutions, adjusting the search direction. are particles selected from the elite layer G, is the particle’s own historical optimal position, is the elite learning factor, used to control the particle The degree of closeness, when the value is larger, the particles are more likely to approach the global optimal direction; is the individual learning factor, which is used to control the particle to move towards its own historical optimal the degree of proximity to facilitate the use of one's own successful experiences; For the Particles in The position in the learning interaction sequence when the round is updated; For the Particles in The position in the learning interaction sequence when the round is updated. Please refer to Figure 3 , which shows a schematic diagram of an iterative update method for a learning interaction sequence provided by an embodiment of the present application, Figure 3In

[15] , the subsequence to be explained x is a learning interaction sequence, the model to be explained is a knowledge tracking model, and the attribution quality measurement module is constructed by the attribution quality function.

[0079] The present application adopts the above method to obtain the historical answer records and questions to be predicted of the target student, and constructs a learning interaction sequence corresponding to the target student based on multiple historical answer records; inputs the learning interaction sequence into the knowledge tracking model to output the target student's answer probability prediction value to the predicted question; generates multiple explanation subsequences based on the learning interaction sequence and particle swarm coding; calculates multiple causal contributions of multiple explanation subsequences to the answer probability prediction value; obtains the target causal contribution that meets the preset ranking conditions among the multiple causal contributions, and tracks and explains the answer probability prediction value based on particle swarm iteration and the explanation subsequence corresponding to the target causal contribution, thereby generating an explanation subsequence through particle swarm coding, and combining causal inference to calculate the causal contribution of each subsequence to the answer probability prediction value, which can effectively solve the problem that the DLKT model cannot intuitively present which interaction sequences mainly drive the probability prediction of students correctly answering a certain question, thereby making the answer probability prediction value more transparent and providing a clear basis for personalized teaching.

[0080] Please refer to Figure 4 , which shows a module diagram of a deep knowledge tracking and interpretation device provided by an embodiment of the present application, the device includes an acquisition module 41 and a processing module 42, wherein,

[0081] The acquisition module 41 is used to obtain the historical answer records and the questions to be predicted of the target student, and construct a learning interaction sequence corresponding to the target student according to the historical answer records and the answer time sequence.

[0082] The processing module 42 is used to input the learning interaction sequence into the knowledge tracking model to output the answer probability prediction value of the target student correctly answering the predicted question; based on the preset strategy, the learning interaction sequence is pre-screened to obtain a candidate subsequence; the causal contribution of the candidate subsequence to the answer probability prediction value is calculated in combination with the educational constraint method, the educational constraint method includes the cognitive load reduction method and the time decay weighted method; based on the causal contribution, a multi-objective particle swarm optimization algorithm is used to generate the target explanation subsequence, and the answer probability prediction value is tracked and explained through the target explanation subsequence.

[0083] In one possible implementation, the processing module 42 is used to perform a pre-screening operation on the learning interaction sequence based on knowledge in the field of pedagogy to obtain candidate subsequences, specifically including: using particle swarm encoding to assign corresponding particles to each subsequence in the learning interaction sequence; obtaining a target candidate subsequence, where the target candidate subsequence is any subsequence in multiple learning interaction sequences; obtaining the position vector of the particle corresponding to the target candidate subsequence, and confirming whether the values ​​of each dimension of the position vector meet the preset screening conditions; if the values ​​of each dimension of the position vector meet the preset screening conditions, then using the target candidate subsequence as the candidate subsequence.

[0084] In a possible implementation, the preset strategy includes a hidden state observation-based generation strategy and a gradient prior-based generation strategy.

[0085] In one possible embodiment, the processing module 42 is used to calculate the causal contribution of the candidate subsequence to the answer probability prediction value in combination with the educational constraint method, specifically including: calculating the predictive ability score corresponding to the explanation subsequence, the predictive ability score is used to evaluate the degree of predictive influence of the interactive information contained in the explanation subsequence on the answer probability prediction value; calculating the sufficiency score and necessity score corresponding to the explanation subsequence, the sufficiency score and necessity score are used to evaluate the internal information redundancy in the explanation subsequence; based on the predictive ability score, sufficiency score and necessity score, and based on the time exponential decay function, calculating the causal contribution of the explanation subsequence to the answer probability prediction value.

[0086] In one possible implementation, the processing module 42 is configured to calculate a prediction ability score corresponding to the interpretation subsequence, specifically including: calculating the prediction ability score using the following formula:

[0087] ;

[0088] in, For the The predictive ability score corresponding to the explanatory subsequence, For the Explanation subsequence, Indicates an intervention operation. Indicates that the knowledge tracking model is based only on To perform prediction operations, Indicates that the knowledge tracking model is based only on The output result obtained by making predictions is is the original predicted value of the answer probability, Used for judgment and Whether they are in the same decision direction, the decision direction includes correct prediction and wrong prediction.

[0089] In a possible implementation, the processing module 42 is configured to calculate the sufficiency score and the necessity score corresponding to the interpretation subsequence, specifically including: calculating the sufficiency score using the following formula:

[0090] ;

[0091] in, is the sufficiency score, Indicates the The predictive ability score corresponding to the explanatory subsequence, Indicates removal The prediction ability score corresponding to the remaining explanation subsequence is calculated by the following formula:

[0092] ;

[0093] in is the necessity score, Represents any interaction data point in the explanation subsequence, express The minimum necessary score, Indicates the summation of the necessity scores of all interaction data points whose necessity scores are less than 0.

[0094] In one possible embodiment, the processing module 42 is used to generate a target explanation subsequence based on the causal contribution using a multi-objective particle swarm optimization algorithm, and track and interpret the answer probability prediction value through the target explanation subsequence, specifically including: dynamically stratifying the candidate subsequences according to the causal contribution to generate a hierarchical explanation subsequence set, the hierarchical explanation subsequence set including a first set layer, a second set layer, a third set layer and a fourth set layer; updating the first set layer, the second set layer and the third set layer through elite learning to obtain a first target explanation subsequence; updating the fourth set layer through reverse learning to obtain a second target explanation subsequence; using the first target explanation subsequence and the second target explanation subsequence as target explanation subsequences, and tracking and interpreting the answer probability prediction value through the target explanation subsequence.

[0095] It should be noted that the above embodiments provide devices that implement their functions using only the division of the above functional modules as examples. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the device and method embodiments provided in the above embodiments are based on the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0096] This application also provides an electronic device. Figure 5 , Figure 5 5 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. The electronic device may include: at least one processor 501, at least one communication bus 502, a user interface 503, at least one network interface 504, and a memory 505.

[0097] The communication bus 502 is used to implement the connection and communication between these components.

[0098] The user interface 503 may include a display screen (Display) and a camera (Camera). Optionally, the user interface 503 may also include a standard wired interface and a wireless interface.

[0099] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).

[0100] The processor 501 may include one or more processing cores. Using various interfaces and circuits, the processor 501 connects to various components within the server. It executes instructions, programs, code sets, or instruction sets stored in the memory 505, as well as accesses data stored in the memory 505, to perform various server functions and process data. Optionally, the processor 501 may be implemented using at least one of the following hardware forms: a digital signal processing (DSP), a field-programmable gate array (FPGA), or a programmable logic array (PLA). The processor 501 may integrate one or a combination of a central processing unit (CPU), a graphics processing unit (GPU), and a modem. The CPU primarily processes the operating system, user interface, and application programs; the GPU is responsible for rendering and drawing content displayed on the display screen; and the modem handles wireless communications. It is understood that the modem may also be implemented as a separate chip, rather than integrated into the processor 501.

[0101] Among them, the memory 505 may include a random access memory (RAM) or a read-only memory (Read-Only Memory). Optionally, the memory 505 includes a non-transitory computer-readable storage medium. The memory 505 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 505 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area may store data involved in the above-mentioned various method embodiments, etc. The memory 505 may also be optionally at least one storage device located away from the aforementioned processor 501. Reference Figure 5 , as a computer storage medium, the memory 505 may include an operating system, a network communication module, a user interface module, and a deep knowledge tracking interpretation application.

[0102] exist Figure 5 In the electronic device shown, the user interface 503 is mainly used to provide an input interface for the user and obtain the data input by the user; and the processor 501 can be used to call the deep knowledge tracking and interpretation application stored in the memory 505. When executed by one or more processors 501, the electronic device executes one or more of the methods described in the above embodiments. It should be noted that for the aforementioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should know that this application is not limited to the order of the actions described, because according to this application, certain 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 and modules involved are not necessarily required for this application.

[0103] The present application also provides a computer-readable storage medium storing instructions, which, when executed by one or more processors, enable an electronic device to execute one or more of the methods described in the above embodiments.

[0104] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0105] In the several embodiments provided in this application, it should be understood that the disclosed devices can be implemented in other ways. For example, the device embodiments described above are merely schematic, such as the division of units, which is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some service interface, and the indirect coupling or communication connection of devices or units can be electrical or other forms.

[0106] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0107] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0108] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of this application, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the method of this application. The aforementioned memory includes various media that can store program code, such as USB flash drives, mobile hard drives, magnetic disks, or optical disks.

[0109] The above descriptions are merely exemplary embodiments disclosed in this application and are not intended to limit the scope of this application. That is, any equivalent changes and modifications made based on the teachings disclosed in this application are still within the scope of this application.

[0110] This application is intended to cover any modifications, uses or adaptations disclosed in this application, which follow the general principles disclosed in this application and include common knowledge or customary technical means in the technical field not disclosed in this application.

Claims

1. A deep knowledge tracing and interpretation method, characterized in that: The method comprises: Obtain the historical answer records and the questions to be predicted of the target student, and construct a learning interaction sequence corresponding to the target student according to the historical answer records in the order of answer time; Inputting the learning interaction sequence into a knowledge tracking model to output a predicted value of the probability of the target student correctly answering the question to be predicted; Based on a preset strategy, a pre-screening operation is performed on the learning interaction sequence to obtain a candidate subsequence; Calculating the causal contribution of the candidate subsequence to the predicted answer probability value in combination with a pedagogical constraint method, wherein the pedagogical constraint method includes a cognitive load reduction method and a time decay weighting method; According to the causal contribution, a multi-objective particle swarm optimization algorithm is used to generate a target explanation subsequence, and the answer probability prediction value is tracked and explained through the target explanation subsequence; The method of combining the educational constraint method to calculate the causal contribution of the candidate subsequence to the answer probability prediction value specifically includes: calculating the prediction ability score corresponding to the explanation subsequence, the prediction ability score is used to evaluate the degree of predictive influence of the interactive information contained in the explanation subsequence on the answer probability prediction value; calculating the sufficiency score and necessity score corresponding to the explanation subsequence, the sufficiency score and the necessity score are used to evaluate the internal information redundancy in the explanation subsequence; based on the prediction ability score, the sufficiency score and the necessity score, and based on the time exponential decay function, calculating the causal contribution of the explanation subsequence to the answer probability prediction value.

2. The method according to claim 1, characterized in that The pre-screening operation is performed on the learning interaction sequence based on a preset strategy to obtain a candidate subsequence, specifically including: Using particle swarm coding to assign corresponding particles to each subsequence in the learning interaction sequence; Acquire a target candidate subsequence, where the target candidate subsequence is any subsequence in the plurality of learning interaction sequences; Obtaining the position vector of the particle corresponding to the target candidate subsequence, and confirming whether the values ​​of each dimension of the position vector meet the preset screening conditions; If the dimension values ​​of the position vector meet the preset screening condition, the target candidate subsequence is used as the candidate subsequence.

3. The method according to claim 2, characterized in that The preset strategies include a hidden state observation-based generation strategy and a gradient prior-based generation strategy.

4. The method according to claim 1, wherein Calculating the prediction ability score corresponding to the explanation subsequence specifically includes: The predictive ability score is calculated by the following formula: ； in, For the The predictive ability score corresponding to each of the explanation subsequences, For the said interpreter sequence, Indicates an intervention operation. Indicates that the knowledge tracking model is based only on To perform prediction operations, Indicates that the knowledge tracking model is based only on The output result obtained by making predictions is is the original predicted value of the answer probability, Used for judgment and Whether they are in the same decision direction, the decision direction includes correct prediction and wrong prediction.

5. The method according to claim 4, characterized in that The calculating of the sufficiency score and the necessity score corresponding to the interpretation subsequence specifically includes: The adequacy score is calculated by the following formula: ； in, is the sufficiency score, Indicates the The predictive ability score corresponding to each of the explanation subsequences, Indicates removal The predictive ability score corresponding to the remaining explanatory subsequence; The necessity score is calculated by the following formula: ； in is the necessity score, represents any interaction data point in the explained subsequence, express The minimum necessary score, Indicates summing the necessity scores of all the interaction data points whose necessity scores are less than 0; It is used to measure the necessity of each factor in the explanatory subsequence based on the redundancy effect. , represents the time step of the target prediction, represent In the time step of the original interaction sequence, λ is the decay factor, Indicates that when Remove back, The predictive ability score of .

6. The method according to claim 1, characterized in that The method of generating a target explanation subsequence based on the causal contribution using a multi-objective particle swarm optimization algorithm and tracking and interpreting the answer probability prediction value through the target explanation subsequence specifically includes: Dynamically stratifying the candidate subsequences according to the causal contribution to generate a hierarchical interpretation subsequence set, the hierarchical interpretation subsequence set comprising a first set layer, a second set layer, a third set layer, and a fourth set layer; updating the first set layer, the second set layer, and the third set layer through elite learning to obtain a first target explanation subsequence; Updating the fourth set layer by reverse learning to obtain a second target explanation subsequence; The first target explanation subsequence and the second target explanation subsequence are used as the target explanation subsequence, and the answer probability prediction value is tracked and interpreted through the target explanation subsequence.

7. A deep knowledge tracking and interpretation device, characterized in that: The device includes an acquisition module and a processing module, wherein: The acquisition module is used to obtain the historical answer records and the questions to be predicted of the target student, and construct a learning interaction sequence corresponding to the target student according to the historical answer records in the order of answer time; The processing module is used to input the learning interaction sequence into the knowledge tracking model to output the answer probability prediction value of the target student correctly answering the to-be-predicted question; based on the preset strategy, pre-screening operation is performed on the learning interaction sequence to obtain a candidate subsequence; the causal contribution of the candidate subsequence to the answer probability prediction value is calculated in combination with the pedagogical constraint method, and the pedagogical constraint method includes a cognitive load reduction method and a time decay weighted method; according to the causal contribution, a multi-objective particle swarm optimization algorithm is used to generate a target explanation subsequence, and the answer probability prediction value is tracked and explained through the target explanation subsequence; the pedagogical constraint method is used to calculate the causal contribution of the candidate subsequence to the answer probability prediction value. The method calculates the causal contribution of the candidate subsequence to the answer probability prediction value, specifically including: calculating the prediction ability score corresponding to the explanation subsequence, the prediction ability score is used to evaluate the degree of predictive influence of the interactive information contained in the explanation subsequence on the answer probability prediction value; calculating the sufficiency score and the necessity score corresponding to the explanation subsequence, the sufficiency score and the necessity score are used to evaluate the internal information redundancy in the explanation subsequence; based on the prediction ability score, the sufficiency score and the necessity score, and based on a time exponential decay function, calculating the causal contribution of the explanation subsequence to the answer probability prediction value.

8. An electronic device, characterized in that: The electronic device comprises a processor, a communication bus, a user interface, a network interface and a memory, wherein the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device executes the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores instructions, and when the instructions are executed, the method according to any one of claims 1 to 6 is performed.

Images