A student answer prediction method and device based on multi-scale knowledge state decoupling

Abstract

The application provides a student answer prediction method and device based on multi-scale knowledge state decoupling, and relates to the field of artificial intelligence education. The method comprises the following steps: obtaining a historical answer sequence of a target student, and obtaining a knowledge state based on the historical answer sequence; performing multi-scale state decoupling processing on the knowledge state to obtain a long-term knowledge memory state, a short-term answer fluctuation state and an overall learning trend state; constructing a long-term causal expert model, a short-term causal expert model and a trend causal expert model respectively, and outputting corresponding long-term answer prediction results, short-term answer prediction results and trend answer prediction results; and performing fusion processing on the prediction results through a multi-expert prediction fusion mechanism to output corresponding answer prediction results of the target student. The application solves the problem that the existing student answer prediction system is difficult to accurately predict student answer results.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence education, and in particular to a method and apparatus for predicting student responses based on multi-scale knowledge state decoupling. Background Technology

[0002] Today, with the development of intelligent education and online learning platforms, students generate a large amount of historical interaction data during practice and answering. Knowledge tracking technology is gradually using deep learning, attention mechanisms and time series models to dynamically model students' knowledge status and predict their answering performance in subsequent practice based on historical answering sequences.

[0003] However, existing student answer prediction systems typically model students' knowledge status as a single whole, making it difficult to distinguish between changes in status at different time scales, such as long-term knowledge accumulation, recent fluctuations in answers, and overall learning trends. This leads to information confusion between different types of knowledge status, causing short-term changes in answers to interfere with long-term knowledge mastery, and thus reducing the accuracy of student answer prediction results.

[0004] Therefore, there is an urgent need for a student response prediction method and device based on multi-scale knowledge state decoupling. Summary of the Invention

[0005] This application provides a student answer prediction method and apparatus based on multi-scale knowledge state decoupling, which solves the problem that existing student answer prediction systems are unable to accurately predict student answer results.

[0006] The first aspect of this application provides a student answer prediction method based on multi-scale knowledge state decoupling. The method includes: obtaining the historical answer sequence of a target student and obtaining the corresponding knowledge state of the target student based on the historical answer sequence; performing multi-scale state decoupling processing on the knowledge state to obtain the decoupled long-term knowledge memory state, short-term answer fluctuation state, and overall learning trend state; constructing a long-term causal expert model based on the long-term knowledge memory state, a short-term causal expert model based on the short-term answer fluctuation state, and a trend causal expert model based on the overall learning trend state; outputting the long-term answer prediction result corresponding to the target student based on the long-term causal expert model, the short-term answer prediction result corresponding to the target student based on the short-term causal expert model, and the trend answer prediction result corresponding to the target student based on the trend causal expert model; and fusing the long-term answer prediction result, short-term answer prediction result, and trend answer prediction result through a multi-expert prediction fusion mechanism and outputting the answer prediction result corresponding to the target student.

[0007] Optionally, multi-scale state decoupling processing is performed on the knowledge state, specifically including: constructing a knowledge tracing causal graph; the nodes of the knowledge tracing causal graph represent variables, and the edges represent causal relationships; wherein, the nodes include historical answer sequences, knowledge states, answer results, target exercises, and confounding variables; constructing a first causal path from the historical answer sequence to the knowledge state; the first causal path is used to describe the process by which the target student forms the current knowledge mastery state based on the historical answer sequence; constructing a second causal path from the knowledge state and target exercises to the answer results; the second causal path is used to describe the process by which the target student generates answer results for the target exercises based on the knowledge state; constructing a third causal path in which confounding variables simultaneously affect both the historical answer sequence and the answer results; the third causal path is used to describe the process by which unobservable influencing factors interfere with the historical answer sequence and the answer results; and performing multi-scale state decoupling processing on the knowledge state based on the first causal path, the second causal path, and the third causal path.

[0008] Optionally, before constructing a long-term causal expert model based on long-term knowledge memory state, a short-term causal expert model based on short-term answer fluctuation state, and a trend causal expert model based on overall learning trend state, the method further includes: obtaining practice features, knowledge concept features, and answer result features corresponding to historical answer sequences; fusing knowledge concept features and answer result features, and adjusting the fused interaction features based on the practice difficulty information corresponding to the practice features to obtain an interaction feature representation of fused practice attributes; performing feature enhancement on the practice sequences corresponding to the practice features and the interaction sequences corresponding to the interaction feature representations based on an attention mechanism to obtain enhanced practice feature representations and enhanced interaction feature representations; and performing feature modeling on long-term knowledge memory state, short-term answer fluctuation state, and overall learning trend state based on the enhanced practice feature representations and enhanced interaction feature representations.

[0009] Optionally, the long-term answer prediction results for the target student are output based on a long-term causal expert model. Specifically, this includes: generating query vectors and key vectors based on the enhanced practice feature representation, and generating value vectors based on the enhanced interaction feature representation; performing weighted modeling based on the query vector, key vector, and value vector through an attention mechanism, and adjusting the attention weights corresponding to the attention mechanism based on the distance relationship between historical time steps to obtain long-term knowledge memory state features; performing nonlinear feature extraction on the correlation between the long-term knowledge memory state features and the enhanced interaction feature representation to obtain long-term answer response features; and fusing the long-term knowledge memory state features and the long-term answer response features to obtain the long-term answer prediction results.

[0010] Optionally, the short-term answer prediction results for the target student are output based on the short-term causal expert model. Specifically, this includes: extracting interaction change features within a local time window through a temporal convolutional network based on the enhanced practice feature representation and the enhanced interaction feature representation to obtain short-term answer fluctuation state features; performing nonlinear feature extraction based on the correlation between the short-term answer fluctuation state features and the enhanced interaction feature representation to obtain short-term answer response features; and fusing the short-term answer fluctuation state features and the short-term answer response features to obtain the short-term answer prediction results.

[0011] Optionally, the trend-based causal expert model outputs the trend-based answer prediction results for the target student, specifically including: based on the enhanced practice feature representation and the enhanced interaction feature representation, performing state evolution modeling on the interaction sequences corresponding to multiple historical time steps through a gated recurrent unit to obtain the overall learning trend state features; performing nonlinear feature extraction based on the correlation between the overall learning trend state features and the enhanced interaction feature representation to obtain the trend-based answer response features; and fusing the overall learning trend state features and the trend-based answer response features to obtain the trend-based answer prediction results.

[0012] Optionally, before fusing the long-term, short-term, and trend-based prediction results using a multi-expert prediction fusion mechanism, the method further includes: calculating the state similarity of long-term knowledge memory state features, short-term response fluctuation state features, and overall learning trend state features at the same time using a similarity calculation method; constraining the feature differences among long-term knowledge memory state features, short-term response fluctuation state features, and overall learning trend state features based on state similarity; adjusting the prediction losses corresponding to long-term, short-term, and trend-based prediction results using a homoscedastic uncertainty weighting method based on feature differences; and optimizing the long-term causal expert model, short-term causal expert model, and trend-based causal expert model based on the adjusted prediction losses.

[0013] A second aspect of this application provides a student answer prediction device based on multi-scale knowledge state decoupling. The device includes a knowledge state construction module, a state decoupling module, and a multi-expert prediction module, wherein... The knowledge state construction module is used to obtain the target student's historical answer sequence and obtain the corresponding knowledge state of the target student based on the historical answer sequence.

[0014] The state decoupling module is used to perform multi-scale state decoupling processing on knowledge states, and obtain the decoupled long-term knowledge memory state, short-term answer fluctuation state, and overall learning trend state.

[0015] The multi-expert prediction module is used to construct long-term causal expert models based on long-term knowledge memory states, short-term causal expert models based on short-term response fluctuation states, and trend causal expert models based on overall learning trend states.

[0016] The multi-expert prediction module is used to output long-term answer prediction results for the target student based on a long-term causal expert model, short-term answer prediction results for the target student based on a short-term causal expert model, and trend answer prediction results for the target student based on a trend causal expert model.

[0017] The multi-expert prediction module is used to fuse long-term, short-term, and trend-based answer prediction results through a multi-expert prediction fusion mechanism, and output the corresponding answer prediction results for the target student.

[0018] A third aspect of this application provides an electronic device 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 to cause the electronic device to perform the method as described above.

[0019] A fourth aspect of this application provides a non-transitory computer-readable storage medium storing a computer program, the computer program being executed by a processor using any of the methods described above.

[0020] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. Obtain the target student's historical answer sequence and acquire the knowledge state based on the historical answer sequence; perform multi-scale state decoupling processing on the knowledge state to obtain the long-term knowledge memory state, short-term answer fluctuation state, and overall learning trend state; construct long-term causal expert models, short-term causal expert models, and trend causal expert models based on different states, and output the corresponding long-term answer prediction results, short-term answer prediction results, and trend answer prediction results; fuse the prediction results through a multi-expert prediction fusion mechanism to output the corresponding answer prediction results for the target student, thereby combining knowledge state information at different time scales to improve the accuracy of student answer prediction.

[0021] 2. By constructing a knowledge tracing causal graph and based on multiple causal paths from historical answer sequences to knowledge states, from knowledge states to answer results, and confounding variables affecting historical answer sequences and answer results, multi-scale state decoupling processing is performed on knowledge states. This reduces unobservable factors and mutual interference between different state factors, enabling long-term knowledge mastery, short-term answer changes, and learning trends to be independently represented, thereby improving the accuracy of knowledge state modeling.

[0022] 3. By integrating practice features, knowledge concept features, and answer result features, and adjusting the interaction feature representation in conjunction with practice difficulty information, while utilizing attention mechanisms to enhance the feature expression of practice sequences and interaction sequences, we can fully explore the correlation between different practice attributes and historical interaction behaviors, thereby improving the completeness and effectiveness of extracting student knowledge state features. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating a student response prediction method based on multi-scale knowledge state decoupling provided in an embodiment of this application; Figure 2 This is a knowledge tracing cause-effect graph provided in an embodiment of this application; Figure 3 This is a schematic diagram of a student answer prediction device based on multi-scale knowledge state decoupling provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0024] Explanation of reference numerals in the attached diagram: 31. Knowledge state construction module; 32. State decoupling module; 33. Multi-expert prediction module; 401. Processor; 402. Communication bus; 403. User interface; 404. Network interface; 405. Memory. Detailed Implementation

[0025] 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 with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0026] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to and includes any or all possible combinations of one or more of the listed items.

[0027] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0028] 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.

[0029] Please refer to Figure 1 The flowchart illustrates a student response prediction method based on multi-scale knowledge state decoupling provided in this application embodiment. The flowchart mainly includes the following steps: S101 to S105.

[0030] Step S101: Obtain the target student's historical answer sequence and obtain the target student's corresponding knowledge status based on the historical answer sequence.

[0031] Specifically, in intelligent learning systems, student interactions with exercises are typically recorded in discrete time. More specifically, it is assumed that there exists a set of learners. Practice Set and the set of knowledge concepts covered in the exercises For students At any moment The interactive behavior can be abstracted into a triple. ,in Indicates the exercise index. This represents the knowledge concept corresponding to this exercise. This indicates the result of the answer in the exercise. Among them, This means the answer is correct. This indicates an incorrect answer.

[0032] Based on this, a student's historical interaction sequence can be represented as: in, Indicates the end time Historical interaction sequence, These represent the practice interaction information of students at different historical moments.

[0033] The objective of knowledge tracing is to, within a given historical sequence and practice of the next moment's goal Corresponding knowledge concepts In this case, by establishing a student knowledge state model, we can predict the student's answer after completing the exercise with the target objective. The final prediction target can be expressed as:

[0034] in, This represents the probability that a student answers the target exercise correctly, given the student's historical interaction sequence, the target exercise at the next moment, and the corresponding knowledge concept. This indicates the student's response to the exercise aimed at the next time step; Indicates the target practice index to be predicted in the next moment; This indicates the knowledge concepts corresponding to the target practice.

[0035] Based on the aforementioned historical interaction sequence, models are built to assess students' completion of exercises, mastery of knowledge concepts, and response feedback during the continuous learning process, thereby obtaining the target student's corresponding knowledge state. This knowledge state describes the target student's level of knowledge mastery after historical exercise interactions and reflects the impact of historical response behavior on subsequent target exercise performance.

[0036] When acquiring historical interaction sequences, the system records the exercise index, knowledge concepts, and response results for each practice interaction to form a time-series data representation that reflects the student's learning process. The exercise index distinguishes different practice tasks, the knowledge concepts describe the knowledge content involved in the exercise, and the response results indicate the student's mastery of the corresponding exercise. By comprehensively analyzing the above interaction information, the system can extract changes in students' knowledge mastery at different time stages, thereby generating a corresponding knowledge state representation. This provides a foundation for subsequent knowledge state analysis and response result prediction.

[0037] Step S102 involves performing multi-scale state decoupling on the knowledge state to obtain the decoupled long-term knowledge memory state, short-term answer fluctuation state, and overall learning trend state.

[0038] Specifically, a multi-scale analysis of knowledge status is conducted, dividing students' knowledge mastery during the learning process into different time scales. The analysis extracts long-term knowledge memory status to represent long-term accumulated knowledge mastery, short-term answer fluctuation status to represent recent changes in answering, and overall learning trend status to represent the overall learning development trend.

[0039] Among them, the long-term knowledge memory state is used to describe the stable knowledge mastery characteristics formed by the target student after multiple historical learning stages, reflecting the long-term knowledge accumulation in the historical learning process; the short-term answer fluctuation state is used to describe the state changes of the target student in the recent continuous practice process, reflecting the fluctuation of answer performance in a short period of time; the overall learning trend state is used to describe the knowledge development trend of the target student that continues to change with the learning process, reflecting the direction of improvement or change in overall learning ability.

[0040] In one possible implementation, step S102 further includes: constructing a knowledge tracing causal graph; the nodes of the knowledge tracing causal graph represent variables, and the edges represent causal relationships; wherein, the nodes include historical answer sequences, knowledge states, answer results, target exercises, and confounding variables; constructing a first causal path from the historical answer sequence to the knowledge state; the first causal path is used to describe the process by which the target student forms the current knowledge mastery state based on the historical answer sequence; constructing a second causal path from the knowledge state and target exercises to the answer result; the second causal path is used to describe the process by which the target student generates answer results for the target exercises based on the knowledge state; constructing a third causal path that simultaneously influences the historical answer sequence and the answer result; the third causal path is used to describe the process by which unobservable influencing factors interfere with the historical answer sequence and the answer result; and performing multi-scale state decoupling processing on the knowledge state based on the first causal path, the second causal path, and the third causal path.

[0041] For details, please refer to Figure 2 This application presents a knowledge tracing causal graph provided in its embodiments. To describe the changes in a target student's knowledge state during the learning process and the influence of historical answering behavior on future answering performance, a knowledge tracing causal graph is constructed to represent the causal relationships between variables. Nodes in the knowledge tracing causal graph represent key variables in the learning process, and edges represent the causal relationships between different variables. Specifically, historical answer sequences represent the practice interaction process of the target student at multiple historical moments; knowledge state represents the knowledge mastery state formed by the target student based on the historical learning process; answer results represent the target student's answering performance for the practice task; target practice represents the current practice task to be predicted; and confounding variables represent unobservable influencing factors that simultaneously affect historical answering behavior and answer results.

[0042] Furthermore, based on the knowledge tracing causal graph, a first causal path from the historical answer sequence to the knowledge state is constructed: The first causal path describes how the target student's historical response sequence influences the formation process of knowledge state at different time scales. Since students' knowledge mastery level is not determined by a single practice interaction, but rather gradually accumulated through multiple historical practice behaviors over a long learning process, knowledge state is further divided into long-term knowledge memory states. Short-term fluctuations in response time and the overall learning trend status .

[0043] Among them, long-term knowledge memory state Used to characterize the stable knowledge mastery of a target student on a specific knowledge point over a longer period, reflecting the continuous accumulation of knowledge and experience during the history learning process; short-term fluctuations in answer patterns. Used to characterize changes in cognitive state during recent consecutive responses by target students, reflecting the impact of recent response behavior on current learning status; overall learning trend status. It is used to characterize the changing trend of students' knowledge mastery level in the long-term learning process and to reflect the development law of students' overall learning ability over time.

[0044] Furthermore, a second causal path is constructed between the knowledge state and the answer result: The second causal path describes the process by which knowledge states at different scales interact with target exercises and influence response results. Specifically, when students answer target exercises, their performance is influenced not only by their current knowledge level but also by long-term knowledge accumulation, recent response status, and overall learning trends. Therefore, the second causal path can characterize the relationship between knowledge states at different scales and target exercises, and reflect the comprehensive impact of different knowledge states on students' response results.

[0045] Furthermore, considering that some influencing factors in the student learning process cannot be directly observed, a third causal path corresponding to the confounding variables is constructed: The third causal path is used to describe the process by which confounding variables simultaneously affect both the historical response sequence and the response results. Since some unobservable factors may simultaneously alter students' historical practice behavior and final response performance, creating a false correlation between the historical response sequence and the response results, the third causal path corresponding to the confounding variables is introduced to describe such potential confounding factors.

[0046] Based on the first, second, and third causal paths described above, a multi-scale state decoupling process is performed on the knowledge state, breaking down the original knowledge state into long-term knowledge memory states. Short-term fluctuations in response time and the overall learning trend status This allows for the separate characterization of knowledge change features at different time scales. Specifically, long-term knowledge memory status is used to construct a long-term causal expert model, short-term response fluctuation status is used to construct a short-term causal expert model, and overall learning trend status is used to construct a trend causal expert model, thereby achieving the modeling of the knowledge change process of the target students at different scales.

[0047] Step S103: Construct a long-term causal expert model through long-term knowledge memory state, construct a short-term causal expert model through short-term response fluctuation state, and construct a trend causal expert model through overall learning trend state.

[0048] Specifically, based on the different knowledge states obtained from multi-scale state decoupling, corresponding causal expert models are constructed. Among them, for the long-term knowledge memory state, a long-term causal expert model is constructed to capture the knowledge accumulation of the target student in the long-term learning process and the influence of historical knowledge on the current answer performance; for the short-term answer fluctuation state, a short-term causal expert model is constructed to capture the state changes of the target student in the recent continuous answer process; and for the overall learning trend state, a trend causal expert model is constructed to characterize the trend of knowledge level changes of the target student in the long-term learning process.

[0049] In one possible implementation, step S103 further includes: acquiring practice features, knowledge concept features, and answer result features corresponding to historical answer sequences; fusing the knowledge concept features and answer result features, and adjusting the fused interaction features based on the practice difficulty information corresponding to the practice features to obtain an interaction feature representation of fused practice attributes; performing feature enhancement on the practice sequences corresponding to the practice features and the interaction sequences corresponding to the interaction feature representations based on an attention mechanism to obtain enhanced practice feature representations and enhanced interaction feature representations; and performing feature modeling on long-term knowledge memory state, short-term answer fluctuation state, and overall learning trend state based on the enhanced practice feature representations and enhanced interaction feature representations.

[0050] Specifically, to reduce the interference of confounding variables on the causal relationship modeling between historical response sequences and response results, when performing feature modeling based on enhanced practice feature representations and enhanced interaction feature representations, knowledge state is used as an intermediate variable affecting response results from historical response sequences. Furthermore, the causal relationship between historical response sequences, knowledge state, target practice, and response results is modeled based on the front-door adjustment relation. The historical response sequence is denoted as... The knowledge state is recorded as The target practice is denoted as The answer result is recorded as .

[0051] First, based on the historical response sequence and the target practice, a conditional probability model is performed on the response results, yielding: in, This indicates the student's response to the target exercise in the next time step; This represents the historical sequence of responses up to the current moment; This indicates the target practice for the next moment; This represents the knowledge concept corresponding to the target exercise. Probability is used to represent the probability that the target student will produce a correct answer given the historical answer sequence, the target exercise, and its corresponding knowledge concept.

[0052] To standardize the representation of target exercises and their corresponding knowledge concepts, target exercises and their corresponding knowledge concepts will be uniformly represented as "target exercises". The conditional probability corresponding to the answer result is expressed as: in, Indicates based on historical response sequence and target practice Predicting the answer result The probability of.

[0053] Furthermore, since confounding variables may simultaneously affect both the historical response sequence and the response results, directly predicting the response results based on the historical response sequence may be affected by confounding variables. To reduce the impact of confounding variables on causal relationship estimation, an intervention operation is introduced. By intervening in the historical response sequence, we obtained:

[0054] in, Indicates a sequence of responses to history. Perform interventions to cut off the influence of confounding variables on the historical response sequence; This represents the probability of obtaining a response result when the historical response sequence is interfered with and combined with target practice.

[0055] Furthermore, based on the causal relationship between historical response sequences, knowledge states, and response results in the knowledge tracing causal graph, and taking the knowledge state as a mediating variable influencing the response results from the historical response sequences, the conditional probability after the above intervention is expanded to obtain: in, Indicates the historical sequence of responses; Indicates the state of knowledge; Indicate the goal; Indicates the answer result; This represents the probability of determining the knowledge state based on the historical response sequence; It represents the probability of obtaining a response result under the combined effects of knowledge state, the intervened historical response sequence, and target practice; This represents the summation of the different state values ​​of the knowledge state.

[0056] Based on the front-door adjustment condition, there is no unblocked back-door path between the historical answer sequence and the knowledge state. Therefore, the above expression can be further simplified to: in, This represents the probability of obtaining a response after intervening in the knowledge state and combining it with target practice. This process uses the knowledge state as an intermediate means to pass on the influence of the historical response sequence on the response result, thereby avoiding the direct use of the association between the historical response sequence and the response result, which may be affected by confounding variables.

[0057] Furthermore, based on the causal relationship between knowledge state and answer results, and assuming that knowledge state can block the direct influence of historical answer sequences on answer results, the post-intervention probability is converted into a conditional probability based on knowledge state and target practice, resulting in: in, This represents the probability of determining the knowledge state based on the historical sequence of responses; This represents the probability of obtaining a response under the combined influence of knowledge state and target practice.

[0058] Furthermore, to further explore the different historical observation states in the historical response sequence, possible states corresponding to the historical response sequence are introduced. ,get: in, This indicates the historical observation state within the historical response sequence; Indicates historical observation status The probability of occurrence; Indicates the state of knowledge Historical observation status and target practice The answer is produced by the combined effect. The probability of the outer layer; Used to iterate through the different state values ​​of the knowledge state, inner layer Used to traverse the different historical observation states corresponding to the historical response sequence.

[0059] Step S104: Output the long-term answer prediction results of the target student based on the long-term causal expert model, output the short-term answer prediction results of the target student based on the short-term causal expert model, and output the trend answer prediction results of the target student based on the trend causal expert model.

[0060] Specifically, the long-term answer prediction results for the target students are output based on the long-term causal expert model, the short-term answer prediction results for the target students are output based on the short-term causal expert model, and the trend answer prediction results for the target students are output based on the trend causal expert model.

[0061] Specifically, based on the enhanced practice feature representation and enhanced interaction feature representation obtained in step S103, the long-term causal expert model, the short-term causal expert model and the trend causal expert model are respectively input to model the changes in the knowledge state of the target students at different time scales, and output the answer prediction results at the corresponding time scales respectively.

[0062] For long-term causal expert models, based on enhanced practice feature representations and enhanced interaction feature representations, an attention mechanism is used to model long-term dependencies in historical interactions to capture the knowledge mastery features formed by target students over a longer period. Specifically, the enhanced practice feature representations are mapped to query vectors, key vectors, and value vectors:

[0063] in, Indicates the first The practice query vector corresponding to each attention head; Indicates the first The practice key vector corresponding to each attention head; Indicates the first The interaction value vector corresponding to each attention head; This represents the enhanced practice feature representation; This represents the enhanced interaction feature representation; , as well as These represent the corresponding learnable linear projection matrices.

[0064] Long-term dependencies are calculated using a multi-head attention mechanism based on the query vector, key vector, and value vector. in, Indicates the first Long-term dependency features of the attention head output; This represents the feature dimension corresponding to each attention head; This represents the upper triangular mask matrix, used to restrict the attention mechanism to only utilize historical information up to the current time step.

[0065] Furthermore, the features output by multiple attention heads are concatenated and fused using an output mapping matrix: in, This indicates the characteristics of practice after multi-head attention fusion; Indicates the number of attention heads; Indicates feature concatenation operation; This represents the final output mapping matrix.

[0066] Based on the fused features, residual connections and normalization are performed to obtain the long-term scale features corresponding to the long-term causal expert model: in, This indicates the enhanced long-term dependency practice characteristics; Indicates random deactivation; This indicates normalization processing.

[0067] Based on the aforementioned long-term scale features and enhanced interaction feature representations, the historical knowledge accumulation process of the target students is modeled to obtain long-term knowledge memory state features. Combined with the causal prediction process, the long-term answer prediction results are output to represent the target students' answer performance for target exercises based on long-term knowledge accumulation.

[0068] For the short-term causal expert model, based on the enhanced practice feature representation and the enhanced interaction feature representation, a temporal convolutional network is used to model the historical interaction changes within the local time window, extract the state change information of the target student in the recent continuous answering process, and obtain the short-term answer fluctuation state features.

[0069] Specifically, the enhanced interactive feature representations are input into a temporal convolutional network in chronological order. Convolutional operations capture local correlations between adjacent time steps, enabling the model to identify patterns of change in the target student's recent answering behavior. For example, consecutive correct answers may indicate an improvement in recent knowledge mastery, while consecutive incorrect answers may indicate a decline. Nonlinear feature extraction is performed based on the correlation between short-term answer fluctuation features and the enhanced interactive feature representations, and this is combined with the corresponding prediction process to obtain short-term answer prediction results.

[0070] For the trend causal expert model, based on the enhanced practice feature representation and the enhanced interaction feature representation, the state evolution model is performed on the interaction sequence corresponding to multiple historical time steps through the gated recurrent unit, and the overall change trend of the target student in the long-term learning process is extracted to obtain the overall learning trend state features.

[0071] Specifically, the enhanced interactive feature representations are input into the gated loop unit in chronological order, and the hidden state update process describes the continuous changes in the target student's knowledge mastery level as the learning process progresses. Among them, the overall learning trend state feature is used to reflect the development trend of the target student's general ability over time, and is distinguished from the knowledge accumulation information corresponding to the long-term knowledge memory state and the recent change information corresponding to the short-term answer fluctuation state.

[0072] Nonlinear feature extraction is performed based on the correlation between the overall learning trend state features and the enhanced interactive feature representation, and the trend response prediction result is obtained by combining the corresponding prediction process.

[0073] Through the above processing, the long-term causal expert model, the short-term causal expert model, and the trend causal expert model model the knowledge status of the target students from three different time scales: long-term knowledge accumulation, short-term answer changes, and overall learning development trends. They also output long-term answer prediction results, short-term answer prediction results, and trend answer prediction results, respectively, providing prediction results at different scales for the subsequent multi-expert prediction fusion mechanism.

[0074] In one possible implementation, step S104 further includes: generating a query vector and a key vector based on the enhanced practice feature representation, and generating a value vector based on the enhanced interaction feature representation; performing weighted modeling based on the query vector, key vector, and value vector using an attention mechanism, and adjusting the attention weights corresponding to the attention mechanism based on the distance relationship between historical time steps to obtain long-term knowledge memory state features; performing nonlinear feature extraction on the correlation between the long-term knowledge memory state features and the enhanced interaction feature representation to obtain long-term response features; and fusing the long-term knowledge memory state features and the long-term response features to obtain a long-term response prediction result.

[0075] Specifically, the long-term causal expert model is used to capture the historical knowledge dependencies of target students over a long period of time. By modeling the enhanced practice feature representation and the enhanced interaction feature representation, it extracts long-term knowledge memory state features that reflect the long-term knowledge accumulation of target students. Long-term knowledge memory state features are used to represent the knowledge mastery status formed by target students in the interaction process of multiple historical time steps. For example, the stable knowledge mastery formed by target students in the early learning process will continue to affect their performance in subsequent related exercises.

[0076] First, query vectors and key vectors are generated based on the enhanced practice feature representations, and value vectors are generated based on the enhanced interaction feature representations. Specifically, the enhanced practice feature representations and enhanced interaction feature representations are mapped to their corresponding feature spaces using trainable linear mapping matrices, resulting in:

[0077] in, Indicates the first The query vector corresponding to each attention head; Indicates the first The key vector corresponding to each attention head; Indicates the first The value vector corresponding to each attention head; This represents the enhanced practice feature representation; This represents the enhanced interaction feature representation; , as well as This represents a trainable weight matrix; This indicates the number of the attention head.

[0078] Furthermore, to capture the relationships between different time steps in the historical interaction process, attention weights are calculated using query vectors, key vectors, and value vectors, and adjusted based on the distance relationships between historical time steps. Since students' knowledge acquisition process has a long-term dependency characteristic, the influence of knowledge states formed in earlier time steps on current performance gradually weakens as time intervals increase. Therefore, a time distance decay mechanism is introduced to correct the attention weights:

[0079] in, Indicates the first Attention is focused on the time step The corresponding adjusted attention weights; This represents the feature dimension corresponding to each attention head; This represents a learnable attenuation coefficient, used to control the degree of attenuation of historical information as time distance increases; Indicates the current time step Step by step with historical time The time distance between them.

[0080] Based on the adjusted attention weights, the final attention distribution is obtained through normalization: in, Indicates the first The attention weights corresponding to each attention head; This represents the upper triangular mask matrix, used to restrict the model to access only historical time step information; Indicates the historical time step index; This represents the normalization function, used to convert the adjusted attention weights into a probability distribution.

[0081] Furthermore, the value vector is weighted and aggregated using attention weights, and the outputs of multiple attention heads are concatenated and fused: in, Indicates the first The output features corresponding to each attention head; Indicates the first A focus of attention on historical time steps The corresponding attention weights; Indicates the first Each attention head corresponds to a value vector of historical time steps.

[0082] By concatenating the outputs of multiple attention heads and combining them with residual connections and normalization, the long-term knowledge memory state characteristics are obtained: in, This represents the long-term knowledge memory state characteristics after fusing the outputs of multiple attention heads; Indicates feature concatenation operation; Indicates the output mapping matrix; This represents the enhanced training feature representation corresponding to the current time step; This indicates normalization processing.

[0083] Furthermore, to enhance the nonlinear expressive power of long-term knowledge memory state features, a high-dimensional representation corresponding to the long-term knowledge memory state features is extracted through nonlinear mapping: in, This represents the final characteristics of the long-term knowledge memory state. and This represents a trainable weight matrix; and Indicates the bias parameter; This represents a non-linear activation function.

[0084] Through the above processing, long-term knowledge memory state features are obtained, which are used to represent the long-term knowledge mastery state of the target student based on the historical interaction process.

[0085] Furthermore, to predict students' performance on target exercises based on long-term knowledge memory state features, nonlinear feature extraction is performed on the correlation between long-term knowledge memory state features and enhanced interaction feature representations to obtain long-term response features. Specifically, long-term knowledge memory state features, historical interaction features, and the interaction features between the two are fused:

[0086] in, This indicates long-term response characteristics; This indicates the characteristics of long-term knowledge memory state; The enhanced interactive feature representation corresponding to a historical moment; , as well as This represents the trainable weight matrix used to extract long-term knowledge memory state features, enhanced interaction feature representations, and the interaction features between the two. This indicates element-wise multiplication; Represents the bias vector; This represents the Sigmoid activation function.

[0087] Finally, the long-term knowledge memory state features and long-term response features are fused, and combined with the target practice information to obtain the long-term response prediction results: in, This indicates the prediction results for long-term responses; This represents the output layer weight matrix; This represents the element-by-element fusion result between long-term knowledge memory state characteristics and long-term response characteristics; This represents a vector concatenation operation; This indicates the practice features corresponding to the current target practice; Indicates the output layer bias parameters; This represents the Sigmoid activation function.

[0088] Through the above process, the long-term causal expert model first generates long-term knowledge memory state features representing the long-term knowledge accumulation process using enhanced practice feature representations and enhanced interaction feature representations. Then, based on the correlation between the long-term knowledge memory state features and the enhanced interaction feature representations, it extracts long-term response features and finally fuses them to obtain the long-term response prediction results. This process can use long-term historical interaction information to characterize the impact of students' knowledge mastery status accumulated over time, improving the accuracy of response prediction on a long-term scale.

[0089] In one possible implementation, step S104 further includes: outputting short-term answer prediction results corresponding to the target student based on a short-term causal expert model, specifically including: extracting interaction change features within a local time window through a temporal convolutional network based on the enhanced practice feature representation and the enhanced interaction feature representation to obtain short-term answer fluctuation state features; performing nonlinear feature extraction based on the correlation between the short-term answer fluctuation state features and the enhanced interaction feature representation to obtain short-term answer response features; and fusing the short-term answer fluctuation state features and the short-term answer response features to obtain short-term answer prediction results.

[0090] Specifically, to obtain information on the recent learning status changes of the target student, the input features of the short-term causal expert model are first constructed based on the enhanced practice feature representation and the enhanced interaction feature representation. Since short-term response performance is influenced not only by the current target practice but also by recent historical interaction behavior, the enhanced practice feature representation corresponding to the current time step is fused with the enhanced interaction feature representation corresponding to the previous historical time step to obtain a joint feature representation:

[0091] in, Indicates the first Joint feature representation corresponding to time step; This represents the enhanced practice feature representation; Indicates the first The enhanced interaction feature representation corresponding to each time step; This represents a vector concatenation operation; and This represents the learnable weight matrix used for feature mapping; and Indicates the bias term; This represents a non-linear activation function.

[0092] Through the above processing, the input features simultaneously include current practice attribute information and recent historical interaction information, providing a feature basis for the subsequent temporal convolutional network to extract short-term change patterns.

[0093] Furthermore, the joint feature representation is input into a temporal convolutional network. This network performs local temporal modeling on the interaction sequence within a finite time window, thereby extracting the changing features of the target student's recent responses. The temporal convolutional network employs a causal convolutional approach with an expanded structure, ensuring that the output at the current moment depends only on the current moment and historical time step information, thus preventing future information from influencing the current prediction result.

[0094] For the Layer-1 temporal convolutional network, in the 1st... The convolution process at time step 1 is represented as: in, Indicates the first Layer-time convolutional networks in the 1st layer The convolutional features output at each time step; Indicates the first Features output by a multi-layered temporal convolutional network; Indicates the first The first layer of temporal convolutional networks The learnable weight matrix corresponding to each convolutional kernel; Indicates the kernel size; Indicates the first The dilation coefficients corresponding to the layer-time convolutional network; This indicates the bias term.

[0095] Among them, the expansion coefficient This is used to control the historical time range that the temporal convolutional network can perceive. By using different levels of dilated convolution, the model can simultaneously capture continuous changes in response behavior within a short time range and local change trends within a slightly longer time range.

[0096] Furthermore, to enhance the nonlinear expressive power of convolutional output features and improve model training stability, weight normalization, nonlinear activation, and random deactivation are applied to the temporal convolutional network output: in, Indicates the first Enhanced features output by a multi-layered temporal convolutional network; This indicates weight normalization processing; Indicates random deactivation; This represents a non-linear activation function.

[0097] To avoid the loss of original short-term interaction information caused by deep temporal convolutions, a residual concatenation is performed between the output of the previous temporal convolution and the output of the current layer: in, This represents the short-term temporal characteristics after residual fusion; Indicates the first Layer-time convolutional networks output features; Indicates the first The output features of a multilayer temporal convolutional network.

[0098] Furthermore, a nonlinear mapping is performed on the short-term time-series features after residual fusion to obtain the short-term response fluctuation state features: in, This indicates the characteristics of short-term fluctuations in response patterns; and Represents the learnable weight matrix; and This represents the bias parameter.

[0099] The short-term response fluctuation features obtained through the above temporal convolution processing can describe the changes in the target student's recent continuous interaction process. For example, consecutive correct answers can increase the short-term response fluctuation, while consecutive incorrect answers can decrease the short-term response fluctuation.

[0100] Furthermore, in order to predict the target student's response performance to the current target practice based on the short-term response fluctuation state characteristics, nonlinear feature extraction is performed based on the correlation between the short-term response fluctuation state characteristics and the enhanced interaction feature representation to obtain short-term response characteristics.

[0101] Specifically, the short-term response fluctuation state features, the enhanced interaction feature representation, and the interaction features between the two are fused: in, This indicates short-term response characteristics; This indicates the characteristics of short-term fluctuations in response patterns; This represents the enhanced interactive feature representation corresponding to the historical time step. , and This represents a learnable weight matrix used to extract different associated features; This indicates the output mapping weight matrix; This represents the element-wise multiplication operation; Represents the bias vector; This represents the Sigmoid activation function.

[0102] Through the above processing, the nonlinear correlation between recent state change information and historical interaction information is modeled to obtain short-term response characteristics.

[0103] Finally, the short-term response fluctuation features and short-term response features are fused together, and combined with the practice features corresponding to the target practice, to obtain the short-term response prediction results: in, This indicates the short-term prediction result; This represents the learnable weight matrix of the output layer; This represents the element-wise fusion result between short-term response fluctuation characteristics and short-term response characteristics; This represents a vector concatenation operation; This indicates the practice characteristics corresponding to the target practice; Indicates the output layer bias term; This represents the Sigmoid activation function.

[0104] Through the above process, the short-term causal expert model first uses a temporal convolutional network to extract recent local time window change information from the enhanced practice feature representation and the enhanced interaction feature representation to obtain short-term answer fluctuation state features; then it combines the enhanced interaction feature representation to extract short-term answer response features; finally, it merges the two to obtain short-term answer prediction results, thereby achieving the prediction of the recent answer performance trend of the target student.

[0105] In one possible implementation, step S104 further includes: outputting the trend response prediction result corresponding to the target student based on the trend causal expert model, specifically including: based on the enhanced practice feature representation and the enhanced interaction feature representation, performing state evolution modeling on the interaction sequence corresponding to multiple historical time steps through a gated recurrent unit to obtain the overall learning trend state feature; performing nonlinear feature extraction based on the correlation between the overall learning trend state feature and the enhanced interaction feature representation to obtain the trend response feature; and fusing the overall learning trend state feature and the trend response feature to obtain the trend response prediction result.

[0106] Specifically, to construct the input features corresponding to the trend causal expert model, the enhanced practice feature representation and the enhanced interaction feature representation are first fused together, using the current target practice information and historical interaction information as the basis for trend state evolution modeling: in, Indicates the first The trend state input features corresponding to each moment; This represents the enhanced practice feature representation; Indicates the first The enhanced interaction feature representation corresponding to each time step; This represents a vector concatenation operation; and Represents the learnable weight matrix; and Indicates the bias term; This represents a non-linear activation function.

[0107] Through the above processing, the enhanced practice feature representation and the enhanced interaction feature representation are fused, enabling the trend causal expert model to simultaneously obtain information on current practice attributes and state changes during historical interactions.

[0108] Furthermore, the trend state input features are input into the gated recurrent unit, and state evolution modeling is performed through the interaction sequences corresponding to multiple historical time steps. The gated recurrent unit describes the changing pattern of the target student's knowledge state as the learning process occurs through update gates, reset gates, and candidate hidden states.

[0109] First, based on the trend state input features and the hidden state corresponding to the previous time step, calculate the update gate: in, Indicates the first The corresponding update gate at any given time; Input features that indicate trend status; Indicates the first The hidden state corresponding to each moment; and This represents the learnable weight matrix corresponding to the update gate; Indicates the bias term; This represents the Sigmoid activation function.

[0110] The update gate is used to control the fusion ratio between historical trend state information and current input information, enabling the model to choose to retain existing trend information or introduce new interaction information based on changes in the learning process.

[0111] Furthermore, a reset gate is calculated based on the trend state input features and the hidden state of the previous time step: in, Indicates the first The corresponding reset door at any given time; and This represents the learnable weight matrix corresponding to the reset gate; This indicates the bias term.

[0112] The reset gate is used to control the degree to which historical trend state information participates in the current state update, enabling the model to adjust the impact of information from different historical time steps on the current trend state.

[0113] Furthermore, candidate trend states are generated based on the trend state input features and the reset historical hidden states: in, Indicates the first The candidate trend state corresponding to the given moment; and This represents the learnable weight matrix calculated for each candidate state. Indicates the bias term; This represents the element-wise multiplication operation; This represents the hyperbolic tangent activation function.

[0114] Furthermore, by fusing the historical trend states and the current candidate trend states according to the update gate, the trend state corresponding to the current time step is obtained: in, This indicates the trend state characteristics corresponding to the current time step; Indicates the first The hidden state of the output of the time-gated loop unit; This represents the hidden state corresponding to the previous time step; Indicates an update to the door; Indicates the candidate trend status.

[0115] Through the above-mentioned gated loop unit processing, state evolution modeling can be performed on the interaction sequence corresponding to multiple historical time steps, extracting the overall ability change pattern of the target student in the continuous learning process, and obtaining the overall learning trend state characteristics.

[0116] Furthermore, to enhance the nonlinear representation of the overall learning trend state characteristics, a nonlinear mapping is applied to the output of the gated recurrent unit: in, This indicates the overall learning trend and state characteristics; and Represents the learnable weight matrix; and This indicates the bias term.

[0117] Through the above processing, we obtain the overall learning trend characteristics that reflect the gradual changes in the general abilities of target students as they learn.

[0118] Furthermore, nonlinear feature extraction is performed based on the correlation between the overall learning trend state features and the enhanced interaction feature representation to obtain trend response features. Specifically, the overall learning trend state features, historical interaction features, and the interaction features between the two are fused:

[0119] in, Indicates trend-based response characteristics; This indicates the overall learning trend and state characteristics; Indicates the first The enhanced interaction feature representation corresponding to each time step; , , as well as Represents the learnable weight matrix; Represents the bias vector; This represents the Sigmoid activation function.

[0120] Through the above processing, the overall trend state formed during the long-term learning process is correlated with historical interaction behavior to obtain trend response characteristics that can reflect the impact of the target student's trend changes on the current answer.

[0121] Finally, the overall learning trend state features and trend response features are fused together, and combined with the practice features corresponding to the target practice, to obtain the trend response prediction results: in, Indicate the trend in the prediction results; This represents the learnable weight matrix corresponding to the output layer; The features represent the result of element-wise fusion of overall learning trend state characteristics and trend response characteristics; This represents a vector concatenation operation; This indicates the practice features corresponding to the current target practice; Indicates the output layer bias term; This represents the Sigmoid activation function.

[0122] Through the above process, the trend causal expert model, based on the enhanced practice feature representation and the enhanced interaction feature representation, performs state evolution modeling on the interaction sequence corresponding to multiple historical time steps through a gated recurrent unit to obtain the overall learning trend state features; further, it extracts the trend response features by combining the enhanced interaction feature representation; finally, it integrates the overall learning trend state features and the trend response features to obtain the trend response prediction results, thereby realizing the prediction of the overall learning ability change trend of the target student.

[0123] Step S105: The long-term answer prediction results, short-term answer prediction results and trend answer prediction results are fused through a multi-expert prediction fusion mechanism, and the answer prediction results corresponding to the target student are output.

[0124] Specifically, long-term answer prediction results, short-term answer prediction results, and trend answer prediction results are obtained separately. Based on the differences between the prediction results at the three different time scales, the long-term answer prediction results, short-term answer prediction results, and trend answer prediction results are adaptively fused so that the final answer prediction results can simultaneously reflect the target student's long-term knowledge mastery, recent answer fluctuations, and overall learning trends, thereby outputting the corresponding answer prediction results for the target student.

[0125] In one possible implementation, step S105 further includes: calculating the state similarity of long-term knowledge memory state features, short-term answer fluctuation state features, and overall learning trend state features at the same time using a similarity calculation method; constraining the feature differences among long-term knowledge memory state features, short-term answer fluctuation state features, and overall learning trend state features based on state similarity; adjusting the prediction loss corresponding to long-term answer prediction results, short-term answer prediction results, and trend answer prediction results using a homoscedastic uncertainty weighting method based on the feature differences; and optimizing the long-term causal expert model, short-term causal expert model, and trend causal expert model based on the adjusted prediction loss.

[0126] Specifically, after obtaining the long-term response prediction results from the long-term causal expert model, the short-term response prediction results from the short-term causal expert model, and the trend response prediction results from the trend causal expert model, the prediction results at the three different time scales are first fused. Since the long-term, short-term, and trend response prediction results reflect the target student's performance from three perspectives—long-term knowledge accumulation, recent response changes, and overall learning trends—the three expert outputs are fused to obtain the final response prediction result.

[0127] in, This indicates the predicted final answer for the target student; This indicates the prediction results for long-term responses; This indicates the short-term prediction result; The prediction result indicates the trend.

[0128] Furthermore, to ensure that the long-term knowledge memory state features, short-term answer fluctuation state features, and overall learning trend state features retain the feature information at their respective time scales, and to avoid different expert modules extracting similar or redundant state representations, the similarity relationship between the three state features is constrained.

[0129] For the same moment For any two state features, calculate the state similarity between long-term knowledge memory state features, short-term response fluctuation state features, and overall learning trend state features using the cosine similarity calculation method: in, and They represent the first Any two state features corresponding to time; ,and ; This represents the state similarity between two state features; Indicates the calculation of the second norm of a vector; This indicates a preset positive number, used to avoid the denominator being zero.

[0130] The above method can be used to measure the degree of information overlap of different experts' corresponding state characteristics at the same time scale.

[0131] Furthermore, based on the calculated state similarity, constraints are imposed on the differences between long-term knowledge memory state characteristics, short-term response fluctuation state characteristics, and overall learning trend state characteristics to construct a state decoupling constraint loss: in, This represents the loss due to state decoupling constraints. Indicates the length of the historical time series; Indicates the first Characteristics of long-term knowledge memory state corresponding to a given moment; Indicates the first The short-term fluctuation characteristics of responses at any given moment; Indicates the first The overall learning trend and state characteristics corresponding to each moment.

[0132] The state decoupling constraint loss minimizes the cosine similarity between different state features, enabling the three expert modules to learn student knowledge state information at different time scales. For example, long-term knowledge memory state features are used to describe the long-term knowledge accumulation process, short-term answer fluctuation state features are used to describe recent changes in answer behavior, and overall learning trend state features are used to describe the overall trend of student ability changes, thereby reducing information redundancy among the three expert modules.

[0133] Furthermore, to enable the three expert models to automatically adjust their optimization level based on the uncertainty of their own prediction tasks, adaptive weighting is applied to the prediction loss corresponding to long-term response prediction results, the prediction loss corresponding to short-term response prediction results, and the prediction loss corresponding to trend response prediction results.

[0134] First, the prediction loss for each of the three expert models is calculated. The prediction loss for the long-term response prediction results is calculated using the binary cross-entropy loss function.

[0135] in, This represents the prediction loss corresponding to the long-term causal expert model. Indicates the target student in the first The actual answer result corresponding to the given moment; This indicates the long-term prediction result.

[0136] The prediction loss for the short-term causal expert model is: in, This represents the prediction loss corresponding to the short-term causal expert model; This indicates the short-term prediction result.

[0137] The prediction loss for the trend causal expert model is: in, This represents the prediction loss corresponding to the trend-causal expert model. The prediction result indicates the trend.

[0138] Furthermore, based on the homoscedastic uncertainty weighting method, the prediction losses of the three expert models are weighted and adjusted to obtain the overall training loss: in, This represents the overall training loss of the model; Indicate the categories of expert models, including , as well as ; Indicates the first The prediction loss corresponding to each expert model; Indicates the first The homoscedasticity uncertainty parameter corresponding to each expert model is used to represent the degree of uncertainty of the corresponding prediction task; This represents the decoupling constraint weight parameter, which is used to adjust the degree of influence of the state decoupling constraint loss on the overall training process; This represents the loss due to state decoupling constraints.

[0139] Among them, the homoscedastic uncertainty weighting method learns the uncertainty parameters corresponding to different expert tasks. It automatically adjusts the contribution of different prediction losses to the overall optimization process. When a certain expert model has significant uncertainty in its prediction task, its corresponding loss weight is reduced; when a certain expert model's prediction is more stable, its corresponding loss weight is increased, thereby achieving a dynamic balance among the three expert models.

[0140] Based on the above process, the information redundancy between long-term knowledge memory state features, short-term answer fluctuation state features, and overall learning trend state features is reduced by state decoupling constraint loss. The prediction loss corresponding to long-term answer prediction results, short-term answer prediction results, and trend answer prediction results is adaptively adjusted by homoscedastic uncertainty weighting. Finally, the long-term causal expert model, short-term causal expert model, and trend causal expert model are optimized so that the three expert models can learn knowledge state features at different scales and improve the accuracy of the target student's answer prediction results.

[0141] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0142] Please refer to Figure 3 This illustration shows a schematic diagram of a student answer prediction device based on multi-scale knowledge state decoupling provided in an embodiment of this application. The device includes a knowledge state construction module 31, a state decoupling module 32, and a multi-expert prediction module 33, wherein... The knowledge state construction module 31 is used to obtain the historical answer sequence of the target student and obtain the corresponding knowledge state of the target student based on the historical answer sequence.

[0143] The state decoupling module 32 is used to perform multi-scale state decoupling processing on the knowledge state to obtain the decoupled long-term knowledge memory state, short-term answer fluctuation state, and overall learning trend state.

[0144] The multi-expert prediction module 33 is used to construct a long-term causal expert model based on the long-term knowledge memory state, a short-term causal expert model based on the short-term response fluctuation state, and a trend causal expert model based on the overall learning trend state.

[0145] The multi-expert prediction module 33 is used to output the long-term answer prediction results of the target student based on the long-term causal expert model, the short-term answer prediction results of the target student based on the short-term causal expert model, and the trend answer prediction results of the target student based on the trend causal expert model.

[0146] The multi-expert prediction module 33 is used to fuse long-term answer prediction results, short-term answer prediction results and trend answer prediction results through a multi-expert prediction fusion mechanism, and output the answer prediction results corresponding to the target student.

[0147] In one possible implementation, the state decoupling module 32 is used to perform multi-scale state decoupling processing on the knowledge state, specifically including: constructing a knowledge tracing causal graph; the nodes of the knowledge tracing causal graph represent variables, and the edges represent causal relationships; wherein, the nodes include historical answer sequences, knowledge states, answer results, target exercises, and confounding variables; constructing a first causal path from the historical answer sequence to the knowledge state; the first causal path is used to describe the process by which the target student forms the current knowledge mastery state based on the historical answer sequence; constructing a second causal path from the knowledge state and target exercises to the answer result; the second causal path is used to describe the process by which the target student generates answer results for the target exercises based on the knowledge state; constructing a third causal path in which confounding variables simultaneously affect the historical answer sequence and the answer result; the third causal path is used to describe the process by which unobservable influencing factors interfere with the historical answer sequence and the answer result; and performing multi-scale state decoupling processing on the knowledge state based on the first causal path, the second causal path, and the third causal path.

[0148] In one possible implementation, before constructing a long-term causal expert model through long-term knowledge memory state, a short-term causal expert model through short-term answer fluctuation state, and a trend causal expert model through overall learning trend state, the multi-expert prediction module 33 further includes: acquiring practice features, knowledge concept features, and answer result features corresponding to historical answer sequences; fusing the knowledge concept features and answer result features, and adjusting the fused interaction features based on the practice difficulty information corresponding to the practice features to obtain an interaction feature representation of fused practice attributes; performing feature enhancement on the practice sequences corresponding to the practice features and the interaction sequences corresponding to the interaction feature representations based on an attention mechanism to obtain enhanced practice feature representations and enhanced interaction feature representations; and performing feature modeling on the long-term knowledge memory state, short-term answer fluctuation state, and overall learning trend state based on the enhanced practice feature representations and enhanced interaction feature representations.

[0149] In one possible implementation, the multi-expert prediction module 33 is used to output the long-term answer prediction result corresponding to the target student based on the long-term causal expert model. Specifically, it includes: generating query vectors and key vectors based on the enhanced practice feature representation, and generating value vectors based on the enhanced interaction feature representation; performing weighted modeling through an attention mechanism based on the query vector, key vector, and value vector, and adjusting the attention weights corresponding to the attention mechanism based on the distance relationship between historical time steps to obtain long-term knowledge memory state features; performing nonlinear feature extraction on the correlation between the long-term knowledge memory state features and the enhanced interaction feature representation to obtain long-term answer response features; and fusing the long-term knowledge memory state features and the long-term answer response features to obtain the long-term answer prediction result.

[0150] In one possible implementation, the multi-expert prediction module 33 is used to output the short-term answer prediction result corresponding to the target student based on the short-term causal expert model. Specifically, it includes: extracting interaction change features within a local time window through a temporal convolutional network based on the enhanced practice feature representation and the enhanced interaction feature representation to obtain short-term answer fluctuation state features; performing nonlinear feature extraction based on the correlation between the short-term answer fluctuation state features and the enhanced interaction feature representation to obtain short-term answer response features; and fusing the short-term answer fluctuation state features and the short-term answer response features to obtain the short-term answer prediction result.

[0151] In one possible implementation, the multi-expert prediction module 33 is used to output the trend response prediction result corresponding to the target student based on the trend causal expert model. Specifically, it includes: based on the enhanced practice feature representation and the enhanced interaction feature representation, performing state evolution modeling on the interaction sequence corresponding to multiple historical time steps through a gated loop unit to obtain the overall learning trend state feature; performing nonlinear feature extraction based on the correlation between the overall learning trend state feature and the enhanced interaction feature representation to obtain the trend response feature; and fusing the overall learning trend state feature and the trend response feature to obtain the trend response prediction result.

[0152] In one possible implementation, the multi-expert prediction module 33 is used to further include, before fusing the long-term response prediction results, short-term response prediction results, and trend response prediction results through the multi-expert prediction fusion mechanism: calculating the state similarity of long-term knowledge memory state features, short-term response fluctuation state features, and overall learning trend state features at the same time using a similarity calculation method; constraining the feature differences among long-term knowledge memory state features, short-term response fluctuation state features, and overall learning trend state features based on state similarity; adjusting the prediction loss corresponding to the long-term response prediction results, the short-term response prediction results, and the trend response prediction results using a homoscedastic uncertainty weighting method based on the feature differences; and optimizing the long-term causal expert model, the short-term causal expert model, and the trend causal expert model based on the adjusted prediction loss.

[0153] It should be noted that the above embodiments of the apparatus are only illustrated by the division of the above functional modules. In practical 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 apparatus and method embodiments provided above belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0154] This application also provides an electronic device. (See reference...) Figure 4 , Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include: at least one processor 401, at least one communication bus 402, a user interface 403, at least one network interface 404, and a memory 405.

[0155] The communication bus 402 is used to enable communication between these components.

[0156] The user interface 403 may include a display screen and a camera. Optionally, the user interface 403 may also include a standard wired interface and a wireless interface.

[0157] The network interface 404 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0158] The processor 401 may include one or more processing cores. The processor 401 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory 405, and by calling data stored in memory 405. Optionally, the processor 401 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 401 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip without being integrated into the processor 401.

[0159] The memory 405 may include random access memory (RAM) or read-only memory. Optionally, the memory 405 may include a non-transitory computer-readable storage medium. The memory 405 may be used to store instructions, programs, code, code sets, or instruction sets. The memory 405 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 touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 405 may also be at least one storage device located remotely from the aforementioned processor 401. (Refer to...) Figure 4 The memory 405, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a student answer prediction application based on multi-scale knowledge state decoupling.

[0160] exist Figure 4 In the illustrated electronic device, the user interface 403 is primarily used to provide an input interface for the user and acquire user input data; while the processor 401 can be used to call the student answer prediction application based on multi-scale knowledge state decoupling stored in the memory 405. When executed by one or more processors 401, the electronic device performs one or more of the methods described in the above embodiments. It should be noted that, for the foregoing method embodiments, for the sake of simplicity, they are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0161] This application also provides a non-transitory computer-readable storage medium storing instructions. When executed by one or more processors, these instructions cause an electronic device to perform one or more of the methods described in the above embodiments.

[0162] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0163] In the various embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some service interface; the indirect coupling or communication connection between apparatuses or units may be electrical or other forms.

[0164] The units described as separate components may or may not be physically separate. The 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 the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0165] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0166] 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 storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part 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 to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0167] The above description is merely an exemplary embodiment disclosed in this application and should not be construed as limiting the scope of this application. Any equivalent changes and modifications made in accordance with the teachings of this application shall still fall within the scope of this application.

[0168] This application is intended to cover any variations, uses, or adaptations disclosed herein that follow the general principles disclosed herein and include common knowledge or customary technical means in the art that are not described in this application.

Claims

1. A student response prediction method based on multi-scale knowledge state decoupling, characterized in that, The method includes: Obtain the target student's historical answer sequence, and obtain the target student's corresponding knowledge status based on the historical answer sequence; The knowledge state is subjected to multi-scale state decoupling processing to obtain the decoupled long-term knowledge memory state, short-term answer fluctuation state and overall learning trend state. A long-term causal expert model is constructed based on the long-term knowledge memory state, a short-term causal expert model is constructed based on the short-term response fluctuation state, and a trend causal expert model is constructed based on the overall learning trend state. Based on the long-term causal expert model, the system outputs the long-term answer prediction results for the target student; based on the short-term causal expert model, the system outputs the short-term answer prediction results for the target student; and based on the trend causal expert model, the system outputs the trend answer prediction results for the target student. The long-term answer prediction results, the short-term answer prediction results, and the trend answer prediction results are fused through a multi-expert prediction fusion mechanism, and the answer prediction results corresponding to the target student are output.

2. The method according to claim 1, characterized in that, The multi-scale state decoupling process for the knowledge state specifically includes: Construct a knowledge tracing causal graph; the nodes of the knowledge tracing causal graph represent variables, and the edges represent causal relationships; wherein, the nodes include the historical answer sequence, the knowledge state, the answer result, the target practice, and confounding variables; Construct a first causal path from the historical answer sequence to the knowledge state; the first causal path is used to describe the process by which the target student forms the current knowledge mastery state based on the historical answer sequence; A second causal path is constructed from the knowledge state and the target practice to the answer result; the second causal path is used to describe the process by which the target student generates the answer result based on the knowledge state for the target practice; A third causal path is constructed in which the confounding variables simultaneously affect the historical response sequence and the response results; the third causal path is used to describe the process by which unobservable influencing factors interfere with the historical response sequence and the response results; The knowledge state is subjected to multi-scale state decoupling processing based on the first causal path, the second causal path, and the third causal path.

3. The method according to claim 1, characterized in that, Before constructing a long-term causal expert model through the long-term knowledge memory state, a short-term causal expert model through the short-term response fluctuation state, and a trend causal expert model through the overall learning trend state, the method further includes: Obtain the practice features, knowledge concept features, and answer result features corresponding to the historical answer sequence; The knowledge concept features and the answer result features are fused together, and the fused interactive features are adjusted based on the practice difficulty information corresponding to the practice features to obtain the interactive feature representation of the fused practice attributes; Based on the attention mechanism, feature enhancement is performed on the practice sequence corresponding to the practice feature and the interaction sequence corresponding to the interaction feature representation, respectively, to obtain the enhanced practice feature representation and the enhanced interaction feature representation; Based on the enhanced practice feature representation and the enhanced interaction feature representation, feature modeling is performed on the long-term knowledge memory state, the short-term response fluctuation state, and the overall learning trend state.

4. The method according to claim 3, characterized in that, The step of outputting the long-term answer prediction results for the target student based on the long-term causal expert model specifically includes: Query vectors and key vectors are generated based on the enhanced practice feature representation, and value vectors are generated based on the enhanced interaction feature representation; Based on the query vector, the key vector, and the value vector, a weighted model is performed using an attention mechanism, and the attention weights corresponding to the attention mechanism are adjusted based on the distance relationship between historical time steps to obtain long-term knowledge memory state features. Nonlinear feature extraction is performed on the correlation between the long-term knowledge memory state features and the enhanced interaction feature representation to obtain long-term response features; The long-term knowledge memory state features and the long-term response features are fused to obtain the long-term response prediction result.

5. The method according to claim 4, characterized in that, The step of outputting the short-term answer prediction results for the target student based on the short-term causal expert model specifically includes: Based on the enhanced practice feature representation and the enhanced interaction feature representation, the interaction change features within a local time window are extracted through a temporal convolutional network to obtain short-term response fluctuation state features. Nonlinear feature extraction is performed based on the correlation between the short-term response fluctuation state features and the enhanced interaction feature representation to obtain short-term response features; The short-term response fluctuation features and the short-term response features are fused to obtain the short-term response prediction result.

6. The method according to claim 5, characterized in that, The step of outputting the trend-based causal expert model for the target student's predicted answer trend specifically includes: Based on the enhanced practice feature representation and the enhanced interaction feature representation, the state evolution model of the interaction sequence corresponding to multiple historical time steps is performed by a gated loop unit to obtain the overall learning trend state features. Nonlinear feature extraction is performed based on the correlation between the overall learning trend state features and the enhanced interaction feature representation to obtain trend response features; The overall learning trend state features and the trend response features are fused to obtain the trend response prediction result.

7. The method according to claim 6, characterized in that, Before fusing the long-term response prediction results, the short-term response prediction results, and the trend response prediction results through a multi-expert prediction fusion mechanism, the method further includes: The similarity of the long-term knowledge memory state features, the short-term answer fluctuation state features, and the overall learning trend state features at the same time is calculated using a similarity calculation method. Based on the state similarity constraint, the feature differences between the long-term knowledge memory state features, the short-term answer fluctuation state features, and the overall learning trend state features; Based on the aforementioned feature differences, a homoscedastic uncertainty weighting method is used to adjust the weights of the prediction loss corresponding to the long-term response prediction result, the prediction loss corresponding to the short-term response prediction result, and the prediction loss corresponding to the trend response prediction result. Based on the adjusted prediction loss, the long-term causal expert model, the short-term causal expert model, and the trend causal expert model are optimized.

8. A student answer prediction device based on multi-scale knowledge state decoupling, characterized in that, The device includes a knowledge state construction module, a state decoupling module, and a multi-expert prediction module, wherein... The knowledge state construction module is used to obtain the target student's historical answer sequence and obtain the knowledge state corresponding to the target student based on the historical answer sequence. The state decoupling module is used to perform multi-scale state decoupling processing on the knowledge state to obtain the decoupled long-term knowledge memory state, short-term answer fluctuation state and overall learning trend state. The multi-expert prediction module is used to construct a long-term causal expert model based on the long-term knowledge memory state, a short-term causal expert model based on the short-term response fluctuation state, and a trend causal expert model based on the overall learning trend state. The multi-expert prediction module is used to output the long-term answer prediction result of the target student based on the long-term causal expert model, the short-term answer prediction result of the target student based on the short-term causal expert model, and the trend answer prediction result of the target student based on the trend causal expert model. The multi-expert prediction module is used to fuse the long-term answer prediction results, the short-term answer prediction results, and the trend answer prediction results through a multi-expert prediction fusion mechanism, and output the answer prediction results corresponding to the target student.

9. An electronic device, characterized in that, It includes a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the student response prediction method based on multi-scale knowledge state decoupling as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1 to 7.

Images