Transfer Knowledge Tracing Method, System and Storage Medium in Sparse Data Scenarios

Abstract

The present invention relates to the field of intelligent education technology, and discloses a method, system and storage medium for migrating knowledge tracking in sparse data scenarios. The construction process of the knowledge tracking model adopted by the method includes: selecting the optimal dense data space by calculating the correlation between the sparse data space and different dense data spaces; performing sequence difference analysis on the question embedding and interaction embedding to obtain guiding knowledge variables; modeling the interaction records of students to generate knowledge state variables of students; after fusing the guiding knowledge variables and knowledge state variables, jointly inputting them into a classifier together with the question embedding at the next moment to predict the probability that the student will correctly answer the question at the next moment; freezing the parameters of the knowledge state navigation module in the optimal dense data space and migrating them to the sparse data space, so as to perform cross-training. In sparse data scenarios, the model can use migrated knowledge to make up for the shortage of local data and significantly improve the prediction accuracy.

Description

Technical Field

[0001] The present invention relates to the field of intelligent education technology, and particularly relates to a method, system and storage medium for migrating knowledge tracking in sparse data scenarios. Background Art

[0002] Knowledge tracking is a method for evaluating and predicting students' understanding levels, mainly achieved by analyzing students' performance in practice questions. The basic goal of knowledge tracking is to provide more accurate evaluations for individual learners, and then provide customized personalized educational materials, which is of great significance to the development of adaptive learning environments. In the field of sequence-based knowledge tracking, the Attentive Knowledge Tracing (AKT) method is regarded as a key technology. This method is widely popular due to three major advantages: fewer parameters, fast processing speed, and excellent performance. Recently, many advanced methods have emerged continuously. For example, contrastive learning techniques are used to reveal semantic similarities or differences in learning history to deepen the understanding of the relationship between students and questions; or two innovative attention-based models are used to clarify the associations between elements in the input sequence, and the prediction instability problem in Deep Knowledge Tracing (DKT) is solved through a Finite State Automaton (FSA).

[0003] Transfer learning applies the knowledge learned in one task to another task, while domain adaptation (DA) is dedicated to solving the differences between the source domain and the target domain. It enables the model in the data-rich source domain to adapt to the data-scarce target domain through parameter transfer.

[0004] Early domain adaptation methods used statistics such as Maximum Mean Discrepancy (MMD) and CORrelation ALignment (CORAL) to align feature distributions. Deep learning techniques have enhanced the domain adaptation ability through means such as domain adversarial neural networks and adaptive batch normalization, and use adversarial training to adjust the deep network layers to adapt to the statistical characteristics of the source domain and the target domain.

[0005] Recently, contrastive learning and self-supervised learning have been introduced into domain adaptation, enabling the model to quickly adapt to a new domain under unsupervised training and accelerating this process with the help of pre-trained knowledge, ensuring that the model has good generalization ability in the target domain.

[0006] The technical problem to be solved by the present invention is that in sparse data scenarios, knowledge tracking models face challenges such as scarce data and short interaction sequences, resulting in low prediction accuracy of knowledge tracking models. Most existing knowledge tracking methods rely on rich data to train models, but in sparse data environments, these knowledge tracking methods are difficult to be effectively applied. Summary of the Invention

[0007] To solve the above technical problems, the present invention provides a transfer knowledge tracking method, system and storage medium in a sparse data scenario, which transfers knowledge in a dense data space to a sparse data space through domain transfer to improve the performance of a knowledge tracking model in the sparse data space, thereby solving the challenges of knowledge tracking in a sparse data scenario.

[0008] To solve the above technical problems, the present invention adopts the following technical solutions:

[0009] A transfer knowledge tracking method in a sparse data scenario, and the construction process of the adopted knowledge tracking model includes:

[0010] Encode the questions and interaction records in the sparse data set and multiple dense data sets into a sparse data space and multiple dense data spaces respectively. Both the sparse data space and the dense data spaces contain question embeddings and interaction embeddings;

[0011] Calculate the correlation between the sparse data space and different dense data spaces through the weighted average difference of the question embeddings, and select the optimal dense data space for the sparse data space;

[0012] Perform knowledge tracking tasks in the optimal dense data space jointly through a knowledge state navigation module and a knowledge sequence analysis module; freeze the knowledge state navigation module of the optimal dense data space, and perform sequence difference analysis on the question embeddings and interaction embeddings to obtain guiding knowledge variables; use the knowledge sequence analysis module of the sparse data space to model the interaction records of students to generate knowledge state variables of students;

[0013] After fusing the guiding knowledge variables and the knowledge state variables, jointly input them with the question embeddings at the next moment into a classifier to predict the probability that a student will correctly answer a question at the next moment;

[0014] By freezing the parameters of the knowledge state navigation module of the optimal dense data space and migrating them to the sparse data space, cross-training is carried out.

[0015] In one embodiment, the encoding of the questions and interaction records in the sparse data set and multiple dense data sets into a sparse data space and multiple dense data spaces respectively, where both the sparse data space and the dense data spaces contain question embeddings and interaction embeddings, specifically includes:

[0016] Encode the questions and interaction records in the sparse data set and the dense data sets through a selected knowledge tracking model to obtain a sparse data space and dense data spaces; within the sparse data space and the dense data spaces, each question and interaction record is respectively represented as a question embedding and an interaction embedding, and the question embeddings and interaction embeddings capture the characteristics of the questions and the information related to the interaction with students.

[0017] In one embodiment, calculating the correlation between the sparse data space and different dense data spaces through the weighted average difference embedded by the topic, and selecting the optimal dense data space for the sparse data space specifically includes:

[0018] Subtract all the topic embeddings in the sparse data space or the dense data space pairwise to obtain difference embeddings, obtain the central characteristics of the sparse data space or the dense data space through weighted average difference embeddings, and use cosine similarity to compare the correlation between the central characteristics of the sparse data space and the central characteristics of different dense data spaces, and select the dense data space corresponding to the highest correlation as the optimal dense data space of the sparse data space.

[0019] In one embodiment, performing sequence difference analysis on the topic embedding and the interaction embedding to obtain a guiding knowledge variable, specifically including:

[0020] In the optimal dense data space, perform misalignment subtraction operations on the topic embedding sequence and the interaction embedding sequence respectively to generate a topic difference sequence and an interaction difference sequence;

[0021] Input the topic difference sequence into the first decoder to output topic understanding features;

[0022] Input the interaction difference sequence into the second decoder to output learning behavior features;

[0023] Use the topic understanding features as queries and keys, and the learning behavior features as values, and input them into the third decoder to generate a guiding knowledge variable.

[0024] In one embodiment, the first decoder, the second decoder, and the third decoder all adopt a monotonic multi-head attention mechanism.

[0025] In one embodiment, using the knowledge sequence analysis module of the sparse data space to model the interaction records of students to generate the knowledge state variables of students, specifically including:

[0026] Model the interaction record sequence of students in the sparse data space ; respectively represent the questions done by the student at time t and the variable used to indicate whether the answer given by the student is correct; input the interaction record sequence into the pre-trained knowledge tracking model to generate the knowledge state variables of the student.

[0027] In one embodiment, the classifier adopts a multi-layer perceptron.

[0028] A transfer knowledge tracking system in a sparse data scenario, including:

[0029] Data Space Construction Module: Encode the questions and interaction records in the sparse dataset and multiple dense datasets into a sparse data space and multiple dense data spaces respectively. Both the sparse data space and the dense data spaces contain question embeddings and interaction embeddings;

[0030] Dense Data Space Selection Module: Calculate the correlation between the sparse data space and different dense data spaces through the weighted average difference of the question embeddings, and select the optimal dense data space for the sparse data space;

[0031] Knowledge Tracing Task Module: jointly perform the knowledge tracing task in the optimal dense data space through the knowledge state navigation module and the knowledge sequence analysis module; freeze the knowledge state navigation module of the optimal dense data space, and perform sequence difference analysis on the question embeddings and interaction embeddings to obtain guiding knowledge variables; use the knowledge sequence analysis module of the sparse data space to model the interaction records of students and generate the knowledge state variables of students;

[0032] Transfer Prediction Module: After fusing the guiding knowledge variables and the knowledge state variables, input them together with the question embeddings at the next moment into a classifier to predict the probability that the student will correctly answer the question at the next moment;

[0033] Model Training Module: Freeze the parameters of the knowledge state navigation module of the optimal dense data space and transfer them to the sparse data space for cross-training.

[0034] In one embodiment, the dense data space selection module subtracts the question embeddings of all the sparse data spaces or dense data spaces pairwise to obtain difference embeddings, obtains the central characteristics of the sparse data space or dense data space through weighted average of the difference embeddings, and compares the central characteristics of the sparse data space with the central characteristics of different dense data spaces using cosine similarity. Select the dense data space corresponding to the highest correlation as the optimal dense data space of the sparse data space.

[0035] The system of the present invention corresponds to the method, and the preferred embodiments applicable to the method are also applicable to the system.

[0036] A computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method in any one of the embodiments are implemented.

[0037] Compared with the prior art, the beneficial technical effects of the present invention are:

[0038] By screening the optimal dense data space, the present invention extracts cross-domain common knowledge, captures dynamic knowledge changes, and generates guiding knowledge variables. In the sparse data scenario, the model can use transferred knowledge to make up for the lack of local data and significantly improve the prediction accuracy (such as the correct answer rate). Description of the Drawings

[0039] Figure 1 This is the flowchart of the knowledge tracing method in the embodiments of the present invention;

[0040] Figure 2 This is the schematic diagram of the architecture of the knowledge tracing model in the embodiments of the present invention;

[0041] Figure 3 This is the schematic diagram showing that the central characteristics for analyzing the item difference embedding in the embodiments of the present invention are more suitable for migration. Detailed Embodiment

[0042] A preferred embodiment of the present invention will be described in detail below with reference to the accompanying drawings.

[0043] As Figure 1 shown, the present invention provides a migration knowledge tracing method in a sparse data scenario. The construction process of the knowledge tracing model adopted includes the following steps:

[0044] S1, Encode the items and interaction records in the sparse data set and multiple dense data sets into a sparse data space and multiple dense data spaces respectively. Both the sparse data space and the dense data spaces contain item embeddings and interaction embeddings;

[0045] S2, Calculate the correlation between the sparse data space and different dense data spaces through the weighted average difference of the item embeddings, and select the optimal dense data space for the sparse data space;

[0046] S3, jointly perform knowledge tracing tasks in the dense data space through the knowledge state navigation module and the knowledge sequence analysis module; freeze the knowledge state navigation module of the optimal dense data space, and perform sequence difference analysis on the item embeddings and interaction embeddings to obtain guiding knowledge variables; use the knowledge sequence analysis module of the sparse data space to model the interaction records of students to generate knowledge state variables of students;

[0047] S4, After fusing the guiding knowledge variables and the knowledge state variables, jointly input them into the classifier together with the item embeddings at the next moment to predict the probability that the student will correctly answer the question at the next moment;

[0048] S5, Freeze the parameters of the knowledge state navigation module of the optimal dense data space and migrate them to the sparse data space, thereby performing cross-training to improve the performance of the knowledge tracing model in the sparse data space.

[0049] The trained knowledge tracing model in the sparse data space can generate the probability that the student will correctly answer the question at the next moment according to the item sequence and interaction records in the actual learning process of the student.

[0050] Dense datasets can use open-source educational datasets (such as large-scale question banks and student answer records). However, the data volume of a single student or the answer records of students in a certain class or school is small, and the prediction accuracy of the knowledge tracing model trained using this sparse data is not high. Through a parameter freezing and transfer mechanism, the present invention transfers the parameters in the knowledge state guidance module in the dense data space to the sparse data space, and dynamically fuses the transferred knowledge variables (i.e., guiding knowledge variables) with the local knowledge state variables to adapt to the personalized characteristics of the target domain. This enables the knowledge tracing model to effectively handle unseen questions or knowledge points and reduces the dependence on the local data distribution.

[0051] In one embodiment, in step S1, the questions and interaction records in the sparse dataset and multiple dense datasets are respectively encoded into a sparse data space and multiple dense data spaces. Both the sparse data space and the dense data spaces include question embeddings and interaction embeddings. Specifically, it includes:

[0052] Through a selected knowledge tracing model, the questions and interaction records in the sparse dataset and the dense datasets are encoded to obtain a sparse data space and dense data spaces. In the sparse data space and the dense data spaces, each question and interaction record is respectively represented as a question embedding and an interaction embedding. The question embedding and the interaction embedding capture the characteristics of the question and the relevant information of the interaction with the student.

[0053] In this way, the questions and interaction records are transformed from the original data form into an embedding representation suitable for model processing, providing a basis for subsequent analysis and prediction.

[0054] A student's learning record includes a question sequence and interaction records. Each question is associated with one or more knowledge points, and there may be dependencies between knowledge points (such as one knowledge point being a sub-knowledge point of another knowledge point).

[0055] In one embodiment, in step S2, the correlation between the sparse data space and different dense data spaces is calculated by the weighted average difference of the question embeddings, and the optimal dense data space is selected for the sparse data space. Specifically, it includes:

[0056] Subtract the question embeddings in the sparse data space or the dense data spaces pairwise to obtain difference embeddings. The central characteristics of the sparse data space or the dense data spaces are obtained by weighted averaging the difference embeddings. The cosine similarity is used to compare the correlation between the central characteristics of the sparse data space and the central characteristics of different dense data spaces, and the dense data space corresponding to the highest correlation is selected as the optimal dense data space for the sparse data space.

[0057] The central feature is obtained through the above approach. The reason for directly obtaining the central characteristics of the original data space is that when using difference embedding in different data spaces, the central characteristics have a greater correlation and are more conducive to subsequent migration.

[0058] The dense data space provides rich question and interaction embedding information. The Knowledge State Navigation module (KSN) identifies the differences in the sequence by learning these embeddings, and then extracts knowledge variables that can guide the learning process. The Knowledge Sequence Analysis module (KSA) assists the KSN by analyzing these knowledge variables to optimize and correct the knowledge tracking process. Through this collaboration, the system can more effectively understand and predict the changes in the student's knowledge state.

[0059] Considering the characteristics of knowledge transfer, the present invention believes that the selection method of the dense data space should meet two conditions: (1) A processed dense data space should be generated such that the correlation between this dense data space is greater than the correlation between the original feature spaces; (2) The correlation between the processed dense data space and the original feature space should be less than the correlation between the original feature spaces themselves. As Figure 3 shown in (a) of and ), after pre-training in different dense data spaces, the difference embedding space Common-diff and the aggregated embedding space Common-aggr can be obtained by adding and subtracting any two question embeddings ( Figure 3 ). After weighted average embedding is performed on the difference embedding space Common-diff and the aggregated embedding space Common-aggr respectively, the cosine similarity between these three data spaces is calculated, as shown in (b) of Figure 3 . In some preferred embodiments, the difference embedding space Common-diff better meets the first condition. As shown in (c) of Figure 3 , by comparing the similarity between the difference embedding space Common-diff and the aggregated embedding space Common-aggr, and their similarities with the corresponding original feature spaces, it can be found that the difference embedding space Common-diff better meets the second condition. Therefore, the difference embedding space can be used as the optimal dense data space.

[0060] In one of the embodiments, the sequence difference analysis of the question embedding and the interaction embedding to obtain the guiding knowledge variables specifically includes:

[0061] In the optimal dense data space, perform a dislocation subtraction operation on the question embedding sequence and the interaction embedding sequence respectively to generate a question difference sequence and an interaction difference sequence;

[0062] Input the question difference sequence into the first decoder to output the question understanding feature;

[0063] Input the interaction difference sequence into the second decoder to output learning behavior features;

[0064] Take the question understanding feature as the query and key, and the learning behavior feature as the value, and input them into the third decoder to generate a guiding knowledge variable.

[0065] The knowledge state guiding process aims to process the difference information and provide transfer parameters for transfer learning. As Figure 2 shown, the present invention performs a misalignment subtraction operation on the question embedding sequence and the interaction embedding sequence to obtain their respective difference sequences. The knowledge tracking model of the present invention performs this operation on the question and interaction embeddings respectively by subtracting the embedding of the previous moment from the embedding of the current moment, so as to obtain the corresponding difference sequences. Logically and intuitively, it can be naturally understood that the information processed from the question embedding is information related to "understanding" the question, while the information processed from the interaction embedding is information related to "learning". In order to process these two types of information, the present invention adopts a first decoder and a second decoder to process these two types of information respectively. The question understanding feature output by the first decoder is used as the query and key, and the learning behavior feature output by the second decoder is used as the value, and is input into the third decoder to obtain the guiding knowledge variable .

[0066] In the dense data space, the parameters of the knowledge state guiding process are learnable, so as to obtain information corresponding to the dense data space. In the sparse data space, the parameters of the knowledge state guiding process are frozen, and the parameters of the knowledge state guiding process in the sparse data space are replaced, so as to achieve knowledge transfer.

[0067] In one embodiment, the first decoder, the second decoder and the third decoder all adopt a monotonic multi-head attention mechanism.

[0068] In one embodiment, the use of the questions and interaction records in the sparse data space to model the knowledge state of the student to generate the knowledge state variable of the student specifically includes:

[0069] Model the interaction record sequence of the student in the sparse data space ; respectively represent the questions done by the student at time t and the variable used to indicate whether the answer given by the student is correct; input the interaction record sequence into the pre-trained knowledge tracking model to generate the student knowledge state variable, which characterizes the student's mastery state of the relevant knowledge points.

[0070] In the sparse data space, a Knowledge Sequence Analysis module (KSA) is used to model the interaction records of students, aiming to extract variables that can reflect the current knowledge level of students from limited and incomplete learning data. By analyzing the interaction records between students and the learning system, KSA identifies their learning behaviors and knowledge mastery status, helping the education system to effectively evaluate and track students' learning progress even in the case of incomplete data. This process enables the system to understand the knowledge state of students, thereby providing personalized learning support.

[0071] In one embodiment, the classifier uses a multi-layer perceptron. The guiding knowledge variables provided by the knowledge state guiding process in the dense data space , and the knowledge state variables of students provided by the knowledge sequence analysis process in the sparse data space , are added to obtain a fusion variable , and combined with the embedding of the question at the next moment , and input into a multi-layer perceptron (MLP) for prediction, outputting the probability that the student correctly answers the question at the next moment .

[0072] During the training process, a binary cross-entropy loss function is used to optimize the knowledge tracing model, reducing the error between the prediction result and the actual label. The entire process is completed alternately in the dense data space and the sparse data space, finally achieving the transfer of knowledge from the dense data space to the sparse data space, and significantly improving the knowledge tracing performance in the sparse data space.

[0073] In some preferred embodiments, the present invention proposes a knowledge tracing method combining knowledge transfer, which transfers the knowledge in the dense data space to the sparse data space through domain transfer, thereby improving the performance of the knowledge tracing model in the sparse data space. First, by calculating the cosine similarity in the dense data space, the most suitable dense data space is selected, and guiding information is extracted from it. Then, by freezing the parameters of the knowledge state guiding process and migrating them to the sparse data space, the model can be trained in a sparse data environment without manually adjusting the transfer function. This method effectively solves the challenges of knowledge tracing in the sparse data space, especially in the case of short interaction records and sparse data, providing an innovative solution and improving the prediction accuracy of the model in a sparse data environment.

[0074] The present invention demonstrates a complete technical process from data preprocessing, knowledge transfer to prediction through the above embodiments, and can be applied to intelligent learning evaluation and personalized recommendation in various educational scenarios.

[0075] In one embodiment, the present invention also discloses a transfer knowledge tracing system in a sparse data scenario, including:

[0076] Data Space Construction Module: Encode the questions and interaction records in the sparse dataset and multiple dense datasets into a sparse data space and multiple dense data spaces respectively. Both the sparse data space and the dense data spaces contain question embeddings and interaction embeddings;

[0077] Dense Data Space Selection Module: Calculate the correlation between the sparse data space and different dense data spaces through the weighted average difference of the question embeddings, and select the optimal dense data space for the sparse data space;

[0078] Knowledge Tracing Task Module: jointly perform the knowledge tracing task in the dense data space through the knowledge state navigation module and the knowledge sequence analysis module; freeze the knowledge state navigation module of the optimal dense data space, and perform sequence difference analysis on the question embeddings and interaction embeddings to obtain guiding knowledge variables; use the knowledge sequence analysis module of the sparse data space to model the interaction records of students and generate the knowledge state variables of students;

[0079] Transfer Prediction Module: After fusing the guiding knowledge variables and the knowledge state variables, jointly input them with the question embeddings at the next moment into a classifier to predict the probability that a student will correctly answer a question at the next moment;

[0080] Model Training Module: Freeze the parameters of the knowledge state navigation module of the optimal dense data space and transfer them to the sparse data space for cross-training.

[0081] The system of the present invention corresponds to the method, and the specific embodiments applicable to the method are equally applicable to the system. The specific embodiments of the method have been described in detail above and will not be repeated here.

[0082] In an exemplary embodiment, the present invention also provides a computer-readable storage medium including instructions, such as a memory including instructions. The above instructions can be executed by a processor to complete the above method. The storage medium can be a computer-readable storage medium. For example, the computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, and an optical data storage device, etc.

[0083] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, it is intended to embrace all changes within the meaning and scope of the equivalent elements of the claims in the present invention, and any reference signs in the claims should not be regarded as limiting the claims involved.

[0084] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative manner of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A migration knowledge tracking method in a sparse data scenario, characterized in that: The construction process of the adopted knowledge tracing model includes: Encode the topics and interaction records in the sparse data set and multiple dense data sets into a sparse data space and multiple dense data spaces respectively, and both the sparse data space and the dense data space contain topic embedding and interaction embedding; The correlation between the sparse data space and different dense data spaces is calculated through the weighted average difference of the question embedding, and the optimal dense data space is selected for the sparse data space; The knowledge tracking task is performed in the optimal dense data space through the knowledge state navigation module and the knowledge sequence analysis module; the knowledge state navigation module of the optimal dense data space is frozen, and the sequence difference analysis of the question embedding and the interaction embedding is performed to obtain the guiding knowledge variable: in the optimal dense data space, the dislocation subtraction operation is performed on the question embedding sequence and the interaction embedding sequence respectively to generate the question difference sequence and the interaction difference sequence; the question difference sequence is input into the first decoder to output the question understanding feature; the interaction difference sequence is input into the second decoder to output the learning behavior feature; the question understanding feature is used as the query and key, and the learning behavior feature is used as the value, and is input into the third decoder to generate the guiding knowledge variable; The knowledge sequence analysis module of sparse data space is used to model the students' interaction records and generate students' knowledge state variables; After the guiding knowledge variable and the knowledge state variable are fused, they are input into the classifier together with the embedding of the question at the next moment to predict the probability that the student will answer the question correctly at the next moment. Cross-training is performed by freezing the parameters of the knowledge state navigation module in the optimal dense data space and migrating it to the sparse data space.

2. The migration knowledge tracking method in sparse data scenario according to claim 1 is characterized in that: The encoding of the topics and interaction records in the sparse data set and the plurality of dense data sets into a sparse data space and a plurality of dense data spaces, respectively, wherein both the sparse data space and the dense data space contain topic embedding and interaction embedding, specifically includes: Through the selected knowledge tracking model, the questions and interaction records in the sparse data set and dense data set are encoded to obtain sparse data space and dense data space; in the sparse data space and dense data space, each question and interaction record is represented as question embedding and interaction embedding respectively, which capture the characteristics of the question and the relevant information of the interaction with the students.

3. The migration knowledge tracking method in a sparse data scenario according to claim 1 is characterized in that: The weighted average difference of the embedded questions is used to calculate the correlation between the sparse data space and different dense data spaces, and the optimal dense data space is selected for the sparse data space, specifically including: All the topic embeddings in the sparse data space or dense data space are subtracted pairwise to obtain difference embeddings, and the central characteristics of the sparse data space or dense data space are obtained by weighted average difference embedding. The correlation between the central characteristics of the sparse data space and the central characteristics of different dense data spaces is compared by cosine similarity, and the dense data space corresponding to the highest correlation is selected as the optimal dense data space of the sparse data space.

4. The method for tracking migration knowledge in a sparse data scenario according to claim 1, characterized in that: The first decoder, the second decoder and the third decoder all adopt a monotonic multi-head attention mechanism.

5. The method for tracking migration knowledge in a sparse data scenario according to claim 1, characterized in that: The knowledge sequence analysis module using the sparse data space models the student's interaction records and generates the student's knowledge state variables, specifically including: Modeling student interaction record sequences in sparse data space ; They represent the questions answered by the students at time t and the variables indicating whether the answers given by the students are correct. The interaction record sequence is input into the pre-trained knowledge tracking model to generate the student knowledge state variable.

6. The method for tracking migration knowledge in a sparse data scenario according to claim 1, characterized in that: The classifier adopts a multi-layer perceptron.

7. A migration knowledge tracking system in a sparse data scenario, characterized in that: include: Data space construction module: Encode the topics and interaction records in the sparse data set and multiple dense data sets into a sparse data space and multiple dense data spaces respectively. Both the sparse data space and the dense data space contain topic embedding and interaction embedding. Dense data space selection module: calculates the correlation between sparse data space and different dense data spaces through the weighted average difference of question embedding, and selects the optimal dense data space for the sparse data space; Knowledge tracking task module: The knowledge state navigation module and the knowledge sequence analysis module jointly perform knowledge tracking tasks in the optimal dense data space; freeze the knowledge state navigation module in the optimal dense data space, and perform sequence difference analysis on question embedding and interaction embedding to obtain guiding knowledge variables; The knowledge sequence analysis module of sparse data space is used to model the students' interaction records and generate students' knowledge state variables; Transfer prediction module: After fusing the guiding knowledge variables and knowledge state variables, they are input into the classifier together with the embedding of the question at the next moment to predict the probability that the student will answer the question correctly at the next moment; Model training module: Cross-training is performed by freezing the parameters of the knowledge state navigation module in the optimal dense data space and migrating it to the sparse data space.

8. The migration knowledge tracking system in a sparse data scenario according to claim 7, characterized in that: The dense data space selection module subtracts all the topic embeddings in the sparse data space or dense data space from each other to obtain difference embeddings, obtains the central characteristics of the sparse data space or dense data space through weighted average difference embeddings, uses cosine similarity to compare the correlation between the central characteristics of the sparse data space and the central characteristics of different dense data spaces, and selects the dense data space corresponding to the highest correlation as the optimal dense data space of the sparse data space.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images