A knowledge tracking method and system for cross-domain knowledge transfer

Abstract

The present invention discloses a knowledge tracking method and system for cross-domain knowledge transfer, including: constructing a student learning behavior representation knowledge base, obtaining a knowledge state vector based on a high-dimensional vector representation of question-level perception and a joint question-answer interaction representation extracted from the student learning behavior representation knowledge base; constructing an autoregressive information processing module to optimize the knowledge state vector to obtain an optimized knowledge state vector; constructing a student answer prediction module based on a two-layer nonlinear fully connected network, inputting the optimized knowledge state vector into the student answer prediction module to obtain a student answer prediction result, thereby realizing knowledge tracking. By generating a representation knowledge base with domain adaptability, the present invention enables the model to transfer knowledge from data-rich fields to data-scarce fields, thereby solving the performance problems of existing methods in data-scarce fields; and reducing sensitivity to specific hyperparameter selection through an adaptive knowledge state extraction network module, significantly saving resource computing power.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent education technology, and in particular relates to a knowledge tracking method and system for cross-domain knowledge transfer. Background Art

[0002] Knowledge tracking technology, a key technology in intelligent education and personalized learning systems, predicts students' mastery of new knowledge points by analyzing their historical learning behaviors, which are converted into digital data. This technology dynamically monitors changes in students' knowledge status, providing a basis for planning personalized learning paths. Based on the prediction results, the knowledge tracking system can recommend targeted learning content for students to strengthen, avoiding repetitive learning of already mastered knowledge points. Teachers can also use this technology to understand students' learning status in real time, identify problems promptly, and adjust teaching strategies. In addition, by providing targeted learning support for students of different levels, knowledge tracking technology helps narrow learning gaps and promote educational equity.

[0003] While existing knowledge tracing methods (such as DKT, AKT, and simpleKT) have achieved significant performance improvements, they generally suffer from two key issues: 1. Model performance largely depends on the quantity and quality of training data. However, due to privacy concerns, sufficient data cannot be obtained for all disciplines. For example, in some fields, there may only be a hundred student interaction sequences, while in others, there may be hundreds of thousands of interaction sequences. 2. Model performance is heavily dependent on hyperparameter combinations. For example, with one hyperparameter combination, the model performance is 0.56, while with another hyperparameter combination, the model performance can reach 0.75. These two issues restrict the application of knowledge tracing methods in real-world educational scenarios to achieve large-scale personalized education. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention proposes a knowledge tracking method and system for cross-domain knowledge migration to solve the problems existing in the above-mentioned prior art.

[0005] To achieve the above objectives, the present invention provides a knowledge tracking method for cross-domain knowledge transfer, comprising:

[0006] Constructing a student learning behavior representation knowledge base, wherein the student learning behavior representation knowledge base includes a question-level high-dimensional vector representation library and a joint question-answer interaction representation library;

[0007] Obtaining a knowledge state vector based on a high-dimensional vector representation of question-level perception extracted from the student learning behavior representation knowledge base and a joint question-answer interaction representation;

[0008] Constructing an autoregressive information processing module, and optimizing the knowledge state vector based on the autoregressive information processing module; obtaining an optimized knowledge state vector;

[0009] A student answer prediction module is constructed based on a two-layer nonlinear fully connected network, and the optimized knowledge state vector is input into the student answer prediction module to obtain the student answer prediction result and realize knowledge tracking.

[0010] Optionally, the process of building a problem-level high-dimensional vector representation library includes:

[0011] Perform feature mapping on the question sequence to obtain a high-dimensional vector representation of question-level perception. Based on the high-dimensional vector representation of question-level perception, a question-level high-dimensional vector representation library is constructed. The process of obtaining the high-dimensional vector representation of question-level perception is as follows:

[0012]

[0013] Where, is the level of questions that can be trained, is the average feature representation of all knowledge points corresponding to the question, Represents bitwise addition.

[0014] Optionally, the process of building a joint question-answer interaction representation library includes:

[0015] Based on the student answer sequence, the embedding layer of the answer behavior representation extractor is used to model the high-order interaction characteristics between the question knowledge points and the student answers, generating a joint question-answer interaction representation. Based on the joint question-answer interaction representation, a joint question-answer interaction representation library is constructed. The process of obtaining the joint question-answer interaction representation is as follows:

[0016]

[0017] Where, is the level of questions that can be trained, is the average feature representation of all knowledge points corresponding to the question, represents bitwise addition, Indicates the unique hot encoding of the knowledge points corresponding to the joint students' questions.

[0018] Optionally, a high-dimensional vector representation of question-level perception is extracted from the question-level high-dimensional vector representation library, and a joint question-answer interaction representation is extracted from the joint question-answer interaction representation library; the high-dimensional vector representation of question-level perception and the obtained query vector and joint question-answer interaction representation are used as the query vector, key vector and value vector, and the query vector and key vector are processed using a scaled dot product attention mechanism to obtain an attention score, which is converted into a probability distribution through a Softmax function, and the value vector is weighted summed according to the probability distribution to obtain a knowledge state vector. The process is expressed as follows:

[0019]

[0020] h t+1 =Concat(Head1,Head2,…,Head n )

[0021] Where Head represents the multi-head representation in the multi-head attention mechanism, Softmax represents the softmax function, represents the transpose of the key vector, represents the scaling factor of the scaled dot product attention, d is the dimension of the key vector, Q is the query vector, and K is the key vector.

[0022] Optionally, the autoregressive information processing module includes several autoregressive information processing units, each autoregressive information processing unit includes a knowledge state extraction module, a double normalization layer, a nonlinear transformation network layer and an adaptive residual connection; the knowledge state extraction module, the normalization layer, the residual connection, the nonlinear transformation network layer and the normalization layer are connected in sequence.

[0023] Optionally, the student answer prediction module is as follows:

[0024]

[0025] Among them, h l t+1 Represents the optimized knowledge state vector, σ represents the Sigmoid function, b1, b2, b, W1, W2, W are learnable parameters, Represents the predicted result of the student’s answer, and the historical learning behavior includes T interactive processes, t∈T.

[0026] Optionally, it also includes: the student answer prediction module is optimized through a binary cross entropy loss function.

[0027] The present invention also provides a knowledge tracking system for cross-domain knowledge migration, comprising:

[0028] Student learning behavior representation and cross-domain transfer module, used to generate a student behavior representation knowledge base;

[0029] A knowledge state extraction module, configured to obtain a knowledge state vector based on a high-dimensional vector representation of question-level perception and a joint question-answer interaction representation extracted from the student learning behavior representation knowledge base using a multi-head self-attention mechanism and a scaled dot-product attention mechanism;

[0030] An autoregressive information processing module is used to construct a multi-layer cascade dynamic information interaction network, optimize the knowledge state vector, and obtain an optimized knowledge state vector;

[0031] The student answer prediction module is used to perform a high-dimensional to low-dimensional projection transformation on the optimized knowledge state vector, optimize parameters based on the binary cross entropy loss function, and finally output the prediction results.

[0032] Compared with the prior art, the present invention has the following advantages and technical effects:

[0033] This invention constructs a multidimensional feature embedding space through a student learning behavior representation and cross-domain transfer module, enabling cross-domain knowledge transfer and addressing the performance issues of existing methods in data-scarce domains. By generating a domain-adaptive representational knowledge base, the model can transfer knowledge from data-rich domains to data-scarce domains. The knowledge state extraction module and autoregressive information processing module reduce sensitivity to specific hyperparameter selection, significantly saving resources and computing power. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0035] Figure 1 Schematic diagram of the technical solution of an embodiment of the present invention. DETAILED DESCRIPTION

[0036] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0037] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0038] Example 1

[0039] like Figure 1 As shown, this embodiment provides a knowledge tracking method for cross-domain knowledge transfer, including:

[0040] Constructing a student learning behavior representation knowledge base, wherein the student learning behavior representation knowledge base includes a question-level high-dimensional vector representation library and a joint question-answer interaction representation library;

[0041] Specifically, for student S, the historical learning behavior includes T interaction processes, each of which is expressed as t ,{c t}, r t >, where t∈T. Specifically, q t represents the question answered by the student at time step t, c t represents the question q answered by the student at time step t t The corresponding knowledge point. If a question corresponds to multiple knowledge points, then it is a set of multiple knowledge points {c t}, r t Indicates that the student answers question q at time step t t The answer, r t =1 means the student answered correctly, r t = 0 means the student answered incorrectly. For the questions answered by students, the question behavior representation extractor E(·) is used to perform feature mapping on the question sequence to obtain a high-dimensional vector representation of the question level perception:

[0042]

[0043] in, is the level of questions that can be trained, is the average feature representation of all knowledge points corresponding to the question, Represents bitwise addition.

[0044] For students’ answers, the answer behavior representation extractor is used to model the high-order interaction characteristics between question knowledge points and students’ answers through the embedding layer, thereby generating a joint question-answer interaction representation:

[0045]

[0046] in Indicates the unique hot encoding of the knowledge points corresponding to the joint students' questions.

[0047] To improve the transferability of students' learning behaviors, we use a cross-domain transfer representation extractor to construct a student learning behavior representation knowledge base based on the high-dimensional vector representation of question-level perception obtained by the question behavior representation extractor and the answer behavior representation extractor and the joint question-answer interaction representation:

[0048]

[0049] in ​Represents a high-dimensional vector representation library at the problem level, represents a joint question-answer interaction representation library, It represents the characteristic representation of the corresponding student problem and the corresponding subject area.

[0050] The student learning behavior representation and cross-domain transfer module effectively enhances the representation capability of cross-domain feature embedding through the cross-domain transfer representation extractor, question behavior representation extractor and answer behavior representation extractor, providing more accurate feature input for subsequent cross-domain knowledge state extraction.

[0051] Obtaining a knowledge state vector based on a high-dimensional vector representation of question-level perception extracted from the student learning behavior representation knowledge base and a joint question-answer interaction representation;

[0052] Furthermore, a high-dimensional vector representation of question-level perception is extracted from the question-level high-dimensional vector representation library, and a joint question-answer interaction representation is extracted from the joint question-answer interaction representation library; the high-dimensional vector representation of question-level perception and the obtained query vector and joint question-answer interaction representation are used as the query vector, key vector and value vector, and the query vector and key vector are processed using the scaled dot product attention mechanism to obtain an attention score, which is converted into a probability distribution through the Softmax function. The value vector is weighted and summed according to the probability distribution to obtain a knowledge state vector. The process is expressed as follows:

[0053]

[0054] h t+1 =Concat(Head1, Head2,..., Head n )

[0055] Where Head represents the multi-head representation in the multi-head attention mechanism, Softmax represents the softmax function, represents the transpose of the key vector, represents the scaling factor of the scaled dot product attention, d is the dimension of the key vector, Q is the query vector, and K is the key vector.

[0056] Specifically, we will use the problem-level high-dimensional vector representation library The high-dimensional vector representation x of the question-level perception extracted from t and joint question-answer interaction representation library The extracted y t As query vector (Q), key vector (K) and value vector (V), respectively, the vector x extracted from the question-level high-order vector representation library is used as the query vector and key vector, and y extracted from the joint question-answer interaction representation library is used as the value vector. That is, the query vector (Q) and key vector (K) are extracted from Indices, the value vector (V) from The index is:

[0057] Q = x t+1 ; K={x1,…,x t}; V = {y1,…,y t}

[0058] A multi-head attention mechanism integrating scaled dot product attention is used to extract the knowledge state h from the student's historical answer sequence. t+1 :

[0059]

[0060] h t+1 =Concat(Head1, Head2,..., Head n )

[0061] Where Head represents the multi-head representation in the multi-head attention mechanism, Softmax represents the softmax function, represents the transpose of the key vector, represents the scaling factor of the scaled dot-product attention, and d is the dimension of the key vector. This embodiment combines two major advantages of the attention mechanism: first, the characteristics of distributed computing and hierarchical data extraction, and second, the scaled dot-product mechanism effectively processes the order and contextual information in the learner's behavior sequence. Furthermore, the use of a multi-head attention strategy can capture deep correlation patterns between data from different dimensions.

[0062] Constructing an autoregressive information processing module, and optimizing the knowledge state vector based on the autoregressive information processing module; obtaining an optimized knowledge state vector;

[0063] Specifically, this module employs an innovative iterative deep state optimization architecture, achieving layer-by-layer refinement of the knowledge state through an auto-regressive information processing module. Each auto-regressive information processing unit comprises a knowledge state extraction module, a dual normalization layer, a nonlinear transformation network layer, and an adaptive residual connection, forming a complete knowledge state optimization loop. This progressive deep architecture not only captures subtle feature changes during the learning process but also projects the knowledge state into a more discriminative feature space through inter-layer feature reorganization and nonlinear mapping.

[0064] The knowledge state extraction module, normalization layer, residual connection, nonlinear transformation network layer, and normalization layer in the autoregressive information processing unit are connected in sequence.

[0065] Furthermore, this architecture significantly reduces parameter redundancy while maintaining the model's expressiveness through feature reconstruction and dimensionality adaptation. This multi-level feature extraction and optimization process gives the model stronger generalization capabilities, allowing it to better adapt to learning tasks of varying difficulty and types while reducing sensitivity to specific hyperparameter choices.

[0066] A student answer prediction module is constructed based on a two-layer nonlinear fully connected network, and the optimized knowledge state vector is input into the student answer prediction module to obtain the student answer prediction result and realize knowledge tracking.

[0067] Furthermore, the student answer prediction module is as follows:

[0068]

[0069] Among them, h l t+1 represents the optimized knowledge state vector, σ represents the Sigmoid function, b1, b2, b, W1, W2 are learnable parameters, Represents the predicted result of the student’s answer, and the historical learning behavior includes T interactive processes, t∈T.

[0070] Furthermore, it also includes: the student answer prediction module is optimized through a binary cross entropy loss function.

[0071] Specifically, this module is based on a two-layer nonlinear fully connected network to realize the knowledge state h l t+1 The high-dimensional to low-dimensional projection of Optimize and dynamically adjust parameters to obtain the final prediction results of students' answers

[0072]

[0073] σ represents the Sigmoid function, b1, b2, b, W1, W2, W are learnable parameters, r t+1 The actual answer results of the students.

[0074] This embodiment also provides a knowledge tracking system for cross-domain knowledge transfer, including:

[0075] Student learning behavior representation and cross-domain transfer module, used to generate a student behavior representation knowledge base;

[0076] A knowledge state extraction module, configured to obtain a knowledge state vector based on a high-dimensional vector representation of question-level perception and a joint question-answer interaction representation extracted from the student learning behavior representation knowledge base using a multi-head self-attention mechanism and a scaled dot-product attention mechanism;

[0077] An autoregressive information processing module is used to construct a multi-layer cascade dynamic information interaction network, optimize the knowledge state vector, and obtain an optimized knowledge state vector;

[0078] The student answer prediction module is used to perform a high-dimensional to low-dimensional projection transformation on the optimized knowledge state vector, optimize parameters based on the binary cross entropy loss function, and finally output the prediction results.

[0079] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A knowledge tracking method for cross-domain knowledge transfer, characterized by: The following steps are involved: Constructing a student learning behavior representation knowledge base, wherein the student learning behavior representation knowledge base includes a question-level high-dimensional vector representation library and a joint question-answer interaction representation library; Obtaining a knowledge state vector based on a high-dimensional vector representation of question-level perception extracted from the student learning behavior representation knowledge base and a joint question-answer interaction representation; Constructing an autoregressive information processing module, and optimizing the knowledge state vector based on the autoregressive information processing module; obtaining an optimized knowledge state vector; Constructing a student answer prediction module based on a two-layer nonlinear fully connected network, inputting the optimized knowledge state vector into the student answer prediction module, obtaining the student answer prediction result, and realizing knowledge tracking; The process of building a problem-level high-dimensional vector representation library includes: Perform feature mapping on the question sequence to obtain a high-dimensional vector representation of question-level perception. Based on the high-dimensional vector representation of question-level perception, a question-level high-dimensional vector representation library is constructed. The process of obtaining the high-dimensional vector representation of question-level perception is as follows: Where, is the level of questions that can be trained, is the average feature representation of all knowledge points corresponding to the question, Represents bitwise addition; The process of building a joint question-answer interaction representation library includes: Based on the student answer sequence, the embedding layer of the answer behavior representation extractor is used to model the high-order interaction characteristics between the question knowledge points and the student answers, generating a joint question-answer interaction representation. Based on the joint question-answer interaction representation, a joint question-answer interaction representation library is constructed. The process of obtaining the joint question-answer interaction representation is as follows: Where, is the level of questions that can be trained, is the average feature representation of all knowledge points corresponding to the question, represents bitwise addition, Indicates the unique hot encoding of the knowledge points corresponding to the joint students' questions; A high-dimensional vector representation of problem-level perception is extracted from the problem-level high-dimensional vector representation library, and a joint question-answer interaction representation is extracted from the joint question-answer interaction representation library; the high-dimensional vector representation of problem-level perception and the obtained query vector and joint question-answer interaction representation are used as the query vector, key vector and value vector, and the query vector and key vector are processed using the scaled dot product attention mechanism to obtain an attention score, which is converted into a probability distribution through a Softmax function, and the value vector is weightedly summed according to the probability distribution to obtain a knowledge state vector.

2. The knowledge tracking method for cross-domain knowledge transfer according to claim 1 is characterized in that: The process of obtaining the knowledge state vector is expressed as follows: h t+1 =Concat(Head1,Head2,...,Head n ) In the formula, Head represents the multi-head representation in the multi-head attention mechanism, Sofemax represents the softmax function, and K T represents the transpose of the key vector, represents the scaling factor of the scaled dot product attention, d is the dimension of the key vector, Q is the query vector, and K is the key vector.

3. The knowledge tracking method for cross-domain knowledge transfer according to claim 1 is characterized in that: The autoregressive information processing module includes several autoregressive information processing units, each of which includes a knowledge state extraction module, a double normalization layer, a nonlinear transformation network layer and an adaptive residual connection; the knowledge state extraction module, the normalization layer, the residual connection, the nonlinear transformation network layer and the normalization layer are connected in sequence.

4. The knowledge tracking method for cross-domain knowledge transfer according to claim 1 is characterized in that: The student answer prediction module is as follows: Among them, h l t+1 Represents the optimized knowledge state vector, σ represents the Sigmoid function, b1, b2, b, W1, W2, W are learnable parameters, Represents the predicted result of the student’s answer, and the historical learning behavior includes T interactive processes, t∈T.

5. The knowledge tracking method for cross-domain knowledge transfer according to claim 1 is characterized in that: Also includes: The student answer prediction module is optimized by a binary cross entropy loss function.

6. A system for implementing the knowledge tracking method for cross-domain knowledge transfer according to any one of claims 1 to 5, characterized in that: include: Student learning behavior representation and cross-domain transfer module, used to generate a student behavior representation knowledge base; A knowledge state extraction module, configured to obtain a knowledge state vector based on a high-dimensional vector representation of question-level perception and a joint question-answer interaction representation extracted from the student learning behavior representation knowledge base using a multi-head self-attention mechanism and a scaled dot-product attention mechanism; An autoregressive information processing module is used to construct a multi-layer cascade dynamic information interaction network, optimize the knowledge state vector, and obtain an optimized knowledge state vector; The student answer prediction module is used to perform a high-dimensional to low-dimensional projection transformation on the optimized knowledge state vector, optimize parameters based on the binary cross entropy loss function, and finally output the prediction results.

Images