Cross-domain cognitive diagnosis method based on prompt technology

Through the cross-domain cognitive diagnosis method based on prompt technology, using soft prompt transfer technology, a cognitive diagnostic model with soft prompt is built, which solves the problem of poor performance of existing models in cross-domain scenarios and improves diagnostic accuracy and generalization ability, suitable for cross-domain cognitive diagnostic tasks with diverse exercise content.

CN119066203AActive Publication Date: 2024-12-03HEFEI UNIV OF TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411322300.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-23
Publication Date
2024-12-03
Estimated Expiration
2044-09-23

Smart Images

  • Figure CN119066203A_ABST
    Figure CN119066203A_ABST
Patent Text Reader

Abstract

The invention discloses a prompting technology-based cross-domain cognitive diagnosis method, which comprises the following steps of: 1, constructing cross-domain data which comprises data sets of a source domain and a target domain, an exercise-knowledge point correlation relationship and the like; 2, constructing a cognitive diagnosis model with soft prompts for the source fields, and learning universal representations among different source fields by using the soft prompts; 3, constructing a cognitive diagnosis model with a soft prompt for the target domain, and migrating the soft prompt from the source domain model to the target domain model; and 4, performing source domain pre-training and target domain fine tuning on the cognitive diagnosis model until convergence, thereby obtaining a trained target domain cognitive diagnosis model for practice recommendation and knowledge tracking of students. According to the method, the thought of cross-domain soft prompt is utilized, the prompt is migrated to the target domain after training is carried out in the source domain, and a small amount of data is used for fine adjustment in the target domain, so that the accuracy of cross-domain diagnosis can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of cognitive diagnosis of smart education, and specifically is a cross-domain cognitive diagnosis method based on prompt technology. Background Art

[0002] In the field of smart education, cognitive diagnosis plays a fundamental role, providing a basis for exercise recommendation and learning guidance. In cognitive diagnosis, student and exercise representations are essential to obtain the cognitive state of students. Therefore, the core challenge lies in obtaining high-quality student and exercise representations. With the booming development of deep learning technology, great efforts have been made to design complex models that learn these embeddings. For example, the neurocognitive diagnosis model (NCDM) sets the dimensions of student and exercise embeddings to correspond to the number of knowledge points. During the interaction process, NCDM uses the association matrix (Q matrix) of exercises and knowledge points for alignment. The relational graph-based cognitive diagnosis model (RCD) constructs a multi-layer graph structure containing student, exercise, and knowledge concepts to obtain their embeddings. It then combines diagnostic and prediction modules to capture the interactions between these three entities and facilitate cognitive diagnosis.

[0003] Despite the significant achievements of existing research, most of the models developed so far are domain-specific. These models assume that the exercises and knowledge points are from the same subject and that the students are from a specific region or school. However, these assumptions limit the generalizability of the models in practical applications. In real-world scenarios, when encountering a new subject or a new group of students from different schools (called a new domain), the model needs to be retrained to make it applicable. However, if the data in this new domain is scarce, the retraining process becomes challenging. Therefore, cross-domain cognitive diagnosis is a challenging and promising research direction with important practical value in the field of intelligent education.

[0004] However, only a few studies have involved cross-domain cognitive diagnosis models. One of the studies is TechCD, which proposes a cross-domain cognitive diagnosis method using transferable knowledge concept graph embedding in cold start scenarios. Another method, ZeroCD, addresses this challenge by leveraging a batch of early participating student data. However, the TechCD method has high computational cost and is difficult to apply to real-time or large-scale scenarios. It also relies heavily on high-quality knowledge concept graphs, and incomplete or inaccurate data will seriously affect the algorithm effect. In the ZeroCD method, the quality of the simulation log depends on the reliability of early student behavior. If there is noise or bias in the behavioral data, it may mislead subsequent cognitive diagnosis. Therefore, how to use the historical answer records of students in the source field to accurately evaluate the mastery of knowledge points and answer prediction of students in the target field is still a problem worthy of in-depth study. Summary of the invention

[0005] In order to overcome the shortcomings of the prior art, the present invention proposes a cross-domain cognitive diagnosis method based on prompt technology, aiming to solve the problem that the existing cognitive diagnosis model performs poorly in cross-domain scenarios. By establishing a connection between the source domain and the target domain, the transfer of student and exercise features learned from the source domain to the target domain is promoted, thereby improving the accuracy of cross-domain diagnosis.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:

[0007] The cross-domain cognitive diagnosis method based on prompting technology of the present invention is characterized in that it is carried out according to the following steps:

[0008] Step 1: Construct cross-domain data;

[0009] Define the source domain collection and target areas ,in, Indicates Source fields, Indicates the number of source fields;

[0010] definition The set of students on , the exercise set is , the knowledge point set is ;in, Any student in , The Exercises are recorded as , The Knowledge points are recorded as ; and the exercises in each source domain are different;

[0011] Define target areas The set of students on , the exercise set is , the knowledge point set is ;in, Any student in , No. Exercises are recorded as , The Knowledge points are recorded as ; and the exercises between the source domain set S and the target domain T are different;

[0012] Using formula (1), we can get the source domain set and target areas Overlapping student sets between , and Middle Students are recorded as :

[0013] (1)

[0014] Defining the Source Domain Pre-training dataset on is a set of triples, where Any triple in express right The answer result is, if ,express Correct answer , , indicating other situations;

[0015] from Randomly select a certain proportion of students’ answer data as the fine-tuning dataset ;in, Any triple in express right The answer result is, if ,express Correct answer , , indicating other situations;

[0016] Will The relationship matrix between the previous exercises and knowledge points is recorded as , where Middle Row vector express and The correlation between each knowledge point above, where 1 or 0 indicates and relevant or irrelevant; express The number of exercises in express The number of knowledge points in

[0017] Will The relationship matrix between the previous exercises and knowledge points is recorded as , where Middle Row vector express and The correlation between each knowledge point above, where 1 or 0 indicates and relevant or irrelevant; express The number of exercises in express The number of knowledge points in

[0018] Step 2: Construct a source domain cognitive diagnosis model with soft prompts and exist The original embedding and The original embedding of exist Top pair Predicted answer results , thereby constructing the binary cross entropy loss , used to train the source domain cognitive diagnosis model, and obtain The best personalized tips on any source field and The best shared hint for all exercises in ;

[0019] Step 3: Construct a target domain cognitive diagnosis model with soft prompts and and Process and obtain exist Top pair Predicted answer results , thus constructing exist Top pair The binary cross entropy loss , which is used to train the target domain cognitive diagnosis model to achieve exercise recommendation and knowledge tracking for students.

[0020] The cross-domain cognitive diagnosis method based on prompting technology of the present invention is also characterized in that step 2 is performed as follows:

[0021] Step 2.1: Through formula (3), we can get exist Composite representation on as well as Composite representation :

[0022] (3)

[0023] In formula (3), express Personalized tips on any source field, express Shared hints for all exercises in and They represent the random initialization exist The original embedding and The original embedding of Represents horizontal concatenation of vectors;

[0024] Step 2.2: Through formula (4), we can get exist The final representation of as well as The final representation :

[0025] (4)

[0026] In formula (4), represents a linear layer for personalized hints on the source domain, represents a linear layer over the source domain for shared cues;

[0027] Step 2.3: Get by formula (5) exist Top pair Predicted answer results :

[0028] (5)

[0029] In formula (5), represents the interactive function that predicts the answer result based on the student and exercise representations;

[0030] Step 2.4: Construct through formula (6) exist Top pair The binary cross entropy loss :

[0031] (6)

[0032] In formula (6), yes exist Top pair The actual answer results;

[0033] Step 2.5: Use formula (7) to construct the source domain cognitive diagnosis model for the objective function of all source domains :

[0034] (7)

[0035] Step 2.6: All pre-training datasets based on the source domain , the source domain cognitive diagnosis model is trained using the Adam optimizer to update the model parameters until the objective function Until convergence, the cognitive diagnosis model trained into the source domain and The best personalized tips on any source field and The best shared hint for all exercises in .

[0036] Further, the step 3 is performed as follows:

[0037] Step 3.1: Get by formula (8): In the target area Personalized tips on :

[0038] (8)

[0039] Step 3.2: Get the target domain through formula (9) Shared hints for all exercises in :

[0040] (9)

[0041] In formula (9), represents the vertical concatenation operation of shared cues in the source domain, Represents a dimension The row vector of ;

[0042] Step 3.3: Use formula (10) to get exist The original embedding on :

[0043] (10)

[0044] Step 3.4: Through formula (11), we can get exist Composite representation on as well as Composite representation :

[0045] (11)

[0046] In formula (11), Represents the random initialization The original embedding of

[0047] Step 3.5: Through formula (12), we can get exist The final representation of as well as The final representation :

[0048] (12)

[0049] In formula (12), represents a linear layer for personalized hints on the target domain, represents a linear layer for shared cues on the target domain;

[0050] Step 3.6: Use formula (13) to get exist Top pair Predicted answer results :

[0051] (13)

[0052] Step 3.7: Construct through formula (14) exist Top pair The binary cross entropy loss :

[0053] (14)

[0054] Step 3.8: Construct the objective function of the source domain cognitive diagnosis model for the target domain through formula (15): :

[0055] (15)

[0056] Step 3.9: Fine-tune the dataset based on the dataset in the target domain , the target domain cognitive diagnosis model is trained based on the Adam optimizer to update the model parameters until the objective function Convergence is achieved, thus obtaining the trained target domain cognitive diagnosis model.

[0057] The electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the cross-domain cognitive diagnosis method, and the processor is configured to execute the program stored in the memory.

[0058] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the cross-domain cognitive diagnosis method when executed by a processor.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] 1. The present invention effectively solves the problem of model performance degradation and poor performance of existing models in cross-domain scenarios through soft prompt transfer technology, thereby significantly improving the diagnostic accuracy, especially when there is a large difference between the source domain and the target domain.

[0061] 2. The present invention adopts a two-stage model of pre-training and fine-tuning to enable the diagnostic model to quickly adapt to new fields, avoiding the overfitting and catastrophic forgetting problems that may be caused by retraining the model, thereby enhancing the generalization ability of the diagnostic model between different fields.

[0062] 3. The present invention is applicable to cross-domain cognitive diagnosis tasks with diversified exercise contents, and can effectively cope with scenarios of different subjects and difficulty levels. This multi-dimensional adaptability ensures that the present invention can play a role in various educational environments and situations. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 A specific flow chart of the cross-domain cognitive diagnosis method based on prompting technology proposed in the present invention;

[0064] Figure 2a This is an architecture diagram of the source domain model when the present invention is applied to the NCDM model;

[0065] Figure 2b This is an architecture diagram of the target model when the present invention is applied to the NCDM model. DETAILED DESCRIPTION

[0066] In this embodiment, for the diversified exercise content, a cross-domain cognitive diagnosis method based on prompt technology is: first, construct cross-domain data: including data sets of source domains and target domains, exercises-knowledge point correlation; second, build a common cognitive diagnosis model with soft prompts for all source domains, because relying solely on entity representation may not adapt to the diversified interactions between different domains, so soft prompts are used to learn the common representations between different source domains; then, build a cognitive diagnosis model with soft prompts for the target domain, define the method of migrating soft prompts from the source domain model to the target domain model, and transfer personalized prompts in a one-to-one manner. Shared prompts are mapped back to the original dimension after splicing, so that the information in the source domain can be maximized. In addition, the method of the present invention uses a linear layer to learn the mapping relationship between personalized prompts and the original embedding of entities in the target domain, replacing the random initialization of the original embedding of entities in the target domain, and making full use of existing data to capture potential information. Finally, two stages of pre-training and fine-tuning of the model are performed. In the pre-training stage, personalized and shared prompts are updated through source domain data. In the fine-tuning phase, a small amount of data from the target domain is used to fine-tune the model parameters to achieve knowledge transfer and adaptation from the source domain to the target domain. The two phases are trained sequentially until convergence. Specifically, Figure 1 As shown, the method is performed in the following steps:

[0067] Step 1: Construct cross-domain data;

[0068] Define the source domain collection and target areas ,in, Indicates Source fields, Indicates the number of source fields;

[0069] definition The set of students on , the exercise set is , the knowledge point set is ;in, Any student in , The Exercises are recorded as , The Knowledge points are recorded as ; and the exercises in each source domain are different;

[0070] Define target areas The set of students on , the exercise set is , the knowledge point set is ;in, Any student in , No. Exercises are recorded as , The Knowledge points are recorded as ; and the exercises between the source domain set S and the target domain T are different.

[0071] Using formula (1), we can get the source domain set and target areas Overlapping student sets between , and Middle Students are recorded as :

[0072] (1)

[0073] Defining the Source Domain Pre-training dataset on is a set of triples, where Any triple in express right The answer result is, if ,express Correct answer , , indicating other situations;

[0074] Define target areas Dataset is a set of triples, from A certain proportion of the answer data from each student is randomly selected as the fine-tuning dataset ;in, Any triple in express right The answer result is, if ,express Correct answer , , indicating other situations.

[0075] Will The relationship matrix between the previous exercises and knowledge points is recorded as , where Middle Row vector express and The correlation between each knowledge point above, where 1 or 0 indicates and relevant or irrelevant; express The number of exercises in express The number of knowledge points in

[0076] Will The relationship matrix between the previous exercises and knowledge points is recorded as , where Middle Row vector express and The correlation between each knowledge point above, where 1 or 0 indicates and relevant or irrelevant; express The number of exercises in express The number of knowledge points in .

[0077] Step 2: Build a source domain cognitive diagnosis model with soft prompts:

[0078] Step 2.1: Considering the characteristics of overlapping and non-overlapping entities in cross-domain scenarios, relying solely on entity representation may not be able to adapt to the diverse interactions between different domains. Therefore, the present invention designs two learnable soft prompts: personalized prompts and shared prompts.

[0079] In a cross-domain cognitive diagnosis scenario involving different disciplines, since the students are the same in the source and target domains, each student can be associated with a personalized prompt, thereby achieving more information transfer. In contrast, shared prompts are used for all exercises in the source domain within each discipline, ensuring a common representation of domain-specific knowledge. exist Composite representation on as well as Composite representation :

[0080] (3)

[0081] In formula (3), express Personalized tips on any source field, express Shared hints for all exercises in and They represent the random initialization exist The original embedding and The original embedding of Represents horizontal concatenation of vectors.

[0082] Step 2.2: In order to integrate the splicing features and extract the joint information, we can obtain exist The final representation of as well as The final representation :

[0083] (4)

[0084] In formula (4), represents a linear layer for personalized hints on the source domain, Represents a linear layer for shared cues on the source domain; compared to the original embedding, the final representation more effectively captures information transfer and domain-specific knowledge in cross-domain scenarios by combining personalized cues with shared cues. In addition, the composite representation can enhance the adaptability of the model and improve the accuracy and generalization ability of cognitive diagnosis.

[0085] Step 2.3: Get by formula (5) exist Top pair Predicted answer results :

[0086] (5)

[0087] In formula (5), represents the interactive function that predicts the answer result based on the student and exercise representations.

[0088] Step 2.4: Construct through formula (6) exist Top pair The binary cross entropy loss :

[0089] (6)

[0090] In formula (6), yes exist Top pair The actual answer results;

[0091] Step 2.5: Use formula (7) to construct the source domain cognitive diagnosis model for the objective function of all source domains :

[0092] (7)

[0093] Step 2.6: All pre-training datasets based on the source domain , the source domain cognitive diagnosis model is trained using the Adam optimizer to update the model parameters until the objective function Until convergence, the trained source domain cognitive diagnosis model and its corresponding The best personalized tips on any source field and The best shared hint for all exercises in ,because ,so There is also an optimal personalized prompt in any source field .

[0094] Step 3: Construct a target domain cognitive diagnosis model with soft prompts:

[0095] Step 3.1: Get by formula (8): In the target area Personalized tips on , the personalized prompts corresponding to the students are transferred from the source domain to the target domain in a one-to-one manner to maintain the integrity of the cross-domain connection information, because these prompts are also related to the target domain:

[0096] (8)

[0097] Step 3.2: In order to effectively capture the commonalities between different source domains, the shared prompts corresponding to the exercises are weighted and combined through formula (9) to obtain the target domain Shared hints for all exercises in :

[0098] (9)

[0099] In formula (9), represents the vertical concatenation operation of shared cues in the source domain, Represents a dimension Through weighted combination, different weights can be assigned according to the importance of each source field to the target field, so as to make full use of the relevant information of each source field and improve the generalization ability and effect of the model.

[0100] Step 3.3: Use formula (10) to get exist The original embedding on , using a linear layer to learn personalized hints Target Area middle The original embedding The mapping relationship between them is:

[0101] (10)

[0102] Contains trainable parameters that are optimized with a small amount of interaction data in the target domain. This strategy replaces random initialization and exploits available data to capture the underlying original information.

[0103] Step 3.4: Through formula (11), we can get exist Composite representation on as well as Composite representation :

[0104] (11)

[0105] In formula (11), Represents the random initialization The original embedding of

[0106] Step 3.5: Through formula (12), we can get exist The final representation of as well as The final representation :

[0107] (12)

[0108] In formula (12), represents a linear layer for personalized hints on the target domain, Represents a linear layer for shared cues on the target domain.

[0109] Step 3.6: Use formula (13) to get exist Top pair Predicted answer results :

[0110] (13)

[0111] Step 3.7: Construct through formula (14) exist Top pair The binary cross entropy loss :

[0112] (14)

[0113] Step 3.8: Construct the objective function of the source domain cognitive diagnosis model for the target domain through formula (15): :

[0114] (15)

[0115] Step 3.9: Fine-tune the dataset based on the dataset in the target domain , the target domain cognitive diagnosis model is trained based on the Adam optimizer to update the model parameters until the objective function Convergence is achieved to obtain a trained cognitive diagnosis model for the target domain, which can be used in scenarios such as student exercise recommendations, knowledge tracking, and computer adaptive testing.

[0116] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0117] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.

[0118] Example:

[0119] The cross-domain cognitive diagnosis method based on prompting technology proposed in the present invention is adapted to most existing cognitive diagnosis models. In this embodiment, facing the diversified exercise content, the cross-domain cognitive diagnosis method proposed in the present invention will be demonstrated on the NCDM model, which is a classic neural network-based model in the field of cognitive diagnosis.

[0120] Figure 2a The cross-domain cognitive diagnosis model based on NCDM in the source domain is presented. The model uses personalized prompts and shared prompts to learn the common representations of subjects and students in multiple source domains. Its construction process can be explained by equations (3) and (4): First, the student In the source field The original embedding on With personalized tips After horizontal splicing and mapping, we get student exist The final representation of ; Secondly, the exercises The original embedding Share Tips with After horizontal splicing and mapping, we get The final representation , and then input the final representation into the cognitive diagnosis interaction function In the function The corresponding parameters of the neural network are all positive; the model is based on its own prediction results And the real results ,Finally, the source domain cognitive diagnosis model is trained according to the objective function of formula (7);

[0121] Figure 2b The cross-domain cognitive diagnosis model based on NCDM in the target domain is presented. The model first transfers the personalized prompts and shared prompts trained in the source domain. In formula (5), the personalized prompts are transferred in a one-to-one manner, that is, the student Individualized prompts in target areas Equivalent to the best personalized tips in the source field , in formula (6), each source domain The corresponding optimal sharing prompt is optimal After being spliced ​​and mapped, the target domain is obtained Sharing tips on In addition, this method uses a linear layer to learn the student Personalized tips for In the target area The original embedding in The mapping relationship between them is applied to the original embedding corresponding to the unseen students, and then the representation is enhanced according to equations (3) and (4) to obtain the target domain Upper school student The final representation and Exercises The final representation , and then input the final representation into the cognitive diagnosis interaction function In the function The corresponding parameters of the neural network are all positive; the model is based on its own prediction results And the real results ,Finally, the target domain cognitive diagnosis model is trained according to the objective function of formula (15);

[0122] The above embodiments only express the case where the present invention is applied to the NCDM model. The present invention can be applied to most cognitive diagnosis models. The core idea of ​​the present invention is to use soft prompts to enhance cognitive diagnosis representation in cross-domain scenarios, which is the scope of protection to be applied for by the patent of the present invention.

Claims

1. A cross-domain cognitive diagnosis method based on prompting technology, characterized in that: The steps are as follows: Step 1: Construct cross-domain data; Define the source domain collection and target areas ,in, Indicates Source fields, Indicates the number of source fields; definition The set of students on , the exercise set is , the knowledge point set is ;in, Any student in , The Exercises are recorded as , The Knowledge points are recorded as ; and the exercises in each source domain are different; Define target areas The set of students on , the exercise set is , the knowledge point set is ;in, Any student in , No. Exercises are recorded as , The Knowledge points are recorded as ; and the exercises between the source domain set S and the target domain T are different; Using formula (1), we can get the source domain set and target areas Overlapping student sets between , and Middle Students are recorded as : (1) Defining the Source Domain Pre-training dataset on is a set of triples, where Any triple in express right The answer result is, if ,express Correct answer , , indicating other situations; from Randomly select a certain proportion of students’ answer data as the fine-tuning dataset ;in, Any triple in express right The answer result is, if ,express Correct answer , , indicating other situations; Will The relationship matrix between the previous exercises and knowledge points is recorded as , where Middle Row vector express and The correlation between each knowledge point above, where 1 or 0 indicates and relevant or irrelevant; express The number of exercises in express The number of knowledge points in Will The relationship matrix between the previous exercises and knowledge points is recorded as , where Middle Row vector express and The correlation between each knowledge point above, where 1 or 0 indicates and relevant or irrelevant; express The number of exercises in express The number of knowledge points in Step 2: Construct a source domain cognitive diagnosis model with soft prompts and exist The original embedding and The original embedding of exist Top pair Predicted answer results , thereby constructing the binary cross entropy loss , used to train the source domain cognitive diagnosis model, and obtain The best personalized tips on any source field and The best shared hint for all exercises in ; Step 3: Construct a target domain cognitive diagnosis model with soft prompts and and Process and obtain exist Top pair Predicted answer results , thus constructing exist Top pair The binary cross entropy loss , which is used to train the target domain cognitive diagnosis model to achieve exercise recommendation and knowledge tracking for students.

2. The cross-domain cognitive diagnosis method based on prompting technology according to claim 1 is characterized in that: Described step 2 is carried out as follows: Step 2.1: Through formula (3), we can get exist Composite representation on as well as Composite representation : (3) In formula (3), express Personalized tips on any source field, express Shared hints for all exercises in and They represent the random initialization exist The original embedding and The original embedding of Represents horizontal concatenation of vectors; Step 2.2: Through formula (4), we can get exist The final representation of as well as The final representation : (4) In formula (4), represents a linear layer for personalized hints on the source domain, represents a linear layer over the source domain for shared cues; Step 2.3: Get by formula (5) exist Top pair Predicted answer results : (5) In formula (5), represents the interactive function that predicts the answer result based on the student and exercise representations; Step 2.4: Construct through formula (6) exist Top pair The binary cross entropy loss : (6) In formula (6), yes exist Top pair The actual answer results; Step 2.5: Use formula (7) to construct the source domain cognitive diagnosis model for the objective function of all source domains : (7) Step 2.6: All pre-training datasets based on the source domain , the source domain cognitive diagnosis model is trained using the Adam optimizer to update the model parameters until the objective function Until convergence, the cognitive diagnosis model trained into the source domain is obtained. The best personalized tips on any source field and The best shared hint for all exercises in .

3. The cross-domain cognitive diagnosis method based on prompting technology according to claim 2 is characterized in that: The step 3 is carried out as follows: Step 3.1: Get by formula (8): In the target area Personalized tips on : (8) Step 3.2: Get the target domain through formula (9) Shared hints for all exercises in : (9) In formula (9), represents the vertical concatenation operation of shared cues in the source domain, Represents a dimension The row vector of ; Step 3.3: Use formula (10) to get exist The original embedding on : (10) Step 3.4: Through formula (11), we can get exist Composite representation on as well as Composite representation : (11) In formula (11), Represents the random initialization The original embedding of Step 3.5: Through formula (12), we can get exist The final representation of as well as The final representation : (12) In formula (12), represents a linear layer for personalized hints on the target domain, represents a linear layer for shared cues on the target domain; Step 3.6: Use formula (13) to get exist Top pair Predicted answer results : (13) Step 3.7: Construct through formula (14) exist Top pair The binary cross entropy loss : (14) Step 3.8: Construct the objective function of the source domain cognitive diagnosis model for the target domain through formula (15): : (15) Step 3.9: Fine-tune the dataset based on the target domain dataset , the target domain cognitive diagnosis model is trained based on the Adam optimizer to update the model parameters until the objective function Convergence is achieved, thus obtaining the trained target domain cognitive diagnosis model.

4. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the cross-domain cognitive diagnosis method according to any one of claims 1 to 3, and the processor is configured to execute the program stored in the memory.

5. 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 cross-domain cognitive diagnosis method according to any one of claims 1 to 3 are executed.

Citation Information

Patent Citations

  • Hierarchical cognitive diagnosis method, device and equipment

    CN108665006A

  • General patient personalized diagnosis and treatment scheme recommendation system based on cognitive map

    CN114496234A