Intelligent education personalized recommendation method based on federal learning

Abstract

The invention discloses an intelligent education personalized recommendation method based on federal learning, and relates to the technical field of intelligent education. Comprising the following steps: a plurality of participants obtain feature representation, establish course recommendation models, train the course recommendation models, encrypt gradient information of each course recommendation model and send the gradient information to a collaboration end; the cooperation end updates gradient information, encrypts the gradient information and sends the gradient information to the multiple participation ends; the plurality of participating ends update model parameters of respective course recommendation models after decryption, and encrypt gradient information of the course recommendation models again and send the gradient information to the collaboration end to update the model parameters of the course recommendation models until the model parameters converge; and the plurality of participants obtain historical answer accuracy distribution of current learning subjects of the respective users, the historical answer accuracy distribution is used as input of the course recommendation model for operation, recommended teaching tutoring courses corresponding to the current learning subjects are obtained, and the recommended teaching tutoring courses are recommended to the users of the plurality of participants. According to the method disclosed by the invention, proper teaching tutoring courses can be recommended to the user according to different types of conditions of the user.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent education, and particularly relates to a personalized recommendation method for intelligent education based on federated learning. Background Art

[0002] With the rapid development of information technology and the popularization of Internet applications, intelligent education has become an important trend in the education field. Personalized recommendation to users according to their learning situations is an important function of intelligent education. Currently, for personalized recommendation of users, the more common method is for the server to obtain the personal information and learning situations of each user, such as the answering scores of practice questions in a recent period of time, etc., and then recommend some popular teaching tutoring courses for the subjects in which the user has poor learning based on the user's learning situation, so as to improve the user's mastery of the knowledge of the corresponding subjects.

[0003] However, adopting such a recommendation method requires the server to collect user data uniformly, which easily leads to the leakage of a large amount of user information. At the same time, everyone's learning and mastery of different question types are different. By adopting the method of recommending popular teaching tutoring courses, it is difficult to accurately recommend suitable teaching tutoring courses for the actual learning situation of the user, resulting in a poor learning improvement effect for the user.

[0004] Therefore, how to provide an effective solution to recommend suitable teaching tutoring courses according to the user's learning and mastery of different question types while preventing the leakage of user information has become an urgent problem to be solved in the prior art. Summary of the Invention

[0005] The purpose of the present invention is to provide a personalized recommendation method for intelligent education based on federated learning to solve the above problems existing in the prior art.

[0006] To achieve the above purpose, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a personalized recommendation method for intelligent education based on federated learning, which is applied to a personalized recommendation system for intelligent education based on federated learning. The personalized recommendation system for intelligent education based on federated learning includes a collaboration end and a plurality of participating ends, and includes: The plurality of participating ends obtain feature representations for personalized recommendation, the feature representations indicate feature information for model training, and the feature information includes the teaching tutoring courses learned by historical users, the answering correct rate distributions of the corresponding subjects before the historical users learn the teaching tutoring courses, and the answering correct rate distributions of the corresponding subjects after the historical users learn the teaching tutoring courses. The answering correct rate distribution includes the answering correct rates of different types of questions; The multiple participating terminals establish a course recommendation model and train their respective course recommendation models based on the feature representation by obtaining training data locally; The multiple participating terminals encrypt the gradient information of their respective course recommendation models and send it to the collaboration terminal; The collaboration terminal decrypts the encrypted gradient information sent by the multiple participating terminals and performs gradient aggregation; The collaboration terminal updates the gradient information based on the gradient aggregation result, encrypts the updated gradient information, and sends it to the multiple participating terminals; The multiple participating terminals decrypt the updated gradient information and update the model parameters of their respective course recommendation models based on the updated gradient information; The multiple participating terminals re-encrypt the gradient information of their respective course recommendation models and send it to the collaboration terminal to re-update the model parameters of their respective course recommendation models until the model parameters of the course recommendation models of each participating terminal among the multiple participating terminals converge; The multiple participating terminals obtain the historical answering correct rate distribution of the current learning subjects of their respective users, and use the historical answering correct rate distribution of the current learning subjects of their respective users as the input of the corresponding course recommendation model for calculation to obtain the recommended teaching guidance courses corresponding to the current learning subjects of the users of the multiple participating terminals; The multiple participating terminals recommend teaching guidance courses to the users of the multiple participating terminals based on the recommended teaching guidance courses corresponding to the current learning subjects of their respective users.

[0007] Based on the above - disclosed content, the present invention obtains feature representations for personalized recommendation through multiple participating parties; the multiple participating parties establish a course recommendation model and train their respective course recommendation models based on the feature representations by obtaining training data locally; the multiple participating parties encrypt the gradient information of their respective course recommendation models and send it to the collaboration party; the collaboration party decrypts the encrypted gradient information sent by the multiple participating parties and performs gradient aggregation; the collaboration party updates the gradient information based on the gradient aggregation result, encrypts the updated gradient information, and sends it to the multiple participating parties; the multiple participating parties decrypt the updated gradient information and update the model parameters of their respective course recommendation models based on the updated gradient information; the multiple participating parties re - encrypt the gradient information of their respective course recommendation models and send it to the collaboration party to re - update the model parameters of their respective course recommendation models of the multiple participating parties until the model parameters of the course recommendation models of each participating party among the multiple participating parties converge; the multiple participating parties obtain the historical answering correct rate distributions of their respective users' current learning subjects and use the historical answering correct rate distributions of their respective users' current learning subjects as inputs to the corresponding course recommendation models for calculation, obtaining the recommended teaching and tutoring courses corresponding to the current learning subjects of the users of the multiple participating parties; the multiple participating parties recommend teaching and tutoring courses to the users of the multiple participating parties based on the recommended teaching and tutoring courses corresponding to the current learning subjects of their respective users. In this way, when recommending teaching and tutoring courses to users, the model training process is migrated from the central server to multiple local participating parties in a federated learning manner, thus avoiding the leakage of private data during the data transmission process, protecting user data privacy, and at the same time being able to reduce communication overhead, improve the real - time performance and response speed of the model. In addition, when recommending teaching and tutoring courses to users, teaching and tutoring courses can be recommended based on the answering correct rates of different types of questions, so that appropriate teaching and tutoring courses can be recommended to users according to the user's personal mastery of different types of questions, improving the user's learning improvement effect.

[0008] In a possible design, the multiple participating parties obtaining the feature representations for personalized recommendation includes: The multiple participating parties obtain the feature representations for personalized recommendation sent by the collaboration party.

[0009] In a possible design, the multiple participating parties use the historical answering correct rate distributions of their respective users' current learning subjects as inputs to the corresponding course recommendation models for calculation, obtaining the recommended teaching and tutoring courses corresponding to the current learning subjects of the users of the multiple participating parties, including: The multiple participating parties use the historical answering correct rate distributions of their respective users' current learning subjects as inputs to the corresponding course recommendation models for calculation, obtaining the predicted answering correct rate distributions corresponding to their respective users after learning various teaching and tutoring courses corresponding to the current learning subjects. Based on the predicted answer correct rate distributions corresponding to various teaching and tutoring courses for their respective users in learning the current learning subject, the multiple participating terminals determine the recommended teaching and tutoring courses corresponding to the current learning subject of the users of the multiple participating terminals.

[0010] In a possible design, the multiple participating terminals determine the recommended teaching and tutoring courses corresponding to the current learning subject of the users of the multiple participating terminals based on the predicted answer correct rate distributions corresponding to various teaching and tutoring courses for their respective users in learning the current learning subject, including: Based on the predicted answer correct rate distributions corresponding to various teaching and tutoring courses for their respective users in learning the current learning subject, the multiple participating terminals determine multiple answer correct rate growth parameters corresponding to the various teaching and tutoring courses for their respective users in learning the current learning subject; The multiple participating terminals use the teaching and tutoring course with the highest corresponding answer correct rate growth parameter among the multiple answer correct rate growth parameters corresponding to the various teaching and tutoring courses for their respective users in learning the current learning subject as the recommended teaching and tutoring course corresponding to the current learning subject of the corresponding users of the multiple participating terminals.

[0011] In a possible design, the answer correct rate growth parameter is obtained by weighting based on the answer correct growth rates of multiple question types.

[0012] In a possible design, the multiple participating terminals encrypt the gradient information of their respective course recommendation models and send it to the collaboration terminal, including: The multiple participating terminals convert the gradient information of their respective course recommendation models into a binary coding sequence; Each participating terminal among the multiple participating terminals determines a first binary secret key sequence for encryption based on a set of binary secret key sequences, and encrypts the corresponding binary coding sequence according to the determined first binary secret key sequence and sends it to the collaboration terminal, where multiple binary secret key sequences are recorded in the set of binary secret key sequences.

[0013] In a possible design, each participating terminal among the multiple participating terminals determines a first binary secret key sequence for encryption based on a set of binary secret key sequences, including: Each participating terminal among the multiple participating terminals determines a second binary secret key sequence for encryption from the set of binary secret key sequences based on the current timestamp; Each participating terminal among the multiple participating terminals determines a third binary secret key sequence for encryption based on the coding order of the binary coding sequence; Among the multiple participating parties, each participating party determines a first binary key sequence for encryption based on the corresponding second binary key sequence and third binary key sequence.

[0014] In a second aspect, the present invention provides an intelligent education personalized recommendation system based on federated learning, including a collaboration party and multiple participating parties. The multiple participating parties are used to obtain feature representations for personalized recommendation; establish a course recommendation model, and train their respective course recommendation models based on the feature representations by obtaining training data locally; and encrypt the gradient information of their respective course recommendation models and send it to the collaboration party, where the feature representation indicates feature information for model training, and the feature information includes the teaching assistance courses learned by historical users, the distribution of the correct answer rates of corresponding subjects before historical users learn the teaching assistance courses, and the distribution of the correct answer rates of corresponding subjects after historical users learn the teaching assistance courses. The distribution of the correct answer rates includes the correct answer rates of different types of questions; The collaboration party decrypts the encrypted gradient information sent by the multiple participating parties and performs gradient aggregation; and update the gradient information based on the gradient aggregation result, and encrypt the updated gradient information and send it to the multiple participating parties; The multiple participating parties are also used to decrypt the updated gradient information and update the model parameters of their respective course recommendation models based on the updated gradient information; encrypt the gradient information of their respective course recommendation models and send it to the collaboration party to re-update the model parameters of the respective course recommendation models of the multiple participating parties until the model parameters of the course recommendation models of each participating party among the multiple participating parties converge; obtain the historical correct answer rate distribution of the current learning subject of their respective users, and use the historical correct answer rate distribution of the current learning subject of their respective users as the input of the corresponding course recommendation model for calculation to obtain the recommended teaching assistance courses corresponding to the current learning subjects of the users of the multiple participating parties; and recommend teaching assistance courses to the users of the multiple participating parties based on the recommended teaching assistance courses corresponding to the current learning subjects of their respective users.

[0015] Beneficial effects: When the present invention recommends teaching tutoring courses to users, it adopts the method of federated learning to migrate the model training process from the central server to multiple local participating parties, thereby avoiding the leakage of private data during the data transmission process, protecting the privacy of user data, and at the same time being able to reduce the communication overhead, improve the real-time performance and response speed of the model. In addition, when recommending teaching tutoring courses to users, it is possible to recommend teaching tutoring courses to users based on the correct answering rates of different types of questions, so as to be able to recommend appropriate teaching tutoring courses to users according to the user's personal mastery of different types of questions, improve the learning improvement effect of users, and facilitate practical application and promotion. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flowchart of the intelligent education personalized recommendation method based on federated learning provided by an embodiment of the present application; Figure 2 It is a block diagram schematic diagram of the intelligent education personalized recommendation system based on federated learning provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the present invention in combination with the drawings and the descriptions of the embodiments or the prior art. Obviously, the following descriptions of the structures of the drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts. It should be noted here that the descriptions of these embodiments are used to help understand the present invention, but do not constitute a limitation to the present invention.

[0018] It should be understood that although terms such as first and second may be used herein to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, the first unit may be called the second unit, and similarly, the second unit may be called the first unit, without departing from the scope of the exemplary embodiments of the present invention.

[0019] It should be understood that for the term "and / or" that may appear in this article, it is only a description of the association relationship of the associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, B exists alone, and A and B exist simultaneously; for the term " / and" that may appear in this article, it is a description of another association object relationship, indicating that two relationships may exist. For example, A / and B may represent: A exists alone, and A and B exist alone; in addition, for the character " / " that may appear in this article, generally, the front and rear associated objects represent an "or" relationship.

[0020] Such as Figure 1As shown in the figure, this embodiment provides a personalized recommendation method for intelligent education based on federated learning, which is applied to a personalized recommendation system for intelligent education based on federated learning. The personalized recommendation system for intelligent education based on federated learning includes a collaboration end and multiple participating ends. The collaboration end is communicatively connected to the multiple participating ends for data interaction or communication. The personalized recommendation method for intelligent education based on federated learning may but is not limited to include the following steps S101-S109.

[0021] Step S101. Multiple participating ends obtain feature representations for personalized recommendation.

[0022] In the embodiments of the present application, the feature representations for personalized recommendation may be sent by the collaboration end to the multiple participating ends, or may be jointly determined by the participating ends through negotiation. The collaboration end may be a server, and the participating ends may be local terminal devices.

[0023] The feature representation indicates the feature information for model training. The feature information may but is not limited to include the teaching tutoring courses learned by historical users, the distribution of the answering correct rates of corresponding subjects before the historical users learned the teaching tutoring courses, and the distribution of the answering correct rates of corresponding subjects after the historical users learned the teaching tutoring courses. The distribution of the answering correct rates includes the answering correct rates of different types of question types.

[0024] The different types of question types may but are not limited to include multiple-choice questions, true or false questions, reading comprehension questions, writing questions, etc. According to different subjects, they can be divided into different question types.

[0025] Step S102. Multiple participating ends establish a course recommendation model and train their respective course recommendation models based on the feature representations by obtaining training data locally.

[0026] After each of the multiple participating ends obtains the feature representations for personalized recommendation, it can establish a course recommendation model and train its respective course recommendation model based on the feature representations by obtaining the training data for training locally. The training data may include various teaching tutoring courses learned by users, the distribution of the answering correct rates of corresponding subjects before (within a period of time) the users learned various teaching tutoring courses, and the distribution of the answering correct rates of corresponding subjects after the users learned various teaching tutoring courses.

[0027] When training, the distribution of the answering correct rates of corresponding subjects before the users learned various teaching tutoring courses and the distribution of the answering correct rates of corresponding subjects after the users learned various teaching tutoring courses can be converted into multi-dimensional vectors, and the distribution of the answering correct rates of corresponding subjects before the users learned various teaching tutoring courses (the corresponding multi-dimensional vectors) can be used as the input, and the distribution of the answering correct rates of corresponding subjects after the users learned various teaching tutoring courses (the corresponding multi-dimensional vectors) can be used as the output for training.

[0028] Among them, the distribution of the correct answer rates for corresponding subjects can be the average correct answer rates for various types of questions. For example, the average correct answer rates for the four types of questions of subject a for user A before learning the teaching tutoring course are 58%, 65%, 80%, and 60%. Then, the distribution of the correct answer rates for subject a for user A before learning the teaching tutoring course can be expressed as (58%, 65%, 80%, 60%), and when converted into a multi-dimensional vector, it can be expressed as (0.58, 0.65, 0.8, 0.6).

[0029] Step S103. Multiple participating ends encrypt the gradient information of their respective course recommendation models and send it to the collaboration end.

[0030] In the embodiments of the present application, each participating end pre-stores a set of binary secret key sequences. The set of binary secret key sequences records multiple binary secret key sequences. The set of binary secret key sequences can be sent by the collaboration end to each participating end, or can be determined through negotiation by each participating end.

[0031] When multiple participating ends encrypt the gradient information of their respective course recommendation models, they can first convert the gradient information of their respective course recommendation models into a binary coding sequence. Then, each participating end determines a first binary secret key sequence for encryption based on the set of binary secret key sequences, and encrypts the corresponding binary coding sequence according to the determined first binary secret key sequence and sends it to the collaboration end.

[0032] Specifically, each participating end can establish a correspondence between time and each binary secret key sequence in the set of binary secret key sequences. For example, different time periods correspond to different binary secret key sequences. When encrypting the gradient information of the course recommendation model, each participating end can determine a second binary secret key sequence for encryption based on the current timestamp from the set of binary secret key sequences.

[0033] Each participating end can also establish a correspondence between the coding order of the binary coding sequence corresponding to the gradient information and each binary secret key sequence in the set of binary secret key sequences. For example, different coding orders of specified multiple-bit binary codings (such as the last four binary codings) in the binary coding sequence corresponding to the gradient information correspond to different binary secret key sequences. When encrypting the gradient information of the course recommendation model, each participating end can also determine a third binary secret key sequence for encryption based on the coding order of the binary coding sequence corresponding to the gradient information.

[0034] Then, each participating end among multiple participating ends determines a first binary key sequence for encryption based on the corresponding second binary key sequence and third binary key sequence. When determining the first binary key sequence, each participating end among the multiple participating ends can perform an exclusive OR operation on its corresponding second binary key sequence and third binary key sequence to determine the first binary key sequence for encryption by each participating end.

[0035] Step S104. The collaborating end decrypts the encrypted gradient information sent by the multiple participating ends and then performs gradient aggregation.

[0036] Among them, gradient aggregation can be to perform an average operation or a weighted operation after decrypting the encrypted gradient information sent by the multiple participating ends.

[0037] The decryption process is the reverse process of the encryption process, and thus it will not be elaborated here.

[0038] Step S105. The collaborating end updates the gradient information based on the gradient aggregation result, and encrypts the updated gradient information and sends it to the multiple participating ends.

[0039] Specifically, the collaborating end can use the gradient aggregation result as the updated gradient information, and encrypt the updated gradient information and send it to the multiple participating ends. It can be understood that the collaborating end can adopt the same or different encryption methods as the participating ends.

[0040] Step S106. The multiple participating ends decrypt the updated gradient information, and update the model parameters of their respective course recommendation models based on the updated gradient information.

[0041] Step S107. The multiple participating ends re-encrypt the gradient information of their respective course recommendation models and send it to the collaborating end to re-update the model parameters of their respective course recommendation models of the multiple participating ends until the model parameters of the course recommendation models of each participating end among the multiple participating ends converge.

[0042] The multiple participating ends update the model parameters of their respective course recommendation models, and can repeat the above steps S103 - S106 until the model parameters of the course recommendation models of each participating end among the multiple participating ends converge. That is, the multiple participating ends can encrypt the latest gradient information of their respective course recommendation models and send it to the collaborating end, so that the collaborating end decrypts it and re-performs gradient aggregation and updates the gradient information, and then encrypts it and sends it back to the multiple participating ends, so that the multiple participating ends re-update the model parameters of their respective course recommendation models until the model parameters of the course recommendation models of each participating end among the multiple participating ends converge.

[0043] Step S108. Multiple participating parties obtain the historical answering correct rate distributions of their respective users' current learning subjects, and use the historical answering correct rate distributions of their respective users' current learning subjects as inputs for the corresponding course recommendation models to perform calculations, obtaining the recommended teaching and tutoring courses corresponding to the current learning subjects of the users of the multiple participating parties.

[0044] Specifically, step S108 may but is not limited to including the following steps S1081 - S1082.

[0045] Step S1081. Multiple participating parties use the historical answering correct rate distributions of their respective users' current learning subjects as inputs for the corresponding course recommendation models to perform calculations, obtaining the predicted answering correct rate distributions corresponding to their respective users after learning various teaching and tutoring courses corresponding to the current learning subjects.

[0046] Step S1082. Multiple participating parties determine the recommended teaching and tutoring courses corresponding to the current learning subjects of the users of the multiple participating parties based on the predicted answering correct rate distributions corresponding to their respective users after learning various teaching and tutoring courses corresponding to the current learning subjects.

[0047] Specifically, multiple participating parties can determine multiple answering correct rate growth parameters corresponding to their respective users after learning various teaching and tutoring courses corresponding to the current learning subjects based on the predicted answering correct rate distributions corresponding to their respective users after learning various teaching and tutoring courses corresponding to the current learning subjects.

[0048] Then, multiple participating parties use the teaching and tutoring course with the highest corresponding answering correct rate growth parameter among the multiple answering correct rate growth parameters corresponding to their respective users after learning various teaching and tutoring courses corresponding to the current learning subjects as the recommended teaching and tutoring course corresponding to the current learning subject of the corresponding users of the multiple participating parties. Among them, the answering correct rate growth parameter can be obtained by weighting the answering correct growth rates of various question types.

[0049] Step S109. Multiple participating parties recommend teaching and tutoring courses to the users of the multiple participating parties based on the recommended teaching and tutoring courses corresponding to the current learning subjects of their respective users.

[0050] In summary, the personalized recommendation method for intelligent education based on federated learning provided by the present invention obtains feature representations for personalized recommendation through multiple participating parties; the multiple participating parties establish a course recommendation model and train their respective course recommendation models based on the feature representations by obtaining training data locally; the multiple participating parties encrypt the gradient information of their respective course recommendation models and send it to the collaboration party; the collaboration party decrypts the encrypted gradient information sent by the multiple participating parties and performs gradient aggregation; the collaboration party updates the gradient information based on the gradient aggregation result, encrypts the updated gradient information, and sends it to the multiple participating parties; the multiple participating parties decrypt the updated gradient information and update the model parameters of their respective course recommendation models based on the updated gradient information; the multiple participating parties re-encrypt the gradient information of their respective course recommendation models and send it to the collaboration party to re-update the model parameters of their respective course recommendation models of the multiple participating parties until the model parameters of the course recommendation models of each participating party among the multiple participating parties converge; the multiple participating parties obtain the historical answering correct rate distribution of the current learning subject of their respective users, and use the historical answering correct rate distribution of the current learning subject of their respective users as the input of the corresponding course recommendation model for calculation, and obtain the recommended teaching tutoring courses corresponding to the current learning subject of the users of the multiple participating parties; the multiple participating parties recommend teaching tutoring courses to the users of the multiple participating parties based on the recommended teaching tutoring courses corresponding to the current learning subject of their respective users. In this way, when recommending teaching tutoring courses to users, the model training process is migrated from the central server to multiple local participating parties in a federated learning manner, thereby avoiding the leakage of private data during data transmission, protecting user data privacy, and at the same time being able to reduce communication overhead, improve the real-time performance and response speed of the model. In addition, when recommending teaching tutoring courses to users, teaching tutoring courses can be recommended based on the answering correct rates of different types of questions, so that appropriate teaching tutoring courses can be recommended to users according to the user's personal mastery of different types of questions, improving the learning improvement effect of users and facilitating practical application and promotion.

[0051] Please refer to Figure 2 , the second aspect of the embodiment of the present application provides a personalized recommendation system for intelligent education based on federated learning. The personalized recommendation system for intelligent education based on federated learning includes a collaboration party and multiple participating parties. The collaboration party is communicatively connected to the multiple participating parties for data interaction or communication. The multiple participating parties are used to obtain feature representations for personalized recommendation; establish a course recommendation model and train their respective course recommendation models based on the feature representations by obtaining training data locally; and Encrypt the gradient information of their respective course recommendation models and send it to the collaboration side, where the feature representation indicates the feature information for model training, and the feature information includes the tutoring courses learned by historical users, the distribution of the answering correct rates of the corresponding subjects before the historical users learned the tutoring courses, and the distribution of the answering correct rates of the corresponding subjects after the historical users learned the tutoring courses. The distribution of the answering correct rates includes the answering correct rates of different types of questions; The collaboration side decrypts the encrypted gradient information sent by the multiple participating sides and then performs gradient aggregation; and Update the gradient information based on the gradient aggregation result, encrypt the updated gradient information, and send it to the multiple participating sides; The multiple participating sides are also used to decrypt the updated gradient information and update the model parameters of their respective course recommendation models based on the updated gradient information; Encrypt the gradient information of their respective course recommendation models and send it to the collaboration side to re-update the model parameters of the respective course recommendation models of the multiple participating sides until the model parameters of the course recommendation models of each participating side among the multiple participating sides converge; Obtain the historical answering correct rate distribution of the current learning subjects of their respective users, and use the historical answering correct rate distribution of the current learning subjects of their respective users as the input of the corresponding course recommendation model for calculation to obtain the recommended tutoring courses corresponding to the current learning subjects of the users of the multiple participating sides; and Recommend tutoring courses to the users of the multiple participating sides based on the recommended tutoring courses corresponding to the current learning subjects of their respective users.

[0052] For the working process, working details and technical effects of the intelligent education personalized recommendation system based on federated learning provided in the second aspect of this embodiment, reference can be made to the first aspect of the embodiment, and details are not described herein again.

[0053] It should be understood that specific details are provided in the following description to facilitate a complete understanding of the example embodiments. However, those of ordinary skill in the art should understand that the example embodiments can be implemented without these specific details. For example, the system can be shown in a block diagram to avoid obscuring the example with unnecessary details. In other instances, well-known processes, structures, and technologies can be shown without unnecessary details to avoid obscuring the example embodiments.

[0054] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A smart education personalized recommendation method based on federated learning, applied to a smart education personalized recommendation system based on federated learning, the smart education personalized recommendation system based on federated learning includes a collaboration end and multiple participating ends, characterized in that: include: The multiple participating terminals obtain feature representations for personalized recommendation, wherein the feature representations indicate feature information for model training, wherein the feature information includes the teaching and tutoring courses studied by the historical user, the distribution of correct answer rates of the corresponding subjects before the historical user studied the teaching and tutoring courses, and the distribution of correct answer rates of the corresponding subjects after the historical user studied the teaching and tutoring courses, wherein the distribution of correct answer rates includes the correct answer rates of different types of questions; The multiple participating terminals establish a course recommendation model, and obtain training data from the local area based on the feature representation to train their respective course recommendation models; The multiple participating terminals encrypt the gradient information of their respective course recommendation models and send the encrypted information to the collaborative terminal; The collaboration end decrypts the encrypted gradient information sent by the multiple participating ends and then performs gradient aggregation; The collaboration end updates the gradient information based on the gradient aggregation result, and encrypts the updated gradient information and sends it to the multiple participating ends; The multiple participating terminals decrypt the updated gradient information, and update the model parameters of their respective course recommendation models based on the updated gradient information; The multiple participating terminals re-encrypt the gradient information of their respective course recommendation models and send them to the collaboration terminal, so as to re-update the model parameters of the course recommendation models of the multiple participating terminals, until the model parameters of the course recommendation models of the multiple participating terminals converge; The multiple participating terminals obtain the historical answer accuracy distribution of the current learning subject of each user, and use the historical answer accuracy distribution of the current learning subject of each user as the input of the corresponding course recommendation model for calculation, to obtain the recommended teaching and tutoring courses corresponding to the current learning subject of the users of the multiple participating terminals; The plurality of participating terminals recommend teaching and tutoring courses to the users of the plurality of participating terminals based on the recommended teaching and tutoring courses corresponding to the current learning subjects of the respective users.

2. The method for personalized recommendation of smart education based on federated learning according to claim 1, characterized in that: The multiple participating terminals obtain feature representations for personalized recommendation, including: The multiple participating terminals obtain the feature representation for personalized recommendation sent by the cooperating terminal.

3. The method for personalized recommendation of smart education based on federated learning according to claim 1, characterized in that: The multiple participating terminals use the historical answer accuracy distribution of the current learning subject of each user as the input of the corresponding course recommendation model to perform calculations to obtain the recommended teaching and tutoring courses corresponding to the current learning subject of the users of the multiple participating terminals, including: The multiple participating terminals use the historical answer accuracy distribution of the current learning subject of each user as the input of the corresponding course recommendation model to perform calculations, and obtain the corresponding predicted answer accuracy distribution after each user studies multiple teaching and tutoring courses corresponding to the current learning subject; The multiple participating terminals determine the recommended teaching and tutoring courses corresponding to the current learning subjects of the users of the multiple participating terminals based on the distribution of predicted answer accuracy rates after the respective users have studied multiple teaching and tutoring courses corresponding to the current learning subjects.

4. The method for personalized recommendation of smart education based on federated learning according to claim 3 is characterized in that: The plurality of participating terminals determine the recommended teaching and tutoring courses corresponding to the current learning subjects of the users of the plurality of participating terminals based on the distribution of the predicted correct answer rates of the respective users after they have learned the plurality of teaching and tutoring courses corresponding to the current learning subjects, including: The multiple participating terminals determine the multiple answer accuracy growth parameters corresponding to the multiple teaching and tutoring courses corresponding to the current learning subject based on the predicted answer accuracy distribution of the respective users after learning the multiple teaching and tutoring courses corresponding to the current learning subject; The multiple participating terminals will use the teaching and tutoring course with the highest answering accuracy growth parameter among the multiple teaching and tutoring courses corresponding to the current learning subject after each user has studied the multiple teaching and tutoring courses corresponding to the current learning subject as the recommended teaching and tutoring course corresponding to the current learning subject of the users corresponding to the multiple participating terminals.

5. The method for personalized recommendation of smart education based on federated learning according to claim 4 is characterized in that: The correct answer rate growth parameter is obtained by weighting the correct answer rate growth rates of multiple question types.

6. The method for personalized recommendation of smart education based on federated learning according to claim 1, characterized in that: The multiple participating terminals encrypt the gradient information of their respective course recommendation models and send them to the collaborative terminal, including: The multiple participating terminals convert the gradient information of their respective course recommendation models into binary coding sequences; Each of the multiple participating terminals determines a first binary key sequence for encryption based on a binary key sequence set, encrypts the corresponding binary coding sequence according to the determined first binary key sequence, and sends the encrypted sequence to the collaborative terminal, wherein the binary key sequence set records multiple binary key sequences.

7. The method for personalized recommendation of smart education based on federated learning according to claim 6, characterized in that: Each of the plurality of participating terminals determines a first binary key sequence for encryption based on the binary key sequence set, including: Each of the plurality of participating terminals determines a second binary key sequence for encryption from the set of binary key sequences based on a current timestamp; Each of the plurality of participating terminals determines a third binary key sequence for encryption based on the encoding order of the binary encoding sequence; Each of the multiple participating terminals determines a first binary key sequence for encryption based on the corresponding second binary key sequence and third binary key sequence.

8. The method for personalized recommendation of smart education based on federated learning according to claim 7, characterized in that: Each of the plurality of participating terminals determines a first binary key sequence for encryption based on the corresponding second binary key sequence and third binary key sequence, including: Each of the multiple participating terminals performs an XOR operation based on the corresponding second binary key sequence and third binary key sequence to determine a first binary key sequence used for encryption.

9. A smart education personalized recommendation system based on federated learning, characterized in that: It includes a collaboration end and a plurality of participating ends, wherein the plurality of participating ends are used to obtain feature representations for personalized recommendation; Establishing a course recommendation model, and acquiring training data from a local location based on the feature representation to train the respective course recommendation models; as well as Encrypting the gradient information of the respective course recommendation models and sending them to the collaboration end, wherein the feature representation indicates feature information used for model training, the feature information includes the teaching and tutoring courses studied by the historical user, the distribution of the correct answer rates of the corresponding subjects before the historical user studied the teaching and tutoring courses, and the distribution of the correct answer rates of the corresponding subjects after the historical user studied the teaching and tutoring courses, and the distribution of the correct answer rates includes the correct answer rates of different types of questions; The collaboration end decrypts the encrypted gradient information sent by the multiple participating ends and then performs gradient aggregation; as well as Update the gradient information based on the gradient aggregation result, and encrypt the updated gradient information and send it to the multiple participating terminals; The multiple participating terminals are also used to decrypt the updated gradient information and update the model parameters of their respective course recommendation models based on the updated gradient information; Encrypting the gradient information of each course recommendation model and sending it to the collaboration end, so as to re-update the model parameters of the course recommendation models of each of the multiple participating ends, until the model parameters of the course recommendation models of each of the multiple participating ends converge; Obtaining the historical answer accuracy distribution of the current learning subject of each user, and using the historical answer accuracy distribution of the current learning subject of each user as the input of the corresponding course recommendation model for calculation, to obtain the recommended teaching and tutoring courses corresponding to the current learning subject of the users of the multiple participating terminals; as well as Based on the recommended teaching and tutoring courses corresponding to the current learning subjects of the respective users, teaching and tutoring courses are recommended to the users of the multiple participating terminals.

Images