Push method and system based on joint prediction

Abstract

The present disclosure provides a push method and system based on joint prediction, the method comprising: obtaining first data of a requesting party at a first participant and second data at a second participant, wherein the first data at the first participant is different from the second data at the second participant, and the first data and the second data are not interoperable; performing a joint prediction based on the first data and the second data obtained by the shared learning combination to perform joint prediction-related push to the requesting party; and feeding back the result of the joint prediction to the first participant and the second participant.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence, and more specifically, to a push method and system based on joint prediction. Background Art

[0002] In many scenarios in life, companies and public welfare groups need to push some activities based on user characteristics. However, traditionally, each company or group operates based on its own data. On the one hand, since they operate separately, users may receive push notifications from each company, which may result in a waste of rights and interests. On the other hand, each participant only has data on the user in his own scenario, but not data in other scenarios. This is a partial rather than complete portrayal of user behavior, resulting in poor push accuracy. Therefore, with the increase in operating costs and the reduction in operating budgets, how to push notifications to the most valuable users is a difficult problem to solve.

[0003] In addition, some companies or groups also conduct joint operations through interoperable features. However, since data is interoperable in plain text, there is a risk of leakage of users' personal privacy information. In addition, since each company has a confidentiality agreement, some particularly sensitive data will not be interoperable, which leads to incomplete characterization of user characteristics, resulting in poor recommendation accuracy.

[0004] Therefore, we hope to provide a push method based on joint prediction, which can break down data silos, realize joint operation while protecting the privacy of each party, reduce unnecessary pushes, and better characterize users. Summary of the invention

[0005] This disclosure is provided to introduce some concepts in a simplified form that will be further described in the following detailed description. This disclosure is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used as an aid in determining the scope of the claimed subject matter.

[0006] In response to the above problems, according to one aspect of the present disclosure, a push method based on joint prediction is provided, the method comprising: obtaining first data of a requesting party at a first participant and second data at a second participant, wherein the first data at the first participant is different from the second data at the second participant, and the first data and the second data are not interoperable; performing a joint prediction based on the first data and the second data obtained by sharing learning to perform joint prediction-related push to the requesting party; and feeding back the result of the joint prediction to the first participant and the second participant.

[0007] In the technical solution of the embodiment of the present disclosure, different participants jointly predict related pushes, and there is no longer a situation where a user may receive two pushes at the same time, reducing unnecessary pushes. In addition, the present disclosure also uses user feature data from different participants to solve the data island problem, which makes the user's characterization more complete and the prediction results more accurate. In addition, the present disclosure uses the encryption technology of shared learning to ensure that participants will not be at risk of privacy leakage.

[0008] According to one embodiment of the present disclosure, performing a joint prediction based on the first data and the second data acquired through shared learning to perform a joint prediction-related push to the requesting party further includes: inputting the acquired first data and the second data into a joint prediction model to perform a joint prediction-related push to the requesting party, wherein the joint prediction model is obtained by combining a first participant model at the first participant and a second participant model at the second participant, and the first participant model and the second participant model are obtained by encrypted training based on the shared learning combined with the first data and the second data via a third-party collaborator.

[0009] According to a further embodiment of the present disclosure, the first participant model and the second participant model are trained by the following operations, wherein the training data set includes first data of a common group at the first participant in the same time period, second data at the second participant, and training labels about the joint prediction results: initializing the first model parameters of the first participant model and the second model parameters of the second participant model; the first participant and the second participant each calculate the corresponding first intermediate result and the second intermediate result and use the public key sent by the third-party collaborator to perform encrypted interaction; the first participant and the second participant each calculate the encrypted gradient and send it to the third-party collaborator after adding a mask; the second participant calculates the encrypted loss and sends it to the third-party collaborator; and the third-party collaborator sends the decrypted gradient back to the first participant and the second participant respectively to update the first model parameters and the second model parameters until the loss converges.

[0010] According to a further embodiment of the present disclosure, the first participant model and the second participant model are trained based on homomorphic encryption and linear regression.

[0011] According to a further embodiment of the present disclosure, the common population is obtained by using an encryption-based sample alignment algorithm.

[0012] According to a further embodiment of the present disclosure, the sample alignment algorithm includes an RSA algorithm, wherein the first participant and the second participant use the public key sent by the third-party collaborator to encrypt the sample data, and send the encrypted sample data to the third-party collaborator; and the third-party collaborator uses the private key to decrypt the encrypted sample data, and takes out the common group and returns it to the first participant and the second participant.

[0013] According to a further embodiment of the present disclosure, the result of the joint prediction is multi-classification.

[0014] According to another aspect of the present disclosure, a push system based on joint prediction is provided, the system comprising: a data acquisition module, for acquiring first data of a requesting party at a first participant and second data at a second participant, wherein the first data at the first participant is different from the second data at the second participant, and the first data and the second data are not interoperable; a joint prediction module, for performing a joint prediction based on the first data and the second data acquired through shared learning to perform a joint prediction-related push to the requesting party; and a result feedback module, for feeding back the result of the joint prediction to the first participant and the second participant.

[0015] In the technical solution of the embodiment of the present disclosure, the present disclosure combines user data based on different features from different participants to perform joint predictions using shared learning, so as to push information related to the joint prediction to the requesting party, thereby avoiding unnecessary pushes and enabling different participants to communicate user features while ensuring data security, solving the data island problem and providing a more complete characterization of users.

[0016] According to one embodiment of the present disclosure, the joint prediction module further: inputs the acquired first data and second data into a joint prediction model to perform joint prediction related push to the requesting party, wherein the joint prediction model is obtained by combining a first participant model at the first participant and a second participant model at the second participant, and the first participant model and the second participant model are obtained by combining the first data and the second data through encrypted training by a third party collaborator based on shared learning.

[0017] According to a further embodiment of the present disclosure, the first participant model and the second participant model are trained by the following operations, wherein the training data set includes first data of a common group at the first participant in the same time period, second data at the second participant, and training labels about the joint prediction results: initializing the first model parameters of the first participant model and the second model parameters of the second participant model; the first participant and the second participant each calculate the corresponding first intermediate result and the second intermediate result and use the public key sent by the third-party collaborator to perform encrypted interaction; the first participant and the second participant each calculate the encrypted gradient and send it to the third-party collaborator after adding a mask; the second participant calculates the encrypted loss and sends it to the third-party collaborator; and the third-party collaborator sends the decrypted gradient back to the first participant and the second participant respectively to update the first model parameters and the second model parameters until the loss converges.

[0018] According to a further embodiment of the present disclosure, the first participant model and the second participant model are trained based on homomorphic encryption and linear regression.

[0019] According to a further embodiment of the present disclosure, the common population is obtained by using an encryption-based sample alignment algorithm.

[0020] According to a further embodiment of the present disclosure, the sample alignment algorithm includes an RSA algorithm, wherein the first participant and the second participant use the public key sent by the third-party collaborator to encrypt the sample data, and send the encrypted sample data to the third-party collaborator; and the third-party collaborator uses the private key to decrypt the encrypted sample data, and takes out the common group and returns it to the first participant and the second participant.

[0021] According to a further embodiment of the present disclosure, the result of the joint prediction is multi-classification.

[0022] According to yet another aspect of the present disclosure, a computer-readable storage medium storing instructions is provided. When the instructions are executed, the machine is caused to perform the method described in any one of the aforementioned aspects.

[0023] These and other features and advantages will become apparent by reading the following detailed description and by reference to the associated drawings.It is to be understood that the foregoing general description and the following detailed description are illustrative only and are not restrictive of the aspects of what is claimed. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to understand in detail the manner in which the above features of the present disclosure are used, the above briefly summarized contents may be described in more detail with reference to various embodiments, some of which are shown in the accompanying drawings. However, it should be noted that the accompanying drawings only show certain typical aspects of the present disclosure and should not be considered to limit its scope, as the description may allow for other equally effective aspects.

[0025] Figure 1 It is a schematic architecture diagram of a push system based on joint prediction according to an embodiment of the present disclosure.

[0026] Figure 2 It is a schematic architecture diagram of a training process of a joint prediction model combining a first participant model and a second participant model according to an embodiment of the present disclosure.

[0027] Figure 3 is a schematic diagram of encryption sample alignment according to an embodiment of the present disclosure.

[0028] Figure 4 It is a schematic diagram of encryption model training according to an embodiment of the present disclosure.

[0029] Figure 5 It is a schematic flowchart of a push method based on joint prediction according to an embodiment of the present disclosure.

[0030] Figure 6 It is a schematic architecture diagram of a push system based on joint prediction according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0031] The present disclosure is described in detail below in conjunction with the accompanying drawings, and the features of the present disclosure will be further revealed in the following specific description. For the convenience of explanation, the implementation of the present disclosure will be described below using payment platforms and offline merchants as examples, but those skilled in the art can understand that the present disclosure is also applicable to other participants.

[0032] Figure 1 FIG. 1 shows a schematic architecture diagram of a push system 100 based on joint prediction according to an embodiment of the present disclosure. Figure 1 As shown in , the system 100 may include at least a data collection module 101 , a joint prediction module 102 and a result feedback module 103 .

[0033] The data collection module 101 can obtain the first data of the requester at the first participant and the second data at the second participant, wherein the first data at the first participant and the second data at the second participant are not interoperable, and the first data is different from the second data, that is, the first data and the second data are based on different characteristics of the requester. In one embodiment, in the scenario of smart risk control, the data collection module 101 can obtain different asset data (e.g., credit data, etc.) of the requester at different data owners. In another embodiment, in the scenario of smart payment, the data collection module 101 can obtain different payment data of the user at different data owners (e.g., the first payment data at the payment platform, the second payment data at the offline merchant, etc.). In another embodiment, in the scenario of smart education, the data collection module 101 can obtain different education data of the requester at different data owners.

[0034] The joint prediction module 102 can perform joint prediction based on the first data and the second data obtained by the shared learning combination to push the joint prediction related information to the requesting party. Further, the joint prediction module 102 can input the first data and the second data obtained into the joint prediction model to push the joint prediction related information to the requesting party, wherein the joint prediction model is obtained by combining the first participant model at the first participant and the second participant model at the second participant, and the first participant model and the second participant model are obtained by encrypted training based on the shared learning combination of the first data and the second data through a third-party collaborator. The specific training process will be described below. Figure 2-4 In one embodiment, the joint prediction module 102 may perform joint prediction based on the payment data of the requesting party at each data owner to push the joint prediction related information to the requesting party (for example, each data owner pushes the equity based on the joint prediction result to avoid the waste of equity).

[0035] The result feedback module 103 can feedback the result of the joint prediction to the first participant and the second participant. In the above-mentioned smart payment implementation, the result feedback module 103 can feedback the result of the joint prediction to different data owners, where different data owners can push related joint predictions based on the feedback results (for example, if the result of the joint prediction indicates that the user is more willing to pay, the lowest level of benefits will be pushed).

[0036] Therefore, by combining the data owned by different data owners to jointly push related predictions, it is possible to avoid the situation where a user receives multiple pushes at the same time, thus reducing unnecessary pushes. In addition, by using the data of different data owners at the same time, data silos are broken, the user's characterization is more complete, and the prediction results of the model are more accurate. In addition, through the encryption technology of shared learning, it can be ensured that different data owners will not have the risk of user privacy data leakage.

[0037] The following uses the red envelope code as an example to explain the scenario where the payment platform and offline merchants use the system 100 to jointly push benefits. Figure 1 As shown in , when the user scans the code to receive the red envelope code, the data collection module 101 can read the relevant payment feature information of the user at the payment platform and offline merchants. For example, the data collection module 101 can obtain the basic information of the user at the payment platform, online and offline payment information, and payment behavior information at offline stores. The joint prediction module 102 can then input the acquired user information into a joint prediction model based on shared learning to obtain a predicted user willingness to pay for the offline merchant's low willingness, so that the offline merchant can push the rights of additional subsidies based on the prediction results (for example, offline consumption instant reduction rights, etc.). Subsequently, the result feedback module 103 can feedback the results of the joint prediction to the payment platform and offline merchants.

[0038] Those skilled in the art will appreciate that the system and its modules disclosed herein may be implemented in hardware or software, and the modules may be merged or combined in any appropriate manner.

[0039] Figure 2 A schematic architecture diagram of a training process 200 of a joint prediction model combining a first participant model and a second participant model according to an embodiment of the present disclosure is shown. The training process 200 mainly includes three stages: data preparation, encrypted sample alignment, and encrypted model training.

[0040] In the process of the first stage (i.e., data preparation), the first participant and the second participant each clean the user features required for training. In one embodiment, the first participant is a payment platform, which can clean the user's basic information and online and offline payment information. The user's basic information may include user gender, age, city where the user is located, etc., and the online and offline payment information may include the user's consumption frequency, payment success rate, etc. The second participant is an offline merchant, which can clean the payment behavior information of the user in its offline store and training labels. The payment behavior information of the user in the offline store includes the user's consumption frequency, consumption preferences, etc., and the training labels may include joint prediction results. Note that data in the same time period should be selected to ensure that the samples can be aligned. Thus, the user features cleaned by different participants outline a more complete user portrait, thereby solving the data island problem. In addition, the labels of the model are no longer simple binary classifications, but multi-classifications, thereby improving the applicability of model prediction.

[0041] After preparing the data, the respective raw data from the first participant and the second participant are obtained. Since the user groups at the first participant and the second participant are not completely overlapping, in the process of the second stage (i.e., encrypted sample alignment), the system can use an encryption-based user sample alignment algorithm to confirm the common user group between the first participant and the second participant without the two parties disclosing their respective data, and without exposing the non-overlapping users, and combine the features of these users for modeling. The sample alignment algorithm can select the RSA algorithm. The process of user sample alignment may include: the collaborator generates a public key and distributes it to the first participant and the second participant, encrypts the user samples, the encrypted user samples are sent to the collaborator, and then the collaborator decrypts the encrypted user samples using the private key, and extracts the common user group of the first participant and the second participant, as described in further detail below Figure 3 As Figure 3 shown, to ensure the privacy of the data, a collaborator (e.g., an enclave platform) can be selected to control the distribution of the public key and the private key. The third-party collaborator can use the RSA algorithm to generate a pair of keys (including a public key and a private key), and the specific description of the RSA algorithm is as follows: (1) Randomly find two large prime numbers P and Q, and calculate n = P * Q; (2) Calculate the Euler's totient function m = $(n) of n; (3) Randomly select a positive integer e such that 1 < e < m and e is relatively prime to m d; (4) According to the extended Euclidean algorithm, find d such that the remainder of e * d / m is 1; (5) The public key is (n, e), and the private key is (n, d); and (6) The public key is used for encryption, where the ciphertext is the remainder of the e-th power of the plaintext divided by n; the private key is used for decryption, where the decrypted plaintext is the remainder of the d-th power of the ciphertext divided by n. In Figure 3In the example, the first participant and the second participant each include an authentication module, user data and an encryption module, wherein the authentication module can be used to receive the public key sent by the collaborator and send it to the encryption module to encrypt the user data, and then the encrypted data is sent to the collaborator. The collaborator may include a decryption module, wherein the decryption module can be used to decrypt the encrypted data uploaded by the first participant and the second participant based on the private key. After decryption, the shared user group can be taken out and returned to the first participant and the second participant as a training data set.

[0042] After the training data set is determined, encrypted model training can be performed in the third stage. Figure 4 The encryption model training process is further detailed in . Figure 4 The specific steps include:

[0043] S1: A reliable third-party collaborator creates a public key-private key pair (for example, using the RSA algorithm) and distributes the public key to the first participant and the second participant to encrypt the data that needs to be exchanged during the training process.

[0044] S2: The first participant and the second participant respectively initialize their respective model parameters (i.e., the first model parameters of the first participant's model and the second model parameters of the second participant's model), and respectively encrypt and exchange the intermediate results required for gradient and loss calculation.

[0045] S3: The first participant and the second participant each calculate the encrypted gradient and send it to the third-party collaborator after adding a mask. In addition, the second participant calculates the encrypted loss and sends it to the third-party collaborator.

[0046] S4: The third-party collaborator transmits the decrypted gradient back to the first participant and the second participant respectively to update the first model parameters at the first participant and the second model parameters at the second participant, and iterates the above steps until the loss converges.

[0047] The above training process is described in detail below using linear regression and homomorphic encryption technology as examples, but those skilled in the art will appreciate that the present invention is also applicable to other machine learning methods and encryption technologies.

[0048] Specifically, firstly, given the learning rate η, the regularization parameter λ, and the training data set at the payment platform Training dataset at offline merchants The model parameters θ associated with the model at the first participant and the second participant, respectively A , Θ B , the training label y at the second participant i , and the random mask R at the first and second participants A , R B.

[0049] The objective function can be expressed as:

[0050]

[0051] set up

[0052] Then the objective function after homomorphic encryption can be expressed as:

[0053]

[0054] Additive homomorphic encryption can be denoted as [[·]].

[0055] set up The gradient can be expressed as:

[0056]

[0057]

[0058] It follows that the first participant can calculate and sends it to the second participant. The second participant can calculate and [[L]], and Then, the first and second participants each calculate the encrypted gradient and add the mask to obtain and And will and The third-party collaborator then decrypts L, and And will Send to the first participant Sent to the second participant, the first participant and the second participant update the parameters Θ of their respective models according to the decrypted gradient A , Θ B . Then, iterate the above steps until the loss function converges and the model training is completed.

[0059] In the inference phase, the collaborator may send a user request with a user identifier to the first participant and the second participant, and the first participant and the second participant may read the relevant user features based on the received user identifier to respectively calculate and And then the calculated and Sent to collaborators, who then compute And decrypt it to get the predicted value y (e.g., 0-2).

[0060] Thus, after the first participant model at the first participant and the second participant model at the second participant are trained, the joint prediction model combining the first participant model and the second participant model can be deployed on the model platform. In one example, when a user receives a red envelope code, the model platform reads user features at different participants, performs joint predictions through the deployed models, and then performs corresponding joint push based on the joint prediction results.

[0061] As mentioned above, during the process of encrypted sample alignment and encrypted model training, the data of the first participant and the second participant are saved locally, and the data interaction during training will not lead to data privacy leakage. In order to further prevent the collaborator from learning information from the first participant and the second participant, the first participant and the second participant can further hide their gradients from the collaborator by adding encrypted random masks. Thus, through the encryption technology of shared learning, it is ensured that the first participant and the second participant will not be at risk of privacy leakage. In addition, without leaking user privacy, the user features of the first participant and the second participant can be combined to more completely outline the user portrait, so that the prediction results are more accurate and the waste of equity is avoided.

[0062] Figure 5 It is a schematic flow chart of a push method 500 based on joint prediction according to an embodiment of the present disclosure. The method 500 starts at step 501, and the data acquisition module 101 can obtain the first data of the requesting party at the first participant and the second data at the second participant, wherein the first data at the first participant is different from the second data at the second participant, and the first data and the second data are not interoperable. In one embodiment, the data acquisition module 101 can obtain the credit data, invoice billing data, etc. of the requesting party at different data owners for subsequent risk prediction. In another embodiment, the data acquisition module 101 can obtain the payment data of the requesting party at different data owners for subsequent payment prediction.

[0063] In step 502, the joint prediction module 102 may perform a joint prediction based on the first data and the second data obtained by the shared learning combination to perform a joint prediction related push to the requesting party. Further, the joint prediction module 102 may input the first data and the second data obtained into a joint prediction model to perform a joint prediction related push to the requesting party, wherein the joint prediction model is obtained by combining the first participant model at the first participant and the second participant model at the second participant, and the first participant model and the second participant model are obtained by encrypted training based on the shared learning combination of the first data and the second data via a third-party collaborator. For example, the training process of the first participant model and the second participant model may include encrypted sample alignment and encrypted model training. When aligning the encrypted samples, the RSA algorithm may be used to align user samples at the collaborator to extract the common user group, including: the first participant and the second participant use the public key sent by the collaborator to encrypt the user sample, and send the encrypted user sample to the third-party collaborator, and then the third-party collaborator uses the private key to decrypt the encrypted user sample, and extract the common user group and return it to the first participant and the second participant. The following steps may be included during encryption model training: the collaborator sends a public key to the first participant and the second participant, the first participant and the second participant calculate and exchange encrypted intermediate results, the first participant and the second participant calculate encrypted gradients and losses based on the intermediate results and send them to the collaborator, and the collaborator then decrypts the gradients and sends them back to the first participant model and the second participant model to update the gradients. In one embodiment, the first participant model and the second participant model are trained based on homomorphic encryption and linear regression.

[0064] In step 503, the result feedback module 103 may feedback the result of the joint prediction to the first participant and the second participant. In one embodiment, the result of the joint prediction may be multi-classified. In one embodiment, the result feedback module 103 may feedback the risk prediction result of the joint prediction to different data owners for intelligent risk control.

[0065] Therefore, on the one hand, different data owners jointly push together, and it is no longer possible for a user to receive multiple unnecessary pushes at the same time, thus reducing unnecessary pushes. On the other hand, this method uses the data of different participants at the same time. Since different participants have data based on different user characteristics, it solves the problem of data islands, and the characterization of users is more complete, making the prediction results more accurate, so as to better formulate push strategies for different joint prediction scenarios. On the other hand, this method ensures that there is no risk of privacy leakage among multiple different participants through shared learning encryption technology.

[0066] Figure 6FIG. 6 is a schematic diagram of a push system 600 based on joint prediction according to an embodiment of the present disclosure. Figure 6 As shown in , the system 600 may include a memory 601 and at least one processor 602. The memory 601 may store a trained joint prediction model. The memory 601 may include RAM, ROM, or a combination thereof. The memory 601 may store computer executable instructions, which, when executed by at least one processor 602, cause the at least one processor to perform various functions described herein, including: obtaining the first data of the requesting party at the first participant and the second data at the second participant, wherein the first data at the first participant is different from the second data at the second participant, and the first data and the second data are not interoperable; performing a joint prediction based on the first data and the second data obtained by the shared learning union to perform a joint prediction related push to the requesting party; and feeding back the result of the joint prediction to the first participant and the second participant. In some cases, the memory 601 may particularly include a BIOS, which may control basic hardware or software operations, such as interaction with peripheral components or devices. The processor 602 may include an intelligent hardware device (e.g., a general-purpose processor, a DSP, a CPU, a microcontroller, an ASIC, an FPGA, a programmable logic device, a discrete gate or transistor logic component, a discrete hardware component, or any combination thereof).

[0067] The various illustrative blocks and modules described in conjunction with the disclosure herein may be implemented or executed with a general purpose processor, a DSP, an ASIC, an FPGA or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general purpose processor may be a microprocessor, but in an alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices (e.g., a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration).

[0068] The functions described herein may be implemented in hardware, software executed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, the functions may be stored on or transmitted by a computer-readable medium as one or more instructions or codes. Other examples and implementations fall within the scope of the present disclosure and the appended claims. For example, due to the nature of software, the functions described herein may be implemented using software executed by a processor, hardware, firmware, hardwiring, or any combination thereof. Features that implement the functions may also be physically located in various locations, including being distributed so that parts of the functions are implemented at different physical locations.

[0069] What has been described above includes examples of various aspects of the claimed subject matter. Of course, it is not possible to describe every conceivable combination of components or methods for the purpose of describing the claimed subject matter, but one of ordinary skill in the art will recognize that many further combinations and permutations of the claimed subject matter are possible. Thus, the disclosed subject matter is intended to encompass all such alterations, modifications, and variations that fall within the spirit and scope of the appended claims.

Claims

1. A push method based on joint prediction, the method comprising: Acquire first payment-related data of the requesting party at a first participant and second payment-related data at a second participant, wherein the first data at the first participant is different from the second data at the second participant, and the first data and the second data are not interoperable; Inputting the acquired first data and second data into a joint prediction model to push the joint prediction related information to the requesting party, wherein the joint prediction model is obtained by combining a first participant model at the first participant and a second participant model at the second participant, and the first participant model and the second participant model are obtained by combining the first data and the second data through encrypted training by a third party collaborator based on shared learning; as well as Feedback the result of the joint prediction to the first participant and the second participant.

2. The method according to claim 1, wherein the first participant model and the second participant model are trained by the following operations, wherein the training data set includes first data of a common population at the first participant in the same time period, second data at the second participant, and training labels about the joint prediction result: Initializing first model parameters of the first participant model and second model parameters of the second participant model; The first participant and the second participant each calculate a corresponding first intermediate result and a second intermediate result and perform encryption interaction using a public key sent by the third-party collaborator; The first participant and the second participant each calculate an encrypted gradient and send it to the third-party collaborator after adding a mask; The second participant calculates the encrypted loss and sends it to the third party collaborator; as well as The third-party collaborator transmits the decrypted gradient back to the first participant and the second participant respectively to update the first model parameters and the second model parameters until the loss converges.

3. In the method as claimed in claim 2, the first participant model and the second participant model are trained based on homomorphic encryption and linear regression.

4. The method according to claim 2, wherein the common population is obtained by using an encryption-based sample alignment algorithm.

5. The method of claim 4, wherein the sample alignment algorithm comprises an RSA algorithm, wherein: The first participant and the second participant encrypt the sample data using the public key sent by the third-party collaborator, and send the encrypted sample data to the third-party collaborator; as well as The third-party collaborator decrypts the encrypted sample data using a private key, and takes out a common group and returns it to the first participant and the second participant. The method according to claim 1 , wherein the result of the joint prediction is multi-classification.

7. A push system based on joint prediction, the system comprising: A data collection module, which obtains first data related to payment of the requesting party at a first participant and second data related to payment at a second participant, wherein the first data at the first participant is different from the second data at the second participant, and the first data and the second data are not interoperable; a joint prediction module, inputting the acquired first data and second data into a joint prediction model to push the joint prediction related information to the requesting party, wherein the joint prediction model is obtained by combining a first participant model at the first participant and a second participant model at the second participant, and the first participant model and the second participant model are obtained by combining the first data and the second data through encrypted training by a third party collaborator based on shared learning; as well as A result feedback module is used to feed back the result of the joint prediction to the first participant and the second participant.

8. The system of claim 7, wherein the first participant model and the second participant model are trained by the following operations, wherein the training data set includes first data of a common population at the first participant in the same time period, second data at the second participant, and training labels about the joint prediction result: Initializing first model parameters of the first participant model and second model parameters of the second participant model; The first participant and the second participant each calculate a corresponding first intermediate result and a second intermediate result and perform encryption interaction using a public key sent by the third-party collaborator; The first participant and the second participant each calculate an encrypted gradient and send it to the third-party collaborator after adding a mask; The second participant calculates the encrypted loss and sends it to the third party collaborator; as well as The third-party collaborator transmits the decrypted gradient back to the first participant and the second participant respectively to update the first model parameters and the second model parameters until the loss converges.

9. In the system as claimed in claim 8, the first participant model and the second participant model are trained based on homomorphic encryption and linear regression.

10. The system of claim 8, wherein the common population is obtained by using a sample alignment algorithm based on encryption.

11. The system of claim 10, wherein the sample alignment algorithm comprises an RSA algorithm, wherein: The first participant and the second participant encrypt the sample data using the public key sent by the third-party collaborator, and send the encrypted sample data to the third-party collaborator; as well as The third-party collaborator decrypts the encrypted sample data using a private key, and takes out a common group and returns it to the first participant and the second participant.

12. The system of claim 7, wherein the result of the joint prediction is multi-classification.

13. A computer-readable storage medium storing instructions, which, when executed, causes a machine to perform the method according to any one of claims 1 to 6.

Images