Method and apparatus for joint prediction based on target model

Abstract

This specification provides a method and apparatus for joint prediction based on a target model. The target model includes a first sub-model based on a factorization machine (FM). The joint prediction involves a first party and a second party. The first party holds several first features of the target object, and the second party holds other features of the target object. In this method, the first party can obtain the k-dimensional parameters for each first feature in the second-order combination parameters of the FM model. Then, it locally calculates a first intermediate term and a second intermediate term. The first intermediate term includes k intermediate values, where any j-th intermediate value is the sum of the products of the j-th dimension parameters corresponding to each first feature. The second intermediate term includes the sum of the squares of the products of each first feature and the values ​​of each dimension parameter. The first party can then send the first intermediate result obtained based on the first and second intermediate terms to the target party, allowing them to be fused to obtain the output of the FM model.

Description

Technical Field

[0001] This specification relates to one or more embodiments in the field of machine learning, and more particularly to a method and apparatus for joint prediction based on a target model while protecting the privacy of all parties. Background Technology

[0002] In the context of internet big data, various platforms have accumulated massive amounts of data. For example, electronic payment platforms possess merchants' transaction data, e-commerce platforms store merchants' sales data, and banking institutions possess merchants' loan data. To increase the value of this data, there is a need for multiple platforms to collaborate on data processing. For instance, the aforementioned electronic payment platforms, e-commerce platforms, and banking institutions intend to jointly conduct machine learning to train a merchant classification model.

[0003] However, the data stored by various parties often includes users' private information. Currently, the industry is paying increasing attention to data security and personal privacy, and China has recently introduced a number of data protection-related policies and regulations. Therefore, protecting the privacy of all parties involved in collaborative data processing has become a key concern.

[0004] In scenarios involving joint model training and joint prediction, depending on the different model structures and algorithms, appropriate methods are needed to protect the privacy of the data of each participating component. In recent years, models utilizing the Factorization Machine (FM) mechanism have achieved excellent results in many scenarios. For such models, there is a desire for improved solutions that can protect data security and enhance performance during joint prediction. Summary of the Invention

[0005] This specification describes one or more embodiments of a method and apparatus for joint prediction based on a target model, which can perform accurate joint prediction for a target model based on an FM model while protecting the data privacy and security of all parties.

[0006] According to a first aspect, a method for joint prediction based on a target model is provided, the target model including a first sub-model based on a factorization machine, the joint prediction involving a first party and a second party, the first party holding several first features of the target object, and the second party holding other features of the target object, the method being executed by the first party, including:

[0007] Obtain the k-dimensional parameters for each first feature in the second-order combination parameters of the first sub-model;

[0008] The first intermediate term and the second intermediate term are computed locally. The first intermediate term includes k intermediate values. Any j-th intermediate value is the sum of the products of the feature value of each first feature and the j-th dimension parameter of that feature. The second intermediate term includes the sum of the squares of the products of each first feature and the value of each dimension parameter.

[0009] The first intermediate result obtained based on the first intermediate term and the second intermediate term is sent to the target party, so that the first intermediate result is fused with the second intermediate result from the second party to obtain the output of the first sub-model.

[0010] According to one embodiment, the method further includes obtaining a first linear parameter for each first feature and a first offset parameter for the first side in the linear combination parameters of the first sub-model; wherein the second intermediate term further includes the first offset parameter and the linear combination result of the feature value of each first feature and its corresponding first linear parameter.

[0011] In one embodiment, the first intermediate result includes the first intermediate item and the second intermediate item; sending the first intermediate result to the target party specifically means sending the first intermediate result to the second party.

[0012] In another embodiment, sending the first intermediate result to the target party specifically includes: adding noise data that satisfies differential privacy based on the first intermediate item and the second intermediate item, as the first intermediate result; and sending the first intermediate result to the second party.

[0013] In another embodiment, sending the first intermediate result to the target party specifically includes: secretly sharing a fragment of the first intermediate item and the second intermediate item, and using the fragment allocated to other parties as the first intermediate result; sending the first intermediate result to the second party, thereby performing a secure multi-party computation (mpc) with the second party under secret sharing.

[0014] According to one implementation, sending a first intermediate result obtained based on the first intermediate item and the second intermediate item to a target party includes: sending the first intermediate result to a third party, wherein the third party also receives the second intermediate result from the second party.

[0015] In one implementation, the target model further includes a second sub-model, and the method further includes: performing target processing on the first feature corresponding to the second sub-model to obtain a first sub-result; and sending the first sub-result to the target party.

[0016] According to a second aspect, an apparatus for joint prediction based on a target model is provided, the target model including a first sub-model based on a factorization machine, the joint prediction involving a first party and a second party, the first party holding several first features of a target object, and the second party holding other features of the target object, the apparatus being deployed in the first party, comprising:

[0017] The acquisition unit is configured to acquire the k-dimensional parameters of each first feature in the second-order combination parameters of the first sub-model;

[0018] The computing unit is configured to locally compute a first intermediate term and a second intermediate term, wherein the first intermediate term includes k intermediate values, and any j-th intermediate value is the sum of the products of the feature value of each first feature and the j-th dimension parameter of that feature; the second intermediate term includes the sum of the squares of the products of each first feature and the value of each dimension parameter.

[0019] The sending unit is configured to send a first intermediate result obtained based on the first intermediate item and the second intermediate item to the target party, so that the target party can fuse the first intermediate result with the second intermediate result from the second party to obtain the output of the first sub-model.

[0020] According to a third aspect, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method described in the first aspect.

[0021] According to a fourth aspect, a computing device is provided, including a memory and a processor, characterized in that the memory stores executable code, and when the processor executes the executable code, it implements the method of the first aspect.

[0022] In the embodiments of this specification, a scheme for joint prediction of a target model based on the FM model is proposed. In this scheme, each participating party obtains the k-dimensional parameters of its own features from the second-order combination parameters of the FM model, and then locally calculates a first intermediate term and a second intermediate term. Both the first and second intermediate terms are further operations on the product of the participating party's feature value and the k-dimensional parameters. The intermediate results obtained based on the first and second intermediate terms can then be sent to the target party, allowing it to fuse the first intermediate result with the intermediate result from the other party to obtain the output of the FM model. This approach enables accurate joint prediction while protecting the data privacy and security of all parties, thus solving the data security problem caused by feature cross-operations during the FM model processing. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 A schematic diagram of the structure of an FM model according to one embodiment is shown;

[0025] Figure 2 This diagram illustrates the structure of a DeepFM model according to one embodiment.

[0026] Figure 3 A flowchart illustrating a method for joint prediction based on a target model according to one embodiment is shown.

[0027] Figure 4 This diagram illustrates the dimensional changes during the DeepFM model computation process.

[0028] Figure 5a and Figure 5b Two interaction scenarios are shown respectively;

[0029] Figure 6 A schematic diagram of the structure of a joint prediction device according to one embodiment is shown. Detailed Implementation

[0030] The solution provided in this specification will now be described with reference to the accompanying drawings.

[0031] In various technical scenarios, the Factorization Machine (FM) model has achieved excellent results in multiple prediction tasks due to its consideration of the correlation between features. The FM model is based on a second-order polynomial regression model, which adds second-order combinations of features and pairwise combination terms to the conventional linear regression. The calculation method of the second-order polynomial regression model is shown in the following formula (1):

[0032]

[0033] Where n represents the number of feature terms in the sample, x i The eigenvalues ​​of the i-th feature; w0, w i ,w ij These are the model parameters. Therefore, the total number of parameters for the combined features is n. 2 indivual.

[0034] Based on this, the FM model is obtained by vectorizing the feature relationships. Specifically, the binomial parameter w in formula (1) can be... ijThis forms a symmetric matrix W. According to the Cholesky decomposition, matrix W is decomposed as: W = VV T The j-th column of matrix V is the latent vector corresponding to the j-th feature. Thus, each parameter w can be... ij Decomposed into:

[0035] w ij = <v i ,v j > (2)

[0036] in, <v i ,v j > represents vector v i With v j The inner product can also be written as v i v j T .

[0037] According to formula (2), formula (1) can be rewritten as formula (3), which is the calculation formula for the FM model:

[0038]

[0039] (3) In the formula, v i This is the latent vector, or parameter vector, corresponding to the i-th feature. The dimension of the latent vector is k, where k < 0. <n。

[0040] Figure 1 A schematic diagram of the structure of an FM model according to one embodiment is shown. Figure 1 As shown, multiple feature terms of a sample can be encoded into vector form (e.g., through one-hot encoding), which generally yields sparse encoded vectors. Then, on the one hand, a first-order linear combination operation can be performed on each feature term, i.e., the first two terms in formula (3) are calculated. On the other hand, in the latent vector space, i.e., the aforementioned k-dimensional space, a second-order combination operation is performed on each feature, i.e., the last term in formula (3) is calculated. Next, the results of the first-order operation and the second-order operation are summed to obtain the model's calculation result y. Subsequently, this result y can be directly output as the model's prediction result, or further processing such as classification can be performed based on this result y.

[0041] The FM model can be used alone or combined with other models to form a comprehensive model. For example, combining the FM model with a deep neural network (DNN) can form the DeepFM model.

[0042] Figure 2 A schematic diagram of the structure of a DeepFM model according to one embodiment is shown. (Comparison) Figure 1 It can be seen that the DeepFM model includes Figure 1The FM model shown is one sub-model, and it also includes a deep neural network (DNN) as another sub-model. The encoded vectors of each feature term of the sample are concatenated and input into the DNN for processing. The result is then fused with the calculation result of the FM model for classification or other predictions of the samples. The DeepFM model described above combines the advantages of the FM model and the deep neural network model, which can simultaneously improve low-dimensional and high-dimensional features, and has excellent comprehensive predictive performance.

[0043] Depending on the characteristics of the scene and the samples, the FM model can also be combined with other models to obtain other comprehensive models based on the FM model. For example, when the samples contain some spatiotemporal comprehensive features, convolutional neural networks (CNNs) and attention-based neural networks can be combined with the FM model, without any restrictions.

[0044] For target models incorporating the FM model, the processing requires second-order feature cross operations. This introduces additional challenges when multiple parties collaborate on the model. Specifically, in typical vertical data distributions, multiple participants, such as Alice and Bob, each possess different feature subsets of the same sample. The second-order feature cross operations in the FM model require combining the feature data from both parties. However, the sample feature data often involves user privacy and is confidential data that the participants must keep secret, preventing direct computation. Therefore, the FM model-based model processing, especially for joint operations involving multiple participants in a vertical data distribution, presents data security challenges and difficulties.

[0045] To address the aforementioned issues, an embodiment of this specification proposes a solution that enables accurate joint prediction of a target model based on the FM model while protecting the data privacy and security of all parties.

[0046] The principle analysis of the embodiments in this specification is described below.

[0047] To more clearly explore the operational characteristics of the second-order feature interaction term, the last term in formula (3), namely the second-order feature interaction term, can be simplified as follows:

[0048]

[0049] The second step involves the vector inner product. <v i ,v j Expand into the form of the sum of vector element-wise products, where v i,f Represents a k-dimensional vector v i The f-th dimension element in.

[0050] Through the above simplification, the calculation formula of the FM model can be transformed into the following formula (5):

[0051]

[0052] As can be seen, v is eliminated in the calculation method of formula (5). j .

[0053] When two participants jointly predict, each participant, A, possesses a subset of features. For example, participant A has 'a' features, and participant B has 'b' features. These two subsets together constitute the aforementioned 'n' features, i.e., n = a + b. In this case, formula (5) can be further transformed into formula (6):

[0054]

[0055] Therefore, in the second-order feature cross-combination calculation of the FM model, the feature calculation part of participant A can be obtained through the above formula (6). Feature calculation part with participant B Separate it.

[0056] Based on the above analysis, a scheme for joint model prediction by two parties is proposed. The specific implementation of this scheme is described in detail below.

[0057] Figure 3 A flowchart illustrating a method for joint prediction based on a target model according to one embodiment is shown. It is understood that this method can be executed by any device, apparatus, platform, or cluster of devices with computing and processing capabilities. As previously mentioned, the target model includes a first sub-model based on a factorization machine (FM). Furthermore, this joint prediction method involves two parties, hereinafter referred to as Party A and Party B. Each party holds a subset of features of the target object. Specifically, Party A holds several (denoted as 'a') first features of the target object, and Party B holds at least one (denoted as 'b') second feature of the target object. Typically, the features of the target object belong to the privacy data of each party.

[0058] In one specific embodiment, the target object is a user. The first party A may be, for example, an electronic payment platform that holds the user's payment-related characteristics, such as total historical payment amount, the amount of the most recent payment, the maximum payment amount, etc. The second party B may be, for example, a banking institution that holds the user's credit-related characteristics, such as credit rating, number of credit cards, overdraft limit, etc.

[0059] In other embodiments, the target object described above can also be other analysis objects, such as goods or items, events on the Internet, content to be recommended, etc. Furthermore, the target object can be a single object or a group of multiple objects.

[0060] Figure 3 The illustrated method flow is executed by either of the two parties; the following description uses the first party, A, as the executing entity. To perform model prediction in conjunction with the second party, such as... Figure 3 As shown, in step 31, the k-dimensional parameters for each first feature in the second-order combination parameters of the first sub-model (FM model) are obtained.

[0061] As mentioned earlier, the FM model calculates the predicted value y corresponding to the sample object using formula (3) or formula (5), where w0, w i , and each v i (where i ranges from 1 to n) are model parameters. The parameter values ​​of these parameters can be determined through model training. Therefore, in the prediction stage, the already determined parameter values ​​can be obtained directly. In step 31 above, the first party A needs to obtain the parameter values ​​of each first feature in the FM model. These parameter values ​​include the k-dimensional parameters of each first feature in the second-order combination parameters, that is, V in formula (6). i , where i ranges from 1 to a, to represent the first feature for term a. Of course, the first party A also needs to obtain the first-order parameters for the first feature, including the linear parameters and offset parameters for the first feature, which are w in formula (6). a0 ,w i , where i ranges from 1 to a.

[0062] Then, in step 33, the first party calculates the first intermediate term and the second intermediate term locally, wherein the first intermediate term includes k intermediate values, and any j-th intermediate value is the sum of the product of the feature value of each first feature and the j-th dimension parameter of that feature, and the second intermediate term includes the sum of the squares of the products of each first feature and its corresponding dimension parameter values.

[0063] As an example, the first intermediate term M1 above can be represented as:

[0064]

[0065] Where, x i V is the eigenvalue of the first feature of the i-th term. i,f It is the k-dimensional parameter vector V corresponding to the first feature of the i-th term. i The value of the f-th dimension parameter in the equation. Therefore, M1 includes the k sum values ​​obtained when f takes k dimension values ​​from 1 to k.

[0066] As an example, the second intermediate term M2 above can be represented as:

[0067]

[0068] According to formula (8), the second intermediate term M2 is the sum of the squares of the first feature of each term and the corresponding parameter values ​​of each dimension.

[0069] In one embodiment, the first party also locally computes a third intermediate term M3, which involves offset parameters for the first features and linear combinations of the eigenvalues ​​of each first feature and their corresponding linear parameters. Specifically, in one example, the third intermediate term can be represented as:

[0070]

[0071] In another embodiment, the first term, based on the combination coefficients, combines the first-order linear result of formula (9) with the sum of squares of formula (8) as the second intermediate term M2. In this embodiment, the second intermediate term can be expressed as:

[0072]

[0073] Therefore, in step 35, the first party obtains a first intermediate result based on the first intermediate term M1 and the second intermediate term M2, and sends the first intermediate result to the target party so that it can merge the first intermediate result with the second intermediate result from the second party to obtain the output of the FM model.

[0074] It is understandable that when the second intermediate term adopts the form of formula (9), the first intermediate result is determined based on the first and second intermediate terms; when the second intermediate term adopts the form of formula (8), the first intermediate result is determined based on the first, second, and third intermediate terms. The following explanation uses the second intermediate term in the form of formula (9) as an example.

[0075] Figure 4 This diagram illustrates the dimensional changes during the DeepFM model computation process. Figure 4 In this context, 'b' represents the batch size, i.e., the number of samples in a batch; 'm' represents the number of feature terms in each sample; and 'k' is the dimension of the latent vector (or parameter vector) during second-order combination. Thus, during the second-order operation, the data processed has a dimension of b*m*k, and the result obtained after the second-order operation has a dimension of b*k*1. This result is combined with the result of the first-order operation (with a dimension of b*1) to form the result of the FM part.

[0076] refer to Figure 4It can be seen that the dimension of the second-order part of the first term A is batchsize*a*k, where batchsize is the number of target objects. After the above calculation, it can be concluded that the dimension of the first intermediate term is batchsize*k; when the second intermediate term adopts the form of formula (9), the dimension of its result is batchsize*1. When the target object is a single object, the dimension of the first intermediate term is k (i.e., including k intermediate values), and the dimension of the second intermediate term is 1.

[0077] Therefore, the first and second intermediate terms obtained in this way not only obscure the information of the eigenvalues ​​but also the information of the eigendimensionality. Within a reasonable timeframe, it is difficult to deduce the individual eigenvalues ​​x from the first and second intermediate terms. i Therefore, directly sending the first and second intermediate items as the first intermediate result to other parties will not leak private data.

[0078] However, in order to further enhance the strength of privacy protection, in some implementations, the first intermediate term and the second intermediate term are subjected to privacy enhancement processing and used as the first intermediate result.

[0079] Specifically, in one embodiment, noise data satisfying differential privacy is added to the first intermediate term M1 and the second intermediate term M2 as the first intermediate result. In one example, noise is added to the first intermediate term M1 to obtain the first noisy intermediate term M′1, i.e.:

[0080]

[0081] In formula (11), Let the Laplace random noise be centered at 0, and its distribution parameters be... Where Δf1 represents the sensitivity of the first intermediate term M1, which can be defined as the difference between the possible maximum and minimum values, and ε1 is the privacy budget.

[0082] Adding noise to the second intermediate term M2 yields the second noisy intermediate term M′2, i.e.:

[0083]

[0084] In formula (12), Δf2 represents the sensitivity of the second intermediate term M2, and ε2 is the privacy budget. Formulas (11) and (12) above are described based on the Laplace mechanism. In other embodiments, noise can also be added based on the Gaussian mechanism.

[0085] With noise added, the first intermediate result includes a first noisy intermediate term M′1 and a second noisy intermediate term M′2.

[0086] In another embodiment, the first intermediate item and the second intermediate item are secretly shared in fragments, and each fragment is used as the first intermediate result.

[0087] Specifically, Party A can deploy a cryptographic engine containing several pre-defined multi-party secure computation (MPC) algorithm protocols, capable of performing secure computations such as secure multiplication and secure addition. In one example, MPC can be implemented based on secret sharing. In this case, Party A feeds a first intermediate item and a second intermediate item into the cryptographic engine, where secure sharding is performed, resulting in multiple shards of the first and second intermediate items, including shards held by Party A and shards allocated to other parties. Party A can then send the shards allocated to other parties as the first intermediate result to the target party. In other examples, the cryptographic engine can also implement MPC based on, for example, homomorphic encryption.

[0088] In this case, the first party can perform homomorphic encryption on the first intermediate item and the second intermediate item, and send them to the target party as the first intermediate result.

[0089] In different implementations, the target party that the first party A interacts with can be the second party B, or an intermediate party independent of the first and second parties, or a third party. Regardless of the implementation, the target party also obtains the second intermediate result obtained by the second party B in a similar manner as described above, and then fuses the first and second intermediate results. Specifically, the target party performs local fusion or MPC operation on the first and second intermediate results according to the aforementioned formula (6) to obtain the output of the FM model.

[0090] Figure 5a and Figure 5b Two interactive scenarios are shown respectively.

[0091] according to Figure 5a In step 51, Party A obtains the first parameters for the first feature in the FM model, and Party B obtains the second parameters for the second feature in the FM model. In step 53, Party A locally calculates the first and second intermediate terms for the first feature based on the first parameters, and obtains the first intermediate result accordingly. Correspondingly, Party B locally calculates the first and second intermediate terms for the second feature based on the second parameters, and obtains the second intermediate result accordingly. In step 55, Party A sends the first intermediate result to Party B. In step 57, Party B merges the first and second intermediate results.

[0092] exist Figure 5b In the middle, steps 51 and 53 are related to Figure 5aThe same applies. In step 55, party A sends the first intermediate result to the third party; party B sends the second intermediate result to the third party. In step 57, the third party merges the first and second intermediate results.

[0093] It can be understood that the fusion of the first and second intermediate results constitutes the output of the FM model. When the target model upon which the joint prediction by the two parties is based includes other model components, namely a second sub-model, the two parties also need to perform joint processing on the second sub-model. For the first party A, it needs to perform target processing on the first feature corresponding to the second sub-model to obtain the first sub-result, which is then sent to the aforementioned target party. The second party B similarly performs the target processing on the second feature to obtain the second sub-result. The fusion of the first and second sub-results is then further fused with the output of the FM model to obtain the final output of the target model. The specific processing procedure of the aforementioned target processing depends on the algorithmic characteristics of the second sub-model. This part is not the focus of this specification and will not be elaborated upon here.

[0094] As can be seen from the above process, in the embodiments of this specification, for the target model based on the FM model, accurate joint prediction can be performed while protecting the data privacy and security of all parties, thus solving the data security problem caused by feature cross-operation in the processing of the FM model.

[0095] According to another embodiment, an apparatus for joint prediction based on a target model is provided. The target model includes a first sub-model based on a factorization machine, and the joint prediction involves a first party and a second party, wherein the first party holds several first features of the target object, and the second party holds other features of the target object. Figure 6 The diagram illustrates a structural schematic of a joint prediction device according to one embodiment. This prediction device can be deployed in a first party, which can be any device, platform, or cluster of devices with data storage, computing, and processing capabilities. Figure 6 As shown, the prediction device 600 includes:

[0096] The acquisition unit 61 is configured to acquire the k-dimensional parameters of each first feature in the second-order combination parameters of the first sub-model;

[0097] The calculation unit 63 is configured to locally calculate the first intermediate term and the second intermediate term. The first intermediate term includes k intermediate values, and any j-th intermediate value is the sum of the products of the feature value of each first feature and the j-th dimension parameter of that feature. The second intermediate term includes the sum of the squares of the products of each first feature and the value of each dimension parameter.

[0098] The sending unit 65 is configured to send a first intermediate result obtained based on the first intermediate item and the second intermediate item to the target party, so that the first intermediate result is fused with the second intermediate result from the second party to obtain the output of the first sub-model.

[0099] According to one embodiment, the acquisition unit 61 is further configured to acquire the first linear parameter for each first feature and the first offset parameter for the first side in the linear combination parameters of the first sub-model; wherein the second intermediate item further includes the first offset parameter and the linear combination result of the feature value of each first feature and its corresponding first linear parameter.

[0100] In one embodiment, the first intermediate result includes the first intermediate item and the second intermediate item; the sending unit 65 is configured to send the first intermediate result to the second party.

[0101] In another embodiment, the sending unit 65 is configured to: add noise data that satisfies differential privacy based on the first intermediate item and the second intermediate item, as the first intermediate result; and send the first intermediate result to the second party.

[0102] In another embodiment, the sending unit 65 is configured to: secretly share fragments of the first intermediate item and the second intermediate item, and use the fragments allocated to other parties as the first intermediate result; send the first intermediate result to the second party, thereby performing multi-party secure computation (mpc) with the second party under secret sharing.

[0103] According to one embodiment, the sending unit 65 is configured to send the first intermediate result to a third party, wherein the third party also receives the second intermediate result from the second party.

[0104] In one implementation, the target model further includes a second sub-model, and the device 600 further includes a second processing unit (not shown), configured to perform target processing on the first feature corresponding to the second sub-model to obtain a first sub-result; the sending unit 65 is further configured to send the first sub-result to the target party.

[0105] For specific examples of how each unit in the above device is implemented, please refer to the previous examples. Figure 3 The above device enables secure joint prediction with another party targeting a specific model.

[0106] According to another embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed in a computer, causes the computer to perform a combination Figure 3 The method described.

[0107] According to another embodiment, a computing device is also provided, including a memory and a processor, wherein executable code is stored in the memory, and when the processor executes the executable code, it implements a combination... Figure 3 The method described.

[0108] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this invention can be implemented using hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium.

[0109] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for joint prediction based on a target model, the target model comprising a first sub-model based on a factorization machine, the joint prediction involving a first party and a second party, the first party holding several first features of a target object, the second party holding second features of the target object, the method being executed by the first party, comprising: Obtain the k-dimensional parameters for each first feature in the second-order combination parameters of the first sub-model; The first intermediate term and the second intermediate term are computed locally. The first intermediate term includes k intermediate values. Any j-th intermediate value is the sum of the products of the feature value of each first feature and the j-th dimension parameter of that feature. The second intermediate term includes the sum of the squares of the products of each first feature and the value of each dimension parameter. The first intermediate result obtained based on the first intermediate term and the second intermediate term is sent to the target party, so that the first intermediate result is fused with the second intermediate result from the second party to obtain the output of the first sub-model; the second intermediate result is obtained based on the fourth intermediate term and the fifth intermediate term, the fourth intermediate term includes k intermediate values, any j-th intermediate value is the sum of the product of the feature value of each second feature and the j-th dimension parameter of that feature, the fifth intermediate term includes the sum of the squares of the products of each second feature and the parameter values ​​of each dimension.

2. The method of claim 1, further comprising obtaining first linear parameters for each first feature in the linear combination parameters of the first sub-model, and a first offset parameter for the first party; wherein, The second intermediate term also includes the first offset parameter, and the linear combination result of the eigenvalues ​​of each first feature and their corresponding first linear parameters.

3. The method of claim 1, wherein, The first intermediate result includes the first intermediate item and the second intermediate item; Sending the first intermediate result obtained based on the first intermediate item and the second intermediate item to the target party includes: The first intermediate result is sent to the second party.

4. The method of claim 1, wherein, Sending the first intermediate result obtained based on the first intermediate item and the second intermediate item to the target party includes: Noise data that satisfies differential privacy is added based on the first intermediate term and the second intermediate term, and this noise is used as the first intermediate result. The first intermediate result is sent to the second party.

5. The method of claim 1, wherein, Sending the first intermediate result obtained based on the first intermediate item and the second intermediate item to the target party includes: The first intermediate item and the second intermediate item are secretly shared in a shard, and the shards allocated to other parties are used as the first intermediate result; The first intermediate result is sent to the second party, thereby enabling secure multi-party computation (mpc) with secret sharing with the second party.

6. The method of claim 1, wherein, Sending the first intermediate result obtained based on the first intermediate item and the second intermediate item to the target party includes: The first intermediate result is sent to a third party, and the third party also receives the second intermediate result from the second party.

7. The method of claim 1, wherein, The target model further includes a second sub-model, and the method further includes: The first feature is subjected to target processing corresponding to the second sub-model to obtain the first sub-result; The first sub-result is sent to the target party.

8. An apparatus for joint prediction based on a target model, the target model including a first sub-model based on a factorization machine, the joint prediction involving a first party and a second party, the first party holding several first features of a target object, the second party holding second features of the target object, the apparatus being deployed in the first party, comprising: The acquisition unit is configured to acquire the k-dimensional parameters of each first feature in the second-order combination parameters of the first sub-model; The computing unit is configured to locally compute a first intermediate term and a second intermediate term, wherein the first intermediate term includes k intermediate values, and any j-th intermediate value is the sum of the products of the feature value of each first feature and the j-th dimension parameter of that feature; the second intermediate term includes the sum of the squares of the products of each first feature and the value of each dimension parameter. The sending unit is configured to send a first intermediate result obtained based on the first intermediate term and the second intermediate term to the target party, so that the first intermediate result is fused with the second intermediate result from the second party to obtain the output of the first sub-model; the second intermediate result is obtained based on the fourth intermediate term and the fifth intermediate term, the fourth intermediate term includes k intermediate values, any j-th intermediate value is the sum of the product of the feature value of each second feature and the j-th dimension parameter of that feature, and the fifth intermediate term includes the sum of the squares of the products of each second feature and the value of each dimension parameter.

9. The apparatus according to claim 8, wherein the acquiring unit is further configured to acquire, in the linear combination parameters of the first sub-model, a first linear parameter for each first feature, and a first offset parameter for the first side; wherein, The second intermediate term also includes the first offset parameter, and the linear combination result of the eigenvalues ​​of each first feature and their corresponding first linear parameters.

10. The apparatus according to claim 8, wherein, The first intermediate result includes the first intermediate item and the second intermediate item; The sending unit is configured to send the first intermediate result to the second party.

11. The apparatus according to claim 8, wherein, The sending unit is configured as follows: Noise data that satisfies differential privacy is added based on the first intermediate term and the second intermediate term, and this noise is used as the first intermediate result. The first intermediate result is sent to the second party.

12. The apparatus according to claim 8, wherein, The sending unit is configured as follows: The first intermediate item and the second intermediate item are secretly shared in a shard, and the shards allocated to other parties are used as the first intermediate result; The first intermediate result is sent to the second party, thereby enabling secure multi-party computation (mpc) with secret sharing with the second party.

13. The apparatus according to claim 8, wherein, The sending unit is configured as follows: The first intermediate result is sent to a third party, and the third party also receives the second intermediate result from the second party.

14. The apparatus according to claim 8, wherein, The target model further includes a second sub-model, and the device further includes a second processing unit configured to perform target processing on the first feature corresponding to the second sub-model to obtain a first sub-result; The sending unit is further configured to send the first sub-result to the target party.

15. A computing device, comprising a memory and a processor, characterized in that, The memory stores executable code, and when the processor executes the executable code, it implements the method of any one of claims 1-7.

Images