Longitudinal federal forecasting methods, systems, electronic devices, media, and program products

Abstract

The application discloses a longitudinal federal prediction method, system, electronic equipment, medium and program product, which is applied to a first participant device and includes the following steps: performing feature transformation on unaligned samples according to a first-party federal feature transformation model to obtain first sample transformed features; uploading the first sample transformed features to a federal server to generate sample fuzzy matching results; performing feature extraction on the unaligned samples based on a first-party federal prediction feature extraction model to obtain first-party prediction sample features; receiving second-party prediction sample features sent by a second participant device; and performing longitudinal federal prediction according to the first-party prediction sample features and the second-party prediction sample features to obtain a longitudinal federal prediction result. The application solves the technical problem that a longitudinal federal learning model constructed based on aligned samples cannot perform longitudinal federal learning prediction on unaligned samples, resulting in low sample prediction accuracy of the unaligned samples.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence technology in financial technology (Fintech), and particularly relates to a vertical federated prediction method and system, an electronic device, a medium and a program product. BACKGROUND

[0002] With the continuous development of financial technology, especially Internet technology finance, more and more technologies (such as distributed, artificial intelligence, etc.) are applied in the financial field, but the financial industry also puts forward higher requirements for technology, such as higher requirements for the distribution of the to-do list of the financial industry.

[0003] As a distributed machine learning method, federated learning can solve the "data island" problem, including vertical federated learning and horizontal federated learning. Among them, the application scenario of vertical federated learning is that the users of each participant overlap, and the features overlap less. Different participants have information of users in different fields / different angles. Commonly used scenarios such as personal lending business of financial institutions, users apply for loans from financial institutions, financial institutions have personal information provided by applicants and credit information of applicants provided by corresponding credit agencies. In order to more accurately assess the qualifications of the applicants, the financial institutions and other institutions (which can provide other aspects of information of the applicants) jointly model to evaluate the applicants, for example, other institutions are shopping platforms, and the shopping platforms can provide consumption record information of the applicants. The two parties participating in the vertical federated learning align the users through commonly agreed applicant identification information such as mobile phone numbers and hardware identification numbers, and perform user alignment under the condition of protecting privacy.

[0004] Therefore, the vertical federated learning model can perform vertical federated learning prediction on the aligned samples between the vertical federated learning participants, and improve the accuracy of sample prediction. The aligned samples are usually aligned based on the identification information of the applicants of each vertical federated learning participant. If the identification information of the applicants needs to be protected, the corresponding samples that need to be aligned through the identification information of the applicants cannot be aligned again for the new applicants or the applicants whose identification information changes. These samples become unaligned samples. For unaligned samples, the current vertical federated learning model constructed based on aligned samples cannot perform vertical federated learning prediction on unaligned samples, which affects the sample prediction accuracy of unaligned samples in the vertical federated learning scenario. SUMMARY

[0005] The main purpose of the present application is to provide a vertical federated prediction method, system, electronic device, medium and program product, which aims to solve the technical problem that the vertical federated learning model constructed based on aligned samples cannot perform vertical federated learning prediction on unaligned samples, resulting in low sample prediction accuracy of unaligned samples.

[0006] To achieve the above object, the application provides a longitudinal federated prediction method applied to a first participant device, which comprises the following steps:

[0007] According to a first-party federated feature transformation model, feature transformation is performed on the unaligned samples to obtain first sample transformed features;

[0008] The first sample transformed features are uploaded to a federated server, wherein the federated server is configured to generate a fuzzy matching result between the unaligned samples and each second-party sample according to the first sample transformed features and each second sample transformed feature uploaded by a second participant device, wherein the second sample transformed feature is obtained by performing feature transformation on a second-party sample according to a second-party federated feature transformation model by the second participant device;

[0009] Based on a first-party federated prediction feature extraction model, feature extraction is performed on the unaligned samples to obtain first-party prediction sample features;

[0010] Second-party prediction sample features sent by the second participant device are received, and longitudinal federated prediction is performed according to the first-party prediction sample features and the second-party prediction sample features to obtain a longitudinal federated prediction result, wherein the second-party prediction sample features are obtained by constructing a model prediction output of each second-party sample and the fuzzy matching result based on a second-party federated prediction feature extraction model, and the model prediction output is obtained by performing feature extraction on the second-party sample by a second-party federated prediction feature extraction model.

[0011] To achieve the above object, the application further provides a longitudinal federated prediction method applied to a second participant device, which comprises the following steps:

[0012] According to a second-party federated feature transformation model, feature transformation is performed on each second-party sample to obtain each second sample transformed feature;

[0013] Each second sample transformed feature is uploaded to a federated server, wherein the federated server is configured to generate a fuzzy matching result between the unaligned samples and each second-party sample according to each second sample transformed feature and first sample transformed features uploaded by a first participant device, wherein the first sample transformed features are obtained by performing feature transformation on the unaligned samples according to a first-party federated feature transformation model by the first participant device;

[0014] The fuzzy matching result fed back by the federated server is received, and a second-party prediction sample feature is constructed according to a model prediction output of each second-party sample and the fuzzy matching result based on a second-party federated prediction feature extraction model;

[0015] sending the second-party predicted sample features to the first-party device for the first-party device to perform longitudinal federated prediction on the first-party predicted sample features of the unaligned samples and the second-party predicted sample features based on a first-party federated prediction feature extraction model.

[0016] To achieve the above object, the application further provides a longitudinal federated prediction method applied to a federated server, comprising:

[0017] receiving first sample transformed features sent by a first-party device and each second sample transformed feature sent by a second-party device, wherein the first sample transformed features are obtained by the first-party device from feature transformation on unaligned samples according to a first federated feature transformation model, and each second sample transformed feature is obtained by the second-party device from feature transformation on a second-party sample according to a second federated feature transformation model;

[0018] generating a fuzzy matching result between the unaligned samples and each second-party sample according to the first sample transformed features and each second sample transformed feature;

[0019] downloading the fuzzy matching result to the second-party device, wherein the second-party device is configured to construct second-party predicted sample features from model prediction outputs of each second-party sample and the fuzzy matching result according to a second-party federated prediction feature extraction model, and the first-party device is configured to perform longitudinal federated prediction according to first-party predicted sample features and the second-party predicted sample features sent by the second-party device, wherein the first-party predicted sample features are obtained by feature extraction on the unaligned samples according to a first-party federated prediction feature extraction model.

[0020] The application further provides a longitudinal federated model construction optimization method applied to a first-party device, comprising:

[0021] extracting first-party training samples from a training sample set, wherein the training sample set is composed of unaligned samples and first-party aligned samples;

[0022] if the first-party training sample is an unaligned sample, performing feature transformation on the first-party training sample according to a first-party federated feature transformation model to obtain first sample transformed features;

[0023] upload the first sample transformed feature to a federal server, where the federal server is configured to generate a fuzzy matching result between the first-party training sample and each second-party sample according to the first sample transformed feature and each second sample transformed feature uploaded by a second-party device, the second sample transformed feature being obtained by performing feature transformation on a second-party sample by the second-party device according to a second-party federated feature transformation model;

[0024] perform feature extraction on the first-party training sample based on a first-party federated prediction feature extraction model to obtain a first-party prediction sample feature;

[0025] receive a second-party prediction sample feature sent by the second-party device, where the second-party prediction sample feature is constructed by the second-party device based on a model prediction output of each second-party sample and the fuzzy matching result, the model prediction output being obtained by performing feature extraction on the second-party sample by a second-party federated prediction feature extraction model;

[0026] perform longitudinal federated learning modeling according to the first-party prediction sample feature and the second-party prediction sample feature to iteratively update the first-party federated prediction feature extraction model and the second-party federated prediction feature extraction model.

[0027] The application also provides a longitudinal federated model construction optimization method applied to a second-party device, the longitudinal federated model construction optimization method comprising:

[0028] perform feature transformation on each second-party sample according to a second-party federated feature transformation model to obtain each second sample transformed feature;

[0029] upload each second sample transformed feature to a federal server, where the federal server is configured to generate a fuzzy matching result between each second-party sample and an unaligned sample according to each second sample transformed feature and a first sample transformed feature uploaded by a first-party device, the first sample transformed feature being obtained by performing feature transformation on the unaligned sample by the first-party device according to a first-party federated feature transformation model;

[0030] receive the fuzzy matching result fed back by the federal server, and construct a second-party prediction sample feature according to a model prediction output of each second-party sample by a second-party federated prediction feature extraction model and the fuzzy matching result;

[0031] send the second-party prediction sample feature to the first-party device, so that the first-party device performs longitudinal federated learning modeling on the first-party prediction sample feature of the unaligned sample and the second-party prediction sample feature based on a first-party federated prediction feature extraction model.

[0032] The application also provides a longitudinal federated model construction optimization method applied to a federated server, and the longitudinal federated model construction optimization method comprises the following steps of:

[0033] receiving first sample transformation features sent by the first participant device and each second sample transformation feature sent by the second participant device, wherein the first sample transformation features are obtained by performing feature transformation on the unaligned samples according to a first federated feature transformation model by the first participant device, and the second sample transformation features are obtained by performing feature transformation on second-party samples according to a second federated feature transformation model by the second participant device;

[0034] generating a fuzzy matching result between the unaligned samples and each second-party sample according to the first sample transformation features and each second sample transformation feature;

[0035] issuing the fuzzy matching result to the second participant device, wherein the second participant device is configured to construct second-party predicted sample features according to a model prediction output of each second-party sample and the fuzzy matching result according to a second-party federated prediction feature extraction model, the first participant device is configured to perform longitudinal federated learning modeling according to first-party predicted sample features and the second-party predicted sample features sent by the second participant device, and the first-party predicted sample features are obtained by performing feature extraction on the unaligned samples according to a first-party federated prediction feature extraction model.

[0036] The application also provides a longitudinal federated prediction system, and the longitudinal federated prediction system comprises:

[0037] The first participant device is configured to perform feature transformation on unaligned samples according to a first-party federated feature transformation model to obtain first sample transformation features, upload the first sample transformation features to a federated server, perform feature extraction on the unaligned samples according to a first-party federated prediction feature extraction model to obtain first-party predicted sample features, receive second-party predicted sample features sent by the second participant device, and perform longitudinal federated prediction according to the first-party predicted sample features and the second-party predicted sample features to obtain a longitudinal federated prediction result;

[0038] The second participant device is configured to perform feature transformation on each second-party sample according to a second-party federated feature transformation model to obtain each second sample transformation feature, upload each second sample transformation feature to the federated server, receive a fuzzy matching result fed back by the federated server, construct second-party predicted sample features according to a model prediction output of each second-party sample and the fuzzy matching result according to a second-party federated prediction feature extraction model, and send the second-party predicted sample features to the first participant device;

[0039] The federal server receives the first sample transformation feature sent by the first participant device and each second sample transformation feature sent by the second participant device, generates a fuzzy matching result between the unaligned sample and each second party sample according to the first sample transformation feature and each second sample transformation feature, and issues the fuzzy matching result to the second participant device.

[0040] The application also provides a longitudinal federal model construction optimization system, which comprises:

[0041] The first participant device extracts a first party training sample from a training sample set, wherein the training sample set is composed of unaligned samples and first party aligned samples; if the first party training sample is an unaligned sample, the first party training sample is subjected to feature transformation according to a first party federal feature transformation model to obtain a first sample transformation feature; the first sample transformation feature is uploaded to a federal server; the first party training sample is subjected to feature extraction based on a first party federal prediction feature extraction model to obtain a first party prediction sample feature; the second participant device sends a second party prediction sample feature; and longitudinal federal learning modeling is performed according to the first party prediction sample feature and the second party prediction sample feature to iteratively update the first party federal prediction feature extraction model and the second party federal prediction feature extraction model.

[0042] The second participant device subjects each second party sample to feature transformation according to a second party federal feature transformation model to obtain each second sample transformation feature; uploads each second sample transformation feature to a federal server; receives a fuzzy matching result fed back by the federal server, and constructs a second party prediction sample feature according to a model prediction output of each second party sample based on a second party federal prediction feature extraction model and the fuzzy matching result; and sends the second party prediction sample feature to the first participant device.

[0043] The federal server receives the first sample transformation feature sent by the first participant device and each second sample transformation feature sent by the second participant device, generates a fuzzy matching result between the unaligned sample and each second party sample according to the first sample transformation feature and each second sample transformation feature, and issues the fuzzy matching result to the second participant device.

[0044] The application also provides an electronic device, which comprises a memory, a processor, and a program of the longitudinal federal prediction method stored in the memory and executable on the processor, and the program of the longitudinal federal prediction method can implement the steps of the longitudinal federal prediction method when executed by the processor.

[0045] The application further provides an electronic device, comprising a memory, a processor, and a program of the longitudinal federal model construction optimization method stored in the memory and executable on the processor, which can realize the steps of the longitudinal federal model construction optimization method when executed by the processor.

[0046] The application further provides a computer readable storage medium, which stores a program of the longitudinal federal prediction method, which can realize the steps of the longitudinal federal prediction method when executed by a processor.

[0047] The application further provides a computer readable storage medium, which stores a program of the longitudinal federal model construction optimization method, which can realize the steps of the longitudinal federal model construction optimization method when executed by a processor.

[0048] The application further provides a computer program product, which comprises a computer program, which can realize the steps of the longitudinal federal prediction method when executed by a processor.

[0049] The application further provides a computer program product, which comprises a computer program, which can realize the steps of the longitudinal federal model construction optimization method when executed by a processor.

[0050] The application provides a longitudinal federated prediction method, system, electronic device, medium and program product. In a longitudinal federated learning scene, for unaligned samples in a first participant device, first, a first-party federated feature transformation model built based on longitudinal federated learning modeling based on aligned samples is used to perform feature transformation on the unaligned samples to obtain first sample transformed features, and then the first sample transformed features are uploaded to a federated server. The second participant device also uses a second-party federated feature transformation model built based on longitudinal federated learning modeling based on aligned samples to perform feature transformation on each second-party sample to obtain each second sample transformed feature, and the second participant device uploads each second sample transformed feature to the federated server. In this way, after receiving the first sample transformed features and each second sample transformed feature, the federated server can perform fuzzy matching based on the similarity between the first sample transformed features and each second sample transformed feature to obtain a fuzzy matching result. Thus, after receiving the fuzzy matching result, the second participant device can use the fuzzy matching result to select a corresponding model output result from the model prediction output of the second-party federated prediction feature extraction model for each second-party sample to construct a second-party prediction sample feature, thereby realizing the construction of the second-party prediction sample feature corresponding to the first-party prediction sample feature at the second participant device. In this way, after performing feature extraction on the unaligned samples based on the first-party federated prediction feature extraction model to obtain first-party prediction sample features, and receiving the second-party prediction sample features sent by the second participant device, the first participant device can perform longitudinal federated prediction using the first-party prediction sample features and the second-party prediction sample features. Therefore, the embodiments of the application realize that, in the first participant device with unaligned samples, the second participant device constructs the second-party prediction sample feature corresponding to the unaligned sample in conjunction with the trusted third party (federated server), thereby realizing longitudinal federated prediction for the unaligned samples in the first participant device. Therefore, the problem of low prediction accuracy of unaligned samples due to the fact that the longitudinal federated learning model constructed based on aligned samples cannot perform longitudinal federated learning prediction on unaligned samples is solved. BRIEF DESCRIPTION OF DRAWINGS

[0051] The accompanying drawings, which are incorporated herein and form part of the specification, illustrate embodiments consistent with the present application and, together with the description, further serve to explain the principles of the application.

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, for those skilled in the art, other drawings can also be obtained based on these drawings without creative labor.

[0053] Figure 1A flowchart of a first embodiment of the longitudinal federated prediction method of the present application;

[0054] Figure 2 A flowchart of a second embodiment of the longitudinal federated prediction method of the present application;

[0055] Figure 3 A flowchart of a third embodiment of the longitudinal federated prediction method of the present application;

[0056] Figure 4 A schematic diagram of an embodiment of the longitudinal federated learning framework in the present application;

[0057] Figure 5 A schematic diagram of another embodiment of the longitudinal federated learning framework in the present application;

[0058] Figure 6 A flowchart of a fourth embodiment of the longitudinal federated prediction method of the present application;

[0059] Figure 7 A flowchart of a fifth embodiment of the longitudinal federated prediction method of the present application;

[0060] Figure 8 A flowchart of a sixth embodiment of the longitudinal federated prediction method of the present application;

[0061] Figure 9 A schematic diagram of the device structure of the hardware operating environment involved in the longitudinal federated prediction method in the embodiments of the present application.

[0062] The purposes, functional features and advantages of the present application will be further described with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0063] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0064] Embodiment One

[0065] The embodiments of the present application provide a longitudinal federated prediction method. In a first embodiment of the longitudinal federated prediction method of the present application, with reference to Figure 1 The method of the present embodiment is applied to a first participant device, and the longitudinal federated prediction method comprises:

[0066] In step S10, the first-party federated feature transformation model is used to perform feature transformation on the unaligned sample to obtain first sample transformed features.

[0067] In step S20, the first sample transformed features are uploaded to a federated server. The federated server is configured to generate a fuzzy matching result between the unaligned sample and each second-party sample according to the first sample transformed features and each second sample transformed features uploaded by a second-party device. The second sample transformed features are obtained by performing feature transformation on a second-party sample by the second-party device according to a second-party federated feature transformation model.

[0068] In step S30, the first-party federated prediction feature extraction model is used to perform feature extraction on the unaligned sample to obtain first-party predicted sample features.

[0069] In step S40, the second-party predicted sample features sent by the second-party device are received, and longitudinal federated prediction is performed according to the first-party predicted sample features and the second-party predicted sample features to obtain a longitudinal federated prediction result. The second-party predicted sample features are constructed by the second-party device based on model prediction outputs of each second-party sample and the fuzzy matching result. The model prediction outputs are obtained by performing feature extraction on the second-party sample by a second-party federated prediction feature extraction model.

[0070] In the embodiment, it is to be noted that the first participant device and the second participant device are participant devices of vertical federated learning, the federated server is a trusted third party, the first participant device is deployed with a first-party federated feature transformation model and a first-party federated prediction feature extraction model, the second participant device is deployed with a second-party federated feature transformation model and a second-party federated prediction feature extraction model, the unaligned sample is a sample in the first participant device that has not been aligned, and the unaligned sample is unknown to the second participant device. The second participant device does not know whether there is a sample aligned with the unaligned sample or which sample is aligned with the unaligned sample. The unaligned sample can be a sample of a new user in the first participant device, or a new sample formed by updating or changing user identification information of an existing user. The user identification information is identification information for sample alignment between the first participant device and the second participant device. In conventional vertical federated learning, the first participant device can align the second-party model prediction result sent by the second participant device with the locally generated first-party model prediction result by using the user identification information, thereby realizing vertical federated prediction. The user identification information is user identity information, which can be an ID card number or a mobile phone number, etc. As an example, the second participant device sends the second-party model prediction result of the second-party sample and the identification information corresponding to the second-party sample to the first participant device. The first participant device can query and determine the corresponding first-party sample aligned with the second-party sample based on the identification information, so that the first participant device aggregates the first-party model prediction result corresponding to the aligned first-party sample and the second-party model prediction result, that is, the vertical federated prediction result, thereby realizing vertical federated prediction.

[0071] The first-party federated feature transformation model and the second-party federated feature transformation model are obtained by the first participant device and the second participant device according to the similarity between the transformation features of the two-party samples through vertical federated learning modeling; the first-party federated prediction feature extraction model and the second-party federated prediction feature extraction model are obtained by the first participant device and the second participant device according to the aligned samples through vertical federated learning modeling. The aligned samples include each first-party aligned sample of the first participant device and each second-party aligned sample of the second participant device, that is, the intersection sample owned by the first participant and the second participant.

[0072] If the first-party sample A and the second-party sample B are aligned samples, the second sample transformed features output by the first-party federated feature transformation model for the first-party sample A and the second sample transformed features output by the second-party federated feature transformation model for the second-party sample B are close in the same feature dimension space, that is, the similarity between the sample features is greater than a preset similarity threshold, or tends to be consistent. The first-party federated prediction feature extraction model and the second-party federated prediction feature extraction model are both used to extract prediction sample features of the aligned samples, and the prediction sample features are used for a federated prediction task. For example, if the federated prediction task is a classification task, the classification task can be completed by further classifying the prediction sample features.

[0073] As an example, the first-party federated feature transformation model includes a first-party local feature extraction model and a first-party local feature mapping model. The first-party local feature extraction model is used to map samples in the first-party device to a preset first feature dimension space, and the first-party local feature mapping model is used to map sample features output by the first-party local feature extraction model from the preset first feature dimension space to a preset second feature dimension space. The second-party federated feature transformation model includes a second-party local feature extraction model and a second-party local feature mapping model. The second-party local feature extraction model is used to map samples in the second-party device to the preset second feature dimension space, and the second-party local feature mapping model is used to map sample features output by the second-party local feature extraction model from the preset second feature dimension space to the preset first feature dimension space.

[0074] As an example, steps S10 to S40 include: mapping the unaligned sample to a preset second feature dimension space by inputting the unaligned sample into the first-party federated feature transformation model to obtain first sample transformed features; uploading the first sample transformed features to the federated server; the federated server receiving the first sample transformed features sent by the first participant device and each second sample transformed feature sent by the second participant device, the federated server generating a fuzzy matching result between the unaligned sample and each second-party sample according to the similarity between the first sample transformed features and each second sample transformed feature, and sending the fuzzy matching result to the second participant device, wherein the second sample transformed feature is obtained by mapping the second-party sample to the preset second feature dimension space according to the second-party federated feature transformation model by the second participant device; the second participant device receives the fuzzy matching result sent by the federated server, and selects at least one target model prediction output from the model prediction outputs of the second-party federated prediction feature extraction model on each second-party sample according to the fuzzy matching result to construct a second-party predicted sample feature; the first participant device inputs the unaligned sample into the first-party federated prediction feature extraction model for feature extraction to obtain a first-party predicted sample feature, receives the second-party predicted sample feature sent by the second participant device, and performs longitudinal federated prediction according to the first-party predicted sample feature and the second-party predicted sample feature to obtain a longitudinal federated prediction result.

[0075] As an example, the step of performing longitudinal federated prediction according to the first-party predicted sample feature and the second-party predicted sample feature to obtain a longitudinal federated prediction result includes:

[0076] aggregating the first-party predicted sample feature and the second-party predicted sample feature to obtain a federated predicted sample feature; classifying the federated predicted sample feature to obtain a federated classification result, and taking the federated classification result as the longitudinal federated prediction result.

[0077] As an example, the first-party sample identifier can be transmitted between the first participant device and the federated server for sample synchronization between the first participant device and the federated server, and the first-party sample identifier can not include user identity information; the second-party sample identifier can be transmitted between the second participant device and the federated server for sample synchronization between the second participant device and the federated server, and the second-party sample identifier can not include user identity information, thereby ensuring no risk of user identity information leakage.

[0078] As an example, the federal server calculates the similarity between the first sample transformed feature and each second sample transformed feature; selects a preset number of features with higher similarity from each second sample transformed feature as target sample transformed features; normalizes each target sample transformed feature to obtain each sample weight, and issues each sample weight and the second-party sample identifier corresponding to each sample weight to the second-party device as a fuzzy matching result. The second-party device receives each sample weight and the second-party sample identifier corresponding to each sample weight, and uses each second-party sample identifier to filter corresponding samples from each second-party sample, inputs the filtered samples into the second-party federal prediction feature extraction model for feature extraction to obtain each target model prediction output; and uses each sample weight to aggregate each target model prediction output to obtain a second-party predicted sample feature.

[0079] Before the step of transforming the unaligned sample according to the first-party federal feature transformation model to obtain the first sample transformed feature, the longitudinal federal prediction method further includes a training step of the first-party federal feature transformation model and the second-party federal feature transformation model, specifically including:

[0080] Step A10, obtaining a first-party sample, performing feature extraction on the first-party sample based on a first-party local feature extraction model to obtain a first local sample feature;

[0081] Step A20, performing feature mapping on the first local sample feature based on a first-party local feature mapping model to obtain a second local sample feature;

[0082] Step A30, receiving third local sample features sent by the second-party device, wherein the third local sample features are obtained by performing feature extraction on second-party samples by the second-party device according to a second-party local feature extraction model;

[0083] Step A40, updating the first-party local feature extraction model and the first-party local feature mapping model according to the similarity loss between the second local sample feature and the third local sample feature to obtain the first-party federal feature transformation model;

[0084] Step A50, sending the first local sample feature to the second-party device, so that the second-party device updates the second-party local feature extraction model and the second-party local feature mapping model according to the similarity loss between the first local sample feature and a fourth local sample feature to obtain the second-party federal feature transformation model, wherein the fourth local sample feature is obtained by performing feature mapping on the third local sample feature by the second-party device according to the second-party local feature mapping model.

[0085] As an example, steps A10 to A50 include: obtaining a first-party sample, mapping the first-party sample to a preset first feature dimension space by inputting the first-party sample into a first-party local feature extraction model to obtain a first local sample feature; mapping the first local sample feature from the preset first feature dimension space to a preset second feature dimension space by inputting the first local sample feature into a first-party local feature mapping model to obtain a second local sample feature; the second-party device inputs a second-party sample corresponding to the first-party sample into a second-party local feature extraction model to map the second-party sample to the preset second feature dimension space to obtain a third local sample feature, and inputs the third local sample feature into a second-party local feature mapping model to map the third local sample feature from the preset second feature dimension space to the preset first feature dimension space to obtain a fourth local sample feature; the first-party device receives the third local sample feature sent by the second-party device, and iteratively updates the first-party local feature extraction model and the first-party local feature mapping model according to a similarity loss between the second local sample feature and the third local sample feature; and the first local sample feature is sent to the second-party device, which receives the first local sample feature and iteratively updates the second-party local feature extraction model and the second-party local feature mapping model according to a similarity loss between the first local sample feature and the fourth local sample feature.

[0086] As an example, the step of iteratively updating the first-party local feature extraction model and the first-party local feature mapping model according to the similarity loss between the second local sample feature and the third local sample feature includes:

[0087] calculating a similarity loss between the second local sample feature and the third local sample feature, if the similarity loss is greater than a preset similarity loss threshold, updating the first-party local feature extraction model and the first-party local feature mapping model according to a gradient calculated based on the similarity loss, and returning to perform the step of obtaining the first-party sample; and if the similarity loss is greater than the preset similarity loss threshold, determining that the first-party local feature extraction model and the first-party local feature mapping model have been iteratively updated.

[0088] As an example, when a financial institution jointly assesses the qualifications of an applicant with other institutions, a risk control model based on longitudinal federated learning can be constructed with other institutions to assess the qualifications of the applicant. However, in a conventional longitudinal federated learning, the basis for the risk control model based on longitudinal federated learning to assess the qualifications of the applicant is to align samples. For unaligned samples of the financial institution, the risk control model based on longitudinal federated learning cannot be used to assess the qualifications of the applicant. If the risk control model constructed locally by the financial institution is used to assess the qualifications of the applicant, the accuracy of the assessment of the qualifications of the applicant is not high.

[0089] In the embodiments of the present application, the unaligned sample in the above example can be a user information sample of an unaligned user of the financial institution, for example, a sample vector composed of user loan information feature values, deposit information feature values, and repayment information feature values, etc., or a user portrait of the financial institution drawn by the financial institution. The first-party federated prediction feature extraction model can be a first-party federated risk control feature extraction model, and the second-party federated prediction feature extraction model can be a second-party federated risk control feature extraction model. The first-party federated risk control feature extraction model is used to extract a first-party risk control feature corresponding to the first-party user information sample, which can be a feature vector. The second-party federated risk control feature extraction model is used to extract a second-party risk control feature corresponding to the second-party user information sample, which can also be a feature vector. Based on the above, as an example, the longitudinal federated prediction method comprises: performing feature transformation on the unaligned user information sample according to a first-party federated feature transformation model to obtain first sample transformed features; uploading the first sample transformed features to a federated server, wherein the federated server is configured to generate a fuzzy matching result between the unaligned user information sample and each second-party user information sample according to the first sample transformed features and each second sample transformed features uploaded by a second-party device, the second sample transformed features being obtained by performing feature transformation on the second-party user information sample according to a second-party federated feature transformation model; performing feature extraction on the unaligned user information sample based on a first-party federated risk control feature extraction model to obtain a first-party risk control feature; receiving a second-party risk control feature sent by the second-party device, and performing applicant qualification evaluation based on longitudinal federated learning on the unaligned user according to the first-party risk control feature and the second-party risk control feature to obtain an applicant qualification federated evaluation result, wherein the second-party risk control feature is constructed based on a model prediction output of each second-party user information sample and the fuzzy matching result, and the model prediction output is obtained by performing feature extraction on the second-party user information sample by a second-party federated risk control feature extraction model. Therefore, when there is an unaligned user information sample of the unaligned user in the first-party device (the financial institution), the embodiments of the present application can assist the second-party device (the other institution) to construct the second-party risk control feature aligned with the first-party risk control feature corresponding to the unaligned user with the aid of the trusted third party (the federated server), so as to realize the applicant qualification evaluation based on longitudinal federated learning for the unaligned user, and improve the accuracy of the applicant qualification evaluation for the unaligned user in the financial institution.

[0090] The embodiment of the application provides a longitudinal federal prediction method. In a longitudinal federal learning scene, for unaligned samples in a first participant device, first, a first-party federal feature transformation model which is constructed based on longitudinal federal learning modeling based on aligned samples is used to perform feature transformation on the unaligned samples to obtain first sample transformed features, and then the first sample transformed features are uploaded to a federal server. The second participant device also uses a second-party federal feature transformation model which is constructed based on longitudinal federal learning modeling based on aligned samples to perform feature transformation on each second-party sample to obtain each second sample transformed feature, and the second participant device uploads each second sample transformed feature to the federal server. In this way, after the federal server receives the first sample transformed features and each second sample transformed feature, the federal server can perform fuzzy matching between the first sample transformed features and each second sample transformed feature based on similarity, to obtain a fuzzy matching result. Therefore, after the second participant device receives the fuzzy matching result, the second participant device can use the fuzzy matching result to select corresponding model output results from model prediction output of the second-party federal prediction feature extraction model on each second-party sample to construct second-party prediction sample features, so that the second-party prediction sample features corresponding to the first-party prediction sample features are constructed at the second participant device. In this way, after the first participant device performs feature extraction on the unaligned samples based on the first-party federal prediction feature extraction model to obtain first-party prediction sample features, and receives the second-party prediction sample features sent by the second participant device, the first participant device can perform longitudinal federal prediction by using the first-party prediction sample features and the second-party prediction sample features. Therefore, the embodiment of the application realizes that the second participant device constructs the second-party prediction sample features corresponding to the unaligned samples of the first participant device with the help of a trusted third party (the federal server), so that longitudinal federal prediction can be performed on the unaligned samples in the first participant device. Therefore, the problem that a longitudinal federal learning model constructed based on aligned samples cannot perform longitudinal federal learning prediction on unaligned samples and causes low prediction accuracy of unaligned samples is solved.

[0091] Embodiment two

[0092] The embodiment of the application also provides a longitudinal federal prediction method. The method of the embodiment is applied to a second participant device, and the longitudinal federal prediction method comprises the following steps. Figure 2

[0093] Step B10, performing feature transformation on each second-party sample based on the second-party federal feature transformation model to obtain each second sample transformed feature.

[0094] ​Step B20, uploading each second sample transformation feature to a federal server, wherein the federal server is configured to generate a fuzzy matching result between the unaligned sample and each second party sample according to each second sample transformation feature and a first sample transformation feature uploaded by a first participant device, the first sample transformation feature being obtained by the first participant device by performing feature transformation on the unaligned sample according to a first party federal feature transformation model;

[0095] Step B30, receiving the fuzzy matching result fed back by the federal server, and constructing a second party predicted sample feature according to a model prediction output of each second party sample by a second party federal prediction feature extraction model and the fuzzy matching result;

[0096] Step B40, sending the second party predicted sample feature to the first participant device, so as to perform longitudinal federal prediction on the first party predicted sample feature of the unaligned sample by a first party federal prediction feature extraction model and the second party predicted sample feature.

[0097] In this embodiment, it should be noted that the fuzzy matching result includes at least one sample weight and a second party user identifier corresponding to the sample weight.

[0098] As an example, steps B10 to B40 include: mapping each second-party sample to a preset second feature dimension space by inputting each second-party sample into the second-party federated feature transformation model to obtain each second sample transformed feature; uploading each second sample transformed feature to the federated server, the federated server receiving the first sample transformed feature sent by the first participant device and each second sample transformed feature sent by the second participant device, the federated server generating a fuzzy matching result between the unaligned sample and each second-party sample according to the similarity between the first sample transformed feature and each second sample transformed feature, and sending the fuzzy matching result to the second participant device, wherein the first sample transformed feature is obtained by mapping the unaligned sample to the preset second feature dimension space according to the first-party federated feature transformation model by the first participant device; the second participant device receives the fuzzy matching result, filters corresponding samples in each second-party sample according to each second-party user identifier in the fuzzy matching result, inputs the corresponding samples into the second-party federated prediction feature extraction model to obtain each target model output, and performs weighted aggregation on each target model output according to the sample weight corresponding to the second-party user identifier in the fuzzy matching result to obtain the second-party predicted sample feature; the second participant device sends the second-party predicted sample feature to the first participant device, the first participant device extracts features from the unaligned sample according to the first-party federated prediction feature extraction model to obtain the first-party predicted sample feature, and performs longitudinal federated prediction according to the first-party predicted sample feature and the second-party predicted sample feature to obtain a longitudinal federated prediction result. The specific implementation process of the first participant device performing longitudinal federated prediction can refer to the specific implementation process in steps S10 to S40 described above, which will not be described here.

[0099] The fuzzy matching result at least includes a sample feature weight, and the step of constructing the second-party predicted sample feature according to the model prediction output of each second-party sample by the second-party federated prediction feature extraction model and the fuzzy matching result includes:

[0100] Step B31, filtering corresponding samples in each second-party sample based on the second-party sample identifier corresponding to each sample feature weight, inputting the corresponding samples into the second-party federated prediction feature extraction model for feature extraction to obtain each target model prediction output;

[0101] Step B32, performing weighted aggregation on each target model prediction output according to each sample feature weight to obtain the second-party predicted sample feature.

[0102] As an example, the step B31 to the step B32 include: converting each second-party sample identifier into a corresponding alignment sample identifier, and searching for a corresponding target second-party sample in each second-party sample according to the alignment sample identifier; inputting each target second-party sample into a second-party federated prediction feature extraction model for feature extraction to obtain each target model prediction output; and performing weighted aggregation on each target model prediction output according to each sample feature weight to obtain a second-party prediction sample feature, where the weighted aggregation can be weighted summation or weighted average, and the alignment sample identifier is a sample identity identifier after sample alignment between the first participant device and the second participant device, and is used for sample synchronization between the first participant device and the second participant device. The alignment sample identifier can be identity information of a user, such as an ID card number or a mobile phone number.

[0103] Before the step of performing feature transformation on each second-party sample according to the second-party federated feature transformation model to obtain each second sample transformed feature, the longitudinal federated prediction method further includes:

[0104] The step C10 includes: obtaining a second-party sample, and performing feature extraction on the second-party sample based on a second-party local feature extraction model to obtain a third local sample feature.

[0105] The step C20 includes: performing feature mapping on the third local sample feature based on a second-party local feature mapping model to obtain a fourth local sample feature.

[0106] The step C30 includes: receiving a first local sample feature sent by the first participant device, where the first local sample feature is obtained by performing feature extraction on a first-party sample based on a first-party local feature extraction model by the first participant device.

[0107] The step C40 includes: updating the second-party local feature extraction model and the second-party local feature mapping model according to a similarity loss between the fourth local sample feature and the first local sample feature to obtain the second-party federated feature transformation model.

[0108] The step C50 includes: sending the third local sample feature to the first participant device, so that the first participant device updates the first-party local feature extraction model and a first-party local feature mapping model according to a similarity loss between a second local sample feature and the third local sample feature to obtain the first-party federated feature transformation model, where the second local sample feature is obtained by performing feature mapping on the first local sample feature based on the first-party local feature mapping model by the first participant device.

[0109] As an example, steps C10 to C50 include: obtaining a second-party sample, mapping the second-party sample to a preset second feature dimension space by inputting the second-party sample into a second-party local feature extraction model to obtain a third local sample feature; mapping the third local sample feature from the preset second feature dimension space to a preset first feature dimension space by inputting the third local sample feature into a second-party local feature mapping model to obtain a fourth local sample feature; the first-party device inputs a first-party sample corresponding to the second-party sample into a first-party local feature extraction model, maps the first-party sample to the preset first feature dimension space to obtain a first local sample feature, inputs the first local sample feature into a first-party local feature mapping model, maps the first local sample feature from the preset first feature dimension space to the preset second feature dimension space to obtain a second local sample feature; the second-party device receives the first local sample feature sent by the first-party device, and iteratively updates the second-party local feature extraction model and the second-party local feature mapping model according to a similarity loss between the first local sample feature and the fourth local sample feature; and the third local sample feature is sent to the first-party device, which receives the third local sample feature and iteratively updates the first-party local feature extraction model and the first-party local feature mapping model according to a similarity loss between the second local sample feature and the third local sample feature.

[0110] As an example, the step of iteratively updating the second-party local feature extraction model and the second-party local feature mapping model according to a similarity loss between the first local sample feature and the fourth local sample feature includes:

[0111] calculating a similarity loss between the first local sample feature and the fourth local sample feature, if the similarity loss is greater than a preset similarity loss threshold, updating the second-party local feature extraction model and the second-party local feature mapping model according to a gradient calculated based on the similarity loss, and returning to perform the step of obtaining a second-party sample; and if the similarity loss is greater than the preset similarity loss threshold, determining that the second-party local feature extraction model and the second-party local feature mapping model have been iteratively updated.

[0112] This application provides a vertical federated prediction method. In a vertical federated learning scenario, for each second-party sample in a second-party device, feature transformation is first performed according to a second-party federated feature transformation model to obtain transformation features for each second sample. These transformation features are then uploaded to a federated server. The federated server generates fuzzy matching results between misaligned samples and each second-party sample based on the transformation features of each second sample and the first sample transformation features uploaded by the first-party device. The first sample transformation features are obtained by the first-party device performing feature transformation on the misaligned samples according to the first-party federated feature transformation model. The method receives the fuzzy matching results from the federated server and performs feature extraction on each second-party sample according to the second-party federated prediction feature extraction model. The model prediction output and the fuzzy matching result are used to construct second-party prediction sample features. Therefore, this embodiment of the application realizes that the first participating device has unaligned samples, and with the help of a trusted third party (federated server), the second participating device constructs second-party prediction sample features corresponding to the first-party prediction sample features of the unaligned samples. The second-party prediction sample features are then sent to the first participating device, which can then perform vertical federated prediction based on the first-party prediction sample features and the second-party prediction sample features. This enables vertical federated prediction of unaligned samples in the first participating device, thus solving the problem that the vertical federated learning model built based on aligned samples cannot perform vertical federated learning prediction of unaligned samples, resulting in low prediction accuracy of unaligned samples.

[0113] Example 3

[0114] This application also provides a vertical federation prediction method. The method in this embodiment is applied to a federation server, as described below. Figure 3 The vertical federated forecasting method includes:

[0115] Step D10: Receive the first sample transformation feature sent by the first participating device and the second sample transformation features sent by the second participating device, wherein the first sample transformation feature is obtained by the first participating device performing feature transformation on the unaligned sample according to the first federated feature transformation model, and the second sample transformation feature is obtained by the second participating device performing feature transformation on the second sample according to the second federated feature transformation model.

[0116] Step D20: Based on the first sample transformation features and each of the second sample transformation features, generate fuzzy matching results between the unaligned samples and each of the second-party samples;

[0117] Step D30, the fuzzy matching result is sent to the second participant device, wherein the second participant device is configured to construct second-party predicted sample features from the model prediction output of each second-party sample and the fuzzy matching result according to a second-party federated prediction feature extraction model, and the first participant device is configured to perform longitudinal federated prediction according to the first-party predicted sample features and the second-party predicted sample features sent by the second participant device, wherein the first-party predicted sample features are obtained by performing feature extraction on the unaligned samples by a first-party federated prediction feature extraction model.

[0118] As an example, steps D10 to D30 include: receiving first sample transformed features sent by the first participant and each second sample transformed feature sent by the second participant device, wherein the first sample transformed features are obtained by performing feature transformation on unaligned samples by the first participant device according to a first federated feature transformation model, and the second sample transformed features are obtained by performing feature transformation on second-party samples by the second participant device according to a second federated feature transformation model; determining the probability that each unaligned sample and each second-party sample belong to the same sample according to the feature similarity between the first sample transformed features and each second sample transformed feature, obtaining each probability value, and generating a fuzzy matching result based on each probability value, wherein the fuzzy matching result can be each probability value itself, or a feature with higher feature similarity selected from each second sample transformed feature according to each probability value; sending the fuzzy matching result to the second participant device, wherein the second participant device is configured to construct second-party predicted sample features from the model prediction output of each second-party sample and the fuzzy matching result according to a second-party federated prediction feature extraction model, and the first participant device is configured to perform longitudinal federated prediction according to the first-party predicted sample features and the second-party predicted sample features sent by the second participant device, wherein the first-party predicted sample features are obtained by performing feature extraction on the unaligned samples by a first-party federated prediction feature extraction model, and the specific process of the first participant device performing longitudinal federated prediction can refer to the specific process in steps S10 to S40 described above, and the specific process of the second participant device constructing second-party predicted sample features according to the fuzzy matching result can refer to the content in steps B10 to B40 described above, and will not be described here.

[0119] As an example, the fuzzy matching result at least includes a sample feature weight, and the step of generating the fuzzy matching result based on each probability value includes:

[0120] selecting a preset number of features with higher probability values from each second sample transformed feature as target sample transformed features according to each probability value; and normalizing the target sample transformed features to obtain each sample feature weight.

[0121] As an example, the fuzzy matching result includes at least one sample feature weight, and the step of generating the fuzzy matching result based on the probability values includes:

[0122] According to the mapping relationship between the probability values and the sample feature weights, the probability values are mapped to corresponding sample feature weights.

[0123] The fuzzy matching result includes at least one sample feature weight, and the step of generating the fuzzy matching result between the unaligned sample and each second-party sample based on the first sample transformed feature and each second sample transformed feature includes:

[0124] Step a, calculating the feature similarity between the first sample transformed feature and each second sample transformed feature;

[0125] Step b, filtering at least one target sample transformed feature from each second sample transformed feature according to the feature similarity;

[0126] Step c, normalizing each target sample transformed feature to obtain each sample feature weight.

[0127] In this embodiment, it should be noted that the first sample transformed feature and each second sample transformed feature can be a feature vector, and the feature similarity can be a cosine similarity.

[0128] As an example, steps a to c include: calculating the cosine similarity between the first sample transformed feature and each second sample transformed feature; filtering a preset number of features with higher cosine similarity from each second sample transformed feature as target sample transformed features; and normalizing each target sample transformed feature to obtain each sample feature weight.

[0129] As an example, the federation server can directly send the feature similarity between the first sample transformed feature and each second sample transformed feature to the second-party device, and the second-party device can filter a preset number of samples with higher similarity from each second-party sample according to the feature similarity, input the second-party federation prediction feature extraction model for feature extraction to obtain each target model prediction output, and then the second-party device can weight and sum each target model prediction output according to the feature similarity, to obtain the second-party predicted sample feature.

[0130] It should be noted that the second participant device has the possibility to deduce the unaligned sample by using the feature similarity and the second sample transformation feature after obtaining the feature similarity, thereby affecting the data privacy protection in the vertical federated learning process. Therefore, in the embodiment of the present application, the federated server sends the normalized second sample transformation feature to the second participant device, and the second participant device cannot deduce the unaligned sample based on the normalized second sample transformation feature, thereby protecting the data privacy in the vertical federated learning process and improving the data privacy protection level.

[0131] The fuzzy matching result at least includes a sample feature weight, and the step of generating the fuzzy matching result between the unaligned sample and each second-party sample according to the first sample transformation feature and each second sample transformation feature includes:

[0132] Step d: splicing the first sample transformation feature with each second sample transformation feature to obtain each spliced feature;

[0133] Step e: classifying each spliced feature to identify the probability of the first sample transformation feature and each second sample transformation feature corresponding to the same sample, respectively, to obtain each classification probability;

[0134] Step f: selecting at least one target sample transformation feature from each second sample transformation feature according to each classification probability;

[0135] Step g: normalizing each target sample transformation feature to obtain each sample feature weight.

[0136] As an example, steps d to g include: splicing the first sample transformation feature with each second sample transformation feature directly to obtain each spliced feature; inputting each spliced feature into a preset classification model for classification to obtain each classification probability, wherein the classification probability represents the probability of the first sample transformation feature and the second sample transformation feature corresponding to the same sample; selecting a preset number of features with higher classification probability from each second sample transformation feature as the target sample transformation feature; and normalizing each target sample transformation feature to obtain each sample feature weight.

[0137] As an example, the normalization of each target sample transformation feature to obtain each sample feature weight is calculated as follows:

[0138] d norm，i = softmax(|d i |)

[0139] where d norm，i is the sample feature weight, and d i is the target sample transformation feature.

[0140] As an example, refer to Figure 4 , Figure 4 is a schematic diagram of an embodiment of a longitudinal federated learning framework in the embodiments of the present application, where participant A is a first participant device, participant B is a second participant device, and trusted party C is a federated server, X A，new is an unaligned sample, X A，al is a first-party sample, M A is a first-party federated prediction feature extraction model, Encoder A is a first-party federated feature transformation model, h a , new is a first-party predicted sample feature, h b，new is a second-party predicted sample feature, Classifier is a federated learning task module, which can be a classification module, for example, for jointly converting the first-party predicted sample feature and the second-party predicted sample feature into a corresponding longitudinal federated prediction result (Score), or calculating a longitudinal federated learning loss (Lss), Lss is a similarity loss, X B and X B，al is a second-party sample, where X B，al may be an intersection sample of the first participant device in the second participant device, X B may be a sample other than the intersection sample of the first participant device in the second participant device, M B is a second-party federated prediction feature extraction model, Encoder B is a second-party federated feature transformation model, f A，new is a first sample transformed feature, f B is a second sample transformed feature, Distance is used to calculate the feature similarity between the first sample transformed feature and the second sample transformed feature, d is a sample feature weight, and ID is a second-party sample identifier corresponding to the sample feature weight.

[0141] As an example, refer to Figure 5 , Figure 5 is a schematic diagram of an embodiment of a longitudinal federated learning framework in the embodiments of the present application, where participant A is a first participant device, participant B is a second participant device, and trusted party C is a federated server, X A，new is an unaligned sample, M A is a first-party federated prediction feature extraction model, Encoder A is a first-party federated feature transformation model, h a，new is a first-party predicted sample feature, h b，newThe classifier is a federated learning task module, for example, can be a classification module, used to jointly convert the first-party predicted sample feature and the second-party predicted sample feature into a corresponding longitudinal federated prediction result (Score), or calculate a longitudinal federated learning loss (Lss), which is a similarity loss, X B M is a second-party sample B The encoder is a second-party federated prediction feature extraction model B The transformer is a second-party federated feature transformation model a，new The first sample transformed feature is V B The second sample transformed feature is Transformer and classifier are used together to detect the probability value of the first sample transformed feature and the second sample transformed feature corresponding to the same sample, Matchscore is a sample feature weight, and ID is a second-party sample identifier corresponding to the sample feature weight. In the embodiment of the application, the first preset feature vector, the first sample transformed feature, the second sample transformed feature, and the second preset feature vector are spliced to obtain a spliced feature, and then the spliced feature is encoded through the Transformer layer to obtain an encoded feature. The encoded feature is input into the classification layer to obtain a classification probability, which represents the probability of the first sample transformed feature and the second sample transformed feature corresponding to the same sample.

[0142] As an example, the unaligned sample can be an unaligned image sample, the first sample transformed feature can be a first image sample transformed feature, the first-party sample can be a first-party image sample, the second-party sample can be a second-party image sample, the second sample transformed feature can be a second image sample transformed feature, the first-party predicted sample feature can be a first-party predicted image sample feature, the second-party predicted sample feature can be a second-party predicted image sample feature, the first-party federated feature transformation model can be a first-party federated image feature transformation model, the second-party federated feature transformation model can be a second-party federated image feature transformation model, the first-party federated prediction feature extraction model can be a first-party federated prediction image feature extraction model, the second-party federated prediction feature extraction model can be a second-party federated prediction image feature extraction model, and the longitudinal federated prediction process can be a longitudinal federated object detection process. The longitudinal federated prediction result can be an object detection result based on longitudinal federated learning, which solves the technical problem that a longitudinal federated learning object detection model constructed based on aligned image samples cannot perform longitudinal federated learning object detection on unaligned image samples, resulting in low object detection accuracy of unaligned image samples.

[0143] The embodiment of the application provides a longitudinal federated prediction method, system, electronic device, medium and program product. In a longitudinal federated learning scene, after a first sample transformed feature and each second sample transformed feature are received, a federated server can perform fuzzy matching according to the similarity between the first sample transformed feature and each second sample transformed feature, to obtain a fuzzy matching result. Therefore, after the second participant device receives the fuzzy matching result, the second participant device can select a corresponding model output result from the model prediction output of each second-party sample in a second-party federated prediction feature extraction model to construct a second-party prediction sample feature, so as to realize the construction of the second-party prediction sample feature corresponding to the first-party prediction sample feature at the second participant device. In this way, the first participant device can perform longitudinal federated prediction by using the first-party prediction sample feature and the second-party prediction sample feature after performing feature extraction on the unaligned sample based on a first-party federated prediction feature extraction model to obtain the first-party prediction sample feature, and receiving the second-party prediction sample feature sent by the second participant device. Therefore, the application realizes that the first participant device has unaligned samples, and the second participant device constructs the second-party prediction sample feature corresponding to the unaligned sample by means of a trusted third party (federated server), so that longitudinal federated prediction can be realized for the unaligned sample in the first participant device. Therefore, the problem that a longitudinal federated learning model constructed based on aligned samples cannot perform longitudinal federated learning prediction on unaligned samples, and the prediction accuracy of unaligned samples is low, is solved.

[0144] Embodiment four

[0145] The embodiment of the application also provides a longitudinal federated model construction optimization method. The method of the embodiment is applied to a first participant device. Referring to Figure 6 , the longitudinal federated model construction optimization method comprises the following steps.

[0146] Step E10, extracting a first-party training sample from a training sample set, wherein the training sample set is composed of unaligned samples and first-party aligned samples;

[0147] Step E20, if the first-party training sample is an unaligned sample, performing feature transformation on the first-party training sample according to a first-party federated feature transformation model, to obtain a first sample transformed feature;

[0148] Step E30, uploading the first sample transformation feature to a federal server, wherein the federal server is configured to generate a fuzzy matching result between the first-party training sample and each second-party sample according to the first sample transformation feature and each second sample transformation feature uploaded by a second-party device, wherein the second sample transformation feature is obtained by performing feature transformation on a second-party sample by the second-party device based on a second-party federal feature transformation model;

[0149] Step E40, performing feature extraction on the first-party training sample based on a first-party federal prediction feature extraction model to obtain a first-party prediction sample feature;

[0150] Step E50, receiving a second-party prediction sample feature sent by the second-party device, wherein the second-party prediction sample feature is constructed based on a model prediction output of each second-party sample and the fuzzy matching result, wherein the model prediction output is obtained by performing feature extraction on the second-party sample by a second-party federal prediction feature extraction model;

[0151] Step E60, performing longitudinal federal learning modeling according to the first-party prediction sample feature and the second-party prediction sample feature to iteratively update the first-party federal prediction feature extraction model and the second-party federal prediction feature extraction model.

[0152] In this embodiment, it should be noted that the first participant device and the second participant device are participant devices of vertical federated learning, the federated server is a trusted third party, the first participant device is deployed with a first-party federated feature transformation model and a first-party federated prediction feature extraction model, the second participant device is deployed with a second-party federated feature transformation model and a second-party federated prediction feature extraction model, the unaligned sample is a sample in the first participant device that has not been aligned, and the unaligned sample is unknown to the second participant device. The second participant device does not know whether there is a sample aligned with the unaligned sample or which sample is aligned with the unaligned sample. The unaligned sample can be a sample of a new user in the first participant device, or a new sample formed by updating or changing user identification information of an existing user. The user identification information is identification information for sample alignment between the first participant device and the second participant device. In conventional vertical federated learning, the first participant device can align the second-party model training parameters sent by the second participant device with the locally generated first-party model training parameters using the user identification information. The model training parameters can be network parameters of the model itself or gradient information, thereby realizing vertical federated learning modeling. The user identification information is user identity information, which can be an ID card number or a mobile phone number. As an example, the second participant device sends the second-party model training parameters of the second-party training sample and the identification information corresponding to the second-party training sample to the first participant device. The first participant device can query and determine the corresponding first-party training sample aligned with the second-party training sample based on the identification information, so that the first participant device aggregates the first-party model training parameters corresponding to the aligned first-party training sample and the second-party model training parameters, thereby realizing vertical federated learning modeling.

[0153] The first-party federated feature transformation model and the second-party federated feature transformation model are obtained by vertical federated learning modeling of the first participant device and the second participant device according to the similarity between the transformation features of the samples of the two parties. The first-party federated prediction feature extraction model and the second-party federated prediction feature extraction model are obtained by vertical federated learning modeling of the first participant device and the second participant device according to the aligned samples. The aligned samples include each first-party aligned sample of the first participant device and each second-party aligned sample of the second participant device, i.e., the intersection samples owned by the first participant and the second participant.

[0154] As an example, the first-party federated feature transformation model comprises a first-party local feature extraction model and a first-party local feature mapping model, the first-party local feature extraction model is configured to map samples in the first-party device to a preset first feature dimension space, and the first-party local feature mapping model is configured to map sample features output by the first-party local feature extraction model from the preset first feature dimension space to a preset second feature dimension space; the second-party federated feature transformation model comprises a second-party local feature extraction model and a second-party local feature mapping model, the second-party local feature extraction model is configured to map samples in the second-party device to the preset second feature dimension space, and the second-party local feature mapping model is configured to map sample features output by the second-party local feature extraction model from the preset second feature dimension space to the preset first feature dimension space.

[0155] As an example, steps E10 to E60 include: extracting a first-party training sample from a training sample set, wherein the training sample set is composed of unaligned samples and first-party aligned samples; if the first-party training sample is a first-party aligned sample, modeling longitudinal federated learning jointly with a corresponding second-party aligned sample at a second-party device according to the first-party aligned sample, and calculating a longitudinal federated learning loss; if the first-party training sample is an unaligned sample, mapping the unaligned sample to a preset second feature dimension space by inputting the unaligned sample into a first-party federated feature transformation model to obtain first sample transformed features, uploading the first sample transformed features to a federated server, the federated server receiving the first sample transformed features sent by the first-party device and each second sample transformed feature sent by the second-party device, the federated server generating a fuzzy matching result between the unaligned sample and each second-party sample according to the similarity between the first sample transformed features and each second sample transformed feature, and sending the fuzzy matching result to the second-party device, wherein the second sample transformed feature is obtained by mapping a second-party sample to the preset second feature dimension space according to a second-party federated feature transformation model by the second-party device, the second-party device receiving the fuzzy matching result sent by the federated server, and selecting at least one target model prediction output from model prediction outputs of the second-party federated prediction feature extraction model for each second-party sample according to the fuzzy matching result to construct second-party prediction sample features; the first-party device extracts features by inputting the unaligned sample into a first-party federated prediction feature extraction model to obtain first-party prediction sample features, receives the second-party prediction sample features sent by the second-party device, performs longitudinal federated prediction according to the first-party prediction sample features and the second-party prediction sample features to obtain a longitudinal federated training prediction result, and calculates a longitudinal federated learning loss according to the longitudinal federated training prediction result and a sample label corresponding to the unaligned sample; if the longitudinal federated learning loss converges, it is determined that the longitudinal federated learning modeling is complete, and if the longitudinal federated learning loss does not converge, updating the first-party federated prediction feature extraction model according to a training gradient calculated by the longitudinal federated learning loss, returning the training gradient to the second-party device for updating the second-party federated prediction feature extraction model according to the training gradient, and returning to the step of extracting a first-party training sample from a training sample set until the calculated longitudinal federated learning loss converges.

[0156] As an example, aggregating the first-party prediction sample features and the second-party prediction sample features to obtain federated prediction sample features; classifying the federated prediction sample features to obtain a federated classification label, and taking the federated classification label as the longitudinal federated training prediction result.

[0157] As an example, a first-party sample identifier can be transmitted between the first-party device and the federated server for sample synchronization between the first-party device and the federated server, the first-party sample identifier can not include user identity information; a second-party sample identifier can be transmitted between the second-party device and the federated server for sample synchronization between the second-party device and the federated server, the second-party sample identifier can not include user identity information, ensuring that there is no risk of user identity information leakage.

[0158] As an example, the federated server calculates the similarity between the first sample transformed feature and each second sample transformed feature; selects a preset number of features with higher similarity in each second sample transformed feature as target sample transformed features; normalizes each target sample transformed feature to obtain each sample weight, and issues each sample weight and the second-party sample identifier corresponding to each sample weight to the second-party device as a fuzzy matching result. The second-party device receives each sample weight and the second-party sample identifier corresponding to each sample weight, and uses each second-party sample identifier to filter corresponding samples in each second-party sample to input into a second-party federated prediction feature extraction model for feature extraction to obtain each target model prediction output; and uses each sample weight to aggregate each target model prediction output to obtain a second-party prediction sample feature.

[0159] The longitudinal federated model construction optimization method further includes a training step of the first-party federated feature transformation model and the second-party federated feature transformation model, specifically including:

[0160] Step F10, obtaining a first-party sample, performing feature extraction on the first-party sample based on a first-party local feature extraction model to obtain a first local sample feature;

[0161] Step F20, performing feature mapping on the first local sample feature based on a first-party local feature mapping model to obtain a second local sample feature;

[0162] Step F30, receiving third local sample features sent by the second-party device, wherein the third local sample features are obtained by performing feature extraction on second-party samples by the second-party device according to a second-party local feature extraction model;

[0163] Step F40, updating the first-party local feature extraction model and the first-party local feature mapping model according to the similarity loss between the second local sample feature and the third local sample feature to obtain the first-party federated feature transformation model;

[0164] Step F50, the first local sample feature is sent to the second participant device, so that the second participant device updates the second-party local feature extraction model and the second-party local feature mapping model according to the similarity loss between the first local sample feature and a fourth local sample feature, to obtain the second-party federated feature transformation model, and the fourth local sample feature is obtained by the second participant device according to feature mapping of the third local sample feature by the second-party local feature mapping model.

[0165] In the present embodiment, it should be noted that the specific implementation process of steps F10 to F50 can refer to the specific implementation process in steps A10 to A50 described above, which will not be repeated here.

[0166] It should be noted that in the current longitudinal federated learning modeling scenario, the first participant device and the second participant device are usually required to align the samples, and the aligned samples are used for longitudinal federated learning modeling. For unaligned samples, they cannot be used for federated learning modeling. When the participants of longitudinal federated learning add new samples or the original sample identity information changes, they usually cannot join the federated learning modeling process.

[0167] The embodiment of the application provides a longitudinal federal model construction optimization method, that is, a first-party training sample is extracted from a training sample set, wherein the training sample set is composed of unaligned samples and first-party aligned samples; if the first-party training sample is an unaligned sample, then feature transformation is performed on the first-party training sample according to a first-party federal feature transformation model, to obtain first sample transformation features; the first sample transformation features are uploaded to a federal server, wherein the federal server is configured to generate a fuzzy matching result between the first-party training sample and each second-party sample according to the first sample transformation features and each second sample transformation feature uploaded by a second participating device, wherein the second sample transformation feature is obtained by performing feature transformation on a second-party sample according to a second-party federal feature transformation model; feature extraction is performed on the first-party training sample based on a first-party federal prediction feature extraction model, to obtain first-party prediction sample features; second-party prediction sample features sent by the second participating device are received, wherein the second-party prediction sample features are obtained by constructing a model prediction output of each second-party sample and the fuzzy matching result based on the second participating device, and the model prediction output is obtained by performing feature extraction on the second-party sample by a second-party federal prediction feature extraction model; longitudinal federal learning modeling is performed according to the first-party prediction sample features and the second-party prediction sample features, to iteratively update the first-party federal prediction feature extraction model and the second-party federal prediction feature extraction model. In this way, for unaligned samples in the first participating device, the second-party prediction sample features can be constructed by the trusted third party (federal server) to assist the second participating device, and the second-party prediction sample features are aligned with the first-party prediction sample features corresponding to the unaligned samples, so that longitudinal federal learning modeling for the unaligned samples in the first participating device is realized, the problem that the unaligned samples cannot participate in longitudinal federal learning modeling is solved, the longitudinal federal learning modeling no longer depends on aligned samples, the threshold for the samples of each participating device to join the longitudinal federal learning modeling is reduced, and the sample universality and the universality of the longitudinal federal learning modeling are improved.

[0168] Embodiment five

[0169] The embodiment of the application further provides a longitudinal federal model construction optimization method, and the method of the embodiment is applied to a second participating device, and the longitudinal federal model construction optimization method comprises the following steps: Figure 7

[0170] Step G10, performing feature transformation on each second-party sample according to a second-party federal feature transformation model, to obtain each second sample transformation feature;

[0171] ​Step G20, uploading each of the second sample transformation features to a federal server, wherein the federal server is configured to generate a fuzzy matching result between the unaligned samples and each of the second party samples according to each of the second sample transformation features and first sample transformation features uploaded by a first participant device, the first sample transformation features being obtained by the first participant device by performing feature transformation on the unaligned samples according to a first party federal feature transformation model;

[0172] Step G30, receiving the fuzzy matching result fed back by the federal server, and constructing second party predicted sample features according to model prediction outputs of each of the second party samples and the fuzzy matching result according to a second party federal prediction feature extraction model;

[0173] Step G40, sending the second party predicted sample features to the first participant device, so as to perform longitudinal federal learning modeling on the first party predicted sample features of the unaligned samples and the second party predicted sample features based on a first party federal prediction feature extraction model.

[0174] As an example, the fuzzy matching result includes at least one sample weight and a second-party user identifier corresponding to the sample weight, and steps G10 to G40 include: mapping each second-party sample to a preset second feature dimension space by inputting each second-party sample into the second-party federated feature transformation model to obtain each second sample transformed feature; uploading each second sample transformed feature to the federated server, the federated server receiving the first sample transformed feature sent by the first participant device and each second sample transformed feature sent by the second participant device, the federated server generating a fuzzy matching result between the unaligned sample and each second-party sample according to the similarity between the first sample transformed feature and each second sample transformed feature, and sending the fuzzy matching result to the second participant device, wherein the first sample transformed feature is obtained by mapping the unaligned sample to the preset second feature dimension space according to the first-party federated feature transformation model; the second participant device receives the fuzzy matching result, filters corresponding samples in each second-party sample according to each second-party user identifier in the fuzzy matching result to input the second-party federated prediction feature extraction model to obtain each target model prediction output, and performs weighted aggregation on each target model prediction output according to the sample weight corresponding to the second-party user identifier in the fuzzy matching result to obtain the second-party prediction sample feature; the second participant device sends the second-party prediction sample feature to the first participant device, the first participant device extracts features from the unaligned sample according to the first-party federated prediction feature extraction model to obtain the first-party prediction sample feature; and longitudinal federated learning modeling is performed according to the first-party prediction sample feature and the second-party prediction sample feature to iteratively update the first-party federated prediction feature extraction model and the second-party federated prediction feature extraction model. The specific implementation process of the first participant device performing longitudinal federated learning modeling can refer to the specific implementation process in steps E10 to E60 described above, which will not be described here.

[0175] The fuzzy matching result includes at least one sample feature weight, and the step of constructing the second-party prediction sample feature according to the model prediction output of each second-party sample by the second-party federated prediction feature extraction model and the fuzzy matching result includes:

[0176] Step G31, filtering corresponding samples in each second-party sample based on the second-party sample identifier corresponding to each sample feature weight to input the second-party federated prediction feature extraction model for feature extraction to obtain each target model prediction output.

[0177] Step G32, performing weighted aggregation on each target model prediction output according to each sample feature weight to obtain the second-party prediction sample feature.

[0178] The specific implementation process of steps G31 to G32 in the embodiments of the present application can refer to the specific content in steps B31 to B32 described above, which will not be repeated here.

[0179] The longitudinal federal model construction optimization method further includes a training step of the first-party federal feature transformation model and the second-party federal feature transformation model, and specifically includes:

[0180] Step H10: obtaining a second-party sample, performing feature extraction on the second-party sample based on a second-party local feature extraction model to obtain third local sample features;

[0181] Step H20: performing feature mapping on the third local sample features based on a second-party local feature mapping model to obtain fourth local sample features;

[0182] Step H30: receiving first local sample features sent by the first-party device, wherein the first local sample features are obtained by performing feature extraction on first-party samples by the first-party device according to a first-party local feature extraction model;

[0183] Step H40: updating the second-party local feature extraction model and the second-party local feature mapping model according to a similarity loss between the fourth local sample features and the first local sample features to obtain the second-party federal feature transformation model;

[0184] Step H50: sending the third local sample features to the first-party device, so that the first-party device updates the first-party local feature extraction model and the first-party local feature mapping model according to a similarity loss between second local sample features and the third local sample features to obtain the first-party federal feature transformation model, wherein the second local sample features are obtained by performing feature mapping on the first local sample features by the first-party device according to the first-party local feature mapping model.

[0185] The specific implementation process of steps H10 to H50 in the embodiments of the present application can refer to the specific content in steps C10 to C50 described above, which will not be repeated here.

[0186] The embodiment of the application provides a longitudinal federal model construction optimization method, that is, according to a second-party federal feature transformation model, feature transformation is performed on each second-party sample to obtain second sample transformed features; each second sample transformed feature is uploaded to a federal server, wherein the federal server is configured to generate a fuzzy matching result between the unaligned sample and each second-party sample according to each second sample transformed feature and first sample transformed features uploaded by a first-party device, the first sample transformed features are obtained by performing feature transformation on the unaligned sample by the first-party device according to a first-party federal feature transformation model; the fuzzy matching result fed back by the federal server is received, and a second-party predicted sample feature is constructed according to a model prediction output of each second-party sample and the fuzzy matching result according to a second-party federal prediction feature extraction model; and the second-party predicted sample feature is sent to the first-party device, so that the first-party device performs longitudinal federal learning modeling on the first-party predicted sample feature of the unaligned sample and the second-party predicted sample feature based on a first-party federal prediction feature extraction model. In this way, for the unaligned sample in the first-party device, the second-party predicted sample feature can be constructed by the trusted third party (federal server) to assist the second-party device, the second-party predicted sample feature is aligned with the first-party predicted sample feature corresponding to the unaligned sample, so that the longitudinal federal learning modeling for the unaligned sample in the first-party device is realized, the problem that the unaligned sample cannot participate in the longitudinal federal learning modeling is solved, the longitudinal federal learning modeling no longer depends on the aligned sample, the threshold for the samples of each party device to join the longitudinal federal learning modeling is reduced, and the sample generality and universality of the longitudinal federal learning modeling are improved.

[0187] Embodiment six

[0188] The embodiment of the application further provides a longitudinal federal model construction optimization method, and the method of the embodiment is applied to a federal server, and the longitudinal federal model construction optimization method comprises the following steps: Figure 8 , the method comprises the following steps:

[0189] Step I10, receiving first sample transformed features sent by a first-party device and each second sample transformed feature sent by a second-party device, wherein the first sample transformed features are obtained by performing feature transformation on an unaligned sample by the first-party device according to a first federal feature transformation model, and the second sample transformed features are obtained by performing feature transformation on a second-party sample by the second-party device according to a second federal feature transformation model;

[0190] Step I20, generating a fuzzy matching result between the unaligned sample and each second-party sample according to the first sample transformed features and each second sample transformed feature;

[0191] Step I30, sending the fuzzy matching result to the second participant device, wherein the second participant device is configured to construct second-party predicted sample features according to model prediction outputs of second-party sample features and the fuzzy matching result, and the first participant device is configured to perform longitudinal federated learning modeling according to first-party predicted sample features and the second-party predicted sample features sent by the second participant device, wherein the first-party predicted sample features are obtained by performing feature extraction on the unaligned samples by the first-party federated prediction feature extraction model.

[0192] As an example, steps I10 to I30 include: receiving first sample transformed features sent by the first participant and each second sample transformed feature sent by the second participant device, wherein the first sample transformed features are obtained by performing feature transformation on unaligned samples by the first participant device according to a first federated feature transformation model, and the second sample transformed features are obtained by performing feature transformation on second-party sample features by the second participant device according to a second federated feature transformation model; determining the probability that each unaligned sample and each second-party sample belong to the same sample according to the feature similarity between the first sample transformed features and each second sample transformed feature, obtaining each probability value, and generating a fuzzy matching result based on each probability value, wherein the fuzzy matching result can be each probability value itself, or a feature with higher feature similarity selected from each second sample transformed feature according to each probability value; sending the fuzzy matching result to the second participant device, wherein the second participant device is configured to construct second-party predicted sample features according to model prediction outputs of second-party sample features and the fuzzy matching result, and the first participant device is configured to perform longitudinal federated learning modeling according to first-party predicted sample features and the second-party predicted sample features sent by the second participant device, wherein the first-party predicted sample features are obtained by performing feature extraction on the unaligned samples by the first-party federated prediction feature extraction model, and the specific process of the first participant device performing longitudinal federated learning modeling can refer to the specific process in steps E10 to E60, which will not be described here again, and the specific process of the second participant device constructing second-party predicted sample features according to the fuzzy matching result can refer to the content in steps G10 to G40, which will not be described here again.

[0193] The fuzzy matching result includes at least a sample feature weight, and the step of generating the fuzzy matching result between the unaligned samples and each second-party sample according to the first sample transformed features and each second sample transformed feature includes:

[0194] Step J10, calculating the feature similarity between the first sample transformed features and each second sample transformed feature;

[0195] Step J20, screening at least one target sample transformation feature from the second sample transformation features according to the feature similarity;

[0196] Step J30, normalizing the target sample transformation features to obtain the sample feature weights.

[0197] The specific implementation process of steps J10 to J30 in the embodiments of the present application can refer to the specific implementation contents in steps a to c described above, which will not be repeated here.

[0198] The step of generating the fuzzy matching result between the unaligned sample and the second-party sample according to the first sample transformation feature and the second sample transformation features includes:

[0199] Step K10, concatenating the first sample transformation feature with each of the second sample transformation features to obtain each concatenated feature;

[0200] Step K20, identifying the probability of the first sample transformation feature and each of the second sample transformation features corresponding to the same sample by classifying each of the concatenated features to obtain each classification probability;

[0201] Step K30, screening at least one target sample transformation feature from the second sample transformation features according to the classification probability;

[0202] Step K40, normalizing the target sample transformation features to obtain the sample feature weights.

[0203] The specific implementation process of steps K10 to K40 in the embodiments of the present application can refer to the specific implementation contents in steps d to g described above, which will not be repeated here.

[0204] As an example, the unaligned sample can be an unaligned image sample, the first sample transformation feature can be a first image sample transformation feature, the first-party sample can be a first-party image sample, the second-party sample can be a second-party image sample, the second sample transformation feature can be a second image sample transformation feature, the first-party federated feature transformation model can be a first-party federated image feature transformation model, the second-party federated feature transformation model can be a second-party federated image feature transformation model, the first-party federated prediction feature extraction model can be a first-party federated prediction image feature extraction model, the second-party federated prediction feature extraction model can be a second-party federated prediction image feature extraction model, and the longitudinal federated prediction process can be a longitudinal federated object detection process. The longitudinal federated prediction result can be an object detection result based on longitudinal federated learning, which solves the technical problem that unaligned image samples cannot be added to the object detection model construction process based on longitudinal federated learning.

[0205] The embodiment of the application provides a longitudinal federal model construction optimization method, that is, receiving first sample transformed features sent by a first participant device and each second sample transformed feature sent by a second participant device, wherein the first sample transformed feature is obtained by performing feature transformation on unaligned samples according to a first federal feature transformation model by the first participant device, and the second sample transformed feature is obtained by performing feature transformation on second party samples according to a second federal feature transformation model by the second participant device; generating a fuzzy matching result between the unaligned samples and each second party sample according to the first sample transformed feature and each second sample transformed feature; and issuing the fuzzy matching result to the second participant device, wherein the second participant device is configured to construct second party predicted sample features according to a model predicted output of each second party sample and the fuzzy matching result according to a second party federal prediction feature extraction model, and the first participant device is configured to perform longitudinal federal learning modeling according to first party predicted sample features and the second party predicted sample features sent by the second participant device, wherein the first party predicted sample features are obtained by performing feature extraction on the unaligned samples according to a first party federal prediction feature extraction model. In this way, for the unaligned samples in the first participant device, the second party predicted sample features can be constructed by the second participant device with the aid of a trusted third party (federal server), the second party predicted sample features are aligned with the first party predicted sample features corresponding to the unaligned samples, so that longitudinal federal learning modeling for the unaligned samples in the first participant device is realized, the problem that the unaligned samples cannot participate in longitudinal federal learning modeling is solved, the longitudinal federal learning modeling no longer depends on aligned samples, the threshold for the samples of each participant device to join the longitudinal federal learning modeling is reduced, and the sample universality and the universality of the longitudinal federal learning modeling are improved.

[0206] Embodiment seven

[0207] The application further provides a longitudinal federal prediction system, which comprises:

[0208] The first participant device is configured to perform feature transformation on unaligned samples according to a first party federal feature transformation model to obtain first sample transformed features, upload the first sample transformed features to a federal server, perform feature extraction on the unaligned samples based on a first party federal prediction feature extraction model to obtain first party predicted sample features, receive second party predicted sample features sent by the second participant device, and perform longitudinal federal prediction according to the first party predicted sample features and the second party predicted sample features to obtain a longitudinal federal prediction result.

[0209] The second participant device performs feature transformation on each second-party sample according to a second-party federated feature transformation model to obtain each second sample transformed feature; uploads each second sample transformed feature to the federated server; receives the fuzzy matching result fed back by the federated server, and constructs a second-party predicted sample feature according to the model prediction output of each second-party sample and the fuzzy matching result according to a second-party federated prediction feature extraction model; and sends the second-party predicted sample feature to the first participant device.

[0210] The federated server receives the first sample transformed feature sent by the first participant device and each second sample transformed feature sent by the second participant device; generates a fuzzy matching result between the unaligned sample and each second-party sample according to the first sample transformed feature and each second sample transformed feature; and sends the fuzzy matching result to the second participant device.

[0211] Optionally, the first participant device is further configured to:

[0212] obtain a first-party sample, perform feature extraction on the first-party sample based on a first-party local feature extraction model to obtain a first local sample feature;

[0213] perform feature mapping on the first local sample feature based on a first-party local feature mapping model to obtain a second local sample feature;

[0214] receive a third local sample feature sent by the second participant device, wherein the third local sample feature is obtained by the second participant device performing feature extraction on a second-party sample according to a second-party local feature extraction model;

[0215] update the first-party local feature extraction model and the first-party local feature mapping model according to a similarity loss between the second local sample feature and the third local sample feature to obtain the first-party federated feature transformation model;

[0216] send the first local sample feature to the second participant device, so that the second participant device updates the second-party local feature extraction model and a second-party local feature mapping model according to a similarity loss between the first local sample feature and a fourth local sample feature to obtain the second-party federated feature transformation model, wherein the fourth local sample feature is obtained by the second participant device performing feature mapping on the third local sample feature according to the second-party local feature mapping model.

[0217] Optionally, the fuzzy matching result at least includes a sample feature weight, and the second participant device is further configured to:

[0218] based on the second-party sample identifier corresponding to each sample feature weight, filtering the corresponding sample from each second-party sample to perform feature extraction in the second-party federated prediction feature extraction model to obtain each target model prediction output;

[0219] According to each sample feature weight, each target model prediction output is weighted and aggregated to obtain the second-party prediction sample feature.

[0220] Optionally, the second participant device is further configured to:

[0221] Obtain a second-party sample, and perform feature extraction on the second-party sample based on a second-party local feature extraction model to obtain a third local sample feature;

[0222] Perform feature mapping on the third local sample feature based on a second-party local feature mapping model to obtain a fourth local sample feature;

[0223] Receive the first local sample feature sent by the first participant device, wherein the first local sample feature is obtained by the first participant device performing feature extraction on a first-party sample based on a first-party local feature extraction model;

[0224] According to the similarity loss between the fourth local sample feature and the first local sample feature, update the second-party local feature extraction model and the second-party local feature mapping model to obtain the second-party federated feature transformation model;

[0225] Send the third local sample feature to the first participant device, so that the first participant device updates the first-party local feature extraction model and the first-party local feature mapping model according to the similarity loss between the second local sample feature and the third local sample feature, and obtains the first-party federated feature transformation model, wherein the second local sample feature is obtained by the first participant device performing feature mapping on the first local sample feature based on the first-party local feature mapping model.

[0226] Optionally, the fuzzy matching result at least includes a sample feature weight, and the federated server is further configured to:

[0227] Calculate the feature similarity between the first sample transformation feature and each second sample transformation feature;

[0228] According to each feature similarity, at least one target sample transformation feature is filtered from each second sample transformation feature;

[0229] Each target sample transformation feature is normalized to obtain each sample feature weight.

[0230] Optionally, the fuzzy matching result at least comprises a sample feature weight, and the federated server is further configured to:

[0231] The first sample transformation feature is spliced with each of the second sample transformation features respectively to obtain a plurality of spliced features;

[0232] By classifying each of the spliced features, a probability that the first sample transformation feature and each of the second sample transformation features correspond to the same sample is identified respectively to obtain a plurality of classification probabilities;

[0233] According to the classification probabilities, at least one target sample transformation feature is screened from the second sample transformation features;

[0234] Each of the target sample transformation features is normalized to obtain the sample feature weight.

[0235] The longitudinal federated prediction system provided by the embodiments of the present application adopts the longitudinal federated prediction method in the above embodiments, and solves the technical problem that the longitudinal federated learning model constructed based on aligned samples cannot perform longitudinal federated learning prediction on unaligned samples, resulting in low sample prediction accuracy of unaligned samples. Compared with the prior art, the longitudinal federated prediction system provided by the embodiments of the present application has the same beneficial effects as the longitudinal federated prediction method provided by the above embodiments, and other technical features in the longitudinal federated prediction system are the same as the features disclosed in the above embodiments, which will not be repeated here.

[0236] In addition, it should be noted that the current longitudinal federated learning prediction system is constructed based on sample alignment, so each participant device needs to store sample identification information locally, such as an ID card number or a mobile phone number, and the number of samples of each participant device in the federated learning prediction system is usually very high, so these sample identification information will occupy a large amount of memory resources of the participant device. In the embodiments of the present application, longitudinal federated learning prediction for unaligned samples is realized, and sample identification information is not required in the entire longitudinal federated learning prediction process, so the corresponding sample identification information does not need to be stored in the federated learning prediction system, and in the scenario of a large number of samples, the memory resources of the federated learning prediction system can be saved.

[0237] Embodiment Eight

[0238] The present application also provides a longitudinal federated model construction optimization system, which comprises:

[0239] The first participant device is configured to extract a first-party training sample from a training sample set, wherein the training sample set is composed of unaligned samples and first-party aligned samples; if the first-party training sample is an unaligned sample, perform feature transformation on the first-party training sample according to a first-party federated feature transformation model to obtain first sample transformed features; upload the first sample transformed features to a federated server; perform feature extraction on the first-party training sample based on a first-party federated prediction feature extraction model to obtain first-party prediction sample features; receive second-party prediction sample features sent by the second participant device; and perform longitudinal federated learning modeling according to the first-party prediction sample features and the second-party prediction sample features to iteratively update the first-party federated prediction feature extraction model and the second-party federated prediction feature extraction model.

[0240] The second participant device is configured to perform feature transformation on each second-party sample according to a second-party federated feature transformation model to obtain each second sample transformed feature; upload each second sample transformed feature to the federated server; receive feedback of a fuzzy matching result from the federated server, and construct second-party prediction sample features according to model prediction outputs of each second-party sample and the fuzzy matching result based on a second-party federated prediction feature extraction model; and send the second-party prediction sample features to the first participant device.

[0241] The federated server is configured to receive first sample transformed features sent by the first participant device and each second sample transformed feature sent by the second participant device; generate a fuzzy matching result between the unaligned samples and each second-party sample according to the first sample transformed features and each second sample transformed feature; and distribute the fuzzy matching result to the second participant device.

[0242] Optionally, the first participant device further comprises:

[0243] Obtain first-party samples, perform feature extraction on the first-party samples based on a first-party local feature extraction model to obtain first local sample features;

[0244] Perform feature mapping on the first local sample features based on a first-party local feature mapping model to obtain second local sample features;

[0245] Receive third local sample features sent by the second participant device, wherein the third local sample features are obtained by performing feature extraction on second-party samples by the second participant device according to a second-party local feature extraction model;

[0246] update the first-party local feature extraction model and the first-party local feature mapping model according to a similarity loss between the second local sample feature and the third local sample feature, to obtain the first-party federated feature transformation model;

[0247] send the first local sample feature to the second-party device, so that the second-party device updates the second-party local feature extraction model and the second-party local feature mapping model according to a similarity loss between the first local sample feature and a fourth local sample feature obtained by performing feature mapping on the third local sample feature according to the second-party local feature mapping model, to obtain the second-party federated feature transformation model.

[0248] Optionally, the fuzzy matching result at least includes a sample feature weight, and the second-party device further includes:

[0249] based on the second-party sample identifier corresponding to each sample feature weight, filtering a corresponding sample from each second-party sample and inputting the sample into the second-party federated prediction feature extraction model for feature extraction, to obtain each target model prediction output;

[0250] perform weighted aggregation on each target model prediction output according to each sample feature weight, to obtain the second-party prediction sample feature.

[0251] Optionally, the second-party device further includes:

[0252] obtain a second-party sample, perform feature extraction on the second-party sample based on a second-party local feature extraction model, to obtain a third local sample feature;

[0253] perform feature mapping on the third local sample feature based on a second-party local feature mapping model, to obtain a fourth local sample feature;

[0254] receive the first local sample feature sent by the first-party device, wherein the first local sample feature is obtained by performing feature extraction on a first-party sample based on a first-party local feature extraction model by the first-party device;

[0255] update the second-party local feature extraction model and the second-party local feature mapping model according to a similarity loss between the fourth local sample feature and the first local sample feature, to obtain the second-party federated feature transformation model;

[0256] The third local sample feature is sent to the first participant device, so that the first participant device updates the first-party local feature extraction model and the first-party local feature mapping model according to a similarity loss between a second local sample feature and the third local sample feature, to obtain the first-party federated feature transformation model, and the second local sample feature is obtained by the first participant device according to feature mapping of the first local sample feature by the first-party local feature mapping model.

[0257] Optionally, the fuzzy matching result at least includes a sample feature weight, and the federated server further comprises:

[0258] Calculate the feature similarity between the first sample transformation feature and each second sample transformation feature.

[0259] According to each feature similarity, at least one target sample transformation feature is screened from each second sample transformation feature.

[0260] Each target sample transformation feature is normalized to obtain each sample feature weight.

[0261] Optionally, the fuzzy matching result at least includes a sample feature weight, and the federated server further comprises:

[0262] The first sample transformation feature is spliced with each second sample transformation feature respectively to obtain each spliced feature.

[0263] By classifying each spliced feature, the probability that the first sample transformation feature and each second sample transformation feature correspond to the same sample is identified respectively to obtain each classification probability.

[0264] According to each classification probability, at least one target sample transformation feature is screened from each second sample transformation feature.

[0265] Each target sample transformation feature is normalized to obtain each sample feature weight.

[0266] The longitudinal federated model construction optimization system provided by the embodiments of the present application adopts the longitudinal federated model construction optimization method in the above embodiments, and solves the technical problem that unaligned samples cannot be added to the longitudinal federated learning modeling process. Compared with the prior art, the longitudinal federated model construction optimization system provided by the embodiments of the present application has the same beneficial effects as the longitudinal federated model construction optimization method provided by the above embodiments, and other technical features in the longitudinal federated model construction optimization system are the same as the features disclosed in the above embodiments, which will not be repeated here.

[0267] In addition, it should be noted that the current longitudinal federal model construction optimization system is constructed based on sample alignment, and therefore each participant device needs to store sample identification information, such as an ID card number or a mobile phone number, locally. The sample quantity of each participant device in the longitudinal federal model construction optimization system is usually high, and therefore these sample identification information occupies a large amount of memory resources of the participant device. In the embodiment of the present application, longitudinal federal learning modeling for unaligned samples is implemented, and sample identification information is not required in the entire longitudinal federal learning modeling process. Therefore, the corresponding sample identification information does not need to be stored in the longitudinal federal model construction optimization system, and the memory resources of the longitudinal federal model construction optimization system can be saved in the scenario of a large amount of samples.

[0268] Embodiment Nine

[0269] The embodiment of the present application provides an electronic device, which comprises at least one processor, and a memory connected with the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the longitudinal federal prediction method and / or the longitudinal federal model construction optimization method in the above embodiments.

[0270] Reference will be made to the following Figure 9 , which shows a structural schematic diagram of an electronic device suitable for implementing the embodiments of the present disclosure. The electronic device in the embodiments of the present disclosure can include, but is not limited to, mobile terminals such as mobile phones, notebook computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Tablet Personal Computers), PMPs (Portable Multimedia Players), vehicle terminals (such as vehicle navigation terminals), and the like, and fixed terminals such as digital TVs, desktop computers, and the like. Figure 9 The electronic device shown is only an example, and should not impose any limitation on the functions and use range of the embodiments of the present disclosure.

[0271] As shown in Figure 9 , the electronic device can include a processing device (such as a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to programs stored in a read-only memory (ROM) or loaded into a random access memory (RAM) from a storage device. In the RAM, various programs and data required for the operation of the electronic device are also stored. The processing device, the ROM, and the RAM are connected to each other through a bus. An input / output (I / O) interface is also connected to the bus.

[0272] Generally, the following systems can be connected to the I / O interface: input devices including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; output devices including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; storage devices including, for example, a magnetic tape, a hard disk, etc.; and communication devices. The communication devices can allow the electronic device to communicate wirelessly or wiredly with other devices to exchange data. Although the electronic device is illustrated as having various systems, it is understood that all of the illustrated systems are not required to implement or have. More or less systems can alternatively be implemented or have.

[0273] In particular, according to embodiments of the present disclosure, the processes described above with reference to the flowcharts can be implemented as a computer software program. For example, embodiments of the present disclosure include a computer program product comprising a computer program carried on a computer readable medium, the computer program containing program code for performing the methods illustrated by the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via the communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by the processing device, the above-mentioned functions defined in the methods of the embodiments of the present disclosure are performed.

[0274] The electronic device provided in the present application adopts the longitudinal federal prediction method and / or the longitudinal federal model construction optimization method in the above embodiments, which solves the technical problems that the longitudinal federal learning model constructed based on aligned samples cannot perform longitudinal federal learning prediction on unaligned samples, resulting in low sample prediction accuracy of unaligned samples and / or solves the technical problem that unaligned samples cannot join the longitudinal federal learning modeling process. Compared with the prior art, the electronic device provided in the embodiments of the present application has the same beneficial effects as the longitudinal federal prediction method and / or the longitudinal federal model construction optimization method provided in the above embodiments, and other technical features in the electronic device are the same as the features disclosed in the above embodiments, which will not be repeated here.

[0275] It should be understood that parts of the present disclosure can be realized by hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any appropriate manner in one or more embodiments or examples.

[0276] The above describes only the specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0277] Embodiment ten

[0278] The embodiment provides a computer readable storage medium having computer readable program instructions stored thereon, the computer readable program instructions being used to perform the method for longitudinal federated prediction and / or the method for longitudinal federated model construction optimization in the above embodiment.

[0279] The computer readable storage medium provided by the embodiment of the present application may, for example, be a U disk, but is not limited to an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, system, or device, or any combination of the above. More specific examples of the computer readable storage medium can include, but are not limited to, an electrical connection having one or more conductive wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the embodiment, the computer readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer readable storage medium can be transmitted by any appropriate medium, including but not limited to electrical wires, optical cables, RF (radio frequency), and the like, or any suitable combination of the above.

[0280] The above computer readable storage medium can be included in an electronic device; or can exist separately without being assembled into an electronic device.

[0281] The above computer readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to: perform feature transformation on the unaligned sample according to a first-party federated feature transformation model to obtain first sample transformed features; upload the first sample transformed features to a federated server, wherein the federated server is configured to generate a fuzzy matching result between the unaligned sample and each second-party sample according to the first sample transformed features and each second sample transformed features uploaded by a second participating device, the second sample transformed features being obtained by performing feature transformation on a second-party sample according to a second-party federated feature transformation model by the second participating device; perform feature extraction on the unaligned sample based on a first-party federated prediction feature extraction model to obtain first-party predicted sample features; receive second-party predicted sample features sent by the second participating device, and perform longitudinal federated prediction according to the first-party predicted sample features and the second-party predicted sample features to obtain a longitudinal federated prediction result, wherein the second-party predicted sample features are obtained by the second participating device based on model prediction outputs of each second-party sample and the fuzzy matching result, and the model prediction outputs are obtained by performing feature extraction on the second-party sample by a second-party federated prediction feature extraction model.

[0282] Or according to the second party federal feature transformation model, the feature of each second party sample is transformed to obtain each second sample transformed feature; each second sample transformed feature is uploaded to a federal server, wherein the federal server is configured to generate a fuzzy matching result between the unaligned sample and each second party sample according to each second sample transformed feature and the first sample transformed feature uploaded by the first participant device, the first sample transformed feature is obtained by the first participant device according to the first party federal feature transformation model for feature transformation of the unaligned sample; receive the fuzzy matching result fed back by the federal server, and construct the second party predicted sample feature according to the model prediction output of each second party sample and the fuzzy matching result of the second party federal prediction feature extraction model; the second party predicted sample feature is sent to the first participant device, so that the first participant device performs longitudinal federal prediction on the first party predicted sample feature of the unaligned sample and the second party predicted sample feature based on the first party federal prediction feature extraction model.

[0283] Or receive the first sample transformed feature sent by the first participant device and each second sample transformed feature sent by the second participant device, wherein the first sample transformed feature is obtained by the first participant device according to the first federal feature transformation model for feature transformation of the unaligned sample, and the second sample transformed feature is obtained by the second participant device according to the second federal feature transformation model for feature transformation of the second party sample; according to the first sample transformed feature and each second sample transformed feature, generate the fuzzy matching result between the unaligned sample and each second party sample; the fuzzy matching result is issued to the second participant device, wherein the second participant device is configured to construct the second party predicted sample feature according to the model prediction output of each second party sample and the fuzzy matching result of the second party federal prediction feature extraction model, and the first participant device is configured to perform longitudinal federal prediction according to the first party predicted sample feature and the second party predicted sample feature sent by the second participant device, the first party predicted sample feature is obtained by the first party federal prediction feature extraction model for feature extraction of the unaligned sample.

[0284] Or extract the first-party training samples from the training sample set, wherein the training sample set is composed of unaligned samples and first-party aligned samples; if the first-party training sample is an unaligned sample, perform feature transformation on the first-party training sample according to the first-party federated feature transformation model to obtain first sample transformed features; upload the first sample transformed features to a federated server, wherein the federated server is configured to generate a fuzzy matching result between the first-party training sample and each second-party sample according to the first sample transformed features and each second sample transformed features uploaded by a second participating device, wherein the second sample transformed features are obtained by performing feature transformation on a second-party sample according to a second-party federated feature transformation model by the second participating device; perform feature extraction on the first-party training sample based on a first-party federated prediction feature extraction model to obtain first-party prediction sample features; receive second-party prediction sample features sent by the second participating device, wherein the second-party prediction sample features are constructed based on model prediction outputs of each second-party sample and the fuzzy matching result, and the model prediction outputs are obtained by performing feature extraction on the second-party sample by a second-party federated prediction feature extraction model; and perform longitudinal federated learning modeling according to the first-party prediction sample features and the second-party prediction sample features to iteratively update the first-party federated prediction feature extraction model and the second-party federated prediction feature extraction model.

[0285] Or perform feature transformation on each second-party sample according to a second-party federated feature transformation model to obtain each second sample transformed feature; upload each second sample transformed feature to a federated server, wherein the federated server is configured to generate a fuzzy matching result between unaligned samples and each second-party sample according to each second sample transformed feature and first sample transformed features uploaded by a first participating device, wherein the first sample transformed features are obtained by performing feature transformation on unaligned samples by the first participating device according to a first-party federated feature transformation model; receive the fuzzy matching result fed back by the federated server, construct second-party prediction sample features according to model prediction outputs of each second-party sample and the fuzzy matching result based on a second-party federated prediction feature extraction model; and send the second-party prediction sample features to the first participating device for the first participating device to perform longitudinal federated learning modeling on first-party prediction sample features of the unaligned samples and the second-party prediction sample features based on a first-party federated prediction feature extraction model.

[0286] Or receiving the first sample transformation feature sent by the first participant device and each second sample transformation feature sent by the second participant device, wherein the first sample transformation feature is obtained by the first participant device from feature transformation of the unaligned sample according to a first federated feature transformation model, and the second sample transformation feature is obtained by the second participant device from feature transformation of the second-party sample according to a second federated feature transformation model; generating a fuzzy matching result between the unaligned sample and each second-party sample according to the first sample transformation feature and each second sample transformation feature; and issuing the fuzzy matching result to the second participant device, wherein the second participant device is configured to construct a second-party predicted sample feature from a model predicted output of each second-party sample and the fuzzy matching result according to a second-party federated prediction feature extraction model, and the first participant device is configured to perform longitudinal federated learning modeling according to a first-party predicted sample feature and the second-party predicted sample feature sent by the second participant device, wherein the first-party predicted sample feature is obtained by feature extraction of the unaligned sample according to a first-party federated prediction feature extraction model.

[0287] Computer program code for carrying out operations of the present disclosure can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C++ or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider).

[0288] The computer readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process, such that the instructions which execute on the computer, other programmable data processing apparatus, or other device implement the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0289] The modules described in the embodiments of the present disclosure can be implemented by software, or by hardware. In some cases, the name of the module does not constitute a limitation on the module itself.

[0290] The computer readable storage medium provided by the present application stores computer readable program instructions for executing the longitudinal federal prediction method and / or the longitudinal federal model construction optimization method described above, solves the technical problem that the longitudinal federal learning model constructed based on the aligned samples cannot perform longitudinal federal learning prediction on the unaligned samples, resulting in low sample prediction accuracy of the unaligned samples, and / or solves the technical problem that the unaligned samples cannot be added to the longitudinal federal learning modeling process. Compared with the prior art, the beneficial effects of the computer readable storage medium provided by the embodiments of the present application are the same as those of the longitudinal federal prediction method and / or the longitudinal federal model construction optimization method provided by the above embodiments, and are not described here.

[0291] Embodiment Eleven

[0292] The present application also provides a computer program product comprising a computer program which, when executed by a processor, implements the steps of the longitudinal federal prediction method and / or the longitudinal federal model construction optimization method as described above.

[0293] The computer program product provided in the application solves the technical problems that the longitudinal federated learning model constructed based on aligned samples cannot perform longitudinal federated learning prediction on unaligned samples, resulting in low sample prediction accuracy of unaligned samples, and / or solves the technical problem that unaligned samples cannot join the longitudinal federated learning modeling process. Compared with the prior art, the beneficial effects of the computer program product provided in the embodiments of the application are the same as those of the longitudinal federated prediction method and / or the longitudinal federated model construction optimization method provided in the above embodiments, which will not be repeated here.

[0294] The above is only the preferred embodiment of the application, and does not limit the patent scope of the application, and any equivalent structure or equivalent process transformation using the content of the specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent processing scope of the application.

Claims

1. A longitudinal federal forecasting method characterized by, The longitudinal federated prediction method applied to a first participant device comprises: transformation of the first sample features to a federated server, wherein the federated server is configured to generate a fuzzy matching result between the unaligned sample and each second sample according to the first sample transformation features and each second sample transformation features uploaded by a second participant device, wherein the second sample transformation features are obtained by performing feature transformation on each second sample according to a second-party federated feature transformation model by the second participant device; feature extraction of the unaligned sample based on a first-party federated prediction feature extraction model to obtain first-party prediction sample features; receiving second-party prediction sample features sent by the second participant device, and performing longitudinal federated prediction according to the first-party prediction sample features and the second-party prediction sample features to obtain a longitudinal federated prediction result, wherein the second-party prediction sample features are obtained by the second participant device based on model prediction outputs of each second sample and the fuzzy matching result, and the model prediction outputs are obtained by performing feature extraction on the second sample by a second-party federated prediction feature extraction model. Before the step of performing feature transformation on the unaligned sample according to the first-party federated feature transformation model to obtain the first sample transformation features, the longitudinal federated prediction method further comprises:

2. The method of claim 1, wherein, obtaining first-party samples, performing feature extraction on the first-party samples based on a first-party local feature extraction model to obtain first local sample features; performing feature mapping on the first local sample features based on a first-party local feature mapping model to obtain second local sample features; receiving third local sample features sent by the second participant device, wherein the third local sample features are obtained by the second participant device according to a second-party local feature extraction model by performing feature extraction on second-party samples; updating the first-party local feature extraction model and the first-party local feature mapping model according to a similarity loss between the second local sample features and the third local sample features to obtain the first-party federated feature transformation model; sending the first local sample features to the second participant device, so that the second participant device updates the second-party local feature extraction model and a second-party local feature mapping model according to a similarity loss between the first local sample features and fourth local sample features to obtain the second-party federated feature transformation model, wherein the fourth local sample features are obtained by the second participant device according to the second-party local feature mapping model by performing feature mapping on the third local sample features. The longitudinal federated prediction method applied to a second participant device comprises:

3. A longitudinal federal forecasting method characterized by, performing feature transformation on each second sample according to a second-party federated feature transformation model to obtain each second sample transformation feature; ​ uploading each of the second sample transformed features to a federated server, wherein the federated server is configured to generate a fuzzy matching result between the unaligned samples and each of the second party samples according to each of the second sample transformed features and first sample transformed features uploaded by a first participant device, the first sample transformed features being obtained by the first participant device from feature transformation of the unaligned samples according to a first party federated feature transformation model; receiving the fuzzy matching result fed back by the federated server, and constructing a second party predicted sample feature according to a model prediction output of each of the second party samples and the fuzzy matching result according to a second party federated prediction feature extraction model; sending the second party predicted sample feature to the first participant device, so that the first participant device performs longitudinal federated prediction on a first party predicted sample feature of the unaligned samples based on a first party federated prediction feature extraction model and the second party predicted sample feature.

4. The vertical federated forecasting method as described in claim 3, characterized in that, The fuzzy matching result at least includes a sample feature weight, and the step of constructing a second party predicted sample feature according to a model prediction output of each of the second party samples and the fuzzy matching result according to a second party federated prediction feature extraction model comprises: based on a second party sample identifier corresponding to each of the sample feature weights, filtering corresponding sample inputs in each of the second party samples into the second party federated prediction feature extraction model for feature extraction to obtain each target model prediction output; according to each of the sample feature weights, weighting and aggregating each of the target model prediction outputs to obtain the second party predicted sample feature.

5. The method of claim 3, wherein the step of determining the future position of the vehicle comprises the step of: Before the step of performing feature transformation on each of the second party samples according to a second party federated feature transformation model to obtain each of the second sample transformed features, the longitudinal federated prediction method further comprises: ​ obtaining a second party sample, performing feature extraction on the second party sample based on a second party local feature extraction model to obtain a third local sample feature; performing feature mapping on the third local sample feature based on a second party local feature mapping model to obtain a fourth local sample feature; receiving a first local sample feature sent by the first participant device, wherein the first local sample feature is obtained by the first participant device from feature extraction of a first party sample according to a first party local feature extraction model; updating the second party local feature extraction model and the second party local feature mapping model according to a similarity loss between the fourth local sample feature and the first local sample feature to obtain the second party federated feature transformation model; sending the third local sample feature to the first participant device, so that the first participant device updates the first party local feature extraction model and the first party local feature mapping model according to a similarity loss between a second local sample feature and the third local sample feature to obtain the first party federated feature transformation model, the second local sample feature being obtained by the first participant device from feature mapping of the first local sample feature according to the first party local feature mapping model.

6. A longitudinal federal forecasting method characterized by, The longitudinal federated prediction method applied to a federated server comprises: receive the first sample transformation feature sent by the first participant device and each second sample transformation feature sent by the second participant device, wherein the first sample transformation feature is obtained by the first participant device from feature transformation of the unaligned sample according to a first federated feature transformation model, and the second sample transformation feature is obtained by the second participant device from feature transformation of a second-party sample according to a second federated feature transformation model; generate a fuzzy matching result between the unaligned sample and each second-party sample according to the first sample transformation feature and each second sample transformation feature; send the fuzzy matching result to the second participant device, wherein the second participant device is configured to construct a second-party predicted sample feature from a model predicted output of each second-party sample and the fuzzy matching result according to a second-party federated predicted feature extraction model, and the first participant device is configured to perform longitudinal federated prediction according to a first-party predicted sample feature and the second-party predicted sample feature sent by the second participant device, wherein the first-party predicted sample feature is obtained by feature extraction of the unaligned sample by a first-party federated predicted feature extraction model.

7. The vertical federated forecasting method as described in claim 6, characterized in that, The fuzzy matching result at least includes a sample feature weight, and the step of generating the fuzzy matching result between the unaligned sample and each second-party sample according to the first sample transformation feature and each second sample transformation feature includes: calculating feature similarity between the first sample transformation feature and each second sample transformation feature; screening at least one target sample transformation feature from each second sample transformation feature according to each feature similarity; normalizing each target sample transformation feature to obtain each sample feature weight.

8. The vertical federated forecasting method as described in claim 6, characterized in that, The fuzzy matching result at least includes a sample feature weight, and the step of generating the fuzzy matching result between the unaligned sample and each second-party sample according to the first sample transformation feature and each second sample transformation feature includes: splicing the first sample transformation feature with each second sample transformation feature to obtain each spliced feature; identifying the probability of the first sample transformation feature and each second sample transformation feature corresponding to the same sample respectively by classifying each spliced feature to obtain each classification probability; screening at least one target sample transformation feature from each second sample transformation feature according to each classification probability; normalizing each target sample transformation feature to obtain each sample feature weight.

9. A method for constructing a longitudinal federal model optimization, characterized in that, The longitudinal federated model construction optimization method applied to the first participant device includes: extracting a first-party training sample from a training sample set, wherein the training sample set is composed of unaligned samples and first-party aligned samples; if the first-party training sample is an unaligned sample, performing feature transformation on the first-party training sample according to a first-party federated feature transformation model to obtain a first sample transformation feature; The first sample transformation feature is uploaded to the federated server, wherein the federated server is used to generate fuzzy matching results between the first party training sample and each second party sample based on the first sample transformation feature and each second sample transformation feature uploaded by the second participating party device, wherein the second sample transformation feature is obtained by the second participating party device performing feature transformation on the second party sample according to the second party federated feature transformation model; Based on the first-party federated prediction feature extraction model, feature extraction is performed on the first-party training samples to obtain the features of the first-party prediction samples. The system receives second-party predicted sample features sent by the second participating device, wherein the second-party predicted sample features are constructed by the second participating device based on the model prediction output of each second-party sample and the fuzzy matching result, and the model prediction output is obtained by the second-party federated prediction feature extraction model to extract features from the second-party samples; Vertical federated learning modeling is performed based on the features of the first-party predicted samples and the features of the second-party predicted samples to iteratively update the first-party federated prediction feature extraction model and the second-party federated prediction feature extraction model.

10. The method for constructing and optimizing a vertical federation model as described in claim 9, characterized in that, The optimization method for constructing the vertical federation model also includes: First-party samples are obtained, and features are extracted from the first-party samples based on the first-party local feature extraction model to obtain the features of the first local samples. Based on the first-party local feature mapping model, feature mapping is performed on the first local sample features to obtain the second local sample features; The device receives a third local sample feature sent by the second participating party device, wherein the third local sample feature is obtained by the second participating party device extracting features from the second party sample based on the second party local feature extraction model; Based on the similarity loss between the second local sample features and the third local sample features, the first local feature extraction model and the first local feature mapping model are updated to obtain the first federated feature transformation model. The first local sample feature is sent to the second participating device, so that the second participating device can update the second local feature extraction model and the second local feature mapping model based on the similarity loss between the first local sample feature and the fourth local sample feature to obtain the second federated feature transformation model. The fourth local sample feature is obtained by the second participating device performing feature mapping on the third local sample feature based on the second local feature mapping model.

11. A method for constructing a longitudinal federal model optimization, characterized in that, Applied to the second participating party's equipment, the vertical federation model construction optimization method includes: Based on the second-party federated feature transformation model, feature transformation is performed on each second-party sample to obtain the transformed features of each second sample. uploading each of the second sample transformation features to a federation server, wherein the federation server is configured to generate a fuzzy matching result between the unaligned sample and each of the second party samples according to each of the second sample transformation features and first sample transformation features uploaded by a first participant device, the first sample transformation features being obtained by the first participant device by performing feature transformation on the unaligned sample according to a first party federation feature transformation model; receiving the fuzzy matching result fed back by the federation server, and constructing a second party predicted sample feature according to a model prediction output of each of the second party samples and the fuzzy matching result according to a second party federation prediction feature extraction model; sending the second party predicted sample feature to the first participant device, so that the first participant device performs longitudinal federation learning modeling on a first party predicted sample feature of the unaligned sample based on a first party federation prediction feature extraction model and the second party predicted sample feature.

12. The method of claim 11, wherein the longitudinal federal model construction optimization is performed by a computer system. The fuzzy matching result at least includes a sample feature weight, and the step of constructing a second party predicted sample feature according to a model prediction output of each of the second party samples and the fuzzy matching result according to a second party federation prediction feature extraction model includes: performing feature extraction on a corresponding sample input into the second party federation prediction feature extraction model in each of the second party samples based on a corresponding second party sample identifier of each of the sample feature weights, to obtain each target model prediction output; performing weighted aggregation on each of the target model prediction outputs according to each of the sample feature weights, to obtain the second party predicted sample feature.

13. The method of claim 11, wherein the longitudinal federal model construction optimization is performed by a computer system. The longitudinal federation model construction optimization method further includes: obtaining a second party sample, performing feature extraction on the second party sample based on a second party local feature extraction model, to obtain a third local sample feature; performing feature mapping on the third local sample feature based on a second party local feature mapping model, to obtain a fourth local sample feature; receiving a first local sample feature sent by the first participant device, wherein the first local sample feature is obtained by the first participant device by performing feature extraction on a first party sample according to a first party local feature extraction model; updating the second party local feature extraction model and the second party local feature mapping model according to a similarity loss between the fourth local sample feature and the first local sample feature, to obtain the second party federation feature transformation model; sending the third local sample feature to the first participant device, so that the first participant device updates the first party local feature extraction model and a first party local feature mapping model according to a similarity loss between a second local sample feature and the third local sample feature, to obtain the first party federation feature transformation model, the second local sample feature being obtained by the first participant device by performing feature mapping on the first local sample feature according to the first party local feature mapping model.

14. A method for constructing and optimizing a vertical federation model, characterized in that, The longitudinal federation model construction optimization method applied to a federation server includes: receive the first sample transformation feature sent by the first participant device and each second sample transformation feature sent by the second participant device, wherein the first sample transformation feature is obtained by the first participant device from feature transformation of the unaligned sample according to a first federated feature transformation model, and the second sample transformation feature is obtained by the second participant device from feature transformation of a second-party sample according to a second federated feature transformation model; generate a fuzzy matching result between the unaligned sample and each second-party sample according to the first sample transformation feature and each second sample transformation feature; send the fuzzy matching result to the second participant device, wherein the second participant device is configured to construct a second-party predicted sample feature from a model prediction output of each second-party sample and the fuzzy matching result according to a second-party federated prediction feature extraction model, and the first participant device is configured to perform longitudinal federated learning modeling according to a first-party predicted sample feature and the second-party predicted sample feature sent by the second participant device, wherein the first-party predicted sample feature is obtained by feature extraction of the unaligned sample by a first-party federated prediction feature extraction model.

15. The method of claim 14, wherein the longitudinal federal model construction optimization is performed by a computer system. The fuzzy matching result at least includes a sample feature weight, and the step of generating the fuzzy matching result between the unaligned sample and each second-party sample according to the first sample transformation feature and each second sample transformation feature includes: ​ calculating feature similarity between the first sample transformation feature and each second sample transformation feature; selecting at least one target sample transformation feature from each second sample transformation feature according to each feature similarity; normalizing each target sample transformation feature to obtain each sample feature weight.

16. The method of claim 14, wherein the longitudinal federal model construction optimization is performed by a computer system. The fuzzy matching result at least includes a sample feature weight, and the step of generating the fuzzy matching result between the unaligned sample and each second-party sample according to the first sample transformation feature and each second sample transformation feature includes: splicing the first sample transformation feature with each second sample transformation feature to obtain each spliced feature; identifying the probability of the first sample transformation feature and each second sample transformation feature corresponding to the same sample respectively by classifying each spliced feature to obtain each classification probability; selecting at least one target sample transformation feature from each second sample transformation feature according to each classification probability; normalizing each target sample transformation feature to obtain each sample feature weight.

17. A longitudinal federal forecasting system characterized by, The longitudinal federated prediction system includes: a first participant device configured to perform feature transformation on an unaligned sample according to a first-party federated feature transformation model to obtain a first sample transformation feature, upload the first sample transformation feature to a federated server, perform feature extraction on the unaligned sample according to a first-party federated prediction feature extraction model to obtain a first-party predicted sample feature, receive a second-party predicted sample feature sent by a second participant device, and perform longitudinal federated prediction according to the first-party predicted sample feature and the second-party predicted sample feature to obtain a longitudinal federated prediction result. The second participant device performs feature transformation on each second-party sample according to a second-party federated feature transformation model to obtain each second sample transformed feature; uploads each second sample transformed feature to the federated server; receives the fuzzy matching result fed back by the federated server, and constructs a second-party predicted sample feature according to the model prediction output of each second-party sample and the fuzzy matching result based on a second-party federated prediction feature extraction model; The second-party predicted sample feature is sent to the first participant device; The federated server receives the first sample transformed feature sent by the first participant device and each second sample transformed feature sent by the second participant device; generates a fuzzy matching result between the unaligned sample and each second-party sample according to the first sample transformed feature and each second sample transformed feature; and sends the fuzzy matching result to the second participant device.

18. A longitudinal federal model construction optimization system, comprising: The longitudinal federated model construction optimization system comprises: The first participant device is configured to extract a first-party training sample from a training sample set, wherein the training sample set comprises unaligned samples and first-party aligned samples; if the first-party training sample is an unaligned sample, perform feature transformation on the first-party training sample according to a first-party federated feature transformation model to obtain a first sample transformed feature; upload the first sample transformed feature to a federated server; perform feature extraction on the first-party training sample based on a first-party federated prediction feature extraction model to obtain a first-party predicted sample feature; receive a second-party predicted sample feature sent by a second participant device; and perform longitudinal federated learning modeling according to the first-party predicted sample feature and the second-party predicted sample feature to iteratively update the first-party federated prediction feature extraction model and the second-party federated prediction feature extraction model; The second participant device is configured to perform feature transformation on each second-party sample according to a second-party federated feature transformation model to obtain each second sample transformed feature; upload each second sample transformed feature to the federated server; receive the fuzzy matching result fed back by the federated server, and construct a second-party predicted sample feature according to the model prediction output of each second-party sample and the fuzzy matching result based on a second-party federated prediction feature extraction model; and send the second-party predicted sample feature to the first participant device; The federated server is configured to receive the first sample transformed feature sent by the first participant device and each second sample transformed feature sent by the second participant device; generate a fuzzy matching result between the unaligned sample and each second-party sample according to the first sample transformed feature and each second sample transformed feature; and send the fuzzy matching result to the second participant device.

19. An electronic device, comprising: The electronic device comprises: at least one processor; and a memory connected to the at least one processor in communication; wherein The memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the vertical federated prediction method of any one of claims 1 to 8, and / or to perform the steps of the vertical federated model construction optimization method of any one of claims 9 to 16.

20. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a program for implementing a vertical federated prediction method and / or a program for implementing a vertical federated model construction optimization method, the program for implementing a vertical federated prediction method being executed by a processor to implement the steps of the vertical federated prediction method of any one of claims 1 to 8, and / or the program for implementing a vertical federated model construction optimization method being executed by a processor to implement the steps of the vertical federated model construction optimization method of any one of claims 9 to 16.

21. A computer program product comprising a computer program, characterized in that, The computer program, when executed by a processor, implements the steps of the vertical federated prediction method of any one of claims 1 to 8, and / or the steps of the vertical federated model construction optimization method of any one of claims 9 to 16.

Images