CN113688986A - Longitudinal federal prediction optimization method, device, medium, and computer program product - Google Patents

Longitudinal federal prediction optimization method, device, medium, and computer program product Download PDF

Info

Publication number
CN113688986A
CN113688986A CN202110982929.9A CN202110982929A CN113688986A CN 113688986 A CN113688986 A CN 113688986A CN 202110982929 A CN202110982929 A CN 202110982929A CN 113688986 A CN113688986 A CN 113688986A
Authority
CN
China
Prior art keywords
model
sample
prediction
party
training
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202110982929.9A
Other languages
Chinese (zh)
Inventor
万晟
高大山
鞠策
谭奔
杨强
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
WeBank Co Ltd
Original Assignee
WeBank Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by WeBank Co Ltd filed Critical WeBank Co Ltd
Priority to CN202110982929.9A priority Critical patent/CN113688986A/en
Publication of CN113688986A publication Critical patent/CN113688986A/en
Priority to PCT/CN2021/139640 priority patent/WO2023024349A1/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The application discloses a longitudinal federal forecast optimization method, equipment, a medium and a computer program product, which comprise the following steps: extracting a sample to be predicted, and generating a first party model prediction result and corresponding intermediate sample characteristics corresponding to the sample to be predicted based on a target prediction model of local iterative training; sending the intermediate sample characteristics to second equipment so that the second equipment can jointly execute model prediction on the intermediate sample characteristics and the ID matching samples corresponding to the samples to be predicted based on a longitudinal federated residual error lifting model to obtain a second square model prediction result; acquiring a first party model weight, and receiving a second party model prediction result and a second party model weight sent by second equipment; and carrying out weighted aggregation on the prediction result of the first party model and the prediction result of the second party model based on the weight of the first party model and the weight of the second party model to obtain a target federal prediction result. The method and the device solve the technical problem that the prediction accuracy of the whole sample of longitudinal federal prediction is low.

Description

Longitudinal federal prediction optimization method, device, medium, and computer program product
Technical Field
The present application relates to the field of artificial intelligence in financial technology (Fintech), and in particular, to a method, apparatus, medium, and computer program product for longitudinal federal forecast optimization.
Background
With the continuous development of financial science and technology, especially internet science and technology, more and more technologies (such as distributed technology, artificial intelligence and the like) are applied to the financial field, but the financial industry also puts higher requirements on the technologies, for example, higher requirements on the distribution of backlog in the financial industry are also put forward.
With the continuous development of computer software and artificial intelligence computing, the application of artificial intelligence technology is also more and more extensive. At present, an existing longitudinal federal prediction model needs to be constructed by performing longitudinal federal learning modeling based on alignment samples among all participants, and after the modeling is completed, the longitudinal federal prediction model is dispersedly deployed in all the participants, and each participant only holds part of the longitudinal federal prediction model. Therefore, in a vertical federal prediction scenario, for aligned samples, the predictor usually performs vertical federal prediction by combining aligned samples scattered in other participants to accurately perform sample prediction. For misaligned samples, the predictor needs to use a separate local model to predict locally. Therefore, when the predictor performs federal prediction with higher accuracy on aligned samples, the predictor cannot perform sample prediction based on the complete model on unaligned samples, so that the overall sample prediction accuracy becomes lower. Therefore, the existing longitudinal federal prediction method has the problem of low overall sample prediction accuracy.
Disclosure of Invention
The main purpose of the present application is to provide a method, device, medium, and computer program product for optimizing longitudinal federation prediction, which aim to solve the technical problem in the prior art that the accuracy of predicting an entire sample for longitudinal federation prediction is low.
In order to achieve the above object, the present application provides a longitudinal federal prediction optimization method, which is applied to a first device, and includes:
extracting a sample to be predicted, and obtaining an intermediate sample feature generated by a feature extractor of a target prediction model for performing feature extraction on the sample to be predicted and a first party model prediction result generated by the target prediction model for performing model prediction on the sample to be predicted, wherein the target prediction model is obtained by local iterative training of first equipment;
sending the intermediate sample features to second equipment, so that the second equipment jointly executes model prediction on the intermediate sample features and the ID matching samples corresponding to the samples to be predicted based on a longitudinal federated residual error lifting model, and a second-party model prediction result is obtained;
obtaining a first-party model weight corresponding to the target prediction model, and receiving a second-party model prediction result sent by the second device and a second-party model weight corresponding to the longitudinal federal residual error hoisting model;
and carrying out weighted aggregation on the first party model prediction result and the second party model prediction result based on the first party model weight and the second party model weight to obtain a target federal prediction result.
In order to achieve the above object, the present application provides a longitudinal federal prediction optimization method, which is applied to a second device, and includes:
receiving an intermediate sample characteristic sent by first equipment, and searching an ID matching sample corresponding to the intermediate sample characteristic;
performing model prediction on the ID matching sample and the intermediate sample feature together based on a longitudinal federated residual lifting model to obtain a second-party model prediction result;
and obtaining a second square model weight corresponding to the longitudinal federated residual lifting model, and sending the second square model prediction result and the second square model weight to the first device, so that the first device generates a target federated prediction result based on a first square model prediction result generated by the target prediction model for a to-be-predicted sample corresponding to the ID matching sample, a first square model weight corresponding to the target prediction model, the second square model prediction result and the second square model weight, wherein the target prediction model is obtained by local iterative training of the first device.
In order to achieve the above object, the present application provides a longitudinal federated learning modeling optimization method, where the longitudinal federated learning modeling optimization method is applied to a first device, and the longitudinal federated learning modeling optimization method includes:
acquiring the weight of a first party initial model, and extracting training samples and training sample labels corresponding to the training samples;
acquiring intermediate training sample characteristics generated by characteristic extraction of a characteristic extractor of a target prediction model to be trained aiming at the training samples;
iteratively optimizing the target prediction model to be trained by calculating the prediction loss of the first party model corresponding to the target prediction model to be trained based on the training sample label, the training model prediction result corresponding to the intermediate training sample characteristic and the first party initial model weight to obtain the target prediction model;
and sending the training sample label, the intermediate training sample characteristics and the first party model prediction loss to second equipment so that the second equipment can calculate second party model prediction loss, and optimizing the residual lifting model to be trained based on the residual loss calculated by the second party model prediction loss and the first party model prediction loss to obtain a longitudinal federal residual lifting model.
In order to achieve the above object, the present application provides a longitudinal federated learning modeling optimization method, where the longitudinal federated learning modeling optimization method is applied to a second device, and the longitudinal federated learning modeling optimization method includes:
acquiring the weight of a second-party initial model, and receiving the intermediate training sample characteristics, the training sample labels and the first-party model prediction loss sent by first equipment;
acquiring a training sample ID matching sample, and performing model prediction on the training sample ID matching sample and the intermediate training sample feature together based on a residual lifting model to be trained to obtain a second-party training model prediction result;
calculating a second-party model prediction loss based on the training sample labels, the second-party initial model weights and the second-party training model prediction results;
and iteratively optimizing the residual lifting model to be trained based on the residual losses generated by the prediction loss of the first square model and the prediction loss of the second square model to obtain the longitudinal federal residual lifting model.
The application also provides a longitudinal federal forecast optimizing device, which is a virtual device and applied to a first device, and the longitudinal federal forecast optimizing device comprises:
the model prediction module is used for extracting a sample to be predicted, obtaining an intermediate sample feature generated by a feature extractor of a target prediction model aiming at the sample to be predicted and performing feature extraction on the sample to be predicted, and obtaining a first party model prediction result generated by the target prediction model aiming at the sample to be predicted, wherein the target prediction model is obtained by local iterative training of the first equipment;
the sending module is used for sending the intermediate sample characteristics to second equipment so that the second equipment can jointly execute model prediction on the intermediate sample characteristics and the ID matching samples corresponding to the samples to be predicted based on a longitudinal federated residual error lifting model to obtain a second square model prediction result;
the receiving module is used for acquiring a first-party model weight corresponding to the target prediction model, and receiving a second-party model prediction result sent by the second equipment and a second-party model weight corresponding to the longitudinal federal residual error hoisting model;
and the weighted aggregation module is used for carrying out weighted aggregation on the first party model prediction result and the second party model prediction result based on the first party model weight and the second party model weight to obtain a target federal prediction result.
The application also provides a longitudinal federal forecast optimizing device, which is a virtual device and applied to a second device, and the longitudinal federal forecast optimizing device comprises:
the receiving and searching module is used for receiving the intermediate sample characteristics sent by the first equipment and searching the ID matching sample corresponding to the intermediate sample characteristics;
the model prediction module is used for jointly performing model prediction on the ID matching sample and the intermediate sample characteristics based on a longitudinal federal residual error hoisting model to obtain a second-party model prediction result;
a sending module, configured to obtain a second square model weight corresponding to the longitudinal federated residual lifting model, and send the second square model prediction result and the second square model weight to the first device, so that the first device generates a target federated prediction result based on a first square model prediction result generated by the target prediction model for the to-be-predicted sample corresponding to the ID matching sample, a first square model weight corresponding to the target prediction model, the second square model prediction result, and the second square model weight, where the target prediction model is obtained by local iterative training of the first device.
The application also provides a vertical federal learning modeling optimization device, vertical federal learning modeling optimization device is virtual device, just vertical federal learning modeling optimization device is applied to first equipment, vertical federal learning modeling optimization device includes:
the first acquisition module is used for acquiring the initial model weight of a first party and extracting training samples and training sample labels corresponding to the training samples;
the second acquisition module is used for acquiring the characteristics of an intermediate training sample generated by the characteristic extractor of the target prediction model to be trained aiming at the characteristic extraction of the training sample;
the iterative optimization module is used for iteratively optimizing the target prediction model to be trained by calculating the prediction loss of the first-party model corresponding to the target prediction model to be trained based on the training sample labels, the training model prediction result corresponding to the intermediate training sample characteristics and the first-party initial model weight to obtain the target prediction model;
and the sending module is used for sending the training sample label, the intermediate training sample characteristics and the first party model prediction loss to second equipment so that the second equipment can calculate second party model prediction loss, and optimizing the to-be-trained residual lifting model based on the second party model prediction loss and the residual loss calculated by the first party model prediction loss to obtain a longitudinal federal residual lifting model.
The application also provides a vertical federal learning modeling optimization device, vertical federal learning modeling optimization device is virtual device, just vertical federal learning modeling optimization device is applied to the second equipment, vertical federal learning modeling optimization device includes:
the receiving module is used for acquiring the initial model weight of the second party and receiving the intermediate training sample characteristics, the training sample labels and the prediction loss of the first party model sent by the first equipment;
the model prediction module is used for acquiring a training sample ID matching sample, and performing model prediction on the training sample ID matching sample and the intermediate training sample based on a residual lifting model to be trained to obtain a second-party training model prediction result;
a calculation module for calculating a second party model prediction loss based on the training sample labels, the second party initial model weights and the second party training model prediction results;
and the iterative optimization module is used for iteratively optimizing the to-be-trained residual lifting model based on the residual losses generated by the first-party model prediction loss and the second-party model prediction loss to obtain the longitudinal federal residual lifting model.
The application also provides a longitudinal federal forecast optimization device, which is an entity device, and the longitudinal federal forecast optimization device comprises: a memory, a processor, and a program of the longitudinal federal predictive optimization method stored in the memory and executable on the processor, the program of the longitudinal federal predictive optimization method being executable by the processor to implement the steps of the longitudinal federal predictive optimization method as described above.
The present application further provides a longitudinal federated learning modeling optimization apparatus, which is an entity apparatus, the longitudinal federated learning modeling optimization apparatus including: a memory, a processor, and a program of the longitudinal federated learning modeling optimization method stored on the memory and executable on the processor, which when executed by the processor, may implement the steps of the longitudinal federated learning modeling optimization method as described above.
The present application also provides a medium, which is a readable storage medium, and the readable storage medium stores a program for implementing the longitudinal federal prediction optimization method, and the program for implementing the longitudinal federal prediction optimization method implements the steps of the longitudinal federal prediction optimization method as described above when executed by a processor.
The present application also provides a medium, which is a readable storage medium, on which a program for implementing the longitudinal federated learning modeling optimization method is stored, and the program for implementing the longitudinal federated learning modeling optimization method implements the steps of the longitudinal federated learning modeling optimization method as described above when executed by a processor.
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 optimization method as described above.
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 learning modeling optimization methodology as described above.
Compared with the technical means that for aligned samples in a longitudinal federated prediction scene, a predictor jointly performs longitudinal federated prediction on aligned samples scattered in other participants to accurately perform sample prediction, and for unaligned samples, the predictor performs local prediction on the basis of a locally held partial longitudinal federated prediction model, the method comprises the steps of firstly extracting samples to be predicted, obtaining intermediate sample characteristics generated by a characteristic extractor of a target prediction model by performing characteristic extraction on the samples to be predicted, and obtaining a first party model prediction result generated by the target prediction model by performing model prediction on the samples to be predicted, wherein the target prediction model is obtained by local iterative training of first equipment, so that a target prediction model based on local iterative training is realized, therefore, for a sample to be predicted which is a non-aligned sample, the first device can independently and accurately predict the sample to be predicted based on a target prediction model serving as a complete model, and further send the intermediate sample feature to the second device, so that the second device can jointly perform model prediction on the intermediate sample feature and an ID (identity) matching sample corresponding to the sample to be predicted based on a longitudinal federated residual error boosting model, more decision bases are provided for the second device to perform model prediction based on the longitudinal federated residual error boosting model, and the accuracy of a second square model prediction result generated by the second device is improved, wherein the longitudinal federated residual error boosting model is formed by the second device based on a longitudinal federated common sample and combines model prediction loss, model prediction loss and model prediction loss of the target prediction model on the longitudinal federated common sample in the first device, The middle public sample characteristics corresponding to the longitudinal federal public sample and the corresponding sample labels are obtained by residual error learning based on longitudinal federal learning with the first device, and then for the ID matched sample aligned with the sample to be predicted, the second device can generate residual error lifting information with higher accuracy corresponding to the sample to be predicted, namely a second square model prediction result, by performing model prediction on the ID matched sample and the middle sample characteristics based on the longitudinal federal residual error lifting model, further obtain the first square model weight corresponding to the target prediction model, receive the second square model prediction result sent by the second device and the second square model weight corresponding to the longitudinal federal residual error lifting model, and perform weighted aggregation on the first square model prediction result and the second square model prediction result based on the first square model weight and the second square model weight, therefore, for a sample to be predicted of an aligned sample between first equipment and second equipment, the first equipment can improve the accuracy of a first-party model prediction result output by a target prediction model by using the second equipment based on the longitudinal federal residual lifting model aiming at the ID matched sample aligned with the sample to be predicted and the residual lifting information (second-party model prediction result) with higher accuracy generated by the middle sample characteristic, and realize the longitudinal federal prediction based on the residual lifting information with higher accuracy on the sample to be predicted, so that the aim of accurately predicting a sample which is not aligned locally and independently based on a complete model can be achieved under the condition of longitudinally federal prediction based on the residual lifting information with higher accuracy on the aligned sample, and the problem that a predicting party performs the federal prediction with higher accuracy on the aligned sample is overcome, the method has the technical defect that the integral sample prediction accuracy is low because the sample prediction based on the complete model cannot be carried out on the unaligned sample, and the integral sample prediction accuracy of longitudinal federal prediction is improved.
Drawings
The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and together with the description, serve to explain the principles of the application.
In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly described below, and it is obvious for those skilled in the art to obtain other drawings without inventive exercise.
FIG. 1 is a schematic flow chart of a first embodiment of a longitudinal federal prediction optimization method of the present application;
FIG. 2 is a schematic flow chart of a second embodiment of the longitudinal federal prediction optimization method of the present application;
FIG. 3 is a schematic flow chart of a first embodiment of a longitudinal federated learning modeling optimization method of the present application;
FIG. 4 is a schematic flow chart diagram of a second embodiment of the longitudinal federated learning modeling optimization method of the present application;
FIG. 5 is a schematic structural diagram of a hardware operating environment related to a longitudinal federated prediction optimization method in the embodiment of the present application;
fig. 6 is a schematic device structure diagram of a hardware operating environment related to the longitudinal federated learning modeling optimization method in the embodiment of the present application.
The objectives, features, and advantages of the present application will be further described with reference to the accompanying drawings.
Detailed Description
It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application.
In a first embodiment of the longitudinal federal prediction optimization method, referring to fig. 1, the longitudinal federal prediction optimization method is applied to a first device, and includes:
step S10, extracting a sample to be predicted, and obtaining an intermediate sample feature generated by a feature extractor of a target prediction model performing feature extraction on the sample to be predicted, and a first party model prediction result generated by the target prediction model performing model prediction on the sample to be predicted, where the target prediction model is obtained by local iterative training of the first device.
In this embodiment, it should be noted that the longitudinal federated prediction optimization method is applied to a longitudinal federated learning scenario, and both the first device and the second device are participants of the longitudinal federated learning scenario. The first device is provided with a sample carrying a sample label, the second device is provided with a sample without a sample label, the first device is a predictor and used for executing a prediction task, and the second device is an auxiliary data provider and used for providing residual error lifting information for the first device so as to improve the accuracy of a prediction result generated by the first device.
In an implementation manner, the target prediction model may be a multi-layer neural network or a depth factorization model, and is used for classifying the sample to be predicted. For example, assuming that the output of the last fully-connected layer of the target prediction model is z, and the activation function is a sigmoid function, the classification result P is sigmoid (z), where P is the probability that the sample to be predicted belongs to the preset sample class.
Specifically, in step S10, a sample to be predicted is extracted, feature extraction is performed on the sample to be predicted based on a feature extractor in a target prediction model to obtain an intermediate sample feature, full connection is performed on the intermediate sample feature based on a classifier in the target prediction model to obtain a target full connection layer output, and the target full connection layer output is converted into a first-party model prediction result by a preset activation function, where the first-party model prediction result may be a classification probability.
Before the steps of performing feature extraction on the sample to be predicted by the feature extractor for obtaining the target prediction model to generate an intermediate sample feature and performing model prediction on the sample to be predicted by the target prediction model to generate a first-party model prediction result, the longitudinal federated prediction optimization method further includes:
step A10, acquiring the weight of the initial model of the first party, and extracting training samples and training sample labels corresponding to the training samples;
in this embodiment, it should be noted that the number of the training samples is at least 1, the training sample label is an identifier of the training sample, and the first-party initial model weight is an initial value of a model weight that is preset in the first device and represents the prediction accuracy of the model. Wherein, the model weight can be set to be equal to the logarithm value of the ratio of the number of samples with correct classification of the prediction model to the number of samples with wrong classification, and optionally, the initial value of the model weight can be set to be 1.
Step A20, acquiring intermediate training sample characteristics generated by a characteristic extractor of a target prediction model to be trained for performing characteristic extraction on the training samples;
in this embodiment, feature extraction is performed on the training sample based on a feature extractor in the target prediction model to be trained, so as to generate an intermediate training sample feature.
Step A30, based on the training sample labels, the training model prediction results corresponding to the intermediate training sample characteristics and the first-party initial model weight, iteratively optimizing the target prediction model to be trained by calculating the first-party model prediction loss corresponding to the target prediction model to be trained, so as to obtain the target prediction model;
in this embodiment, it should be noted that the iterative training process of the target prediction model to be trained includes a plurality of iterative rounds, where one iterative round needs to be iteratively trained based on a preset number of training samples.
Specifically, in step a30, based on the classifier and the preset activation function in the target prediction model to be trained, the intermediate training sample features are converted into a training model prediction result, and based on the training sample, the training model prediction result, and the first-party initial model weight, a first-party model prediction loss is calculated, so as to determine whether the first-party model prediction loss converges. If the prediction loss of the first party model is converged, taking the target prediction model to be trained as the target prediction model; if the prediction loss of the first party model is not converged, updating the target prediction model to be trained by a preset model optimization method based on the gradient calculated by the prediction loss of the first party model, updating the weight of the first party initial model based on the prediction result of the training model and the training sample label, and returning to the execution step: and extracting training samples and training sample labels corresponding to the training samples, and performing next iteration. The preset model optimization method comprises a gradient descent method, a gradient ascent method and the like, and based on the training samples, the training model prediction result and the first-party initial model weight, a loss function of the first-party model prediction loss is calculated as follows:
Figure BDA0003229552350000101
wherein L isA(θA,αA,XAY) is the first-party model prediction loss, NAFor the number of training samples in a round of iteration, θAFor the target prediction model to be trained, αAFor the first party initial model weight, XAIs NAA training sample set composed of training samples, Y is NAA label set formed by training sample labels corresponding to the training samples, yiFor the ith training sample label, x, in one iterationA,iIs the feature of the ith training sample in a round of iteration.
Further, after step a30, the first device counts the number of samples classified correctly and the number of samples classified incorrectly in the iterative training process of the target prediction model, obtains the number of samples classified correctly by the first party and the number of samples classified incorrectly by the first party, and further generates a first party model weight by calculating the ratio of the number of samples classified correctly by the first party to the number of samples classified incorrectly by the first party, wherein the specific process of calculating the first party model weight is as follows:
Figure BDA0003229552350000102
wherein alpha isAAnd B is the number of the first party classified error samples.
Step A40, the training sample label, the middle training sample feature and the first party model prediction loss are sent to a second device, so that the second device calculates a second party model prediction loss based on a to-be-trained residual lifting model, a training sample ID matching sample corresponding to the training sample, the middle training sample feature, the sample label and the obtained second party initial model weight, optimizes the to-be-trained residual lifting model based on the calculated residual loss of the second party model prediction loss and the first party model prediction loss, and obtains the longitudinal federal residual lifting model.
In this embodiment, it should be noted that the second-party initial model weight is an initial value of a model weight preset in the second device and representing the prediction accuracy of the model, wherein the model weight may be set to be equal to a logarithmic value of a ratio of the number of samples with correct classification to the number of samples with wrong classification of the prediction model, and optionally, the initial value of the model weight may be set to be 1.
Specifically, in step a40, a training sample label corresponding to each training sample, a corresponding first-party model prediction loss, a corresponding training sample, and a corresponding intermediate training sample feature of the target prediction model in the iterative training process are sent to the second device, and then the second device extracts a training sample ID, and searches for an ID matching sample corresponding to the sample to be predicted based on the training sample ID, and then the second device performs model prediction by inputting the training sample ID matching sample into a residual lifting model to be trained, and obtains a second-party training model prediction result corresponding to the training sample ID matching sample. And the second equipment splices the training sample ID matching sample corresponding to the training sample ID and the corresponding intermediate training sample characteristic so as to perform characteristic enhancement on the training sample ID matching sample and obtain a training characteristic enhancement sample. And then performing model prediction on the training feature enhancement sample based on the residual lifting model to be trained to obtain the prediction loss of the second square model. And then calculating the prediction loss of the second-party model based on the prediction result of the second-party training model corresponding to the training sample ID matching sample, the corresponding training sample label and the obtained weight of the second-party initial model.
The process of calculating the predicted loss of the second-party model may specifically refer to the process of calculating the predicted loss of the first-party model by using the first device, and is not described herein again. And further calculating residual loss based on the first square model prediction loss and the second square model prediction loss, and judging whether the residual loss is converged. If the residual loss is converged, taking the residual lifting model to be trained as the longitudinal federal residual lifting model; if the residual loss is not converged, updating the residual lifting model to be trained based on the gradient calculated by the residual loss, updating the weight of the second-party initial model based on the prediction result of the second-party training model corresponding to the training sample ID matching sample and the corresponding sample label, and returning to the execution step: the second device extracts the training sample ID for the next iteration. Wherein the second device calculates a loss function of residual loss as follows:
Figure BDA0003229552350000121
wherein, L (theta)B,αB,XBY) is the residual loss, NCNumber of training sample ID matching samples, θ, for a round of iterationsBFor the residual lifting model to be trained, αBFor the second-party initial model weight, XBIs NCA training sample set consisting of training sample ID matching samples, Y is NCThe ID of each training sample matches with a label set formed by labels of the training samples corresponding to the samples,
Figure BDA0003229552350000125
for the ith training sample label in a round of iteration,
Figure BDA0003229552350000122
for the ith training sample ID in a round of iteration to match the characteristics of the sample,
Figure BDA0003229552350000123
is NCThe ID of each training sample matches the predicted loss of the corresponding first party model of the sample,
Figure BDA0003229552350000124
for the ith training sample in one iterationThe ID matches the intermediate training sample characteristics corresponding to the sample.
Further, the second device counts the number of correctly classified samples and the number of wrongly classified samples of the longitudinal federated residual lifting model in the iterative training process, obtains the number of correctly classified samples of the second party and the number of wrongly classified samples of the second party, and further generates the second-party model weight by calculating the ratio of the number of correctly classified samples of the second party and the number of wrongly classified samples of the second party. The specific process of generating the second-party model weight by the second device may refer to the specific process of generating the first-party model weight by the first device, and is not described herein again.
Step S20, the intermediate sample features are sent to second equipment, so that the second equipment can jointly execute model prediction on the intermediate sample features and the ID matching samples corresponding to the samples to be predicted based on a longitudinal federated residual error lifting model, and a second-party model prediction result is obtained.
In this embodiment, it should be noted that the longitudinal federal residual lift model is obtained by performing residual learning based on longitudinal federal learning with the first device by using the second device based on a longitudinal federal public sample in combination with the model prediction loss of the target prediction model in the first device on the longitudinal federal public sample, the intermediate public sample feature corresponding to the longitudinal federal public sample, and the corresponding sample label. And the specific process of obtaining the longitudinal federal residual lifting model by the second device and the first device through residual learning based on longitudinal federal learning by combining the model prediction loss of the target prediction model in the first device on the longitudinal federal public sample, the middle public sample characteristics corresponding to the longitudinal federal public sample and the corresponding sample label based on the longitudinal federal public sample can refer to steps a10 to a30, and details are not repeated here. The longitudinal federal common sample is a sample aligned with an ID in a first device in a second device, that is, a training sample ID matching sample corresponding to a training sample in the first device.
Specifically, the characteristics of the intermediate sample are sent to second equipment, the second equipment splices the ID matching sample corresponding to the sample to be predicted and the characteristics of the intermediate sample corresponding to the ID matching sample, so as to perform characteristic enhancement on the ID matching sample to obtain a characteristic enhancement sample, then the characteristic enhancement sample corresponding to the sample to be predicted is input into a longitudinal federal residual lifting model, model prediction is performed on the characteristic enhancement sample, a second square model prediction result is obtained, and then the second equipment sends the second square model prediction result and the second square model weight corresponding to the longitudinal federal residual lifting model to first equipment.
Wherein, prior to the step of sending the intermediate sample features to the second device, the longitudinal federal prediction optimization method further comprises:
step B10, sending the sample ID to be predicted corresponding to the sample to be predicted to the second device, so that the second device can search the ID matching sample corresponding to the sample ID to be predicted;
in this embodiment, it should be noted that the sample ID to be predicted is a sample ID of the sample to be predicted.
Step B20, if receiving the search failure information sent by the second device, taking the first party model prediction result as a target prediction result;
in this embodiment, if the search failure information sent by the second device is received, the first-party model prediction result is used as the target prediction result. Specifically, if the search failure information sent by the second device is received, it is proved that the sample to be predicted is not an aligned sample between the first device and the second device, and the first-party model prediction result is used as a target prediction result, so that the purpose of independently predicting the sample to be predicted based on the target prediction model serving as a complete model is achieved.
Step B30, if the search failure information sent by the second device is not received, executing the steps of: the intermediate sample features are sent to a second device.
In this embodiment, if the search failure information sent by the second device is not received, the following steps are executed: and sending the intermediate sample characteristics to second equipment to obtain residual lifting information which is sent by the second equipment and calculated based on the intermediate sample characteristics and has higher accuracy, wherein the residual lifting information is a second-party model prediction result output by a longitudinal federal residual lifting model in the second equipment.
Step S30, obtaining a first-party model weight corresponding to the target prediction model, and receiving a second-party model prediction result sent by the second device and a second-party model weight corresponding to the longitudinal federal residual error hoisting model;
in this embodiment, it should be noted that the first-party model weight is obtained by the first device by calculating a ratio of the number of first-party classified correct samples to the number of first-party classified incorrect samples in the iterative training process of the target prediction model.
And step S40, carrying out weighted aggregation on the first party model prediction result and the second party model prediction result based on the first party model weight and the second party model weight to obtain a target federal prediction result.
In this embodiment, specifically, based on the first-party model weight and the second-party model weight, the first-party model prediction result and the second-party model prediction result are subjected to weighted aggregation through a preset aggregation rule, so as to obtain a target federal prediction result. The preset aggregation rule comprises summation, averaging and the like, so that the purpose of improving the accuracy of the first equipment in predicting the sample to be predicted by using residual error improving information generated by the second equipment is achieved.
It should be noted that, because the residual lifting information is generated by the longitudinal federal residual lifting model based on the feature enhancement sample, and the feature enhancement sample is generated by splicing the ID matching sample corresponding to the sample to be predicted and the corresponding intermediate sample feature, the purpose of performing feature enhancement on the ID matching sample based on the intermediate sample feature in the first device is achieved, so that the input of the longitudinal federal residual lifting model has more feature information, the decision basis for generating the residual lifting information by the longitudinal federal residual lifting model is more, the longitudinal federal residual lifting model can output the residual lifting information with higher accuracy, therefore, the residual lifting information with higher accuracy is generated based on the second device, and on the basis of using the residual lifting information to improve the accuracy of the sample prediction of the sample to be predicted as the aligned sample by the first device, the accuracy of the first device for sample prediction of the aligned samples is further improved.
Further, the target prediction model may be set as a binary classification model used as a recommendation model, that is, the target prediction model determines whether to recommend an article corresponding to the sample to be predicted or not or whether to recommend an article to a user corresponding to the sample to be predicted by performing secondary classification on the sample to be predicted, and because the embodiment of the present application realizes that the accuracy of the aligned samples is higher and the longitudinal federal prediction is performed based on the residual lifting information, the target prediction model can perform accurate prediction on unaligned samples locally and independently based on the complete model, thereby improving the overall sample prediction accuracy of the longitudinal federal prediction and improving the overall recommendation accuracy of the recommendation model.
In addition, for the existing longitudinal federal prediction model, since model prediction can be performed only by combining all the participators in longitudinal federal learning, once the participators have data loss or sending downtime, samples cannot be predicted based on complete models and sample data, and the accuracy of sample prediction is further influenced. In the embodiment of the application, because the target prediction model is independently held by the first device, even if the second device is lost or crashed, the first device can independently perform sample prediction on samples to be predicted by means of the target prediction model serving as a complete model, so that the accuracy of the sample prediction is improved when the longitudinal federal learning participators are lost or crashed.
Compared with the technical means of locally predicting unaligned samples in a longitudinal federated prediction scene by a predictor based on a local held partial longitudinal federated prediction model in the prior art, the method for optimizing the longitudinal federated prediction comprises the steps of firstly extracting the samples to be predicted, obtaining an intermediate sample feature generated by feature extraction of a feature extractor of a target prediction model aiming at the samples to be predicted, and obtaining a first party model prediction result generated by the target prediction model aiming at the samples to be predicted, wherein the target prediction model is obtained by local iterative training of first equipment, and further achieving the purpose of locally predicting the samples to be predicted based on the target prediction model of the local iterative training. Therefore, for the sample to be predicted if the sample is not aligned, the first device can perform accurate sample prediction on the sample to be predicted alone based on the target prediction model as a complete model. Further, the intermediate sample features are sent to a second device, so that the second device jointly executes model prediction on the intermediate sample features and the ID matching samples corresponding to the samples to be predicted based on a longitudinal federal residual lifting model, more decision bases are provided for the second device to execute model prediction based on the longitudinal federal residual lifting model, and the accuracy of a second square model prediction result generated by the second device is improved, wherein the longitudinal federal residual lifting model is obtained by performing residual learning based on longitudinal federal learning with the first device based on the longitudinal federal common samples by combining model prediction loss of the target prediction model on the longitudinal federal common samples, the intermediate common sample features corresponding to the longitudinal federal common samples and corresponding sample labels in the first device, and for the ID matching samples aligned with the samples to be predicted, the second device performs model prediction on the ID matching samples and the intermediate sample characteristics based on the longitudinal federal residual lifting model to generate residual lifting information with higher accuracy corresponding to the samples to be predicted, namely a second square model prediction result, further obtains first square model weight corresponding to the target prediction model, receives the second square model prediction result sent by the second device and second square model weight corresponding to the longitudinal federal residual lifting model, and performs weighted aggregation on the first square model prediction result and the second square model prediction result based on the first square model weight and the second square model weight to obtain the target federal prediction result The method comprises the steps of generating residual lifting information (second-party model prediction results) with higher accuracy by sample and intermediate sample characteristics, improving the accuracy of a first-party model prediction result output by a target prediction model, realizing longitudinal federal prediction based on the residual lifting information with higher accuracy on a sample to be predicted, realizing the aim of accurately predicting unaligned samples locally and independently based on a complete model under the condition of longitudinal federal prediction based on the residual lifting information with higher accuracy on aligned samples, overcoming the technical defect that the prediction party cannot perform sample prediction based on the complete model on the unaligned samples under the condition of performing federal prediction with higher accuracy on aligned samples, so that the overall sample prediction accuracy is lowered, and improving the overall sample prediction accuracy of the longitudinal federal prediction.
Further, referring to fig. 2, in another embodiment of the present application, the longitudinal federal prediction optimization method is applied to a second device, and the longitudinal federal prediction optimization method includes:
step C10, receiving the intermediate sample characteristics sent by the first equipment, and searching for an ID matching sample corresponding to the intermediate sample characteristics;
in this embodiment, it should be noted that the intermediate sample feature is generated by the first device performing feature extraction on a sample to be predicted based on a feature extractor in the target prediction model, and the ID matching sample is a sample in the second device, where the sample ID of the sample to be predicted corresponding to the intermediate sample feature is consistent.
Specifically, in step C10, an intermediate sample feature generated by performing feature extraction on the to-be-predicted sample by a feature extractor based on a target prediction model, which is sent by the first device, and a to-be-predicted sample ID corresponding to the to-be-predicted sample sent by the first device are received, and then an ID matching sample is searched according to the to-be-predicted sample ID.
Wherein after the step of finding ID matching samples, the longitudinal federated prediction optimization method further comprises:
step D10, if the search is successful, executing the steps of: performing model prediction on the ID matching sample and the intermediate sample feature together based on a longitudinal federated residual lifting model to obtain a second-party model prediction result;
in this embodiment, if the search is successful, it is proved that the second device has an aligned sample corresponding to the sample to be predicted, that is, an ID matched sample, and then the following steps are performed: and performing model prediction on the ID matching sample and the intermediate sample characteristic together based on a longitudinal federal residual lifting model to obtain a second-party model prediction result, and generating residual lifting information corresponding to the sample to be predicted with higher accuracy based on the longitudinal federal residual lifting model, the ID matching sample corresponding to the sample to be predicted and the corresponding intermediate sample characteristic. The residual lifting information is the second-party model prediction result, and then is sent to the first device, so that the first device can realize the residual lifting information based on higher accuracy, and the accuracy of the sample prediction result of the sample to be predicted of the first device is improved.
Step D20, if the search fails, feeding back search failure information to the first device, so that after receiving the search failure information, the first device takes a first party model prediction result generated for the sample to be predicted based on the target prediction model as a target prediction result.
In this embodiment, if the search fails, it is proved that the second device does not have an aligned sample corresponding to the sample to be predicted, search failure information is fed back to the first device, and after the first device receives the search failure information, a first party model prediction result generated for the sample to be predicted based on a target prediction model can be directly used as a target prediction result, so as to achieve the purpose of performing sample prediction on the unaligned sample alone.
Step C20, based on a longitudinal federal residual error hoisting model, performing model prediction on the ID matching sample and the intermediate sample feature together to obtain a second-party model prediction result;
in this embodiment, it should be noted that the longitudinal federal residual error improvement model is obtained by performing residual error learning based on longitudinal federal learning with the first device by using the second device based on a longitudinal federal public sample in combination with the model prediction loss of the target prediction model in the first device on the longitudinal federal public sample, the intermediate public sample feature corresponding to the longitudinal federal public sample, and the corresponding sample label. And the specific process of obtaining the longitudinal federal residual error improvement model by the second device and the first device through residual error learning based on longitudinal federal learning by combining the model prediction loss of the target prediction model in the first device on the longitudinal federal public sample, the intermediate public sample characteristics corresponding to the longitudinal federal public sample and the corresponding sample label based on the longitudinal federal public sample can refer to the contents in the steps a10 to a40, and is not described herein again.
Specifically, in step C20, based on the intermediate sample features, the ID matching sample is feature-enhanced to obtain a feature-enhanced sample, and then the feature-enhanced sample is input into a vertical federal residual lifting model to perform model prediction, so as to generate a second-party model prediction result.
The step of performing model prediction on the ID matching sample and the intermediate sample feature together based on the longitudinal federated residual lifting model to obtain a second-party model prediction result comprises the following steps:
step C21, splicing the ID matching sample and the characteristics of the intermediate sample to obtain a characteristic enhanced sample;
in this embodiment, the ID matching sample and the intermediate sample feature are spliced to perform feature enhancement on the ID matching sample based on the intermediate sample feature, so as to obtain a feature enhanced sample.
In another practical way, the ID matching sample and the intermediate sample feature are subjected to weighted splicing to perform feature enhancement on the ID matching sample based on the intermediate sample feature, so as to obtain a feature enhanced sample.
And step C22, performing model prediction on the feature enhancement sample based on the longitudinal federal residual lifting model to obtain a second-party model prediction result.
Specifically, in step C22, the feature enhancement samples are subjected to data processing by inputting the feature enhancement samples into a vertical federal residual lifting model. The data processing process comprises convolution, pooling, full connection and the like, so that a full connection layer output result output by the last full connection layer in the longitudinal federal residual lifting model is obtained, and the full connection layer output result is converted into a second square model prediction result through a preset activation function.
Before the step of jointly performing model prediction on the ID matching sample and the intermediate sample feature based on the longitudinal federated residual lifting model to obtain a second-party model prediction result, the longitudinal federated prediction optimization method further includes:
step E10, obtaining a second-party initial model weight, and receiving an intermediate training sample feature, a training sample label and a first-party model prediction loss sent by the first device, wherein the first-party model prediction loss is calculated by the first device based on a first-party model prediction result of a target prediction model on a training sample corresponding to the training sample ID matching sample and the training sample label, and the intermediate training sample feature is obtained by the first device based on feature extraction of the target prediction model for the training sample;
in this embodiment, it should be noted that, for the specific process of the first device generating the predicted loss and the intermediate training sample feature of the first party model, reference may be made to the specific contents in step a10 to step a30, and details thereof are not repeated herein.
In addition, it should be noted that the first device needs to send all training sample IDs, corresponding training sample labels, corresponding first-party model prediction losses, and corresponding intermediate training sample features corresponding to all training samples of the target prediction model in the iterative training process to the second device.
Step E20, obtaining a training sample ID matching sample, and based on a residual lifting model to be trained, performing model prediction on the training sample ID matching sample and the intermediate training sample together to obtain a second-party training model prediction result;
specifically, in step E20, a training sample ID matching sample is extracted, and the training sample ID matching sample is further spliced with the intermediate training sample feature, so as to perform feature enhancement on the training sample ID matching sample based on the intermediate training sample feature, obtain a training feature enhancement sample, and further perform model prediction by inputting the training feature enhancement sample into a to-be-trained residual lifting model, so as to obtain a second-party training model prediction result.
Step E30, calculating a second-party model prediction loss based on the training sample labels, the second-party initial model weights and the second-party training model prediction results;
in this embodiment, the second device calculates a loss function of the second-party model prediction loss based on the training sample labels, the second-party initial model weights, and the second-party training model prediction result as follows:
Figure BDA0003229552350000191
wherein, L (theta)B,αB,XBY) is the residual loss, NCNumber of training sample ID matching samples, θ, for a round of iterationsBFor the residual lifting model to be trained, αBFor the second-party initial model weight, XBIs NCA training sample set consisting of training sample ID matching samples, Y is NCThe ID of each training sample matches with a label set formed by labels of the training samples corresponding to the samples,
Figure BDA0003229552350000192
for the ith training sample label in a round of iteration,
Figure BDA0003229552350000193
is the first in a round of iterationThe i training sample IDs match the characteristics of the samples,
Figure BDA0003229552350000194
is NCThe ID of each training sample matches the predicted loss of the corresponding first party model of the sample,
Figure BDA0003229552350000195
and matching the ID of the ith training sample in one iteration with the corresponding intermediate training sample characteristic of the sample.
And E40, iteratively optimizing the residual lifting model to be trained based on the residual losses generated by the prediction loss of the first-party model and the prediction loss of the second-party model, and obtaining the longitudinal federal residual lifting model.
Specifically, in step E40, a residual loss is calculated based on the first square model prediction loss and the second square model prediction loss, and it is determined whether the residual loss converges. If the residual lifting model is converged, taking the residual lifting model to be trained as the longitudinal federal residual lifting model; if the residual loss is not converged, updating the residual lifting model to be trained by a preset model optimization method based on the gradient calculated by the residual loss, updating the weight of the second-party initial model based on the prediction result of the second-party training model corresponding to the training sample ID matching sample and the corresponding training sample label, and returning to the execution step: and acquiring a training sample ID matching sample, and performing the next iteration. The specific process of calculating the residual loss based on the first-party model prediction loss and the second-party model prediction loss may refer to the specific content in step a40, and is not described herein again.
Further, the second device counts the number of samples which are classified correctly and the number of samples which are classified incorrectly of the residual lifting model to be trained in the iterative training process, obtains the number of samples which are classified correctly and the number of samples which are classified incorrectly in a second party, and further generates second party model weight by calculating the ratio of the number of samples which are classified correctly and the number of samples which are classified incorrectly in the second party. The specific process of generating the second-party model weight by the second device may refer to the specific process of generating the first-party model weight by the first device, and is not described herein again.
Step C30, obtaining a second square model weight corresponding to the longitudinal federated residual lifting model, and sending the second square model prediction result and the second square model weight to the first device, so that the first device generates a target federated prediction result based on a first square model prediction result generated by the target prediction model for the to-be-predicted sample corresponding to the ID matching sample, a first square model weight corresponding to the target prediction model, the second square model prediction result, and the second square model weight, wherein the target prediction model is obtained by local iterative training of the first device.
In this embodiment, it should be noted that, for the specific process of the first device generating the first party model prediction result for the to-be-predicted sample corresponding to the training sample ID matching sample based on the target prediction model, refer to the specific step in step S10, and details are not repeated here.
Specifically, a second-party model weight corresponding to the longitudinal federal residual lifting model is obtained, and the second-party model prediction result and the second-party model weight are sent to the first device, so that the first device performs weighted aggregation on the first-party model prediction result and the second-party model prediction result based on the first-party model weight and the second-party model weight corresponding to the target prediction model through a preset aggregation rule, and obtains a target federal prediction result, so that the higher-accuracy residual lifting information generated based on the second device is realized, the first-party model prediction result of the sample to be predicted in the first device is optimized, and the accuracy of the sample prediction result of the sample to be predicted by the first device is improved.
Compared with the technical means that a predictor performs local prediction based on a locally held partial longitudinal federal prediction model on unaligned samples in a longitudinal federal prediction scene, which is adopted in the prior art, the method comprises the steps of receiving intermediate sample characteristics sent by first equipment, searching an ID (identity) matching sample corresponding to the intermediate sample characteristics, and performing model prediction on the ID matching sample and the intermediate sample characteristics together based on a longitudinal federal residual error boosting model to obtain a second-party model prediction result. The purpose of generating the residual error promoting information with higher accuracy corresponding to the sample to be predicted by using the characteristics of the intermediate sample sent by the first device is achieved. And further obtaining a second square model weight corresponding to the longitudinal federal residual lifting model, and sending the second square model prediction result and the second square model weight to the first device, so that the first device generates a first square model prediction result, a first square model weight corresponding to the target prediction model, a second square model prediction result and a second square model weight for the target prediction model based on the target prediction model and aiming at the to-be-predicted sample corresponding to the ID matching sample, and generates a target federal prediction result. Wherein the target prediction model is obtained by local iterative training of the first equipment, so that the aim of optimizing a first party model prediction result generated by the first equipment by using residual error promotion information with higher accuracy generated by the second equipment for an aligned sample is realized, to generate a target federated prediction result with higher sample prediction accuracy, and since the target prediction model is trained by the first device local iteration, for the unaligned samples, the first device may also perform sample prediction on the sample to be predicted locally and independently based on the target prediction model as a complete model, therefore, the method overcomes the defect that a predictor can not perform the sample prediction based on the complete model on the unaligned samples under the condition of performing the federal prediction with higher accuracy on the aligned samples, the method has the technical defect that the overall sample prediction accuracy is low, and the overall sample prediction accuracy of longitudinal federal prediction is improved.
Further, referring to fig. 3, in another embodiment of the present application, there is further provided a longitudinal federated learning modeling optimization method, where the longitudinal federated learning modeling optimization method is applied to a first device, and the longitudinal federated learning modeling optimization method includes:
step F10, acquiring the weight of the initial model of the first party, and extracting training samples and training sample labels corresponding to the training samples;
in this embodiment, it should be noted that the longitudinal federated learning modeling optimization method is applied to a longitudinal federated learning scenario, and both the first device and the second device are participants of the longitudinal federated learning scenario. The method comprises the steps that a first device is provided with a sample carrying a sample label, a second device is provided with a sample without the sample label, the first device is a predictor and used for building a prediction model, the second device is an auxiliary data provider and used for building a longitudinal federal residual error improvement model which provides residual error improvement information for the first device, and therefore the accuracy of a prediction result generated by the prediction model in the first device is improved.
In addition, the number of the training samples is at least 1, the training sample label is an identifier of the training sample, and the first-party initial model weight is an initial value of a model weight that is preset in the first device and represents the prediction accuracy of the model. Wherein, the model weight can be set to be equal to the logarithm value of the ratio of the number of samples with correct classification to the number of samples with wrong classification of the prediction model. Alternatively, the initial value of the model weight may be set to 1.
Step F20, acquiring intermediate training sample characteristics generated by the characteristic extractor of the target prediction model to be trained for performing characteristic extraction on the training samples;
in this embodiment, feature extraction is performed on the training sample based on a feature extractor in a target prediction model to be trained, so as to obtain an intermediate training sample feature.
Step F30, based on the training sample labels, the training model prediction results corresponding to the intermediate training sample characteristics and the first-party initial model weight, iteratively optimizing the target prediction model to be trained by calculating the first-party model prediction loss corresponding to the target prediction model to be trained, so as to obtain the target prediction model;
specifically, in step F30, the features of the intermediate training samples are fully connected by the classifier in the target prediction model to be trained to obtain training full-connection layer output, the training full-connection layer output is converted into a training model prediction result by a preset activation function, and the first-party model prediction loss is calculated based on the training samples, the training model prediction result, and the first-party initial model weight. Then judging whether the prediction loss of the first party model is converged, and if the prediction loss of the first party model is converged, taking the target prediction model to be trained as the target prediction model; if the prediction loss of the first party model is not converged, updating the target prediction model to be trained by a preset model optimization method based on the gradient calculated by the prediction loss of the first party model, updating the weight of the first party initial model based on the prediction result of the training model and the training sample label, and returning to the execution step: and extracting training samples and training sample labels corresponding to the training samples, and performing next iteration. The calculation process of calculating the prediction loss of the first-party model based on the training samples, the training model prediction result, and the first-party initial model weight may refer to the specific contents in step a10 to step a30, and is not described herein again.
The step of iteratively optimizing the target prediction model to be trained by calculating a first-party model prediction loss corresponding to the target prediction model to be trained based on the training sample labels, the training model prediction results corresponding to the intermediate training sample features and the first-party initial model weight to obtain the target prediction model includes:
step F31, converting the characteristics of the intermediate training samples into a training model prediction result based on the classifier in the target prediction model to be trained;
in this embodiment, based on the classifier in the target prediction model to be trained, full connection is performed on the characteristics of the intermediate training samples to obtain training full connection layer output, and then the training full connection layer output is converted into a training model prediction result through a preset activation function.
Step F32, calculating a first party model prediction loss based on the training sample label, the training model prediction result and the first party initial model weight;
in this embodiment, it should be noted that the specific calculation process in the step F32 can refer to the content in the step a30, and is not described herein again.
Step F33, updating the weight of the first party initial model based on the training model prediction result and the training sample label;
specifically, in step F33, based on the prediction result of the training model and the training sample label, the current classified error sample number and the current classified correct sample number corresponding to the to-be-trained residual lifting model are updated, and the first-party initial model weight is recalculated by calculating the ratio of the current classified correct sample number and the current classified error sample number. For a process of recalculating the first-party initial model weight, refer to the specific process of calculating the first-party model weight by the first device after step a30, which is not described herein again.
And F34, iteratively optimizing the target prediction model to be trained based on the first party model prediction loss and the updated first party initial model weight to obtain the target prediction model.
Specifically, in step F34, it is determined whether the first-party model prediction loss converges, and if the first-party model prediction loss converges, the target prediction model to be trained is used as the target prediction model, and the updated first-party initial model weight is used as the first-party model weight; if the prediction loss of the first party model is not converged, updating the target prediction model to be trained by a preset model optimization method based on the gradient calculated by the prediction loss of the first party model, and returning to the execution step: and extracting training samples and training sample labels corresponding to the training samples, and performing the next iteration based on the updated target prediction model to be trained and the updated initial model weight of the first party.
After the step of obtaining the target prediction model by calculating the prediction loss of the first-party model corresponding to the target prediction model to be trained and iteratively optimizing the target prediction model to be trained based on the training sample labels, the training model prediction result corresponding to the intermediate training sample features and the first-party initial model weight, the longitudinal federated learning modeling optimization method further includes:
step G10, acquiring the number of first party classified correct samples corresponding to the target prediction model and the number of first party classified error samples corresponding to the target prediction model;
in this embodiment, it should be noted that the first-party classified correct sample number is the number of training samples of which the output classification labels for the training samples in the iterative training process of the target prediction model are consistent with the corresponding training sample labels, and the first-party classified incorrect sample number is the number of training samples of which the output classification labels for the training samples in the iterative training process of the target prediction model are inconsistent with the corresponding training sample labels.
Step G20, generating a first party model weight by calculating the ratio of the number of the first party classified correct samples to the number of the first party classified wrong samples.
In this embodiment, the specific calculation formula of step G20 is as follows:
Figure BDA0003229552350000241
wherein alpha isAAnd B is the number of the first party classified error samples.
Step F40, the training sample labels, the intermediate training sample characteristics and the first party model prediction loss are sent to second equipment, so that the second equipment can calculate second party model prediction loss, and based on the second party model prediction loss and the residual loss calculated by the first party model prediction loss, the residual lifting model to be trained is optimized, and a longitudinal federal residual lifting model is obtained.
In this embodiment, it should be noted that the prediction loss of the model of the second party is calculated based on the residual lifting model to be trained, the training sample ID matching sample corresponding to the training sample, the intermediate training sample feature, the sample label, and the obtained weight of the initial model of the second party.
Specifically, training sample labels corresponding to all training samples of the target prediction model in an iterative training process, corresponding first-party model prediction losses, corresponding training sample IDs and corresponding intermediate training sample features are sent to second equipment, then the second equipment extracts the training sample IDs, searches training sample ID matching samples corresponding to the training sample IDs based on the training sample IDs, and inputs training feature enhancement samples corresponding to the training sample ID matching samples and the intermediate training sample features into a residual error lifting model to be trained to execute model prediction, so that second-party training model prediction results are obtained. And then calculating a second-party model prediction loss based on a second-party training model prediction result corresponding to the training sample ID matching sample, a corresponding training sample label and the obtained second-party initial model weight, calculating a residual loss based on the first-party model prediction loss and the second-party model prediction loss by the second equipment, and iteratively optimizing a to-be-trained residual lifting model based on the residual loss to obtain a longitudinal federal residual lifting model. The specific process of constructing the longitudinal federal residual error hoisting model by the second device may refer to the specific contents in step a10 to step a40, and is not described herein again.
After the step of sending the training sample label, the intermediate training sample feature, and the first-party model prediction loss to a second device, so that the second device calculates a second-party model prediction loss based on a to-be-trained residual lifting model, a training sample ID matching sample corresponding to the training sample, the intermediate training sample feature, and the sample label, and optimizes the to-be-trained residual lifting model based on a residual loss calculated by the second-party model prediction loss and the first-party model prediction loss to obtain a longitudinal federated residual lifting model, the longitudinal federated learning optimization method further includes:
step H10, extracting a sample to be predicted, and acquiring an intermediate sample feature generated by a feature extractor of a target prediction model for performing feature extraction on the sample to be predicted, and a first party model prediction result generated by the target prediction model for performing model prediction on the sample to be predicted;
in this embodiment, specifically, a sample to be predicted is extracted, feature extraction is performed on the sample to be predicted based on a feature extractor in a target prediction model to obtain an intermediate sample feature, full connection is performed on the intermediate sample feature based on a classifier in the target prediction model to obtain a target full connection layer output, and the target full connection layer output is converted into a first-party model prediction result through a preset activation function, where the first-party model prediction result may be a classification probability.
Step H20, sending the intermediate sample feature to a second device, so that the second device jointly executes model prediction on the intermediate sample feature and an ID matching sample corresponding to the sample to be predicted based on a longitudinal federated residual error lifting model, and a second square model prediction result is obtained;
in this embodiment, specifically, the intermediate sample features are sent to the second device, and then the second device splices the ID matching sample corresponding to the sample to be predicted and the corresponding intermediate sample features to perform feature enhancement on the ID matching sample to obtain a feature enhancement sample, and then the feature enhancement sample corresponding to the sample to be predicted is input to the vertical federal residual lifting model, so as to perform model prediction on the feature enhancement sample to obtain a second-party model prediction result, and then the second device sends the second-party model prediction result and the second-party model weight corresponding to the vertical federal residual lifting model to the first device.
Step H30, receiving a second-party model prediction result sent by the second device and a second-party model weight corresponding to the longitudinal federal residual error hoisting model;
and H40, carrying out weighted aggregation on the first party model prediction result and the second party model prediction result based on the first party model weight and the second party model weight corresponding to the target prediction model, and obtaining a target federal prediction result.
In this embodiment, specifically, based on the first-party model weight and the second-party model weight, the first-party model prediction result and the second-party model prediction result are subjected to weighted aggregation through a preset aggregation rule, so as to obtain a target federal prediction result. The preset aggregation rule comprises summation, averaging and the like, so that the purpose of improving the accuracy of the first equipment in predicting the sample to be predicted by using residual error improving information generated by the second equipment is achieved.
The embodiment of the application provides a longitudinal federated learning modeling optimization method, namely, a first-party initial model weight is obtained, a training sample and a training sample label corresponding to the training sample are extracted, an intermediate training sample feature generated by a feature extractor of a target prediction model to be trained by performing feature extraction on the training sample is further obtained, the target prediction model is obtained by calculating a first-party model prediction loss corresponding to the target prediction model to be trained and iteratively optimizing the target prediction model to be trained on the basis of the training sample label, a training model prediction result corresponding to the intermediate training sample feature and the first-party initial model weight, and the purpose of locally constructing the target prediction model serving as a complete model on first equipment is achieved. And then the training sample label, the middle training sample characteristics and the first party model prediction loss are sent to second equipment so that the second equipment can calculate second party model prediction loss, the to-be-trained residual lifting model is optimized based on the residual loss calculated by the second party model prediction loss and the first party model prediction loss, a longitudinal federal residual lifting model is obtained, and residual learning based on longitudinal federal learning is achieved. By utilizing the characteristics of the middle training sample corresponding to the training sample in the first equipment and the purpose of the longitudinal federal residual error hoisting model constructed at the second equipment, the characteristic dimension of the ID matching sample of the training sample in the second equipment is expanded, so that the prediction accuracy of the longitudinal federal residual error hoisting model is higher, and further, the aim of accurately predicting the unaligned sample locally and independently can be realized on the basis of a complete model under the condition of longitudinally federal prediction of the unaligned sample based on residual error hoisting information with higher accuracy based on the target prediction model of the first equipment and the longitudinal federal residual error hoisting model at the second equipment. The method lays a foundation for overcoming the technical defect that the whole sample prediction accuracy is lowered because a prediction party cannot perform the sample prediction based on the complete model on the unaligned sample under the condition of performing the federal prediction with higher accuracy on the aligned sample.
Further, referring to fig. 4, in another embodiment of the present application, there is further provided a longitudinal federated learning modeling optimization method, where the longitudinal federated learning modeling optimization method is applied to a second device, and the longitudinal federated learning modeling optimization method includes:
step R10, obtaining the weight of the initial model of the second party, and receiving the intermediate training sample characteristics, the training sample labels and the prediction loss of the first party model sent by the first equipment;
in this embodiment, it should be noted that the first-party model prediction loss is calculated by the first device based on a first-party model prediction result of a target prediction model on a training sample corresponding to the training sample ID matching sample and the training sample label, and the intermediate training sample feature is obtained by performing feature extraction on the training sample by the first device based on a feature extractor of the target prediction model. The specific process of generating the predicted loss and the intermediate training sample features of the first party model by the first device may refer to the specific contents in step a10 to step a30, and will not be described herein again.
Specifically, the weight of the second-party initial model is obtained, and training sample labels corresponding to all training samples, corresponding first-party model prediction losses and corresponding intermediate training sample features of the target prediction model sent by the first device in the iterative training process are received.
Step R20, obtaining a training sample ID matching sample, and based on a residual lifting model to be trained, performing model prediction on the training sample ID matching sample and the intermediate training sample together to obtain a second-party training model prediction result;
in this embodiment, specifically, a training sample ID is extracted, a training sample ID matching sample corresponding to the training sample ID is searched, the training sample ID matching sample and the intermediate training sample feature are further spliced, so that feature enhancement is performed on the training sample ID matching sample based on the intermediate training sample feature to obtain a training feature enhancement sample, and a second-party training model prediction result is obtained by inputting the training feature enhancement sample into a to-be-trained residual lifting model to perform model prediction.
Step R30, calculating a second-party model prediction loss based on the training sample labels, the second-party initial model weights and the second-party training model prediction results;
in this embodiment, it should be noted that, in the step R30, the specific process of the second device calculating the predicted loss of the second-party model may refer to the specific content in the step E30, and is not described herein again.
And R40, iteratively optimizing the residual lifting model to be trained based on the residual losses generated by the prediction loss of the first square model and the prediction loss of the second square model, and obtaining the longitudinal federal residual lifting model.
In this embodiment, specifically, based on the first-party model prediction loss and the second-party model prediction loss, a residual loss is calculated, and then whether the residual loss converges is determined, and if the residual lifting model converges, the to-be-trained residual lifting model is used as the longitudinal federated residual lifting model; if the longitudinal federal residual lifting model is not converged, updating the residual lifting model to be trained through a preset model optimization method based on the gradient calculated by the residual loss, updating the weight of the second-party initial model based on the prediction result of the second-party training model corresponding to the training sample ID matching sample and the corresponding sample label, and returning to the execution step: and acquiring a training sample ID matching sample, and performing the next iteration. The specific process of calculating the residual loss based on the first-party model prediction loss and the second-party model prediction loss may refer to the specific content in step a40, and is not described herein again.
After the step of iteratively optimizing the to-be-trained residual lifting model based on the residual loss generated by the first-party model prediction loss and the second-party model prediction loss to obtain the longitudinal federal residual lifting model, the longitudinal federal learning modeling optimization method further includes:
step T10, acquiring the number of second-party classified correct samples corresponding to the longitudinal federal residual lifting model and the number of second-party classified error samples corresponding to the longitudinal federal residual lifting model;
in this embodiment, it should be noted that the number of the correct samples for classification of the second party is the number of training sample ID matching samples of the longitudinal federated residual lifting model, where output classification labels of the training sample ID matching samples are consistent with corresponding training sample labels in an iterative training process; the second-party classification error sample number is the number of training sample ID matching samples of which the output classification labels aiming at the training sample ID matching samples are inconsistent with the corresponding training sample labels in the iterative training process of the longitudinal federated residual error lifting model.
And step T20, generating a second square model weight by calculating the ratio of the second square classification correct sample number to the second square classification error sample number.
In this embodiment, the specific calculation formula of step T20 is as follows:
Figure BDA0003229552350000281
wherein alpha isAAnd B is the second party classification error sample number.
After the step of iteratively optimizing the to-be-trained residual lifting model based on the residual loss generated by the first-party model prediction loss and the second-party model prediction loss to obtain a longitudinal federal residual lifting model, the longitudinal federal learning modeling optimization method further includes:
step Y10, receiving the intermediate sample characteristics sent by the first equipment, and searching an ID matching sample corresponding to the intermediate sample characteristics;
in this embodiment, specifically, an intermediate sample feature generated by performing feature extraction on the sample to be predicted by a feature extractor based on a target prediction model and sent by a first device, and a sample ID to be predicted corresponding to the sample to be predicted and sent by the first device are received, and then an ID matching sample is searched according to the sample ID to be predicted.
Step Y20, based on the longitudinal federal residual error hoisting model, performing model prediction on the ID matching sample and the intermediate sample feature together to obtain a second-party model prediction result;
in this embodiment, based on the characteristics of the intermediate sample, the ID matching sample is subjected to characteristic enhancement to obtain a characteristic enhancement sample, and then the characteristic enhancement sample is input into a longitudinal federal residual lifting model to perform model prediction, so as to generate a second-party model prediction result. The specific implementation of the feature enhancement on the ID matching sample may refer to the content in step C21, and is not described herein again.
Step Y30, sending the second-party model prediction result and the second-party model weight corresponding to the longitudinal federal residual error boosting model to the first device, so that the first device generates a target federal prediction result based on a first-party model prediction result generated by a target prediction model for a to-be-predicted sample corresponding to the ID matching sample, a first-party model weight corresponding to the target prediction model, the second-party model prediction result and the second-party model weight, wherein the target prediction model is obtained by local iterative training of the first device.
In this embodiment, it should be noted that, for the specific process of the first device generating the first party model prediction result for the to-be-predicted sample corresponding to the training sample ID matching sample based on the target prediction model, refer to the specific step in step S10, and details are not repeated here.
Specifically, a second-party model weight corresponding to the longitudinal federal residual lifting model is obtained, and the second-party model prediction result and the second-party model weight are sent to the first device, so that the first device performs weighted aggregation on the first-party model prediction result and the second-party model prediction result based on the first-party model weight and the second-party model weight corresponding to the target prediction model through a preset aggregation rule, and obtains a target federal prediction result, so that the higher-accuracy residual lifting information generated based on the second device is realized, the first-party model prediction result of the sample to be predicted in the first device is optimized, and the accuracy of the sample prediction result of the sample to be predicted by the first device is improved.
The embodiment of the application provides a longitudinal federated learning modeling optimization method, namely, the weight of an initial model of a second party is obtained, and the characteristics of an intermediate training sample, a training sample label and the prediction loss of a model of the first party, which are sent by first equipment, are received. The first-party model prediction loss is obtained by calculating a first-party model prediction result of the first device on a training sample corresponding to the training sample ID matching sample based on a target prediction model and the training sample label, the characteristics of the intermediate training sample are obtained by performing characteristic extraction on the training sample by the first device based on a characteristic extractor of the target prediction model, so that a training sample ID matching sample is obtained, and model prediction is performed on the training sample ID matching sample and the intermediate training sample characteristics together based on a residual error lifting model to be trained so as to obtain a second-party training model prediction result. And calculating the prediction loss of the model of the second party based on the training sample label, the weight of the initial model of the second party and the prediction result of the training model of the second party, and iteratively optimizing the residual lifting model to be trained based on the residual loss generated by the prediction loss of the model of the first party and the prediction loss of the model of the second party to obtain the longitudinal federated residual lifting model. The method achieves the purpose that the training sample ID matching sample corresponding to the training sample in the second equipment is subjected to feature enhancement by utilizing the characteristics of the middle training sample in the first equipment, and the first equipment is combined to perform residual learning based on the longitudinal federal learning to construct the longitudinal federal residual lifting model. The method comprises the steps that residual error lifting information with higher accuracy is generated by second equipment for aligned samples based on a longitudinal federal residual error lifting model, a first party model prediction result generated by first equipment is optimized to generate a target federal prediction result with higher sample prediction accuracy, and a foundation is laid for overcoming the technical defect that a prediction party cannot perform sample prediction based on a complete model on unaligned samples under the condition that the prediction party performs federal prediction with higher accuracy on aligned samples, so that the overall sample prediction accuracy is lowered.
Referring to fig. 5, fig. 5 is a schematic device system diagram of a hardware operating environment according to an embodiment of the present application.
As shown in fig. 5, the longitudinal federal prediction optimization device may include: a processor 1001, such as a CPU, a memory 1005, and a communication bus 1002. The communication bus 1002 is used for realizing connection communication between the processor 1001 and the memory 1005. The memory 1005 may be a high-speed RAM memory or a non-volatile memory (e.g., a magnetic disk memory). The memory 1005 may alternatively be a memory device separate from the processor 1001 described above.
Optionally, the vertical federal prediction optimization device may further include a rectangular user interface, a network interface, a camera, an RF (Radio Frequency) circuit, a sensor, an audio circuit, a WiFi module, and the like. The rectangular user interface may comprise a Display screen (Display), an input sub-module such as a Keyboard (Keyboard), and the optional rectangular user interface may also comprise a standard wired interface, a wireless interface. The network interface may optionally include a standard wired interface, a wireless interface (e.g., WI-FI interface).
It will be understood by those skilled in the art that the longitudinal federal predictive optimizer system illustrated in fig. 5 does not constitute a limitation of the longitudinal federal predictive optimizer, and may include more or fewer components than shown, or some components in combination, or a different arrangement of components.
As shown in fig. 5, a memory 1005, which is a type of computer storage medium, may include an operating system, a network communication module, and a vertical federal forecast optimization program. The operating system is a program for managing and controlling hardware and software resources of the longitudinal federal forecast optimization device, and supports the operation of the longitudinal federal forecast optimization program and other software and/or programs. The network communication module is used for realizing communication among components in the memory 1005 and communication with other hardware and software in the longitudinal federal forecast optimization system.
In the longitudinal federated prediction optimization apparatus shown in fig. 5, the processor 1001 is configured to execute a longitudinal federated prediction optimization program stored in the memory 1005 to implement the steps of any one of the longitudinal federated prediction optimization methods described above.
The specific implementation of the longitudinal federal prediction optimization device in the application is basically the same as that of each embodiment of the longitudinal federal prediction optimization method, and is not described herein again.
Referring to fig. 6, fig. 6 is a schematic device system diagram of a hardware operating environment according to an embodiment of the present application.
As shown in fig. 6, the longitudinal federal learning modeling optimization device may include: a processor 1001, such as a CPU, a memory 1005, and a communication bus 1002. The communication bus 1002 is used for realizing connection communication between the processor 1001 and the memory 1005. The memory 1005 may be a high-speed RAM memory or a non-volatile memory (e.g., a magnetic disk memory). The memory 1005 may alternatively be a memory device separate from the processor 1001 described above.
Optionally, the longitudinal federal learning modeling optimization device may further include a rectangular user interface, a network interface, a camera, an RF (Radio Frequency) circuit, a sensor, an audio circuit, a WiFi module, and the like. The rectangular user interface may comprise a Display screen (Display), an input sub-module such as a Keyboard (Keyboard), and the optional rectangular user interface may also comprise a standard wired interface, a wireless interface. The network interface may optionally include a standard wired interface, a wireless interface (e.g., WI-FI interface).
It will be understood by those skilled in the art that the longitudinal federated learning modeling optimization facility system illustrated in FIG. 6 does not constitute a limitation of the longitudinal federated learning modeling optimization facility, and may include more or fewer components than those illustrated, or some components in combination, or a different arrangement of components.
As shown in fig. 6, the memory 1005, which is a type of computer storage medium, may include an operating system, a network communication module, and a longitudinal federal learning modeling optimization program. The operating system is a program for managing and controlling hardware and software resources of the longitudinal federated learning modeling optimization equipment and supports the operation of the longitudinal federated learning modeling optimization program and other software and/or programs. The network communication module is used for realizing communication among components in the memory 1005 and communication with other hardware and software in the longitudinal federal learning modeling optimization system.
In the longitudinal federated learning modeling optimization apparatus shown in fig. 6, the processor 1001 is configured to execute a longitudinal federated learning modeling optimization program stored in the memory 1005 to implement the steps of any one of the longitudinal federated learning modeling optimization methods described above.
The specific implementation of the longitudinal federated learning modeling optimization device in the application is basically the same as that of each embodiment of the longitudinal federated learning modeling optimization method, and is not described herein again.
The embodiment of the present application further provides a longitudinal federal prediction optimization device, where the longitudinal federal prediction optimization device is applied to a first device, and the longitudinal federal prediction optimization device includes:
the model prediction module is used for extracting a sample to be predicted, obtaining an intermediate sample feature generated by a feature extractor of a target prediction model aiming at the sample to be predicted and performing feature extraction on the sample to be predicted, and obtaining a first party model prediction result generated by the target prediction model aiming at the sample to be predicted, wherein the target prediction model is obtained by local iterative training of the first equipment;
and the sending module is used for sending the intermediate sample characteristics to second equipment so that the second equipment can jointly execute model prediction on the intermediate sample characteristics and the ID matching samples corresponding to the samples to be predicted based on a longitudinal federated residual error lifting model to obtain a second-party model prediction result. The longitudinal federal residual error hoisting model is obtained by the second equipment through residual error learning based on longitudinal federal learning by combining model prediction loss of the target prediction model in the first equipment on the longitudinal federal public sample, middle public sample characteristics corresponding to the longitudinal federal public sample and a corresponding sample label on the basis of the longitudinal federal public sample;
the receiving module is used for acquiring a first-party model weight corresponding to the target prediction model, and receiving a second-party model prediction result sent by the second equipment and a second-party model weight corresponding to the longitudinal federal residual error hoisting model;
and the weighted aggregation module is used for carrying out weighted aggregation on the first party model prediction result and the second party model prediction result based on the first party model weight and the second party model weight to obtain a target federal prediction result.
Optionally, the longitudinal federal prediction optimization device is further configured to:
sending the ID of the sample to be predicted corresponding to the sample to be predicted to the second equipment so that the second equipment can search the ID matching sample corresponding to the ID of the sample to be predicted;
if the search failure information sent by the second equipment is received, taking the first party model prediction result as a target prediction result;
if the search failure information sent by the second equipment is not received, executing the following steps: the intermediate sample features are sent to a second device.
Optionally, the longitudinal federal prediction optimization device is further configured to:
acquiring the weight of a first party initial model, and extracting training samples and training sample labels corresponding to the training samples;
acquiring intermediate training sample characteristics generated by characteristic extraction of a characteristic extractor of a target prediction model to be trained aiming at the training samples;
iteratively optimizing the target prediction model to be trained by calculating the prediction loss of the first party model corresponding to the target prediction model to be trained based on the training sample label, the training model prediction result corresponding to the intermediate training sample characteristic and the first party initial model weight to obtain the target prediction model;
and sending the training sample label, the intermediate training sample characteristic and the first party model prediction loss to second equipment, so that the second equipment calculates a second party model prediction loss based on a to-be-trained residual lifting model, a training sample ID matching sample corresponding to the training sample, the intermediate training sample characteristic, the sample label and the obtained second party initial model weight, and optimizes the to-be-trained residual lifting model based on the calculated residual loss of the second party model prediction loss and the first party model prediction loss to obtain the longitudinal federated residual lifting model.
The specific implementation of the longitudinal federal prediction optimization device in the application is basically the same as that of each embodiment of the longitudinal federal prediction optimization method, and is not described herein again.
The embodiment of the present application further provides a longitudinal federal prediction optimization device, where the longitudinal federal prediction optimization device is applied to a second device, and the longitudinal federal prediction optimization device includes:
the receiving and searching module is used for receiving the intermediate sample characteristics sent by the first equipment and searching the ID matching sample corresponding to the intermediate sample characteristics;
and the model prediction module is used for jointly performing model prediction on the ID matching sample and the intermediate sample characteristics based on a longitudinal federal residual error hoisting model to obtain a second-party model prediction result. The longitudinal federal residual error hoisting model is obtained by the second equipment through residual error learning based on longitudinal federal learning by combining model prediction loss of a target prediction model in the first equipment on the longitudinal federal public sample, middle public sample characteristics corresponding to the longitudinal federal public sample and a corresponding sample label;
a sending module, configured to obtain a second square model weight corresponding to the longitudinal federated residual lifting model, and send the second square model prediction result and the second square model weight to the first device, so that the first device generates a target federated prediction result based on a first square model prediction result generated by the target prediction model for the to-be-predicted sample corresponding to the ID matching sample, a first square model weight corresponding to the target prediction model, the second square model prediction result, and the second square model weight, where the target prediction model is obtained by local iterative training of the first device.
Optionally, the model prediction module is further configured to:
splicing the ID matching sample and the characteristics of the intermediate sample to obtain a characteristic enhanced sample;
and performing model prediction on the feature enhancement sample based on the longitudinal federal residual lifting model to obtain a second square model prediction result.
Optionally, the longitudinal federal prediction optimization device is further configured to:
if the search is successful, executing the following steps: performing model prediction on the ID matching sample and the intermediate sample feature together based on a longitudinal federated residual lifting model to obtain a second-party model prediction result;
and if the search fails, feeding back search failure information to the first equipment, so that the first equipment takes a first party model prediction result generated by aiming at the sample to be predicted based on a target prediction model as a target prediction result after receiving the search failure information.
Optionally, the longitudinal federal prediction optimization device is further configured to:
acquiring the weight of a second-party initial model, and receiving an intermediate training sample feature, a training sample label and a first-party model prediction loss sent by the first equipment, wherein the first-party model prediction loss is calculated by the first equipment based on a first-party model prediction result of a target prediction model on a training sample corresponding to the training sample ID matching sample and the training sample label, and the intermediate training sample feature is obtained by performing feature extraction on the training sample by the first equipment based on a feature extractor of the target prediction model;
acquiring a training sample ID matching sample, and performing model prediction on the training sample ID matching sample and the intermediate training sample feature together based on a residual lifting model to be trained to obtain a second-party training model prediction result;
calculating a second-party model prediction loss based on the training sample labels, the second-party initial model weights and the second-party training model prediction results;
and iteratively optimizing the residual lifting model to be trained based on the residual losses generated by the prediction loss of the first square model and the prediction loss of the second square model to obtain the longitudinal federal residual lifting model.
The specific implementation of the longitudinal federal prediction optimization device in the application is basically the same as that of each embodiment of the longitudinal federal prediction optimization method, and is not described herein again.
The embodiment of the present application further provides a longitudinal federated learning modeling optimization apparatus, where the longitudinal federated learning modeling optimization apparatus is applied to a first device, and the longitudinal federated learning modeling optimization apparatus includes:
the first acquisition module is used for acquiring the initial model weight of a first party and extracting training samples and training sample labels corresponding to the training samples;
the second acquisition module is used for acquiring the characteristics of an intermediate training sample generated by the characteristic extractor of the target prediction model to be trained aiming at the characteristic extraction of the training sample;
the iterative optimization module is used for iteratively optimizing the target prediction model to be trained by calculating the prediction loss of the first-party model corresponding to the target prediction model to be trained based on the training sample labels, the training model prediction result corresponding to the intermediate training sample characteristics and the first-party initial model weight to obtain the target prediction model;
and the sending module is used for sending the training sample label, the intermediate training sample characteristics and the first party model prediction loss to second equipment so that the second equipment can calculate second party model prediction loss, and optimizing the to-be-trained residual lifting model based on the second party model prediction loss and the residual loss calculated by the first party model prediction loss to obtain a longitudinal federal residual lifting model. And calculating the prediction loss of the model of the second party based on the residual lifting model to be trained, the training sample ID matching sample corresponding to the training sample, the characteristics of the intermediate training sample, the sample label and the obtained weight of the initial model of the second party.
Optionally, the iterative optimization module is further configured to:
converting the characteristics of the intermediate training samples into a training model prediction result based on a classifier in the target prediction model to be trained;
calculating a first party model prediction loss based on the training sample label, the training model prediction result and the first party initial model weight;
updating the first party initial model weight based on the training model prediction result and the training sample label;
and iteratively optimizing the target prediction model to be trained based on the prediction loss of the first party model and the updated initial model weight of the first party to obtain the target prediction model.
Optionally, the longitudinal federated learning modeling optimization apparatus is further configured to:
acquiring the number of first party classified correct samples corresponding to the target prediction model and the number of first party classified error samples corresponding to the target prediction model;
generating a first party model weight by calculating a ratio of the number of the first party classification correct samples to the number of the first party classification error samples.
Optionally, the longitudinal federated learning modeling optimization apparatus is further configured to:
extracting a sample to be predicted, and acquiring an intermediate sample feature generated by a feature extractor of a target prediction model aiming at the sample to be predicted and a first party model prediction result generated by the target prediction model aiming at the sample to be predicted;
sending the intermediate sample features to second equipment, so that the second equipment jointly executes model prediction on the intermediate sample features and the ID matching samples corresponding to the samples to be predicted based on a longitudinal federated residual error lifting model, and a second-party model prediction result is obtained;
receiving a second-party model prediction result sent by the second device and a second-party model weight corresponding to the longitudinal federal residual error hoisting model;
and performing weighted aggregation on the first party model prediction result and the second party model prediction result based on the first party model weight and the second party model weight corresponding to the target prediction model to obtain a target federal prediction result.
The specific implementation of the longitudinal federated learning modeling optimization device in the application is basically the same as that of each embodiment of the longitudinal federated learning modeling optimization method, and is not described herein again.
The embodiment of the present application further provides a longitudinal federated learning modeling optimization apparatus, where the longitudinal federated learning modeling optimization apparatus is applied to a second device, and the longitudinal federated learning modeling optimization apparatus includes:
and the receiving module is used for acquiring the initial model weight of the second party and receiving the intermediate training sample characteristics, the training sample labels and the prediction loss of the first party model sent by the first equipment. The first-party model prediction loss is calculated by the first equipment based on a first-party model prediction result of a target prediction model on a training sample corresponding to the training sample ID matching sample and the training sample label, and the intermediate training sample features are obtained by feature extraction of the training sample by the first equipment based on a feature extractor of the target prediction model;
the model prediction module is used for acquiring a training sample ID matching sample, and performing model prediction on the training sample ID matching sample and the intermediate training sample based on a residual lifting model to be trained to obtain a second-party training model prediction result;
a calculation module for calculating a second party model prediction loss based on the training sample labels, the second party initial model weights and the second party training model prediction results;
and the iterative optimization module is used for iteratively optimizing the to-be-trained residual lifting model based on the residual losses generated by the first-party model prediction loss and the second-party model prediction loss to obtain the longitudinal federal residual lifting model.
Optionally, the longitudinal federated learning modeling optimization apparatus is further configured to:
acquiring a second party classified correct sample number and a second party classified error sample number corresponding to the longitudinal federal residual lifting model;
generating a second party model weight by calculating a ratio of the second party classification correct sample number and the second party classification error sample number.
Optionally, the longitudinal federated learning modeling optimization apparatus is further configured to:
receiving an intermediate sample characteristic sent by first equipment, and searching an ID matching sample corresponding to the intermediate sample characteristic;
performing model prediction on the ID matching sample and the intermediate sample feature together based on the longitudinal federal residual lifting model to obtain a second square model prediction result;
and sending the second-party model prediction result and the second-party model weight corresponding to the longitudinal federal residual error hoisting model to the first equipment, so that the first equipment generates a first-party model prediction result for the to-be-predicted sample corresponding to the ID matching sample based on a target prediction model, a first-party model weight corresponding to the target prediction model, the second-party model prediction result and the second-party model weight, and generates a target federal prediction result, wherein the target prediction model is obtained by local iterative training of the first equipment.
The specific implementation of the longitudinal federated learning modeling optimization device in the application is basically the same as that of each embodiment of the longitudinal federated learning modeling optimization method, and is not described herein again.
The present application provides a medium, which is a readable storage medium, and the readable storage medium stores one or more programs, and the one or more programs are further executable by one or more processors for implementing the steps of any one of the above-mentioned longitudinal federal forecast optimization methods.
The specific implementation of the readable storage medium of the application is substantially the same as that of each embodiment of the longitudinal federal prediction optimization method, and is not described herein again.
The present application provides a medium, which is a readable storage medium, and the readable storage medium stores one or more programs, and the one or more programs are further executable by one or more processors for implementing the steps of any one of the above methods for longitudinal federal learning modeling optimization.
The specific implementation manner of the readable storage medium of the application is basically the same as that of each embodiment of the longitudinal federated learning modeling optimization method, and is not described herein again.
The present application provides a computer program product, and the computer program product includes one or more computer programs, which can also be executed by one or more processors for implementing the steps of any one of the above-mentioned longitudinal federal prediction optimization methods.
The specific implementation of the computer program product of the present application is substantially the same as each embodiment of the above-described longitudinal federal prediction optimization method, and is not described herein again.
The present application provides a computer program product, and the computer program product includes one or more computer programs, which can also be executed by one or more processors for implementing the steps of any one of the above methods for longitudinal federated learning modeling optimization.
The specific implementation of the computer program product of the present application is substantially the same as each embodiment of the above-described longitudinal federated learning modeling optimization method, and is not described herein again.
The above description is only a preferred embodiment of the present application, and not intended to limit the scope of the present application, and all modifications of equivalent structures and equivalent processes, which are made by the contents of the specification and the drawings, or which are directly or indirectly applied to other related technical fields, are included in the scope of the present application.

Claims (23)

1. A longitudinal federal prediction optimization method is applied to first equipment, and comprises the following steps:
extracting a sample to be predicted, and obtaining an intermediate sample feature generated by a feature extractor of a target prediction model for performing feature extraction on the sample to be predicted and a first party model prediction result generated by the target prediction model for performing model prediction on the sample to be predicted, wherein the target prediction model is obtained by local iterative training of first equipment;
sending the intermediate sample features to second equipment, so that the second equipment jointly executes model prediction on the intermediate sample features and the ID matching samples corresponding to the samples to be predicted based on a longitudinal federated residual error lifting model, and a second-party model prediction result is obtained;
obtaining a first-party model weight corresponding to the target prediction model, and receiving a second-party model prediction result sent by the second device and a second-party model weight corresponding to the longitudinal federal residual error hoisting model;
and carrying out weighted aggregation on the first party model prediction result and the second party model prediction result based on the first party model weight and the second party model weight to obtain a target federal prediction result.
2. The longitudinal federal prediction optimization method as claimed in claim 1, wherein prior to the step of sending the intermediate sample characteristics to a second device, the longitudinal federal prediction optimization method further comprises:
sending the ID of the sample to be predicted corresponding to the sample to be predicted to the second equipment so that the second equipment can search the ID matching sample corresponding to the ID of the sample to be predicted;
if the search failure information sent by the second equipment is received, taking the first party model prediction result as a target prediction result;
if the search failure information sent by the second equipment is not received, executing the following steps: the intermediate sample features are sent to a second device.
3. The longitudinal federal prediction optimization method of claim 1, wherein the longitudinal federal residual lifting model is obtained by performing residual learning based on longitudinal federal learning with the first device by the second device based on a longitudinal federal public sample in combination with a model prediction loss of the target prediction model in the first device on the longitudinal federal public sample, an intermediate public sample feature corresponding to the longitudinal federal public sample and a corresponding sample label.
4. The longitudinal federal prediction optimization method as claimed in claim 3, wherein before the steps of obtaining the intermediate sample features generated by feature extraction of the target prediction model by the feature extractor for the sample to be predicted and obtaining the first-party model prediction result generated by model prediction of the target prediction model for the sample to be predicted, the longitudinal federal prediction optimization method further comprises:
acquiring the weight of a first party initial model, and extracting training samples and training sample labels corresponding to the training samples;
acquiring intermediate training sample characteristics generated by characteristic extraction of a characteristic extractor of a target prediction model to be trained aiming at the training samples;
iteratively optimizing the target prediction model to be trained by calculating the prediction loss of the first party model corresponding to the target prediction model to be trained based on the training sample label, the training model prediction result corresponding to the intermediate training sample characteristic and the first party initial model weight to obtain the target prediction model;
and sending the training sample label, the intermediate training sample characteristic and the first party model prediction loss to second equipment, so that the second equipment calculates a second party model prediction loss based on a to-be-trained residual lifting model, a training sample ID matching sample corresponding to the training sample, the intermediate training sample characteristic, the sample label and the obtained second party initial model weight, and optimizes the to-be-trained residual lifting model based on the calculated residual loss of the second party model prediction loss and the first party model prediction loss to obtain the longitudinal federated residual lifting model.
5. A longitudinal federal prediction optimization method is applied to a second device, and comprises the following steps:
receiving an intermediate sample characteristic sent by first equipment, and searching an ID matching sample corresponding to the intermediate sample characteristic;
performing model prediction on the ID matching sample and the intermediate sample feature together based on a longitudinal federated residual lifting model to obtain a second-party model prediction result;
and obtaining a second square model weight corresponding to the longitudinal federated residual lifting model, and sending the second square model prediction result and the second square model weight to the first device, so that the first device generates a target federated prediction result based on a first square model prediction result generated by the target prediction model for a to-be-predicted sample corresponding to the ID matching sample, a first square model weight corresponding to the target prediction model, the second square model prediction result and the second square model weight, wherein the target prediction model is obtained by local iterative training of the first device.
6. The longitudinal federal prediction optimization method of claim 5, wherein the longitudinal federal residual lift model is obtained by performing residual learning based on longitudinal federal learning with the first device by using the second device based on a longitudinal federal public sample in combination with a model prediction loss of the target prediction model in the first device on the longitudinal federal public sample, an intermediate public sample feature corresponding to the longitudinal federal public sample and a corresponding sample label.
7. The longitudinal federation prediction optimization method of claim 5, wherein the step of performing model prediction on the ID matched sample and the intermediate sample feature together based on a longitudinal federation residual lifting model to obtain a second-party model prediction result comprises:
splicing the ID matching sample and the characteristics of the intermediate sample to obtain a characteristic enhanced sample;
and performing model prediction on the feature enhancement sample based on the longitudinal federal residual lifting model to obtain a second square model prediction result.
8. The longitudinal federal prediction optimization method of claim 5, wherein after the step of finding ID match samples, the longitudinal federal prediction optimization further comprises:
if the search is successful, executing the following steps: performing model prediction on the ID matching sample and the intermediate sample feature together based on a longitudinal federated residual lifting model to obtain a second-party model prediction result;
and if the search fails, feeding back search failure information to the first equipment, so that the first equipment takes a first party model prediction result generated by aiming at the sample to be predicted based on a target prediction model as a target prediction result after receiving the search failure information.
9. The longitudinal federal prediction optimization method as claimed in claim 5, wherein before the step of performing model prediction on the ID matching samples and the intermediate sample features together based on a longitudinal federal residual error boosting model to obtain a second-party model prediction result, the longitudinal federal prediction optimization method further comprises:
acquiring the weight of a second-party initial model, and receiving an intermediate training sample feature, a training sample label and a first-party model prediction loss sent by the first equipment, wherein the first-party model prediction loss is calculated by the first equipment based on a first-party model prediction result of a target prediction model on a training sample corresponding to the training sample ID matching sample and the training sample label, and the intermediate training sample feature is obtained by performing feature extraction on the training sample by the first equipment based on a feature extractor of the target prediction model;
acquiring a training sample ID matching sample, and performing model prediction on the training sample ID matching sample and the intermediate training sample feature together based on a residual lifting model to be trained to obtain a second-party training model prediction result;
calculating a second-party model prediction loss based on the training sample labels, the second-party initial model weights and the second-party training model prediction results;
and iteratively optimizing the residual lifting model to be trained based on the residual losses generated by the prediction loss of the first square model and the prediction loss of the second square model to obtain the longitudinal federal residual lifting model.
10. A longitudinal federated learning modeling optimization method is applied to a first device and comprises the following steps:
acquiring the weight of a first party initial model, and extracting training samples and training sample labels corresponding to the training samples;
acquiring intermediate training sample characteristics generated by characteristic extraction of a characteristic extractor of a target prediction model to be trained aiming at the training samples;
iteratively optimizing the target prediction model to be trained by calculating the prediction loss of the first party model corresponding to the target prediction model to be trained based on the training sample label, the training model prediction result corresponding to the intermediate training sample characteristic and the first party initial model weight to obtain the target prediction model;
and sending the training sample label, the intermediate training sample characteristics and the first party model prediction loss to second equipment so that the second equipment can calculate second party model prediction loss, and optimizing the residual lifting model to be trained based on the residual loss calculated by the second party model prediction loss and the first party model prediction loss to obtain a longitudinal federal residual lifting model.
11. The longitudinal federated learning modeling optimization method of claim 10, wherein the second-party model prediction loss is calculated based on a residual lifting model to be trained, a training sample ID matching sample corresponding to the training sample, the intermediate training sample features, the sample label, and the obtained second-party initial model weight.
12. The longitudinal federated learning modeling optimization method of claim 10, wherein the step of iteratively optimizing the target prediction model to be trained by calculating a first-party model prediction loss corresponding to the target prediction model to be trained based on the training sample labels, the training model prediction results corresponding to the intermediate training sample features, and the first-party initial model weight to obtain the target prediction model comprises:
converting the characteristics of the intermediate training samples into a training model prediction result based on a classifier in the target prediction model to be trained;
calculating a first party model prediction loss based on the training sample label, the training model prediction result and the first party initial model weight;
updating the first party initial model weight based on the training model prediction result and the training sample label;
and iteratively optimizing the target prediction model to be trained based on the prediction loss of the first party model and the updated initial model weight of the first party to obtain the target prediction model.
13. The longitudinal federated learning modeling optimization method of claim 10, wherein after the step of obtaining the target prediction model by iteratively optimizing the target prediction model to be trained by calculating a first-party model prediction loss corresponding to the target prediction model to be trained based on the training sample labels, the training model prediction results corresponding to the intermediate training sample features, and the first-party initial model weight, the longitudinal federated learning modeling optimization method further comprises:
acquiring the number of first party classified correct samples corresponding to the target prediction model and the number of first party classified error samples corresponding to the target prediction model;
generating a first party model weight by calculating a ratio of the number of the first party classification correct samples to the number of the first party classification error samples.
14. The longitudinal federal learning modeling optimization method of claim 10, wherein after the step of sending the training sample labels, the intermediate training sample features and the first party model prediction loss to a second device for the second device to calculate a second party model prediction loss, and optimizing the residual lifting model to be trained based on the second party model prediction loss and a residual loss calculated by the first party model prediction loss to obtain a longitudinal federal residual lifting model, the longitudinal federal learning modeling optimization method further comprises:
extracting a sample to be predicted, and acquiring an intermediate sample feature generated by a feature extractor of a target prediction model aiming at the sample to be predicted and a first party model prediction result generated by the target prediction model aiming at the sample to be predicted;
sending the intermediate sample features to second equipment, so that the second equipment jointly executes model prediction on the intermediate sample features and the ID matching samples corresponding to the samples to be predicted based on a longitudinal federated residual error lifting model, and a second-party model prediction result is obtained;
receiving a second-party model prediction result sent by the second device and a second-party model weight corresponding to the longitudinal federal residual error hoisting model;
and performing weighted aggregation on the first party model prediction result and the second party model prediction result based on the first party model weight and the second party model weight corresponding to the target prediction model to obtain a target federal prediction result.
15. A longitudinal federated learning modeling optimization method is applied to a second device and comprises the following steps:
acquiring the weight of a second-party initial model, and receiving the intermediate training sample characteristics, the training sample labels and the first-party model prediction loss sent by first equipment;
acquiring a training sample ID matching sample, and performing model prediction on the training sample ID matching sample and the intermediate training sample feature together based on a residual lifting model to be trained to obtain a second-party training model prediction result;
calculating a second-party model prediction loss based on the training sample labels, the second-party initial model weights and the second-party training model prediction results;
and iteratively optimizing the residual lifting model to be trained based on the residual losses generated by the prediction loss of the first square model and the prediction loss of the second square model to obtain the longitudinal federal residual lifting model.
16. The longitudinal federated learning modeling optimization method of claim 15, wherein the first-party model prediction loss is calculated by the first device based on a first-party model prediction result of a target prediction model on a training sample corresponding to the training sample ID matching sample and the training sample label, and the intermediate training sample feature is obtained by feature extraction performed by the first device based on a feature extractor of the target prediction model for the training sample.
17. The longitudinal federal learning modeling optimization method of claim 15, wherein after the step of iteratively optimizing the residual lifting model to be trained to obtain the longitudinal federal residual lifting model based on the residual losses generated by the first-party model prediction loss and the second-party model prediction loss, the longitudinal federal learning modeling optimization method further comprises:
acquiring a second party classified correct sample number and a second party classified error sample number corresponding to the longitudinal federal residual lifting model;
generating a second party model weight by calculating a ratio of the second party classification correct sample number and the second party classification error sample number.
18. The longitudinal federated learning modeling optimization method of claim 15, wherein after the step of iteratively optimizing the to-be-trained residual lifting model based on the residual losses to obtain a longitudinal federated residual lifting model, the longitudinal federated learning modeling optimization method further comprises:
receiving an intermediate sample characteristic sent by first equipment, and searching an ID matching sample corresponding to the intermediate sample characteristic;
performing model prediction on the ID matching sample and the intermediate sample feature together based on the longitudinal federal residual lifting model to obtain a second square model prediction result;
and sending the second-party model prediction result and the second-party model weight corresponding to the longitudinal federal residual error hoisting model to the first equipment, so that the first equipment generates a first-party model prediction result for the to-be-predicted sample corresponding to the ID matching sample based on a target prediction model, a first-party model weight corresponding to the target prediction model, the second-party model prediction result and the second-party model weight, and generates a target federal prediction result, wherein the target prediction model is obtained by local iterative training of the first equipment.
19. A longitudinal federal predictive optimization device, characterized in that it comprises: a memory, a processor, and a program stored on the memory for implementing the longitudinal federated prediction optimization method,
the memory is used for storing a program for realizing the longitudinal federal forecast optimization method;
the processor is configured to execute a program implementing the longitudinal federal predictive optimization method to implement the steps of the longitudinal federal predictive optimization method as claimed in any of claims 1 to 9.
20. A longitudinal federated learning modeling optimization apparatus, characterized in that the longitudinal federated learning modeling optimization apparatus comprises: a memory, a processor, and a program stored on the memory for implementing the longitudinal federated learning modeling optimization method,
the memory is used for storing a program for realizing the longitudinal federal learning modeling optimization method;
the processor is configured to execute a program implementing the longitudinal federated learning modeling optimization method to implement the steps of the longitudinal federated learning modeling optimization method recited in any one of claims 10 to 18.
21. A medium which is a readable storage medium, characterized in that the readable storage medium has stored thereon a program for implementing a longitudinal federal prediction optimization method, the program for implementing the longitudinal federal prediction optimization method being executed by a processor to implement the steps of the longitudinal federal prediction optimization method as claimed in any one of claims 1 to 9.
22. A medium being a readable storage medium, characterized in that the readable storage medium has stored thereon a program implementing a longitudinal federated learning modeling optimization method, which is executed by a processor to implement the steps of the longitudinal federated learning modeling optimization method recited in any one of claims 10 to 18.
23. A computer program product comprising a computer program, wherein the computer program when executed by a processor implements the steps of the longitudinal federated prediction optimization method of any one of claims 1 to 9, or the steps of the longitudinal federated learning modeling optimization method of any one of claims 10 to 18.
CN202110982929.9A 2021-08-25 2021-08-25 Longitudinal federal prediction optimization method, device, medium, and computer program product Pending CN113688986A (en)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN202110982929.9A CN113688986A (en) 2021-08-25 2021-08-25 Longitudinal federal prediction optimization method, device, medium, and computer program product
PCT/CN2021/139640 WO2023024349A1 (en) 2021-08-25 2021-12-20 Method for optimizing vertical federated prediction, device, medium, and computer program product

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110982929.9A CN113688986A (en) 2021-08-25 2021-08-25 Longitudinal federal prediction optimization method, device, medium, and computer program product

Publications (1)

Publication Number Publication Date
CN113688986A true CN113688986A (en) 2021-11-23

Family

ID=78582579

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110982929.9A Pending CN113688986A (en) 2021-08-25 2021-08-25 Longitudinal federal prediction optimization method, device, medium, and computer program product

Country Status (2)

Country Link
CN (1) CN113688986A (en)
WO (1) WO2023024349A1 (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2023024349A1 (en) * 2021-08-25 2023-03-02 深圳前海微众银行股份有限公司 Method for optimizing vertical federated prediction, device, medium, and computer program product

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116567652B (en) * 2023-05-19 2024-02-23 上海科技大学 Omnidirectional super-surface-assisted air calculation energized vertical federal learning method
CN116644372B (en) * 2023-07-24 2023-11-03 北京芯盾时代科技有限公司 Account type determining method and device, electronic equipment and storage medium

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109002861B (en) * 2018-08-10 2021-11-09 深圳前海微众银行股份有限公司 Federal modeling method, device and storage medium
US20210089964A1 (en) * 2019-09-20 2021-03-25 Google Llc Robust training in the presence of label noise
CN112001740A (en) * 2020-06-19 2020-11-27 南京理工大学 Combined prediction method based on adaptive neural network
CN112700010A (en) * 2020-12-30 2021-04-23 深圳前海微众银行股份有限公司 Feature completion method, device, equipment and storage medium based on federal learning
CN112785002A (en) * 2021-03-15 2021-05-11 深圳前海微众银行股份有限公司 Model construction optimization method, device, medium, and computer program product
CN113688986A (en) * 2021-08-25 2021-11-23 深圳前海微众银行股份有限公司 Longitudinal federal prediction optimization method, device, medium, and computer program product

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2023024349A1 (en) * 2021-08-25 2023-03-02 深圳前海微众银行股份有限公司 Method for optimizing vertical federated prediction, device, medium, and computer program product

Also Published As

Publication number Publication date
WO2023024349A1 (en) 2023-03-02

Similar Documents

Publication Publication Date Title
CN108898186B (en) Method and device for extracting image
CN113688986A (en) Longitudinal federal prediction optimization method, device, medium, and computer program product
CN108520470B (en) Method and apparatus for generating user attribute information
US20200004815A1 (en) Text entity detection and recognition from images
CN110046706B (en) Model generation method and device and server
CN111860868A (en) Training sample construction method, device, equipment and computer readable storage medium
WO2023024350A1 (en) Longitudinal federal prediction optimization method and device, medium, and computer program product
WO2022048195A1 (en) Longitudinal federation modeling method, apparatus, and device, and computer readable storage medium
CN112785002A (en) Model construction optimization method, device, medium, and computer program product
CN111815169A (en) Business approval parameter configuration method and device
CN114492601A (en) Resource classification model training method and device, electronic equipment and storage medium
CN111428883A (en) Federal modeling method, device and readable storage medium based on backward law
CN113628059A (en) Associated user identification method and device based on multilayer graph attention network
CN111768242A (en) Order-placing rate prediction method, device and readable storage medium
CN110633717A (en) Training method and device for target detection model
CN115359308A (en) Model training method, apparatus, device, storage medium, and program for identifying difficult cases
CN113792892A (en) Federal learning modeling optimization method, apparatus, readable storage medium, and program product
CN112801489A (en) Litigation case risk detection method, litigation case risk detection device, litigation case risk detection equipment and readable storage medium
CN111784402A (en) Multi-channel based order-descending rate prediction method and device and readable storage medium
CN115439700B (en) Image processing method and device and machine-readable storage medium
CN114237182B (en) Robot scheduling method and system
EP4332791A1 (en) Blockchain address classification method and apparatus
CN113673699A (en) Longitudinal federal prediction optimization method, device, medium, and computer program product
CN113793298A (en) Pulmonary nodule detection model construction optimization method, equipment, storage medium and product
CN115205778A (en) Scene recognition method, device, equipment and storage medium

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication