Data Processing Method, Apparatus, Device, and Storage Medium for Federated Transfer Learning

Abstract

The present invention discloses a data processing method, apparatus, device and storage medium for federated transfer learning, wherein the method comprises: determining each first feature extraction model through a first sample, a second sample and a plurality of feature extraction models; determining a trained prediction model and each second feature extraction model based on the first sample, a third sample, a prediction model to be trained and the first feature extraction model; and determining a target prediction model and each target feature extraction model based on the second sample, a fourth sample, a trained prediction model and the second feature extraction model. The present invention can obtain a target model that takes into account both model migration and model interpretability through federated transfer learning, transfers sample knowledge through the target feature extraction model, realizes the federation of sample data based on different participants, reduces the time consumption for the trained model to reach the target performance, and improves the utilization rate of computer computing resources.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a data processing method, apparatus, device, and storage medium for federated transfer learning. Background Art

[0002] Supervised machine learning training on large amounts of labeled data can achieve excellent performance and results. However, large labeled datasets are limited in both quantity and application domains, and manually labeling sufficient training data is often costly. Transfer learning is often used to address this problem. This involves training a discriminator to adjust the parameters of the transfer learning network. This minimizes the distributional offset between the source domain data and the target domain data under these parameters, resulting in better results when the transfer learning network is applied to the target domain to accomplish the desired task.

[0003] However, the lack of interpretability of deep learning models makes them difficult to use for transfer learning in applications that require model interpretability (such as financial risk control). Low-complexity deep learning models have a weak ability to learn transferable knowledge from raw data, and therefore have weak transfer capabilities. This creates a contradiction: deep learning models with strong transfer capabilities lack interpretability, while deep learning models with strong interpretability have weak transfer capabilities, resulting in deep learning models being unable to balance interpretability and transfer capabilities. Furthermore, different user data is currently stored on the servers of different data holders or participants. To protect user data privacy, different servers cannot directly interact with each other. That is, servers cannot share their own user data for joint modeling. As a result, each server can only perform modeling based on a small amount of user data, requiring longer training time for the trained model to achieve the target performance. This in turn requires a large amount of computer resources and computing power, resulting in low utilization of computer computing resources.

[0004] The above content is only used to assist in understanding the technical solution of the present invention and does not constitute an admission that the above content is prior art. Summary of the Invention

[0005] The main purpose of the present invention is to provide a data processing method, device, equipment and storage medium for federated transfer learning, aiming to solve the technical problems that deep learning models cannot take into account both interpretability and migration capabilities, and that different participants cannot directly interact with time series data, resulting in low utilization of computer computing resources.

[0006] To achieve the above objectives, the present invention provides a data processing method for federated transfer learning, which includes the following steps:

[0007] The first participant determines, by the first participant, each first feature extraction model based on a first sample corresponding to a first business scenario of the first participant, a second sample corresponding to a second business scenario, and a plurality of feature extraction models;

[0008] Determining a trained prediction model and each second feature extraction model based on the first sample, a third sample corresponding to the first business scenario of the second participant, the prediction model to be trained, and the first feature extraction model;

[0009] Based on the second sample, the fourth sample corresponding to the second business scenario of the third party, the trained prediction model and the second feature extraction model, a target prediction model and each target feature extraction model are determined.

[0010] Furthermore, the step of determining each first feature extraction model by the first participant based on the first sample corresponding to the first business scenario of the first participant, the second sample corresponding to the second business scenario, and the plurality of feature extraction models includes:

[0011] The first participant groups the features of the first sample based on a preset business requirement to obtain a preset number of first sample feature groups, and groups the features of the second sample based on the preset business requirement to obtain a preset number of second sample feature groups;

[0012] The first participant obtains, through a preset number of feature extraction models, a first feature representation corresponding to each first sample feature group and a second feature representation corresponding to each second sample feature group;

[0013] The first participant determines a plurality of first domain distinction loss values based on the first feature representation and its corresponding domain distinction model, and determines each third feature extraction model based on each first domain distinction loss value and its corresponding feature extraction model;

[0014] The first participant determines multiple second domain distinction loss values based on the second feature representation and its corresponding domain distinction model, and determines each first feature extraction model based on each second domain distinction loss value and the corresponding third feature extraction model.

[0015] Furthermore, the step of obtaining, by the first participant, a first feature representation corresponding to each first sample feature group and a second feature representation corresponding to each second sample feature group using a preset number of feature extraction models includes:

[0016] The first participant inputs each first sample feature group into a corresponding preset number of feature extraction models to obtain a first feature representation, and inputs each second sample feature group into a corresponding preset number of feature extraction models to obtain a second feature representation.

[0017] Furthermore, the step of determining, by the first participant, a plurality of first domain distinction loss values based on the first feature representation and its corresponding domain distinction model, and determining each third feature extraction model based on each first domain distinction loss value and its corresponding feature extraction model includes:

[0018] The first participant inputs each first feature representation into a corresponding domain distinction model to obtain a first domain distinction loss value corresponding to each first feature representation;

[0019] The first participant determines the updated domain distinction models based on the respective first domain distinction loss values, and updates the corresponding feature extraction models through domain adversarial learning based on the respective first domain distinction loss values to obtain respective third feature extraction models.

[0020] Furthermore, the step of determining, by the first participant, a plurality of second domain distinction loss values based on the second feature representation and its corresponding domain distinction model, and determining each first feature extraction model based on each second domain distinction loss value and the corresponding third feature extraction model includes:

[0021] The first participant inputs each second feature representation into the corresponding updated domain distinction model to obtain a second domain distinction loss value corresponding to each second feature representation;

[0022] The first participant determines the target domain discrimination model based on each second domain discrimination loss value, and updates the corresponding third feature extraction model through domain adversarial learning based on each second domain discrimination loss value to obtain each first feature extraction model.

[0023] Furthermore, the step of determining a trained prediction model and each second feature extraction model based on the first sample, the third sample corresponding to the first business scenario of the second participant, the prediction model to be trained, and the first feature extraction model includes:

[0024] Obtain the third sample of the second participant through the second participant, and obtain the first feature representation corresponding to the first sample through the first participant;

[0025] Based on the first feature representation and the third sample, the first prediction loss value of the prediction model to be trained is calculated through the longitudinal federated learning algorithm, and the prediction model to be trained and each first feature extraction model are updated based on the first prediction loss value to obtain the trained prediction model and each second feature extraction model.

[0026] Furthermore, the steps of calculating a first prediction loss value of the prediction model to be trained by a longitudinal federated learning algorithm based on the first feature representation and the third sample, and updating the prediction model to be trained and each first feature extraction model based on the first prediction loss value to obtain a trained prediction model and each second feature extraction model include:

[0027] Based on the first feature representation and the first feature representation set corresponding to the third sample, performing model training on the prediction model to be trained by a longitudinal federated learning algorithm to obtain a first prediction loss value;

[0028] Based on the first prediction loss value, updating the prediction model to be trained by a longitudinal federated learning algorithm to obtain a trained prediction model;

[0029] Based on the first prediction loss value, each first feature extraction model is updated through a longitudinal federated learning algorithm to obtain a second feature extraction model.

[0030] Furthermore, the step of determining a target prediction model and each target feature extraction model based on the second sample, a fourth sample corresponding to the second business scenario of the third party, the trained prediction model, and the second feature extraction model includes:

[0031] Obtain a fourth sample from the third party through the third party, and obtain a second feature representation corresponding to the second sample through the first party;

[0032] Based on the second feature representation and the fourth sample, the second prediction loss value of the trained prediction model is calculated through the longitudinal federated learning algorithm, and the trained prediction model and each second feature extraction model are updated based on the second prediction loss value to obtain the target prediction model and each target feature extraction model.

[0033] Furthermore, the steps of calculating a second prediction loss value of the trained prediction model through a longitudinal federated learning algorithm based on the second feature representation and the fourth sample, and updating the trained prediction model and each second feature extraction model based on the second prediction loss value to obtain a target prediction model and each target feature extraction model include:

[0034] Based on the second feature representation and the feature representation set corresponding to the fourth sample, the trained prediction model is trained by a longitudinal federated learning algorithm to obtain a second prediction loss value;

[0035] Based on the second prediction loss value, updating the trained prediction model to be trained through a longitudinal federated learning algorithm to obtain a target prediction model;

[0036] Based on the second prediction loss value, each second feature extraction model is updated through a longitudinal federated learning algorithm to obtain a target feature extraction model.

[0037] Furthermore, the target prediction model is a fraud score prediction model; the first business scenario includes a user's credit score scenario, the second business scenario includes a user's fraud score scenario, the first sample includes credit score data of a first participant, the second sample includes fraud score data of the first participant, the third sample includes credit score data of a second participant, and the fourth sample includes fraud score data of a third participant;

[0038] The step of determining a trained prediction model and each second feature extraction model based on the first sample, the third sample corresponding to the first business scenario of the second participant, the prediction model to be trained, and the first feature extraction model includes:

[0039] Determining a trained fraud score prediction model and each second feature extraction model based on the credit score data of the first participant, the credit score data of the second participant, the prediction model to be trained, and the first feature extraction model;

[0040] The step of determining a target prediction model and each target feature extraction model based on the second sample, a fourth sample corresponding to the second business scenario of the third party, the trained prediction model, and the second feature extraction model includes:

[0041] The fraud score prediction model and each target feature extraction model are determined based on the fraud score data of the first participant, the fraud score data of the third participant, the trained fraud score prediction model and the second feature extraction model.

[0042] Furthermore, after the step of determining the target prediction model and each target feature extraction model based on the second sample, the fourth sample corresponding to the second business scenario of the third party, the trained prediction model, and the second feature extraction model, the method further includes:

[0043] Obtaining a feature representation to be predicted corresponding to a first sample to be predicted in the first participant, wherein the first participant groups features of the first sample to be predicted based on preset business requirements to obtain feature groups of the sample to be predicted, and obtaining a feature representation to be predicted corresponding to each feature group of the sample to be predicted based on a target feature extraction model, wherein the first sample to be predicted is a sample corresponding to the second business scenario;

[0044] Obtaining a second sample to be predicted from the third party, where the second sample to be predicted is a second business scenario sample;

[0045] Based on each feature representation to be predicted and the feature representation set corresponding to the second sample to be predicted, the prediction result of the target prediction model is calculated through the longitudinal federated learning algorithm.

[0046] In addition, to achieve the above-mentioned object, the present invention further provides a data processing device for federated transfer learning, wherein the data processing device for federated transfer learning comprises:

[0047] A first model determination module is configured to determine, by the first participant, first feature extraction models based on a first sample corresponding to a first business scenario of the first participant, a second sample corresponding to a second business scenario, and a plurality of feature extraction models;

[0048] A second model determination module is configured to determine a trained prediction model and each second feature extraction model based on the first sample, a third sample corresponding to the first business scenario of the second participant, the prediction model to be trained, and the first feature extraction model;

[0049] The target model determination module is used to determine the target prediction model and each target feature extraction model based on the second sample, the fourth sample corresponding to the second business scenario of the third participant, the trained prediction model and the second feature extraction model.

[0050] In addition, to achieve the above-mentioned purpose, the present invention also provides a data processing device for federated transfer learning, wherein the data processing device for federated transfer learning includes: a memory, a processor, and a data processing program for federated transfer learning stored on the memory and runnable on the processor. When the data processing program for federated transfer learning is executed by the processor, the steps of the aforementioned data processing method for federated transfer learning are implemented.

[0051] In addition, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, on which a data processing program for federated transfer learning is stored. When the data processing program for federated transfer learning is executed by a processor, the steps of the aforementioned data processing method for federated transfer learning are implemented.

[0052] In addition, to achieve the above-mentioned objectives, the present invention also provides a computer program product, which, when executed, implements the steps of the aforementioned data processing method for federated transfer learning.

[0053] The present invention determines each first feature extraction model based on a first sample corresponding to a first business scenario in the first participant, a second sample corresponding to a second business scenario and a plurality of feature extraction models by a first participant; then determines a trained prediction model and each second feature extraction model based on the first sample, a third sample corresponding to the first business scenario in the second participant, a prediction model to be trained and a first feature extraction model; then determines a target prediction model and each target feature extraction model based on the second sample, a fourth sample corresponding to the second business scenario in the third participant, a trained prediction model and a second feature extraction model, and determines the target feature extraction model by the prediction model to be trained and the first feature extraction model. The model underwent two rounds of training, achieving the goal of jointly establishing a high-performance target prediction model through federated learning by leveraging the rich features of the first participant, the category labels of the second participant, and the category labels of the third participant. This enabled federated transfer learning to obtain a target model that takes into account both model transfer and model interpretability. The target feature extraction model was used to transfer sample knowledge, and predictions were made using the interpretable target prediction model, which explained the contribution of each feature group to the prediction results. This achieved the goal of balancing model transfer and model interpretability. The federated construction of the model based on sample data from different participants improved the accuracy of predictions such as fraud scoring for the samples being predicted. At the same time, the federated construction of the model based on sample data from different participants overcomes the drawbacks of existing technologies that consume a large amount of computer computing resources, reduces the time required for the trained model to reach target performance, and improves the utilization rate of computer computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 1 is a schematic diagram of the structure of a data processing device for federated transfer learning in a hardware operating environment involved in an embodiment of the present invention;

[0055] Figure 2 This is a flow chart of the first embodiment of the data processing method for federated transfer learning of the present invention;

[0056] Figure 3 This is a sample schematic diagram of the first business scenario in the data processing method for federated transfer learning of the present invention;

[0057] Figure 4 This is a sample schematic diagram of the second business scenario in the data processing method of federated transfer learning of the present invention;

[0058] Figure 5 Schematic diagram of the training process of federated transfer learning in the data processing method of federated transfer learning of the present invention;

[0059] Figure 6 Schematic diagram of the prediction process in the data processing method of federated transfer learning of the present invention;

[0060] Figure 7 This is a schematic diagram of the functional modules of an embodiment of data processing for federated transfer learning of the present invention.

[0061] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0062] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0063] like Figure 1 As shown, Figure 1 It is a structural diagram of a data processing device for federated transfer learning in a hardware operating environment involved in an embodiment of the present invention.

[0064] The data processing device for federated transfer learning in the embodiment of the present invention can be a PC, or it can be a mobile terminal device with display function such as a smart phone, a tablet computer, an e-book reader, an MP3 (Moving Picture Experts Group Audio Layer III, Moving Picture Experts Compression Standard Audio Layer 3) player, an MP4 (Moving Picture Experts Group Audio Layer IV, Moving Picture Experts Compression Standard Audio Layer 4) player, or a portable computer.

[0065] like Figure 1 As shown, the data processing device for federated transfer learning may include: a processor 1001, such as a CPU, a network interface 1004, a user interface 1003, a memory 1005, and a communication bus 1002. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory, or a stable memory (non-volatile memory), such as a disk memory. The memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0066] Optionally, the data processing equipment for federated transfer learning can also include cameras, RF (Radio Frequency) circuits, sensors, audio circuits, WiFi modules, and the like. Sensors include light sensors, motion sensors, and other sensors. Of course, the data processing equipment for federated transfer learning can also be configured with other sensors, such as gyroscopes, barometers, hygrometers, thermometers, and infrared sensors, which are not detailed here.

[0067] Those skilled in the art will understand that Figure 1 The terminal structure shown in does not constitute a limitation on the data processing device for federated transfer learning, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0068] like Figure 1 As shown, the memory 1005 as a computer storage medium may include an operating system, a network communication module, a user interface module, and a data processing program for federated transfer learning.

[0069] exist Figure 1 In the terminal shown, the network interface 1004 is mainly used to connect to the background server and communicate data with the background server; the user interface 1003 is mainly used to connect to the client (user end) and communicate data with the client; and the processor 1001 can be used to call the data processing program of federated transfer learning stored in the memory 1005.

[0070] In this embodiment, the data processing device for federated transfer learning includes: a memory 1005, a processor 1001, and a data processing program for federated transfer learning stored on the memory 1005 and executable on the processor 1001. When the processor 1001 calls the data processing program for federated transfer learning stored in the memory 1005, it executes the steps of the data processing method for federated transfer learning in each of the following embodiments.

[0071] The present invention also provides a data processing method for federated transfer learning, referring to Figure 2 , Figure 2 This is a flow chart of the first embodiment of the data processing method for federated transfer learning of the present invention.

[0072] Step S101, the first participant determines each first feature extraction model based on a first sample corresponding to a first business scenario of the first participant, a second sample corresponding to a second business scenario, and multiple feature extraction models;

[0073] In this embodiment, the first business scenario and the second business scenario are two similar but different business scenarios, and the samples corresponding to the first business scenario and the samples corresponding to the second business scenario have the same feature space. For example, the first business scenario is credit scoring and fraud scoring, and the second business scenario is fraud scoring.

[0074] The data processing method for federated transfer learning in this embodiment includes at least three participants, for example, participant A (first participant), participant B (second participant), and participant C (third participant). Participants A and B have a large number of overlapping samples in the first business scenario s. Participants A and C have a certain number of overlapping samples in the second business scenario t.

[0075] The first sample is the sample corresponding to the first business scenario of the first participant, that is, the data corresponding to the first sample is data from the first business scenario; the second sample is the sample corresponding to the second business scenario of the first participant, that is, the data corresponding to the second sample is data from the second business scenario; the third sample is the sample corresponding to the first business scenario of the second participant, that is, the data corresponding to the third sample is data from the first business scenario; the fourth sample is the sample corresponding to the second business scenario of the third participant, that is, the data corresponding to the fourth sample is data from the second business scenario. The first and third samples are overlapping samples of participants A and B in the first business scenario, respectively, and the first and fourth samples are overlapping samples of participants A and C in the second business scenario, respectively. The number of samples in the first sample is (far) greater than the number of samples in the second sample.

[0076] The data corresponding to the first sample, the data corresponding to the second sample, the data corresponding to the third sample, and the data corresponding to the fourth sample can all be tabular data. In the tabular data, each column represents a feature (such as age, gender, job, salary, etc.); each row represents a sample (such as a user of a certain financial product).

[0077] Participant A provides different domain labels (or business labels) for samples from different business scenarios, i.e., each sample in the first sample has a different domain label than each sample in the second sample, each sample in the first sample has the same domain label, and each sample in the second sample has the same domain label. For example, the domain label of each sample in the first sample is 1, and the domain label of each sample in the second sample is 0. Participant B has a category label for each overlapping sample in the first business scenario s, i.e., each sample in the third sample has a category label; Participant C has a category label for a small number of overlapping samples in business scenario t, i.e., each sample in the fourth sample has a category label.

[0078] In this embodiment, when conducting federated transfer learning, the first participant determines each first feature extraction model based on the first sample corresponding to the first business scenario in the first participant, the second sample corresponding to the second business scenario, and multiple feature extraction models. Specifically, first determine the multiple first feature representations corresponding to the first sample and the multiple second feature representations corresponding to the second sample through each feature extraction model. Since the samples in the first sample and the samples in the second sample both have domain labels, each first feature representation is input into the corresponding domain differentiation model for model training to obtain an updated domain differentiation model and multiple first domain differentiation loss values, and update the corresponding feature extraction model according to the first domain differentiation loss value to obtain a third feature extraction model, wherein each first feature representation input corresponds to a domain differentiation model, wherein each first feature representation input corresponds to a feature extraction model.

[0079] Then, the second feature representation is input into the corresponding updated domain distinction model for model training, the target domain distinction model and multiple second domain distinction loss values are determined, and the corresponding third feature extraction model is updated according to the second domain distinction loss value to obtain the first feature extraction model.

[0080] Step S102: determining a trained prediction model and each second feature extraction model based on the first sample, the third sample corresponding to the first business scenario of the second participant, the prediction model to be trained, and the first feature extraction model;

[0081] In this embodiment, after determining each first feature extraction model, a third sample corresponding to the first business scenario of the second participant is obtained, and the third sample is an overlapping sample corresponding to the first sample in the sample of the second participant. The third sample belongs to the first business scenario, and based on the first sample, the third sample, the prediction model to be trained and the first feature extraction model, the trained prediction model and each second feature extraction model are determined. Specifically, the first feature representation and the first prediction loss value corresponding to the third sample are determined through the prediction model to be trained, that is, the first feature representation and the third sample corresponding to the first sample are model trained by the prediction model to be trained to obtain the first prediction loss value, and the prediction model to be trained and each first feature extraction model are updated according to the first prediction loss value to obtain the trained prediction model and each second feature extraction model.

[0082] Step S103: Determine a target prediction model and each target feature extraction model based on the second sample, the fourth sample corresponding to the second business scenario of the third party, the trained prediction model and the second feature extraction model.

[0083] In this embodiment, after obtaining the trained prediction model and each second feature extraction model, a fourth sample corresponding to the second business scenario of the third participant is obtained. The fourth sample is an overlapping sample corresponding to the second sample in the samples of the third participant, and the fourth sample belongs to the second business scenario.

[0084] Then, based on the second sample, the fourth sample, the trained prediction model and each second feature extraction model, the target prediction model and each target feature extraction model are determined. Specifically, the second prediction loss value corresponding to the second feature representation and the fourth sample is determined through the trained prediction model, that is, the second feature representation and the fourth sample are trained by the trained prediction model to obtain the second prediction loss value, and the trained prediction model and each second feature extraction model are updated according to the second prediction loss value to obtain the target prediction model and each target feature extraction model.

[0085] The prediction model to be trained is an interpretable machine learning model, for example, the prediction model to be trained is a linear regression model or a logistic regression model.

[0086] Reference Figure 3 , Figure 3 In the example, participant A is the first participant and participant B is the second participant. Participants A and B have a large number of overlapping samples in the first business scenario s. i A,s is the first sample feature group corresponding to the first sample, f j B,s is the monomer feature corresponding to the third sample.

[0087] Reference Figure 4 , Figure 4 In the example, participant A is the first participant and participant C is the third participant. Participants A and C have a large number of overlapping samples in the second business scenario t. i A,t is the second sample feature group corresponding to the second sample, f j C,t is the monomer feature corresponding to the fourth sample.

[0088] Preferably, the target prediction model is a fraud scoring prediction model; the first business scenario includes a user's credit scoring scenario, the second business scenario includes a user's fraud scoring scenario, the first sample includes credit scoring data in the first participant, the second sample is fraud scoring data in the first participant, the third sample is credit scoring data in the second participant, and the fourth sample is fraud scoring data in the third participant.

[0089] Step S102 includes: determining a trained fraud score prediction model and each second feature extraction model based on the credit score data of the first participant, the credit score data of the second participant, the prediction model to be trained, and the first feature extraction model;

[0090] Step S103 includes: determining a fraud score prediction model and each target feature extraction model based on the fraud score data of the first participant, the fraud score data of the third participant, the trained fraud score prediction model and the second feature extraction model.

[0091] Among them, credit score data includes the user's age, gender, job, salary, education level, deposit information, fixed asset information, loan information, early repayment record, overdue repayment record, etc.; fraud score data includes the user's age, gender, job, salary, education level, deposit information, fixed asset information, loan information, early repayment record, overdue repayment record, pre-repayment information, pre-loan information, etc.

[0092] Specifically, the credit score data of the first participant and the credit score data of the second participant are input into the prediction model to be trained for federal training, and the prediction model to be trained and the first feature extraction model are updated according to the training results to obtain the trained fraud score prediction model and each second feature extraction model; then, the trained fraud score prediction model is federally trained according to the fraud score data of the first participant and the fraud score data of the third participant, and the trained fraud score prediction model and the second feature extraction model are updated according to the training results to obtain the fraud score prediction model and each target feature extraction model, and then, the features of the samples to be predicted (including the first sample to be predicted of the first participant and the second sample to be predicted of the third participant) can be extracted through the target feature extraction model, and federal prediction can be performed through the fraud score prediction model to obtain the prediction result corresponding to the second sample to be predicted, that is, the fraud score.

[0093] The data processing method of federated transfer learning proposed in this embodiment is that the first participant determines each first feature extraction model based on the first sample corresponding to the first business scenario of the first participant, the second sample corresponding to the second business scenario, and multiple feature extraction models; then determines the trained prediction model and each second feature extraction model based on the first sample, the third sample corresponding to the first business scenario of the second participant, the prediction model to be trained, and the first feature extraction model; then determines the target prediction model and each target feature extraction model based on the second sample, the fourth sample corresponding to the second business scenario of the third participant, the trained prediction model, and the second feature extraction model, and determines the target feature extraction model and each target feature extraction model by the prediction model to be trained. The model and the first feature extraction model are trained in two rounds, and federated learning is used to leverage the rich features of the first participant, the category labels of the second participant, and the category labels of the third participant to jointly establish a high-performance target prediction model. Federated transfer learning can be used to obtain a target model that takes into account both model migration and model interpretability. The knowledge of the sample is transferred through the target feature extraction model, and predictions are made through the interpretable target prediction model. The contribution of each feature group to the prediction results can be explained, thus achieving the goal of taking into account both model migration and model interpretability. Since the model is constructed based on the federation of sample data from different participants, the accuracy of predictions such as fraud scoring for the samples to be predicted is improved. At the same time, the federation of sample data from different participants overcomes the drawbacks of existing technologies that consume a large amount of computer computing resources, reduces the time required for the trained model to reach target performance, and improves the utilization rate of computer computing resources.

[0094] Based on the first embodiment, a second embodiment of the data processing method for federated transfer learning of the present invention is proposed. In this embodiment, step S101 includes:

[0095] Step S201: The first participant groups the features of the first sample based on the preset business requirements to obtain a preset number of first sample feature groups, and groups the features of the second sample based on the preset business requirements to obtain a preset number of second sample feature groups.

[0096] Step S202: The first participant obtains a first feature representation corresponding to each first sample feature group and a second feature representation corresponding to each second sample feature group using a preset number of feature extraction models;

[0097] Step S203: The first participant determines a plurality of first domain distinction loss values based on the first feature representation and its corresponding domain distinction model, and determines each third feature extraction model based on each first domain distinction loss value and its corresponding feature extraction model;

[0098] In step S204, the first participant determines multiple second domain distinction loss values based on the second feature representation and its corresponding domain distinction model, and determines each first feature extraction model based on each second domain distinction loss value and the corresponding third feature extraction model.

[0099] In this embodiment, in order to make the number of first feature representations the same as the number of second feature representations, for the first business scenario and the second business scenario, the first and second business scenarios are first determined to have some features that are completely identical as preset business requirements. Then, the first participant groups the features of the first sample based on the preset business requirements to obtain a first sample feature group corresponding to the first sample, and groups the features of the second sample based on the preset business requirements to obtain a second sample feature group corresponding to the second sample. The first participant sequentially inputs the first sample feature group into the corresponding feature extraction model to obtain the first feature representation, and sequentially inputs the second sample feature group into the corresponding feature extraction model to obtain the second feature representation.

[0100] Furthermore, in one embodiment, step S202 includes:

[0101] The first participant inputs each first sample feature group into a corresponding preset number of feature extraction models to obtain a first feature representation, and inputs each second sample feature group into a corresponding preset number of feature extraction models to obtain a second feature representation.

[0102] Specifically, the first participant may number the first sample feature group, the second sample feature group and the feature extraction model respectively to achieve a one-to-one correspondence between the first sample feature group, the second sample feature group and the feature extraction model according to the numbering, that is, each sample feature group in the first sample feature group (or the second sample feature group) corresponds to a feature extraction model, and then the first participant inputs each sample feature group in the first sample feature group into the corresponding feature extraction model to obtain multiple first feature representations, and the first participant inputs each sample feature group in the second sample feature group into the corresponding feature extraction model to obtain multiple second feature representations.

[0103] Then, the first participant determines multiple first domain distinction loss values based on the first feature representation and the domain distinction model, and determines a third feature extraction model based on the first domain distinction loss values and the feature extraction model. The first feature representation corresponds to the domain distinction model one-to-one, that is, each sample feature group in the first sample feature group (or the second sample feature group) corresponds to a domain distinction model. The first participant inputs each first feature representation into the corresponding domain distinction model for model training to obtain the first domain distinction loss value corresponding to each first feature representation, and updates the corresponding feature extraction model based on each first domain distinction loss value to obtain the third feature extraction model. At the same time, the corresponding domain distinction model is updated based on each first domain distinction loss value to obtain an updated domain distinction model.

[0104] Finally, the first participant determines multiple second domain discrimination loss values based on the second feature representation and the updated domain discrimination model, and determines the first feature extraction model based on the second domain discrimination loss values and the third feature extraction model. Specifically, the first participant inputs each second feature representation into the corresponding updated domain discrimination model for model training to obtain the second domain discrimination loss value corresponding to each second feature representation, and updates the corresponding third feature extraction model based on each second domain discrimination loss value to obtain the first feature extraction model. At the same time, the first participant updates the corresponding updated domain discrimination model based on each second domain discrimination loss value to obtain the target domain discrimination model.

[0105] The data processing method of federated transfer learning proposed in this embodiment is that the first participant groups the features of the first sample based on preset business requirements to obtain a preset number of first sample feature groups, and groups the features of the second sample based on preset business requirements to obtain a preset number of second sample feature groups; then the first participant obtains the first feature representation corresponding to each first sample feature group and the second feature representation corresponding to each second sample feature group through a preset number of feature extraction models; then the first participant determines multiple first domain distinction loss values based on the first feature representation and its corresponding domain distinction model, and determines each third feature extraction model based on each first domain distinction loss value and the corresponding feature extraction model; finally, the first participant determines multiple second domain distinction loss values based on the second feature representation and its corresponding domain distinction model, and determines each first feature extraction model based on each second domain distinction loss value and the corresponding third feature extraction model, thereby updating the domain distinction model and the feature extraction model and improving the accuracy of model training.

[0106] Based on the second embodiment, a third embodiment of the data processing method for federated transfer learning of the present invention is proposed. In this embodiment, step S203 includes:

[0107] Step S301: The first participant inputs each first feature representation into a corresponding domain distinction model to obtain a first domain distinction loss value corresponding to each first feature representation;

[0108] In step S302, the first participant determines updated domain distinction models based on the respective first domain distinction loss values, and updates corresponding feature extraction models through domain adversarial learning based on the respective first domain distinction loss values to obtain respective third feature extraction models.

[0109] In this embodiment, when the first feature representation is obtained, the first participant inputs each first feature representation into the corresponding domain distinction model for model training to obtain the first domain distinction loss value corresponding to each first feature representation, that is, the first participant inputs the first feature representation into the corresponding domain distinction model respectively, obtains the first predicted domain label corresponding to each first feature representation, and calculates the first domain distinction loss value corresponding to each first feature representation based on the first predicted domain label and the domain label corresponding to each first feature representation.

[0110] Then, based on each first domain discrimination loss value, the corresponding feature extraction model is updated through domain adversarial learning to obtain a third feature extraction model, and the corresponding domain discrimination model is updated according to each first domain discrimination loss value to obtain an updated domain discrimination model.

[0111] Furthermore, in one embodiment, step S204 includes:

[0112] Step S303: The first participant inputs each second feature representation into the corresponding updated domain distinction model to obtain a second domain distinction loss value corresponding to each second feature representation;

[0113] In step S304, the first participant determines the target domain discrimination model based on each second domain discrimination loss value, and updates the corresponding third feature extraction model through domain adversarial learning based on each second domain discrimination loss value to obtain each first feature extraction model.

[0114] In this embodiment, each second feature representation is input into the corresponding updated domain distinction model in turn for model training to obtain the second domain distinction loss value corresponding to each second feature representation, that is, the second feature representation is input into the corresponding updated domain distinction model respectively to obtain the second predicted domain label corresponding to each second feature representation, and the second domain distinction loss value corresponding to each second feature representation is calculated according to the second predicted domain label and the domain label corresponding to each second feature representation.

[0115] Then, based on each second domain discrimination loss value, domain adversarial learning is used to update the corresponding third feature extraction model to obtain the first feature extraction model, and the corresponding updated domain discrimination model is updated according to each second domain discrimination loss value to obtain the target domain discrimination model. In the iterative training process of subsequent model training, the target domain discrimination model is used as the domain discrimination model.

[0116] It should be noted that the number of training times for the domain distinction model, i.e., the preset number of training times, can be set in advance. When the target domain distinction model is obtained, the number of training times corresponding to the domain distinction model is updated, that is, the number of training times is added by 1 to obtain a new number of training times. If the updated number of training times reaches the preset number of training times, the current feature extraction model training is completed; if the updated number of training times does not reach the preset number of training times, the target domain distinction model is used as the domain distinction model, the first feature extraction model is used as the feature extraction model, and the process returns to step S301.

[0117] Reference Figure 5 , Figure 5 In the first sample feature group, the feature group F1 A,s And the feature group F1 in the second sample feature group A,t The target feature model R0 is obtained by training, ... through the feature group F in the first sample feature group n A,s And the feature group F in the second sample feature group n A,t Train to get the updated feature model R n-1 , D0...D2 is the domain differentiation model, L d,0 ...L d,3 Discrimination loss value for the domain.

[0118] The data processing method for federated transfer learning proposed in this embodiment is that the first participant inputs each first feature representation into the corresponding domain discrimination model to obtain the first domain discrimination loss value corresponding to each first feature representation; then the first participant determines the updated domain discrimination models based on the each first domain discrimination loss value, and updates the corresponding feature extraction models through domain adversarial learning based on the each first domain discrimination loss value to obtain each third feature extraction model, thereby realizing the update of the domain discrimination model and the feature extraction model, and further improving the accuracy of model training.

[0119] Based on the first embodiment, a fourth embodiment of the data processing method for federated transfer learning of the present invention is proposed. In this embodiment, step S102 includes:

[0120] Step S401: obtaining the third sample of the second participant through the second participant, and obtaining the first feature representation corresponding to the first sample through the first participant;

[0121] Step S402: Based on the first feature representation and the third sample, the first prediction loss value of the prediction model to be trained is calculated through the longitudinal federated learning algorithm, and the prediction model to be trained and each first feature extraction model are updated based on the first prediction loss value to obtain the trained prediction model and each second feature extraction model.

[0122] In this embodiment, when the first feature extraction model is obtained, the third sample of the second participant is first obtained through the second participant, and the first feature representation corresponding to the first sample is obtained through the first participant. The third sample does not need to group features; each sample feature in the third sample is a single feature.

[0123] Then, based on the first feature representation and the third sample, the first prediction loss value of the prediction model to be trained is calculated through the longitudinal federated learning algorithm, that is, based on the longitudinal federated learning, the first prediction loss value corresponding to the first feature representation and the third sample is determined through the prediction model to be trained, and the prediction model to be trained and each first feature extraction model are updated based on the first prediction loss value to obtain the trained prediction model and each second feature extraction model, that is, based on the longitudinal federated learning, the first feature representation and the third sample are input into the interpretable prediction model to be trained for model training to obtain the first prediction loss value. Specifically, the feature representation set corresponding to the first feature representation and the third sample is determined, and based on the longitudinal federated learning, the feature representation set corresponding to the first feature representation and the third sample is input into the interpretable prediction model to be trained for model training. The feature representation set corresponding to the first feature representation and the third sample is a data set including all first feature representations and third samples. The prediction model to be trained is updated according to the first prediction loss value to obtain the trained prediction model, and the first feature extraction model is updated according to the first prediction loss value to obtain each second feature extraction model.

[0124] It should be noted that, in other embodiments, when determining the second feature extraction model, the number of training rounds of the prediction model to be trained is accumulated. If the number of training rounds is less than a preset value, the trained prediction model is used as the prediction model to be trained, the second feature extraction model is used as the first feature extraction model, and step S402 is continued.

[0125] The data processing method of federated transfer learning proposed in this embodiment obtains the third sample of the second participant through the second participant, and obtains the first feature representation corresponding to the first sample through the first participant; then based on the first feature representation and the third sample, the first prediction loss value of the prediction model to be trained is calculated through the longitudinal federated learning algorithm, and the prediction model to be trained and each first feature extraction model are updated based on the first prediction loss value to obtain the trained prediction model and each second feature extraction model. By training the prediction model to be trained according to the first feature representation and the third sample, the accuracy of model training is further improved.

[0126] Based on the fourth embodiment, a fifth embodiment of the data processing method for federated transfer learning of the present invention is proposed. In this embodiment, step S402 includes:

[0127] Step S501: Based on the first feature representation and the first feature representation set corresponding to the third sample, the prediction model to be trained is trained by a longitudinal federated learning algorithm to obtain a first prediction loss value;

[0128] Step S502: Based on the first prediction loss value, the prediction model to be trained is updated by a longitudinal federated learning algorithm to obtain a trained prediction model;

[0129] Step S503: Based on the first predicted loss value, each first feature extraction model is updated through a longitudinal federated learning algorithm to obtain a second feature extraction model.

[0130] In this embodiment, after determining the first feature extraction model, the feature representation set corresponding to the first feature representation and the third sample is determined, and based on the first feature representation and the first feature representation set corresponding to the third sample, the prediction model to be trained is trained through a longitudinal federated learning algorithm to obtain a first prediction loss value, that is, the feature representation set corresponding to the first feature representation and the third sample is input into an interpretable prediction model to be trained based on longitudinal federated learning to perform model training to obtain the corresponding first prediction category label, and the first prediction loss value is calculated based on the first prediction category label and the category label (true category label) corresponding to the third sample. The feature representation set corresponding to the first feature representation and the third sample is a data set including all first feature representations and third samples.

[0131] Then, based on the first prediction loss value, the prediction model to be trained is updated through the longitudinal federated learning algorithm to obtain the trained prediction model, and based on the first prediction loss value, the first feature extraction model is updated through the longitudinal federated learning algorithm to obtain the second feature extraction model.

[0132] The data processing method of federated transfer learning proposed in this embodiment trains the prediction model to be trained through a longitudinal federated learning algorithm based on the first feature representation and the first feature representation set corresponding to the third sample to obtain a first prediction loss value; then, based on the first prediction loss value, the prediction model to be trained is updated through the longitudinal federated learning algorithm to obtain a trained prediction model; and then, based on the first prediction loss value, each first feature extraction model is updated through the longitudinal federated learning algorithm to obtain a second feature extraction model, thereby realizing the updating of the prediction model to be trained and the first feature extraction model according to the first feature representation and the first prediction loss value corresponding to the third sample, thereby further improving the accuracy of model training.

[0133] Based on the first embodiment, a sixth embodiment of the data processing method for federated transfer learning of the present invention is proposed. In this embodiment, step S103 includes:

[0134] Step S601: obtaining a fourth sample from the third participant through the third participant, and obtaining a second feature representation corresponding to the second sample through the first participant;

[0135] Step S602: Based on the second feature representation and the fourth sample, the second prediction loss value of the trained prediction model is calculated by the longitudinal federated learning algorithm, and the trained prediction model and each second feature extraction model are updated based on the second prediction loss value to obtain the target prediction model and each target feature extraction model.

[0136] In this embodiment, when the second feature extraction model is obtained, the third party first obtains the fourth sample from the third party, and then obtains the second feature representation corresponding to the second sample through the first party. The fourth sample does not require feature grouping; each sample feature in the fourth sample is a single feature.

[0137] Then, based on the second feature representation and the fourth sample, the second prediction loss value of the trained prediction model is calculated through the longitudinal federated learning algorithm, that is, the second prediction loss value corresponding to the second feature representation and the fourth sample is determined through the trained prediction model based on the longitudinal federated learning, and the trained prediction model and each second feature extraction model are updated based on the second prediction loss value to obtain the target prediction model and each target feature extraction model, that is, the second feature representation and the fourth sample are input into the trained prediction model for model training based on the longitudinal federated learning to obtain the second prediction loss value. Specifically, the feature representation set corresponding to the second feature representation and the fourth sample is determined, and the feature representation set corresponding to the second feature representation and the fourth sample is input into the trained prediction model for model training based on the longitudinal federated learning. The feature representation set corresponding to the second feature representation and the fourth sample is a data set including all second feature representations and fourth samples. The trained prediction model is updated according to the second prediction loss value to obtain the target prediction model, and the second feature extraction model is updated according to the second prediction loss value to obtain each target feature extraction model.

[0138] Reference Figure 5 , Figure 5 In the first feature representation corresponding to the first sample feature group, r i A,s (r1 A,s 、r2 A,s 、r3 A,s etc.) and monomer characteristics f j B,s (f1 B,s 、f2 B,s 、f3 B,s etc.) Input the prediction model to be trained and obtain the prediction loss value L cls B,s , and finally according to the predicted loss value L cls B,s Each updated feature extraction model is updated to obtain a target feature extraction model, and the prediction model to be trained is updated to obtain a target prediction model G.

[0139] The data processing method of federated transfer learning proposed in this embodiment obtains the fourth sample from the third participant through the third participant, and obtains the second feature representation corresponding to the second sample through the first participant; then based on the second feature representation and the fourth sample, the second prediction loss value of the trained prediction model is calculated through the longitudinal federated learning algorithm, and the trained prediction model and each second feature extraction model are updated based on the second prediction loss value to obtain the target prediction model and each target feature extraction model. The target model is obtained by training according to the second feature representation and the fourth sample, thereby further improving the accuracy of model training.

[0140] Based on the sixth embodiment, a seventh embodiment of the data processing method for federated transfer learning of the present invention is proposed. In this embodiment, step S602 includes:

[0141] Step S701: Based on the second feature representation and the feature representation set corresponding to the fourth sample, the trained prediction model is trained using a longitudinal federated learning algorithm to obtain a second prediction loss value;

[0142] Step S702: Based on the second prediction loss value, the trained prediction model to be trained is updated by a longitudinal federated learning algorithm to obtain a target prediction model;

[0143] Step S703: Based on the second predicted loss value, each second feature extraction model is updated through a longitudinal federated learning algorithm to obtain a target feature extraction model.

[0144] In this embodiment, after determining the second feature extraction model, the feature representation set corresponding to the second feature representation and the fourth sample is determined. Based on the second feature representation and the feature representation set corresponding to the fourth sample, the trained prediction model is trained through a longitudinal federated learning algorithm to obtain a second prediction loss value, that is, the feature representation set corresponding to the second feature representation and the fourth sample is input into the trained prediction model for model training based on longitudinal federated learning to obtain the corresponding first prediction category label, and the second prediction loss value is calculated based on the first prediction category label and the category label (true category label) corresponding to the fourth sample. The feature representation set corresponding to the second feature representation and the fourth sample is a data set including all second feature representations and fourth samples.

[0145] Then, according to the second prediction loss value, the trained prediction model is updated through the longitudinal federated learning algorithm to obtain the target prediction model, and according to the second prediction loss value, the second feature extraction model is updated through the longitudinal federated learning algorithm to obtain each target feature extraction model.

[0146] The data processing method of federated transfer learning proposed in this embodiment trains the trained prediction model through a longitudinal federated learning algorithm based on the second feature representation and the feature representation set corresponding to the fourth sample to obtain a second prediction loss value; then, based on the second prediction loss value, the trained prediction model to be trained is updated through the longitudinal federated learning algorithm to obtain a target prediction model; and then, based on the second prediction loss value, each second feature extraction model is updated through the longitudinal federated learning algorithm to obtain a target feature extraction model, thereby realizing the updating of the trained prediction model and the second feature extraction model according to the second feature representation and the fourth sample, further improving the accuracy of model training.

[0147] Based on the above embodiments, an eighth embodiment of the data processing method for federated transfer learning of the present invention is proposed. In this embodiment, after step S103, the method further includes:

[0148] Step S801: Obtain a feature representation to be predicted corresponding to a first sample to be predicted in the first participant, wherein the first participant groups features of the first sample to be predicted based on preset business requirements to obtain feature groups of the sample to be predicted, and obtains a feature representation to be predicted corresponding to each feature group of the sample to be predicted based on a target feature extraction model, wherein the first sample to be predicted is a sample corresponding to the second business scenario;

[0149] Step S802: obtaining a second sample to be predicted from a third party, where the second sample to be predicted is a second business scenario sample;

[0150] Step S803 , based on each feature representation to be predicted and the feature representation set corresponding to the second sample to be predicted, a prediction result of the target prediction model is calculated by a longitudinal federated learning algorithm.

[0151] It should be noted that after obtaining the target prediction model and the target feature extraction model, the target prediction model and the target feature extraction model can be used to perform federated prediction on the second business scenario based on the first participant and the third participant.

[0152] In this embodiment, the feature representation to be predicted corresponding to the first sample to be predicted in the first participant is obtained, wherein the first participant groups the features of the first sample to be predicted based on preset business requirements to obtain a feature group of the sample to be predicted, and obtains the feature representation to be predicted corresponding to each feature group of the sample to be predicted based on the target feature extraction model, and the first sample to be predicted is the sample corresponding to the second business scenario; specifically, the process of grouping the features of the first sample to be predicted is similar to the scoring process of the first sample, which will not be repeated here. Then, the second sample to be predicted in the third participant is obtained. The first sample to be predicted is the sample corresponding to the second business scenario in the first participant, that is, the data corresponding to the first sample to be predicted is the data of the second business scenario; the second sample to be predicted is the sample corresponding to the second business scenario in the third participant, that is, the data corresponding to the second sample to be predicted is the data of the second business scenario.

[0153] Next, the feature representation set corresponding to each feature representation to be predicted and the second sample to be predicted is determined. Specifically, each feature representation to be predicted and the second sample to be predicted are concatenated to obtain the feature representation set, which includes all features of each feature representation to be predicted and the second sample to be predicted. Then, based on the feature representation set corresponding to each feature representation to be predicted and the second sample to be predicted, the prediction result of the target prediction model is calculated using a longitudinal federated learning algorithm. That is, based on longitudinal federated learning, the feature representation set corresponding to each feature representation to be predicted and the second sample to be predicted is input into the target prediction model to obtain a prediction result, and the target prediction model can be used to accurately perform federated predictions on the samples of the first and third participants.

[0154] Reference Figure 6 , Figure 6 China, F1 A,t 、F2 A,t 、F3 A,t is the feature group of the sample to be predicted corresponding to the first sample to be predicted, f1 C ,t 、f2 C,t 、f3 C,t 、f4 C,t F1 is the single feature to be predicted corresponding to the second sample to be predicted, A,t 、F2 A,t 、F3 A,t Input the corresponding target feature extraction models R1, R2, and R3 respectively to obtain multiple feature representations to be predicted r1 A,t 、r2 A,t 、r3 At Then, multiple feature representations to be predicted and the monomer features to be predicted are input into the target prediction model G to obtain the prediction result L cls C,s , according to the prediction result L cls C,s Determine the class labels.

[0155] The data processing method of federated transfer learning proposed in this embodiment obtains the feature representation to be predicted corresponding to the first sample to be predicted in the first participant, wherein the first participant groups the features of the first sample to be predicted based on preset business requirements to obtain a feature group of the sample to be predicted, and obtains the feature representation to be predicted corresponding to each feature group of the sample to be predicted based on the target feature extraction model, the first sample to be predicted is the sample corresponding to the second business scenario, and then obtains the second sample to be predicted in the third participant, and then calculates the prediction result of the target prediction model based on the vertical federated learning algorithm based on each feature representation to be predicted and the feature representation set corresponding to the second sample to be predicted, and can accurately perform federated prediction on the samples of the first participant and the third participant through the target prediction model, the target feature extraction model obtained through transfer learning can improve the accuracy of the prediction result, and the target prediction model can explain the reason for the prediction result, so as to achieve a balance between model migration and model interpretability.

[0156] The embodiment of the present invention also provides a data processing device for federated transfer learning, referring to Figure 7 , Figure 7 This is a functional module diagram of an embodiment of a data processing device for federated transfer learning according to the present invention. The data processing device for federated transfer learning includes:

[0157] A first model determination module 100 is configured to determine, by a first participant, first feature extraction models based on a first sample corresponding to a first business scenario of the first participant, a second sample corresponding to a second business scenario, and a plurality of feature extraction models;

[0158] A second model determination module 200 is configured to determine a trained prediction model and each second feature extraction model based on the first sample, a third sample corresponding to the first business scenario of the second participant, the prediction model to be trained, and the first feature extraction model;

[0159] The target model determination module 300 is used to determine the target prediction model and each target feature extraction model based on the second sample, the fourth sample corresponding to the second business scenario of the third participant, the trained prediction model and the second feature extraction model.

[0160] Optionally, the first model determination module 100 is further configured to:

[0161] The first participant groups the features of the first sample based on a preset business requirement to obtain a preset number of first sample feature groups, and groups the features of the second sample based on the preset business requirement to obtain a preset number of second sample feature groups;

[0162] The first participant obtains, through a preset number of feature extraction models, a first feature representation corresponding to each first sample feature group and a second feature representation corresponding to each second sample feature group;

[0163] The first participant determines a plurality of first domain distinction loss values based on the first feature representation and its corresponding domain distinction model, and determines each third feature extraction model based on each first domain distinction loss value and its corresponding feature extraction model;

[0164] The first participant determines multiple second domain distinction loss values based on the second feature representation and its corresponding domain distinction model, and determines each first feature extraction model based on each second domain distinction loss value and the corresponding third feature extraction model.

[0165] Optionally, the first model determination module 100 is further configured to:

[0166] The first participant inputs each first sample feature group into a corresponding preset number of feature extraction models to obtain a first feature representation, and inputs each second sample feature group into a corresponding preset number of feature extraction models to obtain a second feature representation.

[0167] Optionally, the first model determination module 100 is further configured to:

[0168] The first participant inputs each first feature representation into a corresponding domain distinction model to obtain a first domain distinction loss value corresponding to each first feature representation;

[0169] The first participant determines the updated domain distinction models based on the respective first domain distinction loss values, and updates the corresponding feature extraction models through domain adversarial learning based on the respective first domain distinction loss values to obtain respective third feature extraction models.

[0170] Optionally, the first model determination module 100 is further configured to:

[0171] The first participant inputs each second feature representation into the corresponding updated domain distinction model to obtain a second domain distinction loss value corresponding to each second feature representation;

[0172] The first participant determines the target domain discrimination model based on each second domain discrimination loss value, and updates the corresponding third feature extraction model through domain adversarial learning based on each second domain discrimination loss value to obtain each first feature extraction model.

[0173] Optionally, the second model determination module 200 is further configured to:

[0174] Obtain the third sample of the second participant through the second participant, and obtain the first feature representation corresponding to the first sample through the first participant;

[0175] Based on the first feature representation and the third sample, the first prediction loss value of the prediction model to be trained is calculated through the longitudinal federated learning algorithm, and the prediction model to be trained and each first feature extraction model are updated based on the first prediction loss value to obtain the trained prediction model and each second feature extraction model.

[0176] Optionally, the second model determination module 200 is further configured to:

[0177] Based on the first feature representation and the first feature representation set corresponding to the third sample, performing model training on the prediction model to be trained by a longitudinal federated learning algorithm to obtain a first prediction loss value;

[0178] Based on the first prediction loss value, updating the prediction model to be trained by a longitudinal federated learning algorithm to obtain a trained prediction model;

[0179] Based on the first prediction loss value, each first feature extraction model is updated through a longitudinal federated learning algorithm to obtain a second feature extraction model.

[0180] Optionally, the target model determination module 300 is further configured to:

[0181] Obtain a fourth sample from the third party through the third party, and obtain a second feature representation corresponding to the second sample through the first party;

[0182] Based on the second feature representation and the fourth sample, the second prediction loss value of the trained prediction model is calculated through the longitudinal federated learning algorithm, and the trained prediction model and each second feature extraction model are updated based on the second prediction loss value to obtain the target prediction model and each target feature extraction model.

[0183] Optionally, the target model determination module 300 is further configured to:

[0184] Based on the second feature representation and the feature representation set corresponding to the fourth sample, the trained prediction model is trained by a longitudinal federated learning algorithm to obtain a second prediction loss value;

[0185] Based on the second prediction loss value, updating the trained prediction model to be trained through a longitudinal federated learning algorithm to obtain a target prediction model;

[0186] Based on the second prediction loss value, each second feature extraction model is updated through a longitudinal federated learning algorithm to obtain a target feature extraction model.

[0187] Optionally, the target prediction model is a fraud scoring prediction model; the first business scenario includes a user's credit scoring scenario, the second business scenario includes a user's fraud scoring scenario, the first sample includes credit scoring data in the first participant, the second sample is fraud scoring data in the first participant, the third sample is credit scoring data in the second participant, and the fourth sample is fraud scoring data in the third participant.

[0188] The second model determination module 200 is further configured to:

[0189] Determining a trained fraud score prediction model and each second feature extraction model based on the credit score data of the first participant, the credit score data of the second participant, the prediction model to be trained, and the first feature extraction model;

[0190] The target model determination module 300 is further configured to:

[0191] The fraud score prediction model and each target feature extraction model are determined based on the fraud score data of the first participant, the fraud score data of the third participant, the trained fraud score prediction model and the second feature extraction model.

[0192] Optionally, the data processing device for federated transfer learning further includes:

[0193] Obtaining a feature representation to be predicted corresponding to a first sample to be predicted in the first participant, wherein the first participant groups features of the first sample to be predicted based on preset business requirements to obtain feature groups of the sample to be predicted, and obtaining a feature representation to be predicted corresponding to each feature group of the sample to be predicted based on a target feature extraction model, wherein the first sample to be predicted is a sample corresponding to the second business scenario;

[0194] Obtaining a second sample to be predicted from the third party, where the second sample to be predicted is a second business scenario sample;

[0195] Based on each feature representation to be predicted and the feature representation set corresponding to the second sample to be predicted, the prediction result of the target prediction model is calculated through the longitudinal federated learning algorithm.

[0196] In addition, an embodiment of the present invention further proposes a storage medium, on which a data processing program for federated transfer learning is stored. When the data processing program for federated transfer learning is executed by a processor, the steps of the data processing method for federated transfer learning described above are implemented.

[0197] Among them, the method implemented when the data processing program of federated transfer learning running on the processor is executed can refer to the various embodiments of the data processing method of federated transfer learning of the present invention, and will not be repeated here.

[0198] In addition, an embodiment of the present invention also proposes a computer program product, which, when executed, implements the steps of the data processing method for federated transfer learning as described above.

[0199] Among them, the method implemented when the computer program product executes the data processing program of federated transfer learning can refer to the various embodiments of the data processing method of federated transfer learning of the present invention, and will not be repeated here.

[0200] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or system comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or system. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or system comprising the element.

[0201] The serial numbers of the above embodiments of the present invention are for description only and do not represent the advantages or disadvantages of the embodiments.

[0202] Through the description of the above embodiments, those skilled in the art can clearly understand that the above embodiment methods can be implemented by means of software plus the necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better embodiment. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for enabling a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present invention.

[0203] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A data processing method for federated transfer learning, characterized in that: The data processing method of federated transfer learning includes the following steps: The first participant determines, by the first participant, each first feature extraction model based on a first sample corresponding to a first business scenario of the first participant, a second sample corresponding to a second business scenario, and a plurality of feature extraction models; Determining a trained prediction model and each second feature extraction model based on the first sample, a third sample corresponding to the first business scenario of the second participant, the prediction model to be trained, and the first feature extraction model; Determining a target prediction model and each target feature extraction model based on the second sample, a fourth sample corresponding to the second business scenario of the third party, the trained prediction model, and the second feature extraction model; Obtaining a feature representation to be predicted corresponding to a first sample to be predicted in the first participant, wherein the first participant groups features of the first sample to be predicted based on preset business requirements to obtain feature groups of the sample to be predicted, and obtaining a feature representation to be predicted corresponding to each feature group of the sample to be predicted based on a target feature extraction model, wherein the first sample to be predicted is a sample corresponding to the second business scenario; Obtaining a second sample to be predicted from the third party, where the second sample to be predicted is a second business scenario sample; Based on the feature representations of each to-be-predicted sample and the feature representation set corresponding to the second to-be-predicted sample, the prediction result of the target prediction model is calculated through a longitudinal federated learning algorithm; Among them, the target prediction model is a fraud scoring prediction model; the first business scenario includes the user's credit scoring scenario, the second business scenario includes the user's fraud scoring scenario, the first sample is the credit scoring data of the first participant, the second sample is the fraud scoring data of the first participant, the third sample is the credit scoring data of the second participant, and the fourth sample is the fraud scoring data of the third participant.

2. The data processing method for federated transfer learning according to claim 1, wherein: The step of determining each first feature extraction model by the first participant based on the first sample corresponding to the first business scenario of the first participant, the second sample corresponding to the second business scenario, and the plurality of feature extraction models includes: The first participant groups the features of the first sample based on a preset business requirement to obtain a preset number of first sample feature groups, and groups the features of the second sample based on the preset business requirement to obtain a preset number of second sample feature groups; The first participant obtains, through a preset number of feature extraction models, a first feature representation corresponding to each first sample feature group and a second feature representation corresponding to each second sample feature group; The first participant determines a plurality of first domain distinction loss values based on the first feature representation and its corresponding domain distinction model, and determines each third feature extraction model based on each first domain distinction loss value and its corresponding feature extraction model; The first participant determines multiple second domain distinction loss values based on the second feature representation and its corresponding domain distinction model, and determines each first feature extraction model based on each second domain distinction loss value and the corresponding third feature extraction model.

3. The data processing method for federated transfer learning according to claim 2, wherein: The step of obtaining, by the first participant, a first feature representation corresponding to each first sample feature group and a second feature representation corresponding to each second sample feature group using a preset number of feature extraction models includes: The first participant inputs each first sample feature group into a corresponding preset number of feature extraction models to obtain a first feature representation, and inputs each second sample feature group into a corresponding preset number of feature extraction models to obtain a second feature representation.

4. The data processing method for federated transfer learning according to claim 2, wherein: The step of determining, by the first participant, a plurality of first domain distinction loss values based on the first feature representation and its corresponding domain distinction model, and determining each third feature extraction model based on each first domain distinction loss value and its corresponding feature extraction model includes: The first participant inputs each first feature representation into a corresponding domain distinction model to obtain a first domain distinction loss value corresponding to each first feature representation; The first participant determines the updated domain distinction models based on the respective first domain distinction loss values, and updates the corresponding feature extraction models through domain adversarial learning based on the respective first domain distinction loss values to obtain respective third feature extraction models.

5. The data processing method for federated transfer learning according to claim 4, wherein: The step of determining, by the first participant, a plurality of second domain distinction loss values based on the second feature representation and its corresponding domain distinction model, and determining each first feature extraction model based on each second domain distinction loss value and the corresponding third feature extraction model includes: The first participant inputs each second feature representation into the corresponding updated domain distinction model to obtain a second domain distinction loss value corresponding to each second feature representation; The first participant determines the target domain discrimination model based on each second domain discrimination loss value, and updates the corresponding third feature extraction model through domain adversarial learning based on each second domain discrimination loss value to obtain each first feature extraction model.

6. The data processing method for federated transfer learning according to claim 1, wherein: The step of determining a trained prediction model and each second feature extraction model based on the first sample, the third sample corresponding to the first business scenario of the second participant, the prediction model to be trained, and the first feature extraction model includes: Obtain the third sample of the second participant through the second participant, and obtain the first feature representation corresponding to the first sample through the first participant; Based on the first feature representation and the third sample, the first prediction loss value of the prediction model to be trained is calculated through the longitudinal federated learning algorithm, and the prediction model to be trained and each first feature extraction model are updated based on the first prediction loss value to obtain the trained prediction model and each second feature extraction model.

7. The data processing method for federated transfer learning according to claim 6, wherein: The steps of calculating a first prediction loss value of the prediction model to be trained by a longitudinal federated learning algorithm based on the first feature representation and the third sample, and updating the prediction model to be trained and each first feature extraction model based on the first prediction loss value to obtain a trained prediction model and each second feature extraction model include: Based on the first feature representation and the first feature representation set corresponding to the third sample, training the prediction model to be trained by a longitudinal federated learning algorithm to obtain a first prediction loss value; Based on the first prediction loss value, updating the prediction model to be trained by a longitudinal federated learning algorithm to obtain a trained prediction model; Based on the first prediction loss value, each first feature extraction model is updated through a longitudinal federated learning algorithm to obtain a second feature extraction model.

8. The data processing method for federated transfer learning according to claim 1, wherein: The step of determining a target prediction model and each target feature extraction model based on the second sample, a fourth sample corresponding to the second business scenario of the third party, the trained prediction model, and the second feature extraction model includes: Obtain a fourth sample from the third party through the third party, and obtain a second feature representation corresponding to the second sample through the first party; Based on the second feature representation and the fourth sample, the second prediction loss value of the trained prediction model is calculated through the longitudinal federated learning algorithm, and the trained prediction model and each second feature extraction model are updated based on the second prediction loss value to obtain the target prediction model and each target feature extraction model.

9. The data processing method for federated transfer learning according to claim 8, wherein: The steps of calculating a second prediction loss value of the trained prediction model by a longitudinal federated learning algorithm based on the second feature representation and the fourth sample, and updating the trained prediction model and each second feature extraction model based on the second prediction loss value to obtain a target prediction model and each target feature extraction model include: Based on the second feature representation and the feature representation set corresponding to the fourth sample, the trained prediction model is trained by a longitudinal federated learning algorithm to obtain a second prediction loss value; Based on the second prediction loss value, updating the trained prediction model to be trained through a longitudinal federated learning algorithm to obtain a target prediction model; Based on the second prediction loss value, each second feature extraction model is updated through a longitudinal federated learning algorithm to obtain a target feature extraction model.

10. The data processing method for federated transfer learning according to claim 1, wherein: The step of determining a trained prediction model and each second feature extraction model based on the first sample, the third sample corresponding to the first business scenario of the second participant, the prediction model to be trained, and the first feature extraction model includes: Determining a trained fraud score prediction model and each second feature extraction model based on the credit score data of the first participant, the credit score data of the second participant, the prediction model to be trained, and the first feature extraction model; The step of determining a target prediction model and each target feature extraction model based on the second sample, a fourth sample corresponding to the second business scenario of the third party, the trained prediction model, and the second feature extraction model includes: The fraud score prediction model and each target feature extraction model are determined based on the fraud score data of the first participant, the fraud score data of the third participant, the trained fraud score prediction model and the second feature extraction model.

11. A data processing device for federated transfer learning, characterized in that: The data processing device of the federated transfer learning includes: A first model determination module is configured to determine, by the first participant, first feature extraction models based on a first sample corresponding to a first business scenario of the first participant, a second sample corresponding to a second business scenario, and a plurality of feature extraction models; A second model determination module is configured to determine a trained prediction model and each second feature extraction model based on the first sample, a third sample corresponding to the first business scenario of the second participant, the prediction model to be trained, and the first feature extraction model; a target model determination module, configured to determine a target prediction model and each target feature extraction model based on the second sample, a fourth sample corresponding to the second business scenario of the third party, the trained prediction model, and the second feature extraction model; The data processing device for federated transfer learning also includes: Obtaining a feature representation to be predicted corresponding to a first sample to be predicted in the first participant, wherein the first participant groups features of the first sample to be predicted based on preset business requirements to obtain feature groups of the sample to be predicted, and obtaining a feature representation to be predicted corresponding to each feature group of the sample to be predicted based on a target feature extraction model, wherein the first sample to be predicted is a sample corresponding to the second business scenario; Obtaining a second sample to be predicted from the third party, where the second sample to be predicted is a second business scenario sample; Based on the feature representations of each to-be-predicted sample and the feature representation set corresponding to the second to-be-predicted sample, the prediction results of the target prediction model are calculated through the longitudinal federated learning algorithm. Among them, the target prediction model is a fraud scoring prediction model; the first business scenario includes the user's credit scoring scenario, the second business scenario includes the user's fraud scoring scenario, the first sample is the credit scoring data of the first participant, the second sample is the fraud scoring data of the first participant, the third sample is the credit scoring data of the second participant, and the fourth sample is the fraud scoring data of the third participant.

12. A data processing device for federated transfer learning, characterized in that: The data processing device for federated transfer learning includes: a memory, a processor, and a data processing program for federated transfer learning stored on the memory and executable on the processor. When the data processing program for federated transfer learning is executed by the processor, the steps of the data processing method for federated transfer learning according to any one of claims 1 to 10 are implemented.

13. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a data processing program for federated transfer learning, and when the data processing program for federated transfer learning is executed by a processor, the steps of the data processing method for federated transfer learning according to any one of claims 1 to 10 are implemented.

14. A computer program product, characterized in that When the computer program product is executed, the steps of the data processing method for federated transfer learning according to any one of claims 1 to 10 are implemented.

Images