Model construction method, device and storage medium based on federated learning

Abstract

The present application discloses a model building method, device and storage medium based on federated learning, the method comprising: when an evaluation model building instruction is detected, determining the business scenario and business goal based on the evaluation model building instruction; based on the business scenario and business goal, determining the target data, and performing feature mining and feature selection on the target data to obtain target feature data; based on the target feature data, executing a preset federated process to achieve federated modeling with other second participants and obtain a target model. In the present application, feature mining and feature selection are performed for specific business scenarios and business goals to accurately obtain target feature data that is consistent with the actual business, and by executing a preset federated process, accurate federated modeling with other second participants is achieved based on massive data and the target model is obtained, thereby improving the evaluation accuracy of enterprises with actual business usage scenarios.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and in particular to a model building method, device and storage medium based on federated learning. Background Art

[0002] With the continuous development of financial technology, especially Internet technology finance, more and more technologies (such as distributed, blockchain, artificial intelligence, etc.) are applied in the financial field, but the financial industry also puts forward higher requirements on technology. For example, the financial industry also has higher requirements for the construction of evaluation models for corporate evaluation.

[0003] As the country vigorously supports small and medium-sized enterprises, assessment models for enterprise assessment have also emerged. The current service providers that provide assessment models include Tianyancha and others. The public data used by Tianyancha to build the model comes from the National Enterprise Credit Information Publicity System, the National Judgment Documents Network, the National Enforcement Information Disclosure Network, the Trademark Office of the State Intellectual Property Office, the Copyright Office and other channels. Although the data for building the model is public, the assessment factors and algorithms used by Tianyancha in enterprise assessment are not disclosed, and for institutions with actual business usage scenarios, there are problems of scenario mismatch and inaccurate assessment. Summary of the invention

[0004] The main purpose of this application is to provide a model construction method, device and storage medium based on federated learning, aiming to solve the technical problems that the existing unified evaluation model is used for each enterprise, which is prone to scenario inconsistency and inaccurate evaluation in practical applications.

[0005] To achieve the above objectives, the present application provides a model building method based on federated learning, which is applied to a first participant. The model building method based on federated learning includes:

[0006] When an evaluation model building instruction is detected, determining a business scenario and a business goal based on the evaluation model building instruction;

[0007] Based on the business scenario and business goal, determine target data, and perform feature mining and feature selection on the target data to obtain target feature data;

[0008] Based on the target characteristic data, a preset federation process is executed to achieve federated modeling with other second participants and obtain a target model.

[0009] Optionally, the step of determining target data based on the business scenario and the business goal, and performing feature mining and feature selection on the target data to obtain target feature data includes:

[0010] Based on the business scenario and business goal, select characteristic factors from a preset characteristic factor library;

[0011] Selecting target data matching the attributes of the characteristic factors from preset own data and public data;

[0012] Performing data preprocessing on the target data to obtain preprocessed data;

[0013] The preprocessed data is subjected to self-feature mining and self-feature selection, and the preprocessed data is subjected to public feature mining and public feature selection to obtain target feature data.

[0014] Optionally, after the step of implementing federated modeling with other second participants and obtaining a target model based on the target feature data by executing a preset federation process, the method includes:

[0015] Evaluate the model stability and accuracy of the target model using preset evaluation data, and determine whether the target model passes the model evaluation;

[0016] If the target model passes the model evaluation, the enterprise data to be processed is input into the target model to obtain the enterprise scoring result.

[0017] Optionally, if the target model passes the model evaluation, the step of inputting the to-be-processed enterprise data into the target model to obtain the enterprise scoring result includes:

[0018] If the target model passes the model evaluation, the target enterprise's to-be-processed enterprise data is input into the target model to obtain a model score for the target enterprise;

[0019] Obtain a third-party rating of the target enterprise, and combine the model rating and the third-party rating to obtain a rating result for the enterprise.

[0020] Optionally, the step of evaluating the target model for model stability and accuracy by using preset evaluation data, and determining whether the target model passes the model evaluation comprises:

[0021] Performing model perturbation on the target model by using the perturbation data in the preset evaluation data, and determining a first model error between a first prediction result obtained after the perturbation and a first preset result;

[0022] If the first model error is less than a first preset value, determining that the target model passes the stability assessment;

[0023] Verifying the model accuracy of the target model using the accuracy evaluation data in the preset evaluation data, and determining a second model error between a second prediction result obtained after verification and a second preset result;

[0024] If the second model error is less than a second preset value, it is determined that the target model passes the accuracy assessment.

[0025] Optionally, the step of implementing federated modeling with other second participants and obtaining a target model by executing a preset federated process based on the target feature data includes:

[0026] Obtaining a preset model to be trained, and obtaining a preset training completion condition of the preset model to be trained;

[0027] Based on the target feature data, in conjunction with other second participants, the model calculation intermediate variables of the preset model to be trained are optimized and updated until the preset model to be trained meets the preset training completion conditions, thereby obtaining the target model.

[0028] Optionally, the step of optimizing and updating the model calculation intermediate variables of the preset model to be trained in conjunction with other second participants based on the target feature data until the preset model to be trained meets the preset training completion condition to obtain the target model includes:

[0029] Inputting the target feature data into a preset model to be trained, so as to perform prediction processing on the target feature data based on the preset model to be trained to obtain a third prediction result, determining a third model error between the third prediction result and a third preset result of the target feature data, and updating the preset model to be trained based on the third model error to iteratively update the model calculation intermediate variables of the preset model to be trained;

[0030] Determine whether the preset model to be trained by iterative training meets the preset replacement and update conditions. If the preset model to be trained meets the preset replacement and update conditions, replace and update the intermediate variables of the model calculation by executing the preset longitudinal federated learning process to obtain the replaced and updated preset model to be trained;

[0031] Continue to iteratively train and replace and update the replaced and updated preset model to be trained until the preset model to be trained meets the preset training completion condition, thereby obtaining the target model.

[0032] The present application also provides a target subject evaluation method based on federated learning, which is applied to a first participant. The target subject evaluation method based on federated learning includes:

[0033] Determine business scenarios and business goals from received evaluation model building instructions;

[0034] Based on the business scenario and business goal, target data is determined, and feature mining and feature selection are performed on the target data to obtain target feature data;

[0035] Based on the target feature data, obtain a target model by conducting federated modeling with other second participants;

[0036] The target model is used to perform scoring processing on the enterprise data to be processed to obtain a scoring result for the enterprise.

[0037] Optionally, before the step of using the target subject evaluation model to score the enterprise data to be processed to obtain the scoring result of the enterprise, the method includes:

[0038] Evaluate the model stability and accuracy of the target model using preset evaluation data, and determine whether the target model passes the model evaluation;

[0039] If the target model passes the model evaluation, the step of using the target model to score the enterprise data to be processed is performed to obtain a scoring result for the enterprise.

[0040] Optionally, the step of determining target data based on the business scenario and the business goal, and performing feature mining and feature selection on the target data to obtain target feature data includes:

[0041] Based on the business scenario and business goal, select characteristic factors from a preset characteristic factor library;

[0042] Selecting target data matching the attributes of the characteristic factors from preset own data and public data;

[0043] Performing data preprocessing on the target data to obtain preprocessed data;

[0044] The preprocessed data is subjected to self-feature mining and self-feature selection, and the preprocessed data is subjected to public feature mining and public feature selection to obtain target feature data.

[0045] The present application also provides a model building device based on federated learning, which is applied to a first participant, and the model building device based on federated learning includes:

[0046] A detection module, configured to determine a business scenario and a business goal based on the evaluation model construction instruction when an evaluation model construction instruction is detected;

[0047] A first determination module is used to determine target data based on the business scenario and business goal, and perform feature mining and feature selection on the target data to obtain target feature data;

[0048] The federation module is used to implement federation modeling with other second parties and obtain a target model by executing a preset federation process based on the target feature data.

[0049] The present application also provides a target subject evaluation device based on federated learning, which is applied to a first participant. The target subject evaluation device based on federated learning includes:

[0050] A receiving module, used to determine a business scenario and a business goal from the received evaluation model building instruction;

[0051] A second determination module is used to determine target data based on the business scenario and business goal, and perform feature mining and feature selection on the target data to obtain target feature data;

[0052] A first acquisition module is used for obtaining a target model through federation modeling with other second participants based on the target feature data;

[0053] The second acquisition module is used to use the target model to perform scoring processing on the enterprise data to be processed to obtain the scoring results of the enterprise.

[0054] The present application also provides a model building device based on federated learning, which is a physical device. The model building device based on federated learning includes: a memory, a processor, and a program of the model building method based on federated learning stored in the memory and executable on the processor. When the program of the model building method based on federated learning is executed by the processor, the steps of the model building method based on federated learning as described above can be implemented.

[0055] The present application also provides a readable storage medium, on which is stored a program for implementing a model building method based on federated learning. When the program for the model building method based on federated learning is executed by a processor, the steps of the model building method based on federated learning as described above are implemented.

[0056] The present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the above-mentioned model building method based on federated learning.

[0057] The present application provides a model building method, device and storage medium based on federated learning. Compared with the prior art that adopts a unified evaluation model for each enterprise, which is prone to the problem of scenario inconsistency and inaccurate evaluation in actual application, the present application determines the business scenario and business goal based on the evaluation model building instruction when the evaluation model building instruction is detected; determines the target data based on the business scenario and business goal, and performs feature mining and feature selection on the target data to obtain target feature data; based on the target feature data, implements federated modeling with other second participants and obtains the target model by executing a preset federated process. In the present application, by first determining the business scenario and business goal when the evaluation model building instruction is detected, and then performing feature mining and feature selection for specific business scenarios and business goals, accurately obtains the target feature data that is consistent with the actual business, and then based on the target feature data, executes the preset federal process to accurately implement federated modeling with other second participants on the basis of massive data and obtain the target model, that is, accurately obtain the evaluation model, thereby improving the evaluation accuracy of enterprises with actual business usage scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0059] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0060] Figure 1 This is a flow chart of the first embodiment of the model building method based on federated learning in this application;

[0061] Figure 2 This is a detailed flow chart of step S10 of the first embodiment of the model building method based on federated learning in this application;

[0062] Figure 3 A schematic diagram of the device structure of the hardware operating environment involved in the embodiment of the present application;

[0063] Figure 4 Schematic diagram of the scenario for building a model based on federated learning for this application.

[0064] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0065] It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0066] The embodiments of the present invention provide embodiments of the model building method based on federated learning. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than here. The execution subjects of each embodiment of the model building method based on federated learning of the present invention can be devices such as smart phones, personal computers and servers. For ease of description, the following embodiments are described with the first participant as the execution subject.

[0067] The present application embodiment provides a model construction method based on federated learning. In the first embodiment of the model construction method based on federated learning of the present application, refer to Figure 1 , applied to the first participant, the model building method based on federated learning includes:

[0068] Step S10, when an evaluation model building instruction is detected, determining a business scenario and a business goal based on the evaluation model building instruction;

[0069] Step S20, based on the business scenario and business goal, determine target data, and perform feature mining and feature selection on the target data to obtain target feature data;

[0070] Step S30, based on the target feature data, by executing a preset federation process, to achieve federation modeling with other second participants and obtain a target model.

[0071] The specific steps are as follows:

[0072] Step S10, when an evaluation model building instruction is detected, determining a business scenario and a business goal based on the evaluation model building instruction;

[0073] In this embodiment, the model building method based on federated learning is applied to a model building device based on federated learning in the first participant, the model building device based on federated learning belongs to a model building system based on federated learning, the model building system based on federated learning belongs to a model building device based on federated learning, when an evaluation model building instruction is detected, a business scenario and a business goal are determined based on the evaluation model building instruction, wherein the triggering method of the evaluation model building instruction includes:

[0074] Method 1: Based on the model building interface, the user triggers the evaluation model building instruction by clicking, touching or speaking;

[0075] Method 2: A trigger program for the evaluation model construction instruction, or a trigger module for the evaluation model construction instruction is pre-set in the first participant. When the preset trigger conditions are met, the trigger program or the trigger module automatically triggers the evaluation model construction instruction.

[0076] It should be noted that the model construction system based on federated learning is provided with a configuration module, through which the model construction based on federated learning can be flexibly performed, wherein the flexible model construction based on federated learning includes at least the following situations:

[0077] Case 1: Flexibly determine the external data and / or internal data to be obtained through the configuration module;

[0078] Specifically, for example, the external data that needs to be obtained may be data from third parties such as the National Enterprise Credit Information Publicity System, the National Judgment Documents Network, the National Enforcement Information Disclosure Network, the Trademark Office of the State Intellectual Property Office, the Copyright Office, etc. The external data that needs to be obtained may also be market public data such as industry market data, industry news information, peer competitors, upstream and downstream companies, etc., and the internal data that needs to be obtained may be product activity, the number of products used by the company, abnormal behavior (including risky behavior, etc.), and other data.

[0079] Case 2: The feature indicators to be obtained are flexibly determined through the configuration module, and the feature indicators are classified into types through the configuration module, that is, the scoring modeling items are determined;

[0080] Specifically, for example, for the financial industry, the scoring modeling items can be financial related items, credit related items, business operation items, etc. For the patent industry, the scoring modeling items can be talent structure items, business scale items, etc.

[0081] Case 3: other participants who need to participate in the federation are flexibly determined through the configuration module, that is, in this embodiment, the participants of the federation can be obtained through configuration.

[0082] Specifically, for example, based on whether they have the same user, other participants who need to participate in the vertical federation can be determined.

[0083] In this embodiment, the method of flexibly constructing a model based on federated learning by setting a configuration module includes:

[0084] Method 1: The user selects the configuration items in the configuration module through the interface to flexibly build a model based on federated learning;

[0085] Method 2: Build the configuration items in the configuration module by importing them to flexibly build a model based on federated learning.

[0086] When an evaluation model building instruction is detected, the business scenario and business goal are determined based on the evaluation model building instruction, wherein the business scenario may refer to a financing scenario, a micro-loan scenario, or a credit loan scenario, a patent business scenario, etc., and the business goal may be whether to obtain a loan, the difficulty of financing, or patent litigation, etc., wherein the business scenario and business goal are carried in the evaluation model building instruction.

[0087] Step S20, based on the business scenario and business goal, determine target data, and perform feature mining and feature selection on the target data to obtain target feature data;

[0088] Based on the business scenario and business goal, the target data is determined. Specifically, based on the business scenario and business goal, the data source is determined, and based on the business scenario and business goal, the target data is selected from the data source. After the target data is determined, feature mining and feature selection are performed on the target data to obtain target feature data. Specifically, based on the evaluation model construction instruction, feature mining and feature selection are performed on the target data to obtain target feature data.

[0089] The step of determining target data based on the business scenario and business goal, and performing feature mining and feature selection on the target data to obtain target feature data includes:

[0090] Step S21, selecting a characteristic factor from a preset characteristic factor library based on the business scenario and business goal;

[0091] In this embodiment, after obtaining the business scenario and business goal, based on the business scenario and business goal, characteristic factors are selected from a preset characteristic factor library (including all banking business factors), or all characteristic factors are directly selected. Specifically, the characteristic factor can be a third party's basic credit score factor, industry trend characteristic factor, and product performance characteristic factor, wherein the characteristic factor can be a modeling item or a superordinate concept of a modeling item.

[0092] For example, the business scenario is a credit loan scenario, and the business goal is to determine whether to issue a credit loan. Specifically, the credit loan scenario and the corresponding characteristic factors that can determine whether to issue a credit loan are selected from a preset characteristic factor library, such as a credit score factor, an expected loan factor, a deposit income factor, an educational background factor, etc. In this embodiment, the mapping relationship between the business scenario and the business goal as a whole and the characteristic factors can be pre-set.

[0093] Step S22, selecting target data matching the attributes of the characteristic factor from the preset own data and public data;

[0094] Select target data that matches the attribute of the feature factor from the preset own data and public data. Specifically, based on the basic credit score factor, select the first target sub-data that matches the attribute of the basic credit score factor from the preset public data (preset third-party data); based on the industry trend feature factor, select the second target sub-data that matches the attribute of the industry trend feature factor from the preset public data; based on the product performance feature factor, select the third target sub-data that matches the attribute of the product performance feature factor from the preset own data. The attribute matching may refer to: first, the name or keyword of the corresponding feature factor in the target (sub) data; second, the data content (keyword value) of the corresponding feature factor in the target data. Combine the first target sub-data, the second target sub-data and the third target sub-data to obtain the target data.

[0095] In this embodiment, based on the business scenario and business objectives, characteristic factors are selected from a preset characteristic factor library, and then target data is selected from preset proprietary data and public data for subsequent preprocessing, rather than directly performing subsequent preprocessing based on preset proprietary data and public data. This can reduce the amount of data processing and improve data processing efficiency.

[0096] Step S23, performing data preprocessing on the target data to obtain preprocessed data;

[0097] The target data is preprocessed to obtain preprocessed data, wherein the preprocessing method includes: outlier correction, missing value removal, etc. In this embodiment, the purpose of preprocessing is to prevent the actual features from affecting the model training results.

[0098] Step S24, performing self-feature mining and self-feature selection on the preprocessed data, and performing public feature mining and public feature selection on the preprocessed data to obtain target feature data.

[0099] The preprocessed data is subjected to self-feature mining and self-feature selection, and the preprocessed data is subjected to public feature mining and public feature selection. Specifically, the preprocessed data is subjected to positive self-feature mining and positive self-feature selection, and the preprocessed data is subjected to positive public feature mining and positive public feature selection to obtain target feature data, wherein positive public features refer to positive public features in public feature data that have an impact on or are intended for business objectives, and positive self-features refer to positive self-features in self-feature data that have an impact on or are intended for business objectives. For example, if the conversion rate of an advertisement on a certain platform is to be obtained, the positive features are features in the user data collected after watching the advertisement, rather than features in all user data. In this embodiment, it should be noted that feature mining refers to the acquisition of model building item data, and feature selection refers to the selection of required feature data from model building item data, wherein feature data and feature selection may involve processing methods such as natural language processing, unstructured data processing, customer behavior sequence processing, and industry periodicity processing.

[0100] In this embodiment, after the preprocessed data is subjected to self-feature mining and self-feature selection, and the preprocessed data is subjected to public feature mining and public feature selection, the target feature data is obtained. Specifically, the target feature data refers to the model building item and the specific value of the model building item. For example, the model building item is product activity, and the specific value of product activity is 50,000 users with daily activity. In this embodiment, the preprocessed data is subjected to self-feature mining and self-feature selection, and the preprocessed data is subjected to public feature mining and public feature selection to obtain the target feature data. Therefore, the data training model that has an impact on or is intended for the business goal can be selected to improve the accuracy of the model.

[0101] Step S30, based on the target feature data, by executing a preset federation process, to achieve federation modeling with other second participants and obtain a target model.

[0102] After obtaining the target characteristic data, other second participants participating in the federation are determined, and based on the target characteristic data, a preset federation process is executed to achieve federation modeling with other second participants and obtain a target model.

[0103] In this embodiment, after obtaining the target model, the results of the enterprise evaluation can be obtained based on the target model. The evaluation results or output items of the enterprise evaluation are scores, labels (such as risky users, potential users, high-value users, dormant users, growing users, etc.) or rankings.

[0104] The step of implementing federated modeling with other second participants and obtaining a target model by executing a preset federated process based on the target feature data includes:

[0105] Step S31, obtaining a preset model to be trained, and obtaining a preset training completion condition of the preset model to be trained;

[0106] In this embodiment, before performing federated training, it is first necessary to obtain a preset model to be trained and obtain a preset training completion condition of the preset model to be trained. The preset model to be trained is a basic model of federated learning.

[0107] Step S32, based on the target feature data, in conjunction with other second participants, optimize and update the model calculation intermediate variables of the preset model to be trained until the preset model to be trained meets the preset training completion conditions, thereby obtaining the target model.

[0108] Based on the target feature data, in conjunction with other second participants, the model calculation intermediate variables of the preset model to be trained are optimized and updated until the preset model to be trained meets the preset training completion conditions, and the target model is obtained, wherein, based on the target feature data, a preset longitudinal federated learning process is executed to obtain aggregated model parameters that meet the preset iterative training termination conditions, so as to obtain the target model.

[0109] The step of optimizing and updating the model calculation intermediate variables of the preset model to be trained in conjunction with other second participants based on the target feature data until the preset model to be trained meets the preset training completion condition to obtain the target model includes:

[0110] Step M1, inputting the target feature data into a preset model to be trained, so as to perform prediction processing on the target feature data based on the preset model to be trained, obtaining a third prediction result, determining a third model error between the third prediction result and a third preset result of the target feature data, and updating the preset model to be trained based on the third model error, so as to iteratively update the model calculation intermediate variables of the preset model to be trained;

[0111] In this embodiment, the target feature data is input into a preset model to be trained, so as to perform prediction processing on the target feature data based on the preset model to be trained to obtain a third prediction result, and a third model error between the third prediction result and a third preset result of the target feature data is determined. The preset model to be trained is updated based on the third model error to train and update the model calculation intermediate variables of the preset model to be trained.

[0112] Specifically, the intermediate variables calculated by the model may be model parameters, or gradients, etc.

[0113] Step M2, determining whether the preset model to be trained by iterative training meets the preset replacement and update conditions. If the preset model to be trained meets the preset replacement and update conditions, the preset vertical federated learning process is executed to replace and update the intermediate variables of the model calculation for training and update, so as to obtain the replaced and updated preset model to be trained;

[0114] Determine whether the preset model to be trained after iterative training has reached a preset replacement and update condition, and the preset replacement and update condition may be reaching the first number of iterative training times, such as iterative training 1 time or 500 times. If the preset replacement and update condition is not reached, return to the step of inputting the target feature data into the preset model to be trained, so as to perform prediction processing on the target feature data based on the preset model to be trained, obtain a third prediction result, and determine the third model error between the third prediction result and the third preset result of the target feature data, until the iteratively updated preset model to be trained reaches the preset replacement and update condition. If the preset model to be trained reaches the preset replacement and update condition, the preset longitudinal federated learning process is executed to replace and update the intermediate variables calculated by the training update model to obtain the replaced and updated preset model to be trained, wherein the replacement and update of the preset model to be trained by executing the preset longitudinal federated learning process refers to: Figure 4 As shown, the intermediate variables of model calculation are sent to the intermediary party, so that the intermediary party can aggregate the intermediate variables of model calculation of each participant to obtain the aggregated variables, and then send each aggregated variable to each participant to obtain the replaced and updated preset model to be trained, so that each participant can continue to train the replaced and updated preset model to be trained.

[0115] Step M3, continuously iteratively train and replace and update the replaced and updated preset model to be trained until the preset model to be trained meets the preset training completion condition, thereby obtaining the target model.

[0116] The first participant continues to iteratively train and replace and update the replaced and updated preset model to be trained until the preset model to be trained meets the preset training completion condition (which may be that the training reaches a second preset number of training times or the preset loss function converges), and obtains the target model.

[0117] After the step of implementing federated modeling with other second participants and obtaining a target model based on the target feature data by executing a preset federated process, the method includes:

[0118] Step S40, evaluating the model stability and accuracy of the target model by using preset evaluation data, and determining whether the target model passes the model evaluation;

[0119] In this embodiment, after obtaining the target model, the model stability and accuracy of the target model are evaluated through preset evaluation data, and it is determined whether the target model passes the model evaluation. Specifically, the model stability of the target model is evaluated through the accuracy evaluation data in the preset evaluation data, and it is determined whether the target model passes the model stability evaluation. The model accuracy of the target model is evaluated through the disturbance data in the preset evaluation data, and it is determined whether the target model passes the model accuracy evaluation. If the target model passes the model accuracy evaluation and the model stability evaluation, it is determined that the target model passes the model evaluation.

[0120] Step S50: If the target model passes the model evaluation, the enterprise data to be processed is input into the target model to obtain a scoring result of the enterprise.

[0121] If the target model passes the model evaluation, the enterprise data to be processed is input into the target model to obtain the enterprise's scoring result (or evaluation result). That is, in this embodiment, it should be noted that the enterprise data to be processed includes third-party data such as credit data. That is, in this embodiment, the target model is a model constructed by internal data and external data (including third-party data). Therefore, the enterprise data to be processed (including third-party data) is input into the target model to accurately obtain the enterprise's scoring result. That is, in this embodiment, third-party data can be part of the model construction item.

[0122] The present application provides a model building method, device and storage medium based on federated learning. Compared with the prior art that adopts a unified evaluation model for each enterprise, which is prone to the problem of scenario inconsistency and inaccurate evaluation in actual application, the present application determines the business scenario and business goal based on the evaluation model building instruction when the evaluation model building instruction is detected; determines the target data based on the business scenario and business goal, and performs feature mining and feature selection on the target data to obtain target feature data; based on the target feature data, implements federated modeling with other second participants and obtains the target model by executing a preset federated process. In the present application, by first determining the business scenario and business goal when the evaluation model building instruction is detected, and then performing feature mining and feature selection for specific business scenarios and business goals, accurately obtains the target feature data that is consistent with the actual business, and then based on the target feature data, executes the preset federal process to accurately implement federated modeling with other second participants on the basis of massive data and obtain the target model, that is, accurately obtain the evaluation model, thereby improving the evaluation accuracy of enterprises with actual business usage scenarios.

[0123] Further, based on the first embodiment of the present application, in another embodiment of the present application, the step of evaluating the model stability and accuracy of the target model by preset evaluation data, and determining whether the target model passes the model evaluation includes:

[0124] Step N1, performing model perturbation on the target model by using the perturbation data in the preset evaluation data, and determining a first model error between a first prediction result obtained after the perturbation and a first preset result;

[0125] Step N2, if the first model error is less than a first preset value, determining that the target model passes the stability assessment;

[0126] In this embodiment, the target model is perturbed by the disturbance data in the preset evaluation data, and the disturbance verification data is processed based on the disturbance model obtained after the disturbance, and a first prediction result obtained after the disturbance model predicts the disturbance verification data is determined, a first preset result of the disturbance verification data is obtained, and a first model error between the first prediction result and the first preset result is determined. If the first model error is less than or equal to a first preset value, it is determined that the target model has passed the stability evaluation. If the first model error is greater than the first preset value, it is determined that the target model has not passed the stability evaluation.

[0127] Step N3, verifying the model accuracy of the target model using the accuracy evaluation data in the preset evaluation data, and determining a second model error between the second prediction result obtained after verification and the second preset result;

[0128] Step N4: if the second model error is less than a second preset value, it is determined that the target model passes the accuracy assessment.

[0129] In this embodiment, the model accuracy of the target model is verified specifically through the accuracy evaluation data in the preset evaluation data, and the second model error between the second prediction result obtained after verification and the second preset result is determined. If the second model error is less than or equal to the second preset value, it is determined that the target model has passed the accuracy evaluation; if the second model error is greater than the second preset value, it is determined that the target model has not passed the accuracy evaluation.

[0130] In this embodiment, after evaluating the stability and accuracy of the target model, data prediction is performed to ensure the long-term effectiveness and accuracy of the model.

[0131] In this embodiment, the target model is perturbed by the perturbation data in the preset evaluation data, and the first model error between the first prediction result obtained after the perturbation and the first preset result is determined; if the first model error is less than the first preset value, it is determined that the target model has passed the stability evaluation; the target model is verified for model accuracy by the accuracy evaluation data in the preset evaluation data, and the second model error between the second prediction result obtained after the verification and the second preset result is determined; if the second model error is less than the second preset value, it is determined that the target model has passed the accuracy evaluation. In this embodiment, before the enterprise evaluation is performed, the accuracy and stability of the target model are also evaluated to ensure the accuracy of the enterprise evaluation.

[0132] Further, based on the first embodiment and the second embodiment of the present application, in another embodiment of the present application, if the target model passes the model evaluation, the step of inputting the enterprise data to be processed into the target model to obtain the scoring result of the enterprise includes:

[0133] Step P1, if the target model passes the model evaluation, the target enterprise's to-be-processed enterprise data is input into the target model to obtain a model score for the target enterprise;

[0134] Step P2, obtaining the third-party rating of the target enterprise, combining the model rating and the third-party rating to obtain the rating result of the enterprise.

[0135] In this embodiment, it should be noted that the model building item may not include third-party data, such as the user's credit score or credit score, etc. At this time, the enterprise's scoring result consists of two parts, one part is the model score, and the other part is the third-party score. Therefore, if the target model passes the model evaluation, the target enterprise's to-be-processed enterprise data is input into the target model to obtain the model score of the target enterprise. After obtaining the model score, the third-party score of the target enterprise is obtained, and the model score and the third-party score are combined to obtain the enterprise's scoring result. The specific method of combination can be: respectively determine the scoring proportions of the model score and the third-party score, and calculate the enterprise's scoring result based on the scoring proportions of the model score and the third-party score.

[0136] In this embodiment, if the target model passes the model evaluation, the target enterprise's to-be-processed enterprise data is input into the target model to obtain the model score of the target enterprise; the third-party score of the target enterprise is obtained, and the enterprise score result is obtained by combining the model score and the third-party score. In this embodiment, the enterprise score result is accurately obtained.

[0137] Further, based on the above first and second embodiments, a fourth embodiment of a model construction method based on federated learning of the present invention is proposed. In this embodiment, the model construction method based on federated learning includes:

[0138] Step A10, determining the business scenario and business goal from the received evaluation model building instruction;

[0139] In this embodiment, the business scenario and the business goal are determined in advance from the received evaluation model building instruction. The specific analysis process can refer to the process of determining the business scenario and the business goal in the first embodiment mentioned above.

[0140] Step A20, based on the business scenario and business goal, determine target data, and perform feature mining and feature selection on the target data to obtain target feature data;

[0141] After obtaining the business scenario and business goal, target data is determined based on the business scenario and business goal, and feature mining and feature selection are performed on the target data to obtain target feature data. The target feature data can be determined according to the method for determining the target feature data in the first embodiment to obtain the target feature data.

[0142] Furthermore, the step A20 includes:

[0143] Step A201, selecting a characteristic factor from a preset characteristic factor library based on the business scenario and business goal;

[0144] First, select feature factors from the preset feature factor library according to the business scenario and business objectives, that is, first select modeling items from the preset feature factor library according to the business scenario and business objectives.

[0145] According to the business scenario and business goal, the process of selecting characteristic factors from the preset characteristic factor library can refer to step S21 in the first embodiment.

[0146] Step A202, selecting target data matching the attributes of the characteristic factor from preset proprietary data and public data;

[0147] The selection of target data matching the attribute of the characteristic factor from the preset own data and the public data may refer to step S22 in the first embodiment.

[0148] Step A203, performing data preprocessing on the target data to obtain preprocessed data;

[0149] The process of obtaining the pre-processed data may refer to step S23 in the above-mentioned first embodiment.

[0150] Step A204, performing self-feature mining and self-feature selection on the preprocessed data, and performing public feature mining and public feature selection on the preprocessed data to obtain target feature data.

[0151] The self-feature mining and self-feature selection are performed on the preprocessed data, and the public feature mining and public feature selection are performed on the preprocessed data to obtain the target feature data, which can refer to step S24 in the first embodiment.

[0152] Step A30, based on the target feature data, perform federation modeling with other second participants to obtain a target model;

[0153] The specific process of obtaining the target model through federation modeling with other second participants based on the target feature data may refer to the first embodiment described above.

[0154] Step A40: Use the target model to score the enterprise data to obtain a scoring result for the enterprise.

[0155] The specific process of using the target model to score the enterprise data to be processed and obtaining the scoring results of the enterprise can refer to the first embodiment and the third embodiment mentioned above.

[0156] Wherein, before the step of using the target subject evaluation model to score the enterprise data to be processed and obtaining the scoring result of the enterprise, the method includes:

[0157] Step A01, evaluating the model stability and accuracy of the target model by using preset evaluation data, and determining whether the target model passes the model evaluation;

[0158] Step A02, if the target model passes the model evaluation, then the step of using the target model to score the enterprise data to be processed is executed to obtain the scoring result of the enterprise.

[0159] Before using the target model to evaluate the enterprise data to be processed, the target model is evaluated for model stability and accuracy using preset evaluation data. The specific process may refer to the second embodiment described above.

[0160] Compared with the use of a unified evaluation model for each enterprise, which is prone to the problems of scenario inconsistency and inaccurate evaluation in actual applications, this embodiment determines the business scenario and business goal from the received evaluation model construction instruction; based on the business scenario and business goal, determines the target data, and performs feature mining and feature selection on the target data to obtain the target feature data; based on the target feature data, performs federation modeling with other second participants to obtain the target model; uses the target model to score the enterprise data to obtain the enterprise's scoring result. In this embodiment, the business scenario and business goal are first determined, and then feature mining and feature selection are performed for specific business scenarios and business goals to accurately obtain target feature data that is consistent with the actual business, and then based on the target feature data, by executing the preset federation process, the federation modeling with other second participants is accurately realized on the basis of massive data and the target model is obtained, that is, the evaluation model is accurately obtained, and then the enterprise data to be processed is scored to obtain the enterprise's scoring result, thereby improving the evaluation accuracy of enterprises with actual business usage scenarios.

[0161] Reference Figure 3 , Figure 3 It is a schematic diagram of the device structure of the hardware operating environment involved in the embodiment of the present application.

[0162] like Figure 3 As shown, the model building device based on federated learning may include: a processor 1001, such as a CPU, a memory 1005, and a communication bus 1002. Among them, the communication bus 1002 is used to realize the connection and communication between the processor 1001 and the memory 1005. The memory 1005 can be a high-speed RAM memory or a stable memory (non-volatile memory), such as a disk memory. The memory 1005 can also be a storage device independent of the aforementioned processor 1001.

[0163] Optionally, the model building device based on federated learning may also include a rectangular user interface, a network interface, a camera, an RF (Radio Frequency) circuit, a sensor, an audio circuit, a WiFi module, etc. The rectangular user interface may include a display screen (Display), an input submodule such as a keyboard (Keyboard), and the optional rectangular user interface may also include a standard wired interface and a wireless interface. The network interface may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface).

[0164] Those skilled in the art will understand that Figure 3 The structure of the model building device based on federated learning shown in the figure does not constitute a limitation of the model building device based on federated 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.

[0165] like Figure 3 As shown, the memory 1005 as a computer storage medium may include an operating system, a network communication module, and a model building method program based on federated learning. The operating system is a program that manages and controls the hardware and software resources of the model building device based on federated learning, and supports the operation of the model building method program based on federated learning and other software and / or programs. The network communication module is used to realize the communication between the components inside the memory 1005, and to communicate with other hardware and software in the model building method system based on federated learning.

[0166] exist Figure 3 In the federated learning-based model building device shown, the processor 1001 is used to execute the federated learning-based model building method program stored in the memory 1005 to implement the steps of the federated learning-based model building method described in any one of the above items.

[0167] The specific implementation of the model building device based on federated learning in the present application is basically the same as the above-mentioned embodiments of the model building method based on federated learning, and will not be repeated here.

[0168] The embodiment of the present application further provides a model construction device based on federated learning, which is applied to a first participant, and the model construction device based on federated learning includes:

[0169] A detection module, configured to determine a business scenario and a business goal based on the evaluation model building instruction when an evaluation model building instruction is detected;

[0170] A first determination module is used to determine target data based on the business scenario and business goal, and perform feature mining and feature selection on the target data to obtain target feature data;

[0171] The federation module is used to implement federation modeling with other second parties and obtain a target model by executing a preset federation process based on the target feature data.

[0172] Optionally, the first determining module includes:

[0173] A first determining unit, configured to select a characteristic factor from a preset characteristic factor library based on the business scenario and the business goal;

[0174] A first selection unit, configured to select target data matching the attributes of the characteristic factor from preset own data and public data;

[0175] A first acquisition unit, configured to perform data preprocessing on the target data to obtain preprocessed data;

[0176] The second selection unit is used to perform self-feature mining and self-feature selection on the preprocessed data, and perform public feature mining and public feature selection on the preprocessed data to obtain target feature data.

[0177] Optionally, the device further comprises:

[0178] An evaluation module, used to evaluate the model stability and accuracy of the target model by using preset evaluation data, and determine whether the target model passes the model evaluation;

[0179] The input module is used to input the enterprise data to be processed into the target model if the target model passes the model evaluation to obtain the enterprise scoring result.

[0180] Optionally, the evaluation module includes:

[0181] An input unit, configured to input the to-be-processed enterprise data of the target enterprise into the target model if the target model passes the model evaluation, so as to obtain a model score of the target enterprise;

[0182] The second acquisition unit is used to obtain a third-party score of the target enterprise, and obtain a score result of the enterprise by combining the model score and the third-party score.

[0183] Optionally, the evaluation module includes:

[0184] A second determination unit is used to perform model perturbation on the target model by using the perturbation data in the preset evaluation data, and determine a first model error between a first prediction result obtained after the perturbation and a first preset result;

[0185] A third determining unit, configured to determine that the target model passes the stability assessment if the first model error is less than a first preset value;

[0186] A fourth determination unit, configured to verify the model accuracy of the target model using the accuracy evaluation data in the preset evaluation data, and determine a second model error between a second prediction result obtained after verification and a second preset result;

[0187] A fifth determining unit is used to determine that the target model passes the accuracy assessment if the second model error is less than a second preset value.

[0188] Optionally, the federation module includes:

[0189] A third acquisition unit is used to acquire a preset model to be trained and acquire a preset training completion condition of the preset model to be trained;

[0190] An execution unit is used to optimize and update the model calculation intermediate variables of the preset model to be trained based on the target feature data in conjunction with other second participants until the preset model to be trained meets the preset training completion conditions to obtain the target model.

[0191] Optionally, the execution unit includes:

[0192] A first training subunit is used to input the target feature data into a preset model to be trained, to perform prediction processing on the target feature data based on the preset model to be trained, to obtain a third prediction result, to determine a third model error between the third prediction result and a third preset result of the target feature data, and to update the preset model to be trained based on the third model error, so as to iteratively update the model calculation intermediate variables of the preset model to be trained;

[0193] A judgment subunit is used to judge whether the preset model to be trained in iterative training meets the preset replacement and update conditions. If the preset model to be trained meets the preset replacement and update conditions, the preset vertical federated learning process is executed to replace and update the intermediate variables of the model calculation updated in training, so as to obtain the replaced and updated preset model to be trained;

[0194] The second training subunit is used to continuously iteratively train and replace and update the replaced and updated preset model to be trained until the preset model to be trained meets the preset training completion condition, thereby obtaining the target model.

[0195] The specific implementation of the model building device based on federated learning in the present application is basically the same as the various embodiments of the model building method based on federated learning mentioned above, and will not be repeated here.

[0196] An embodiment of the present application provides a storage medium, and the storage medium stores one or more programs, and the one or more programs can also be executed by one or more processors to implement the steps of any of the above-mentioned federated learning-based model building methods.

[0197] The specific implementation method of the storage medium of the present application is basically the same as the above-mentioned embodiments of the model building method based on federated learning, and will not be repeated here.

[0198] The present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the above-mentioned model building method based on federated learning.

[0199] The specific implementation of the computer program product of the present application is basically the same as the various embodiments of the above-mentioned model building method based on federated learning, and will not be repeated here.

[0200] The present application also provides a target subject evaluation device based on federated learning, which is applied to a first participant. The target subject evaluation device based on federated learning includes:

[0201] A receiving module, used to determine a business scenario and a business goal from the received evaluation model building instruction;

[0202] A second determination module is used to determine target data based on the business scenario and business goal, and perform feature mining and feature selection on the target data to obtain target feature data;

[0203] A first acquisition module is used for obtaining a target model through federation modeling with other second participants based on the target feature data;

[0204] The second acquisition module is used to use the target model to perform scoring processing on the enterprise data to be processed to obtain the scoring results of the enterprise.

[0205] The specific implementation of the target subject evaluation device based on federated learning in the present application is basically the same as the various embodiments of the target subject evaluation method based on federated learning mentioned above, and will not be repeated here.

[0206] An embodiment of the present application provides a storage medium, and the storage medium stores one or more programs, and the one or more programs can also be executed by one or more processors to implement the steps of any of the above-mentioned target subject evaluation methods based on federated learning.

[0207] The specific implementation method of the storage medium of the present application is basically the same as the above-mentioned embodiments of the target subject evaluation method based on federated learning, and will not be repeated here.

[0208] The present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the above-mentioned target subject evaluation method based on federated learning.

[0209] The specific implementation methods of the computer program product of the present application are basically the same as the above-mentioned embodiments of the target subject evaluation method based on federated learning, and will not be repeated here.

[0210] The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent processing scope of the present application.

Claims

1. A model construction method based on federated learning, characterized in that: Applied to the first participant, the model building method based on federated learning includes: When an evaluation model building instruction is detected, determining a business scenario and a business goal based on the evaluation model building instruction; Based on the business scenario and business goal, target data is determined, and feature mining and feature selection are performed on the target data to obtain target feature data; Based on the target characteristic data, a preset federation process is executed to realize federation modeling with other second participants and obtain a target model; The step of determining target data based on the business scenario and business goal, and performing feature mining and feature selection on the target data to obtain target feature data includes: Based on the business scenario and business goal, select characteristic factors from a preset characteristic factor library; Selecting target data matching the attributes of the characteristic factors from preset own data and public data; Performing data preprocessing on the target data to obtain preprocessed data; The preprocessed data is subjected to self-feature mining and self-feature selection, and the preprocessed data is subjected to public feature mining and public feature selection to obtain target feature data.

2. The model building method based on federated learning as claimed in claim 1, characterized in that: After the step of implementing federated modeling with other second participants and obtaining a target model based on the target feature data by executing a preset federated process, the method includes: Evaluate the model stability and accuracy of the target model using preset evaluation data, and determine whether the target model passes the model evaluation; If the target model passes the model evaluation, the enterprise data to be processed is input into the target model to obtain the enterprise scoring result.

3. The model building method based on federated learning as claimed in claim 2, characterized in that: If the target model passes the model evaluation, the step of inputting the enterprise data to be processed into the target model to obtain the enterprise scoring result includes: If the target model passes the model evaluation, the target enterprise's to-be-processed enterprise data is input into the target model to obtain a model score for the target enterprise; Obtain a third-party rating of the target enterprise, and combine the model rating and the third-party rating to obtain a rating result for the enterprise.

4. The model building method based on federated learning as claimed in claim 2, characterized in that: The step of evaluating the stability and accuracy of the target model by preset evaluation data and determining whether the target model passes the model evaluation comprises: Performing model perturbation on the target model by using the perturbation data in the preset evaluation data, and determining a first model error between a first prediction result obtained after the perturbation and a first preset result; If the first model error is less than a first preset value, determining that the target model passes the stability assessment; Verifying the model accuracy of the target model using the accuracy evaluation data in the preset evaluation data, and determining a second model error between a second prediction result obtained after verification and a second preset result; If the second model error is less than a second preset value, it is determined that the target model passes the accuracy assessment.

5. The model building method based on federated learning as claimed in claim 1, characterized in that: The step of implementing federated modeling with other second participants and obtaining a target model by executing a preset federated process based on the target feature data includes: Obtaining a preset model to be trained, and obtaining a preset training completion condition of the preset model to be trained; Based on the target feature data, in conjunction with other second participants, the model calculation intermediate variables of the preset model to be trained are optimized and updated until the preset model to be trained meets the preset training completion conditions, thereby obtaining the target model.

6. The model building method based on federated learning as claimed in claim 5, characterized in that: The step of optimizing and updating the model calculation intermediate variables of the preset model to be trained in conjunction with other second participants based on the target feature data until the preset model to be trained meets the preset training completion condition to obtain the target model includes: Inputting the target feature data into a preset model to be trained, so as to perform prediction processing on the target feature data based on the preset model to be trained to obtain a third prediction result, determining a third model error between the third prediction result and a third preset result of the target feature data, and updating the preset model to be trained based on the third model error to iteratively update the model calculation intermediate variables of the preset model to be trained; Determine whether the preset model to be trained by iterative training meets the preset replacement and update conditions. If the preset model to be trained meets the preset replacement and update conditions, replace and update the intermediate variables of the model calculation by executing the preset longitudinal federated learning process to obtain the replaced and updated preset model to be trained; Continue to iteratively train and replace and update the replaced and updated preset model to be trained until the preset model to be trained meets the preset training completion condition, thereby obtaining the target model.

7. A target subject evaluation method based on federated learning, characterized in that: Applied to the first participant, the target subject evaluation method based on federated learning includes: Determine business scenarios and business goals from received evaluation model building instructions; Based on the business scenario and business goal, target data is determined, and feature mining and feature selection are performed on the target data to obtain target feature data; Based on the target feature data, obtain a target model by conducting federated modeling with other second participants; The target model is used to score the enterprise data to be processed to obtain a scoring result of the enterprise; Before the step of using the target subject evaluation model to score the enterprise data to be processed and obtaining the scoring result of the enterprise, the method includes: Evaluate the model stability and accuracy of the target model using preset evaluation data, and determine whether the target model passes the model evaluation; If the target model passes the model evaluation, the step of using the target model to score the enterprise data to be processed is performed to obtain a scoring result for the enterprise.

8. The target subject evaluation method based on federated learning according to claim 7, characterized in that: The step of determining target data based on the business scenario and business goal, and performing feature mining and feature selection on the target data to obtain target feature data includes: Based on the business scenario and business goal, select characteristic factors from a preset characteristic factor library; Selecting target data matching the attributes of the characteristic factors from preset own data and public data; Performing data preprocessing on the target data to obtain preprocessed data; The preprocessed data is subjected to self-feature mining and self-feature selection, and the preprocessed data is subjected to public feature mining and public feature selection to obtain target feature data.

9. A model building device based on federated learning, characterized in that: The model building device based on federated learning includes: a memory, a processor, and a program stored in the memory for implementing the model building method based on federated learning, The memory is used to store a program for implementing a model building method based on federated learning; The processor is used to execute a program for implementing the model building method based on federated learning, so as to implement the steps of the model building method based on federated learning as described in any one of claims 1 to 6, or to implement the steps of the target subject evaluation method based on federated learning as described in any one of claims 7 to 8.

10. A readable storage medium, characterized in that: The readable storage medium stores a program for implementing a model building method based on federated learning, and the program for implementing a model building method based on federated learning is executed by a processor to implement the steps of the model building method based on federated learning as described in any one of claims 1 to 6, or to implement the steps of the target subject evaluation method based on federated learning as described in any one of claims 7 to 8.

11. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, it implements the steps of the method for building a model based on federated learning as described in any one of claims 1 to 6, or implements the steps of the target subject evaluation method based on federated learning as described in any one of claims 7 to 8.

Images