Business model training method and device, equipment and storage medium

Abstract

The application discloses a business model training method and device, equipment and a storage medium, and belongs to the field of data processing. The method comprises the following steps: extracting a data domain from original business data to obtain a data domain correlation parameter corresponding to the original business data; calling a generator and a discriminator to perform data processing iteration until the data processing iteration meets an iteration termination condition; training a to-be-trained model obtained from a central server by using target synthetic data; uploading the trained model to the central server to enable the central server to generate a comprehensive model; in the case that the comprehensive model does not meet an expected standard condition, generating synthetic data by calling the generator and the discriminator according to the comprehensive model issued by the central server, training the comprehensive model by using the synthetic data, and uploading the comprehensive model to the central server until the comprehensive model generated by the central server meets the expected standard condition, and determining the comprehensive model as a business model. According to the embodiment of the application, the data security can be improved.

Description

Technical Field

[0001] This application belongs to the field of data processing, and in particular relates to a business model training method, apparatus, device and storage medium. Background Technology

[0002] With the continuous development of digital technology, data has become one of the most important factors of production. More and more business analyses rely on large amounts of data as the foundation for modeling or other operations. Data is often distributed and held by multiple parties. Due to data confidentiality, data is not easily shared between these parties, resulting in business models trained by each party using their own stored data being neither comprehensive nor accurate.

[0003] To improve the comprehensiveness and accuracy of business models, homomorphic encryption can be used to process the original data, and the encrypted data can be used to train the model. However, homomorphic encryption involves multiple addition and multiplication operations on the original data, and the encrypted data still has a strong correlation with the original data. It is possible to reverse engineer the original data from the encrypted data, thereby compromising data confidentiality and reducing data security. Summary of the Invention

[0004] This application provides a business model training method, apparatus, device, and storage medium that can ensure data security during the business model training process.

[0005] In a first aspect, embodiments of this application provide a business model training method, comprising: acquiring original business data; extracting data domains from the original business data; obtaining data domain correlation parameters corresponding to the original business data based on the original business data, wherein the data domain correlation parameters are used to characterize the correlation between data domains; calling a generator and a discriminator to perform data processing iterations until the data processing iterations meet the iteration cutoff conditions; training a model to be trained obtained from a central server using target synthetic data to obtain a trained model; wherein the generator is used to generate synthetic data based on the data domain correlation parameters, the discriminator is used to generate data domain correlation parameters based on the input data, and the target synthetic data is obtained by the generator based on the data domain correlation parameters of the most recent data processing iteration; uploading the trained model to the central server so that the central server can generate a comprehensive model based on the trained model from multiple feedback sources; if the comprehensive model does not meet the expected standard conditions, calling the generator and discriminator to collaboratively generate synthetic data based on the comprehensive model issued by the central server as the model to be trained; using the synthetic data to train the model to be trained; and uploading the trained model to the central server until the comprehensive model generated by the central server based on the trained model from multiple feedback sources meets the expected standard conditions, and determining the comprehensive model as the business model.

[0006] Secondly, embodiments of this application provide a business model training method, comprising: a preprocessing module, used to acquire raw business data, extract data domains from the raw business data, and obtain data domain correlation parameters corresponding to the raw business data, wherein the data domain correlation parameters are used to characterize the correlation relationships between data domains; and a training module, used to call a generator and a discriminator to perform data processing iterations until the data processing iterations meet the iteration cutoff condition, and to train a model to be trained obtained from a central server using target synthetic data to obtain a trained model, wherein the generator is used to generate synthetic data based on the data domain correlation parameters, and the discriminator is used to generate data domain correlation parameters based on the input data. The target synthetic data is obtained by the generator based on the data domain correlation parameters of the most recent data processing iteration; the upload module is used to upload the trained model to the central server so that the central server can generate a comprehensive model based on the trained model from multiple feedbacks; the training module is also used to, when the comprehensive model does not meet the expected standard conditions, call the generator and discriminator to collaboratively generate synthetic data based on the comprehensive model issued by the central server as the model to be trained, use the synthetic data to train the model to be trained, and call the upload module to upload the trained model to the central server until the comprehensive model generated by the central server based on the trained model from multiple feedbacks meets the expected standard conditions, and then determine the comprehensive model as the business model.

[0007] Thirdly, embodiments of this application provide a business model training device, including: a processor and a memory storing computer program instructions; the processor executes the computer program instructions to implement the business model training method of the first aspect.

[0008] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the business model training method of the first aspect.

[0009] This application provides a business model training method, apparatus, device, and storage medium. It extracts data domains from raw business data, and based on the relationships between these data domains, it iterates data processing using a generator and a discriminator. This enables the generator to output synthetic data that closely resembles the original business data in appearance, meaning, and statistical characteristics, but from which the original business data cannot be traced. This synthetic data is used to train a model provided by a central server. The trained model is then fed back to the central server. The central server, based on the trained models provided by multiple model training participants, obtains a comprehensive model. If the comprehensive model does not meet the desired standard conditions, it is distributed as a model to be trained to each model training participant for further training. Each model training participant continues to feed back its trained model to the central server, enabling the central server to obtain a new comprehensive model. This process continues until the comprehensive model meets the desired standard conditions. The model that meets the desired standard conditions is then determined as the business model. During model training, the synthetic data used to train the model is not real data, which can protect data privacy and security. Furthermore, the model to be trained, distributed by the central server, is trained locally by each model training participant using the synthetic data generated locally by the model training participant. The synthetic data of different model training participants will not be exchanged, which can further improve data security. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A flowchart illustrating a business model training method provided in one embodiment of this application;

[0012] Figure 2 A flowchart of a business model training method provided in another embodiment of this application;

[0013] Figure 3 A flowchart illustrating a business model training method provided in yet another embodiment of this application;

[0014] Figure 4 A flowchart illustrating an example of a business model training process provided in an embodiment of this application;

[0015] Figure 5 This is a schematic diagram of the structure of a business model training device provided in an embodiment of this application;

[0016] Figure 6 A schematic diagram of the structure of a business model training device provided in another embodiment of this application;

[0017] Figure 7 This is a schematic diagram of the structure of a business model training device provided in an embodiment of this application. Detailed Implementation

[0018] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0019] With the continuous development of digital technology, data has become a crucial production factor. Increasingly, business analysis relies on large amounts of data as the foundation for modeling and other operations. Data is often distributed and held by multiple parties. Due to data confidentiality, data is not easily shared between these parties, resulting in business models trained by each party using their own data being incomplete and inaccurate. To improve the comprehensiveness and accuracy of business models, homomorphic encryption can be used to process the original data and then use the encrypted data to train the model. However, homomorphic encryption involves multiple addition and multiplication operations on the original data. The encrypted data still retains a strong correlation with the original data, potentially allowing the original data to be deduced from the encrypted data, thus compromising data confidentiality and reducing data security.

[0020] This application provides a business model training method, apparatus, device, and storage medium. Each model training participant can extract data domains from the original business data. Based on the relationships between data domains in the original business data, a generator and a discriminator are used for iterative data processing. This enables the generator to output synthetic data that is similar to the original business data in appearance, meaning, and statistical characteristics, but cannot be traced back to the original business data. This synthetic data is used to train the model to be trained provided by the central server, and the trained model is then fed back to the central server. The central server can interact with multiple model training participants once or multiple times, using the synthetic data provided by the participants to update the comprehensive model until it reaches the desired standard conditions. The synthetic data is not real data, thus protecting data privacy and security. Furthermore, the model to be trained distributed by the central server is trained locally by each participant using the synthetic data generated locally. The synthetic data from different participants is not shared, further ensuring data security.

[0021] The first aspect of this application provides a business model training method, which can be applied to scenarios involving training business models. This business model training method can be executed by devices or equipment of the model training participants, and the type and number of such devices or equipment are not limited herein. For ease of explanation, "model training participant" in the following text refers to the device or equipment of the model training participant. Figure 1 A flowchart illustrating a business model training method provided in one embodiment of this application. Figure 1 As shown, the business model training method may include steps S101 to S104.

[0022] In step S101, the original business data is obtained, the data domain is extracted from the original business data, and the data domain correlation parameters corresponding to the original business data are obtained based on the original business data.

[0023] Raw business data is authentic data and can be considered source data. For example, raw business data may include user age data, transaction amount data, transaction location data, transaction type data, and transaction description data, etc., and is not limited to these. The leakage of raw business data may pose a risk to data security.

[0024] Data domains can be extracted from raw business data. A data domain can be considered as a category into which data items are divided. Data items can be represented as fields or other data types. For example, the raw business data might be <26 years old, 100 yuan, 0, XX city XX district XX avenue credit card spending>. Here, "26 years old," "100 yuan," "0," and "XX city XX district XX avenue credit card spending" are the data items corresponding to user age, transaction amount, whether a coupon was used, and transaction description, respectively. User age, transaction amount, whether a coupon was used, and transaction description are the data domains. In some examples, data items of consecutive character types can be further subdivided; that is, the data domain corresponding to this data item can be further subdivided into two or more data domains. For example, the data item corresponding to the transaction description is "XX city XX district XX avenue credit card spending," which can be further subdivided into four data items: "XX city," "XX district," "XX avenue," and "credit card spending." The data domains corresponding to "XX city," "XX district," "XX avenue," and "credit card spending" can be business region, business region, business region, and business type, respectively. That is, the data domain "business description" can be subdivided into two business domains: business region and business type.

[0025] Based on the data items corresponding to the data domains in the original business data, data domain correlation parameters can be obtained. These parameters characterize the relationships between data domains. These relationships can include information describing the relationship, such as relationship type and degree of correlation. Data domain correlation parameters can be implemented as characters, strings, tables, relational views, etc., and are not limited here.

[0026] In step S102, the generator and discriminator are invoked to perform data processing iterations until the data processing iterations meet the iteration cutoff condition. The target synthetic data is used to train the model to be trained obtained from the central server to obtain the trained model.

[0027] The generator produces synthetic data based on data domain association parameters. The discriminator generates data domain association parameters based on the input data. A single data processing iteration may include: the synthetic data generated by the generator is input into the discriminator, which then generates new data domain association parameters based on the input synthetic data and feeds them back to the generator. If the data processing iteration does not meet the iteration cutoff condition, the iteration can be repeated until the cutoff condition is met. If the data processing iteration meets the cutoff condition, the generator can be invoked to generate synthetic data, i.e., the target synthetic data, based on the data domain association parameters obtained from the most recent data processing iteration. In other words, the target synthetic data is obtained by the generator based on the data domain association parameters from the most recent data processing iteration.

[0028] The iteration cutoff condition is the criterion for determining whether the data domain correlation result parameters generated by the discriminator achieve the desired effect. This condition can be set according to the scenario, requirements, experience, etc., and is not limited here. Meeting the iteration cutoff condition in a data processing iteration indicates that the data domain correlation parameters generated by the discriminator in this iteration can achieve the desired effect; that is, the synthetic data generated using these data domain correlation parameters, which cannot trace the original business data, can meet the needs of training the model.

[0029] In some examples, the iteration cutoff condition may include: the number of data processing iterations has reached a preset number of iterations. The preset number of iterations can be set according to the scenario, requirements, experience, etc., and is not limited here. For example, the preset number of iterations is 10.

[0030] The model to be trained can be obtained from the central server. When a participant trains the model to be trained for the first time, the model to be trained issued by the central server is the initial model. When a participant trains the model to be trained for the i-th time, the model to be trained issued by the central server is a comprehensive model generated based on the trained models fed back by multiple participants for the (i-1)-th time, where i is an integer greater than 1.

[0031] In step S103, the trained model is uploaded to the central server so that the central server can generate a comprehensive model based on the trained model fed back from multiple parties.

[0032] The central server combines the trained models with feedback from multiple sources to generate a comprehensive model. The model parameters of the comprehensive model can be obtained from the model parameters of the trained models with feedback from multiple sources. During the generation of the comprehensive model, the central server can also adjust the model parameters of the trained models as needed to obtain the model parameters of the comprehensive model, thus generating the comprehensive model.

[0033] The central server can use federated learning to generate a comprehensive model based on the trained model from multiple sources, without limiting the specific method of federated learning.

[0034] In step S104, if the integrated model does not meet the expected standard conditions, the generator and discriminator are called to collaboratively generate synthetic data based on the integrated model issued by the central server as the model to be trained. The synthetic data is used to train the model to be trained, and the trained model is uploaded to the central server until the integrated model generated by the central server based on the feedback from multiple parties meets the expected standard conditions, and the integrated model is determined as the business model.

[0035] The expected standard conditions include the criteria for determining whether the integrated model meets the business model standards. These conditions can be set based on scenarios, requirements, experience, etc., and are not limited here. For example, the expected standard conditions may include that the accuracy of the integrated model's data processing is greater than or equal to a preset accuracy threshold. If the integrated model meets the expected standard conditions, it means that the integrated model meets the business model standards and can be used as a business model. If the integrated model does not meet the expected standard conditions, it means that the integrated model does not meet the business model standards and requires further training. The model training participants interact with the central server, using synthetic data generated by the generator locally to train the integrated model issued by the central server. The trained model is then transmitted to the central server, which re-integrates the trained models provided by multiple model training participants to update the integrated model, thereby improving its quality until the integrated model meets the expected standard conditions.

[0036] In this embodiment, data domains are extracted from the original business data. Based on the relationships between data domains in the original business data, a generator and a discriminator are used for iterative data processing. This enables the generator to output synthetic data that is similar to the original business data in appearance, meaning, and statistical characteristics, but cannot be traced back to the original business data. This synthetic data is used to train the model to be trained provided by the central server. The trained model is then fed back to the central server. The central server obtains a comprehensive model based on the trained models provided by multiple model training participants. If the comprehensive model does not meet the expected standard conditions, it is distributed as a model to be trained to each model training participant for further training. Each model training participant continues to feed back its trained model to the central server so that the central server obtains a new comprehensive model, until the comprehensive model meets the expected standard conditions. The model that meets the expected standard conditions is then determined as the business model. During model training, the synthetic data used to train the model is not real data, thus protecting data privacy and security. Furthermore, the model to be trained distributed by the central server is trained locally by each model training participant using the synthetic data generated locally by the participant. The synthetic data from different model training participants does not circulate among themselves, further enhancing data security.

[0037] The synthetic data used by each model training participant is different. By combining the trained models obtained from the synthetic data from all participants, a comprehensive model is obtained. This avoids the bias and incompleteness that may be caused by data provided by a single model training participant, and improves the effectiveness and generalization ability of the final business model. Moreover, in this embodiment, no data encryption is performed. Compared with homomorphic encryption methods that require multiple addition and multiplication operations, this embodiment can save a significant amount of storage and computing resources.

[0038] In some embodiments, the model training participants can update the discriminator and generator based on the model parameters of the comprehensive model issued by the central server as the model to be trained, so that the generator outputs synthetic data that is more suitable for training the model to be trained, and use the synthetic data to train the model to be trained. The trained model is then uploaded to the central server again until the comprehensive model generated by the central server based on the feedback from multiple parties reaches the expected standard conditions. Figure 2 A flowchart illustrating a business model training method provided in another embodiment of this application. Figure 2 and Figure 1 The difference is that, Figure 1 Step S104 can be further refined as follows: Figure 2 Steps S1041 to S1045 in the process.

[0039] In step S1041, if the integrated model does not meet the expected standard conditions, the model parameters of the integrated model issued by the central server as the model to be trained are obtained.

[0040] If the integrated model fails to meet the expected standard conditions, further training is required. The central server distributes the integrated model as a trainable model to the model training participants. These participants can analyze the integrated model to obtain its parameters. In some examples, the model parameters reflect the relationships between the data domains required by the trainable model. In other examples, the model parameters are analyzed to obtain results, which can then be used to determine the relationships between the data domains required by the trainable model.

[0041] In step S1042, the discriminator is updated according to the model parameters, so that the discriminator generates the target data domain correlation parameters.

[0042] Based on the model parameters, the relationships between data domains required for the model to achieve the desired standard conditions can be determined. These relationships can then be used to adjust the discriminator's model parameters and optimize its internal processing logic, thus updating the discriminator. The updated discriminator can then generate target data domain correlation parameters. The correlations between data domains represented by these target data domain correlation parameters are consistent with or closely resemble the correlations between data domains required by the model to be trained.

[0043] In step S1043, the generator is invoked to generate synthetic data based on the target data domain correlation parameters.

[0044] The generator can generate corresponding synthetic data based on the correlation parameters of the target data domain.

[0045] In step S1044, the model to be trained is trained using synthetic data to obtain the trained model, and then step S103 is executed.

[0046] The newly generated synthetic data is used to train the model to be trained, further optimizing the model. The optimized model is the trained model. After obtaining the trained model, it is uploaded to the central server again. The central server combines the trained models fed back by multiple model training participants to generate a comprehensive model and determines whether the comprehensive model meets the expected standard conditions. If the comprehensive model does not meet the expected standard conditions, step S1041 is executed. If the comprehensive model meets the expected standard conditions, step S1045 is executed.

[0047] In step S1045, if the comprehensive model generated by the central server based on feedback from multiple parties after training meets the expected standard conditions, the comprehensive model is determined as the business model.

[0048] Once the comprehensive model meets the expected standard conditions, it can be identified as a usable business model.

[0049] Through multiple interactions between the central server and various model training participants, the comprehensive model undergoes multiple optimizations, enabling it to reach the desired standard conditions as quickly as possible. During each optimization process, each participant trains the previously generated comprehensive model locally using newly generated synthetic data. The locally generated synthetic data is not shared among participants, but the features of the trained models obtained by each participant are aggregated at the central server for comprehensive optimization, further improving the model's efficiency and generalization ability. Furthermore, if the comprehensive model fails to meet the desired standard conditions, the discriminator is updated, ensuring its output more closely matches the target data domain correlation parameters required by the business model. This allows the generator to produce synthetic data more suitable for training the model, compensating for potential biases and incompleteness in the original business data.

[0050] In some embodiments, the generator and discriminator may update and iterate with each other to obtain synthetic data that is more suitable for training the model to be trained. Figure 3 A flowchart illustrating a business model training method provided in yet another embodiment of this application. Figure 3 and Figure 1 The difference is that, Figure 1 Step S102 can be further refined as follows: Figure 3 Steps S1021 to S1027 in the process.

[0051] In step S1021, the generator is invoked to generate synthetic data based on the data domain correlation parameters corresponding to the original business data, and the synthetic data is mixed with the acquired business data and input into the discriminator.

[0052] The first data processing iteration includes steps S1021 and S1022. Here, the data domain correlation parameters corresponding to the original business data are used as the feedback parameters required by the generator, so that the generator can output synthesized data based on the data domain correlation parameters corresponding to the original business data.

[0053] To further improve the adaptability of the generated synthetic data, some real data, i.e., business data, can be acquired and mixed with the generated synthetic data before being input into the discriminator. The acquired business data can be the original business data from step S101, or newly acquired business data different from the original business data; there is no limitation here. By expanding the data input to the discriminator, the data domain correlation parameters generated by the discriminator have more data support.

[0054] In step S1022, the discriminator is invoked to generate corresponding data domain correlation parameters based on the input data.

[0055] The callable discriminator splits the input data into data items according to the data domain. The discriminator randomly swaps these data items, resulting in multiple combined data sets, at least some of which differ from the input data. Based on the data domain correlation parameters corresponding to the original business data, the discriminator scores the combined data sets, obtaining a score result. If the score result does not meet the scoring cutoff condition, the discriminator randomly swaps the data items again, adding more combined data sets, and scores them, repeating this process until the score result meets the scoring cutoff condition. If the score result meets the scoring cutoff condition, the discriminator generates the corresponding data domain correlation result parameters based on the combined data. The score result obtained from scoring the combined data characterizes the degree of closeness between the combined data and the actual data. In some cases, the score result of some combined data sets may even be higher than the score result of the business data, indicating that the combined data is closer to the actual data than the original business data is to the actual data.

[0056] In some examples, after the first data processing iteration, it can be determined whether the data processing iteration meets the iteration cutoff condition; if the iteration cutoff condition is met, steps S1023 to S1026 can be skipped and step S1027 can be executed directly; if the iteration medium condition is not met, steps S1023 to S1027 can continue to be executed.

[0057] In step S1023, the data domain correlation parameters are transmitted to the generator.

[0058] In step S1024, the generator is invoked to generate synthetic data based on the transmitted data domain correlation parameters, and the synthetic data is mixed with the previously generated synthetic data and input into the discriminator.

[0059] The data input to the discriminator here includes the synthesized data generated this time and the synthesized data generated last time.

[0060] In step S1025, the discriminator is invoked to generate corresponding data domain correlation parameters based on the input data.

[0061] By mixing the synthetic data generated in two separate iterations, the amount of data input to the discriminator is increased, and the influence of the synthetic data from the previous data processing iteration is introduced. This allows the data domain correlation parameters generated by the discriminator to have more data support, and the current data processing iteration can be obtained from the previous data processing iteration. This makes the correlation between data domains represented by the data domain correlation parameters generated by the discriminator closer to the correlation between data domains in the real data.

[0062] The discriminator can be invoked to split the input data into data items according to the data domain; the discriminator can then randomly swap the data items to obtain multiple combined data; the discriminator can then score the combined data based on the data domain correlation parameters corresponding to the original business data to obtain a score result; if the score result does not meet the scoring cutoff condition, the discriminator can again randomly swap the data items, add more combined data, and score the combined data until the score result meets the scoring cutoff condition. Finally, the discriminator can then generate the corresponding data domain correlation result parameters based on the combined data. The content of the corresponding data domain correlation parameters generated by the discriminator in step S1025 is basically the same as that generated by the discriminator in step S1022, and will not be repeated here.

[0063] In step S1026, it is determined whether the data processing iteration meets the iteration cutoff condition. If the iteration cutoff condition is met, step S1027 is executed; if the iteration cutoff condition is not met, step S1023 is executed.

[0064] In step S1027, the generator is invoked to generate target synthetic data based on the transmitted data domain correlation parameters. The target synthetic data is then used to train the model to be trained obtained from the central server to obtain the trained model.

[0065] The target synthetic data is the synthetic data generated under the condition that the data processing iterations meet the iteration deadline. Using the target synthetic data to train the model can improve the optimization level of the trained model relative to the original model.

[0066] The scoring cutoff conditions in the above embodiments may include, but are not limited to: the proportion of combined data with a scoring result greater than or equal to the target score is greater than or equal to a preset proportion threshold; and / or, the number of combined data reaches the target number.

[0067] The target score is the threshold for the scoring result that meets the requirements of the synthesized data. It can be set according to the scenario, requirements, experience, etc., and is not limited here. The preset proportion threshold is the proportion threshold that meets the requirements of the synthesized data. It can be set according to the scenario, requirements, experience, etc., and is not limited here. Combined data with a score greater than or equal to the target score is data that meets the requirements in terms of similarity to real business data in terms of appearance, meaning, statistical characteristics, etc. If the proportion of combined data with a score greater than or equal to the target score is higher than or equal to the preset proportion threshold, it indicates that the discriminator has generated a better data domain correlation result parameter that can be generated in this data processing iteration.

[0068] The target number is a threshold quantity to meet the requirements of the synthesized data. In some cases, the target number may be the maximum number of combined data that can be generated. When the number of combined data reaches the target number, it indicates that the discriminator has generated the optimal data domain correlation result parameters that can be produced in this data processing iteration.

[0069] In some embodiments, when the data processing iteration meets the iteration deadline, the target synthetic data can also be quality evaluated. Only target synthetic data that passes the quality evaluation can be used to train the model to be trained, thereby further improving the adaptability of the synthetic data used for training to the requirements. Specifically, when the data processing iteration meets the iteration deadline, the target synthetic data is analyzed to obtain at least one indicator parameter of the target synthetic data; based on the indicator parameter, the quality score of the target synthetic data is obtained; if the quality score is less than a preset standard score, the generator and discriminator are called again for data processing iteration to obtain newly generated target synthetic data, until the quality score is greater than or equal to the preset standard score; if the quality score is greater than or equal to the preset standard score, the model to be trained obtained from the central server is trained using the target synthetic data to obtain the trained model. Corresponding weights can be set for the indicator parameters, and the quality score is obtained using a weight algorithm based on the indicator parameters and their corresponding weights.

[0070] In some examples, the metrics parameters may include one or more of the following: statistical metrics, likelihood metrics, detection metrics, learning effectiveness metrics, and privacy metrics. Statistical metrics characterize the similarity of statistical features between synthetic data and business data. Statistical metrics are obtained by comparing the differences in statistical features between synthetic data and real business data. Likelihood metrics characterize the likelihood values ​​between synthetic data and business data. Likelihood metrics may include likelihood values ​​and / or log-likelihood values. Likelihood metrics are obtained by fitting a probability model of real business data, calculating the likelihood value of the synthetic data, and then using the likelihood value. Detection metrics characterize the probability that the synthetic data will be classified in the same class as the business data through regression classification. Detection metrics are obtained by mixing synthetic data with real business data, using a regression classification model to obtain the classification result, and then using the classification result. Learning effectiveness metrics characterize the effectiveness of synthetic data as the basis for business analysis. Learning effectiveness metrics are used to evaluate whether synthetic data can be used to replace real data for business analysis. Privacy metrics characterize the probability that the real data cannot be traced back to the synthetic data. Privacy metrics data are used to evaluate whether synthetic data could be reverse-engineered to identify relevant information from real data, and to assess the possibility of leaking real data information.

[0071] To facilitate understanding, the following example illustrates the overall process of the business model training method. Figure 4 A flowchart illustrating an example of a business model training process provided in an embodiment of this application. For example... Figure 4 As shown, the business model training process may include steps S201 to S215.

[0072] In step S201, the original business data is preprocessed to obtain the data domain correlation parameters corresponding to the original business data.

[0073] In step S202, the generator is invoked to generate synthetic data based on the data domain correlation parameters.

[0074] In step S203, it is determined whether the data processing iteration meets the iteration cutoff condition. If the iteration cutoff condition is not met, step S204 is executed; if the iteration cutoff condition is met, step S210 is executed.

[0075] In step S204, the synthesized data is mixed with other data and input into the discriminator.

[0076] In step S205, the discriminator is invoked to sample the input data according to the data domain to obtain the data items.

[0077] In step S206, the discriminator is invoked to randomly exchange data items to obtain combined data.

[0078] In step S207, the discriminator is invoked to score the combined data based on the data domain correlation result parameters.

[0079] In step S208, it is determined whether the combined data meets the scoring cutoff criteria. If the scoring cutoff criteria are not met, the process returns to step S206; if the scoring cutoff criteria are met, step S209 is executed.

[0080] In step S209, data domain correlation result parameters corresponding to the combined data are generated, and step S202 is executed.

[0081] In step S210, the training model is obtained from the central server.

[0082] In step S211, the model to be trained is trained using synthetic data to obtain the trained model.

[0083] In step S212, the trained model is uploaded to the central server, where it undergoes global training to obtain a comprehensive model.

[0084] Global training refers to the process of training a comprehensive model by using feedback from multiple participants in the training process.

[0085] In step S213, it is determined whether the integrated model meets the expected standard conditions. If the expected standard conditions are not met, step S214 is executed; if the expected standard conditions are met, step S215 is executed.

[0086] The central server can determine whether the integrated model meets the expected standard conditions and transmit the judgment result to the model training participants so that they can determine whether the integrated model meets the expected standard conditions.

[0087] In step S214, the integrated model is obtained from the central server as the model to be trained, the discriminator is updated according to the model parameters of the model to be trained, the target data domain correlation parameters are output, and step S202 is executed.

[0088] In step S215, a comprehensive model is obtained from the central server and used as a business model.

[0089] The specific details of steps S201 to S215 above can be found in the relevant descriptions in the above embodiments, and will not be repeated here.

[0090] It should be noted that the acquisition, storage, use, and processing of information and data in this embodiment are all authorized by the user or relevant organization and comply with the relevant provisions of national laws and regulations.

[0091] The second aspect of this application provides a business model training device. Figure 5This is a schematic diagram of the structure of a business model training device provided in an embodiment of this application, as shown below. Figure 5 As shown, the business model training device 300 may include a preprocessing module 301, a training module 302, and an upload module 303.

[0092] The preprocessing module 301 can be used to acquire raw business data, extract data fields from the raw business data, and obtain the data field correlation parameters corresponding to the raw business data based on the raw business data.

[0093] Data domain association parameters are used to characterize the association relationships between data domains.

[0094] The training module 302 can be used to call the generator and discriminator to perform data processing iterations until the data processing iterations meet the iteration cutoff conditions. The target synthetic data is used to train the model to be trained obtained from the central server to obtain the trained model.

[0095] In some examples, the iteration cutoff condition includes: the number of data processing iterations reaches a preset number of iterations.

[0096] The generator produces synthetic data based on data domain association parameters. The discriminator generates data domain association parameters based on the input data. The target synthetic data is obtained by the generator based on the data domain association parameters from the most recent data processing iteration.

[0097] The upload module 303 can be used to upload the trained model to the central server so that the central server can generate a comprehensive model based on the feedback from multiple parties.

[0098] Training module 302 can also be used to generate synthetic data in the event that the integrated model does not meet the expected standard conditions. Based on the integrated model issued by the central server as the model to be trained, it calls the generator and discriminator to work together to generate synthetic data. The synthetic data is then used to train the model to be trained. Finally, it calls upload module 303 to upload the trained model to the central server. This process continues until the integrated model generated by the central server based on feedback from multiple parties meets the expected standard conditions, at which point the integrated model is determined as the business model.

[0099] In this embodiment, data domains are extracted from the original business data. Based on the relationships between data domains in the original business data, a generator and a discriminator are used for iterative data processing. This enables the generator to output synthetic data that is similar to the original business data in appearance, meaning, and statistical characteristics, but cannot be traced back to the original business data. This synthetic data is used to train the model to be trained provided by the central server. The trained model is then fed back to the central server. The central server obtains a comprehensive model based on the trained models provided by multiple model training participants. If the comprehensive model does not meet the expected standard conditions, it is distributed as a model to be trained to each model training participant for further training. Each model training participant continues to feed back its trained model to the central server so that the central server obtains a new comprehensive model, until the comprehensive model meets the expected standard conditions. The model that meets the expected standard conditions is then determined as the business model. During model training, the synthetic data used to train the model is not real data, thus protecting data privacy and security. Furthermore, the model to be trained distributed by the central server is trained locally by each model training participant using the synthetic data generated locally by the participant. The synthetic data from different model training participants does not circulate among themselves, further enhancing data security.

[0100] The synthetic data used by each model training participant is different. By combining the trained models obtained from the synthetic data from all participants, a comprehensive model is obtained. This avoids the bias and incompleteness that may be caused by data provided by a single model training participant, and improves the effectiveness and generalization ability of the final business model. Moreover, in this embodiment, no data encryption is performed. Compared with homomorphic encryption methods that require multiple addition and multiplication operations, this embodiment can save a significant amount of storage and computing resources.

[0101] In some embodiments, the training module 302 can be used to: obtain model parameters of the comprehensive model issued by the central server as the model to be trained; update the discriminator according to the model parameters, so that the discriminator generates target data domain correlation parameters, wherein the correlation between data domains represented by the target data domain correlation parameters is consistent with or tends to be consistent with the correlation between data domains required by the model to be trained; and call the generator to generate synthetic data according to the target data domain correlation parameters.

[0102] In some embodiments, the training module 302 can be used to: call the generator to generate synthetic data based on the data domain correlation parameters corresponding to the original business data, and mix the synthetic data with the acquired business data and input it into the discriminator; call the discriminator to generate the corresponding data domain correlation parameters based on the input data, and transmit the data domain correlation parameters to the generator; call the generator to generate synthetic data based on the transmitted data domain correlation parameters, and mix the synthetic data with the previously generated synthetic data and input it into the discriminator; call the discriminator to generate the corresponding data domain correlation parameters based on the input data, until the data processing iteration meets the iteration cutoff condition.

[0103] In some examples, training module 302 can be used to: call the discriminator to split the input data into data items according to the data domain; call the discriminator to randomly swap the data items to obtain multiple combined data; call the discriminator to score the combined data based on the data domain correlation parameters corresponding to the original business data to obtain a scoring result; if the scoring result does not meet the scoring cutoff condition, call the discriminator to randomly swap the data items again, add combined data, and score the combined data until the scoring result meets the scoring cutoff condition, and call the discriminator to generate the corresponding data domain correlation result parameters based on the combined data.

[0104] In some examples, the scoring cutoff criteria include: the proportion of combined data with a score greater than or equal to the target score is higher than or equal to a preset proportion threshold; and / or, the number of combined data reaches the target number.

[0105] Figure 6 This is a schematic diagram of the structure of a business model training device provided in another embodiment of this application. Figure 6 and Figure 5 The difference is that, Figure 6 The business model training device 300 shown may also include a quality assessment module 304.

[0106] The quality assessment module 304 can be used to: analyze the target synthetic data and obtain at least one index parameter of the target synthetic data when the data processing iteration meets the iteration deadline; obtain the quality score of the target synthetic data based on the index parameter; and control the training module 302 to call the generator and discriminator again to perform data processing iteration to obtain newly generated target synthetic data, until the quality score obtained by the quality assessment module 304 is greater than or equal to the preset standard score.

[0107] A third aspect of this application also provides a business model training device. Figure 7 This is a schematic diagram of the structure of a business model training device provided in one embodiment of this application. Figure 7As shown, the business model training device 400 includes a memory 401, a processor 402, and a computer program stored on the memory 401 and capable of running on the processor 402.

[0108] In some examples, the processor 402 described above may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits that may be configured to implement the embodiments of this application.

[0109] Memory 401 may include read-only memory (ROM), random access memory (RAM), disk storage media device, optical storage media device, flash memory device, electrical, optical, or other physical / tangible memory storage device. Therefore, typically, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the business model training method according to embodiments of this application.

[0110] The processor 402 reads the executable program code stored in the memory 401 to run the computer program corresponding to the executable program code, so as to implement the business model training method in the above embodiment.

[0111] In some examples, the business model training device 400 may also include a communication interface 403 and a bus 404. For example, Figure 7 As shown, the memory 401, processor 402, and communication interface 403 are connected through bus 404 and complete communication with each other.

[0112] The communication interface 403 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application. Input devices and / or output devices can also be connected through the communication interface 403.

[0113] Bus 404 includes hardware, software, or both, that couples the components of the business model training device 400 together. For example, and not limitingly, bus 404 may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Micro Channel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-E) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local Bus (VLB) bus, or other suitable buses, or a combination of two or more of these. Where appropriate, bus 404 may include one or more buses. Although specific buses are described and illustrated in the embodiments of this application, this application considers any suitable bus or interconnection.

[0114] A fourth aspect of this application also provides a computer-readable storage medium storing computer program instructions. When these computer program instructions are executed by a processor, they can implement the business model training method described in the above embodiments and achieve the same technical effect. To avoid repetition, further details are omitted here. The aforementioned computer-readable storage medium may include non-transitory computer-readable storage media, such as read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, etc., and is not limited thereto.

[0115] This application also provides a computer program product. When the instructions in the computer program product are executed by the processor of an electronic device, the electronic device executes the business model training method described above and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0116] It should be clarified that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. For the device embodiments, equipment embodiments, computer-readable storage medium embodiments, and computer program product embodiments, the relevant parts can be referred to the description section of the method embodiments. This application is not limited to the specific steps and structures described above and shown in the figures. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application. Furthermore, for the sake of brevity, detailed descriptions of known methods and techniques are omitted here.

[0117] The aspects of this application have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by dedicated hardware performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0118] Those skilled in the art will understand that the above embodiments are exemplary and not restrictive. Different technical features appearing in different embodiments can be combined to achieve beneficial effects. Based on a study of the drawings, specification, and claims, those skilled in the art should be able to understand and implement other variations of the disclosed embodiments. In the claims, the term "comprising" does not exclude other means or steps; the quantifier "a" does not exclude a plurality; the terms "first" and "second" are used to identify names and not to indicate any particular order. No reference numerals in the claims should be construed as limiting the scope of protection. The functionality of multiple parts appearing in the claims can be implemented by a single hardware or software module. The appearance of certain technical features in different dependent claims does not mean that these technical features cannot be combined to achieve beneficial effects.

Claims

1. A business model training method, characterized in that, include: Obtain raw business data, extract data domains from the raw business data, and obtain data domain correlation parameters corresponding to the raw business data based on the raw business data. The data domain correlation parameters are used to characterize the correlation between data domains. A data domain is a category item into which data is divided. The generator and discriminator are invoked to perform data processing iterations until the data processing iteration meets the iteration deadline. The model to be trained, obtained from the central server, is trained using the target synthetic data to obtain the trained model. The generator generates synthetic data based on the data domain correlation parameters. The synthetic data is mixed with business data or with the previously generated synthetic data and input into the discriminator. The discriminator generates the data domain correlation parameters based on the input data and transmits the data domain correlation parameters to the generator for iteration. The target synthetic data is obtained by the generator based on the data domain correlation parameters of the most recent data processing iteration. The trained model is uploaded to the central server so that the central server can generate a comprehensive model based on the trained model fed back from multiple parties. If the integrated model fails to meet the expected standard conditions, the generator and discriminator are called to collaboratively generate synthetic data based on the integrated model issued by the central server as the model to be trained. The synthetic data is used to train the model to be trained, and the trained model is uploaded to the central server. This process continues until the integrated model generated by the central server based on feedback from multiple parties meets the expected standard conditions, at which point the integrated model is determined as the business model.

2. The method according to claim 1, characterized in that, The step of generating synthetic data by calling the generator and discriminator in collaboration based on the comprehensive model issued by the central server as the model to be trained includes: Obtain the model parameters of the integrated model issued by the central server as the model to be trained; Based on the model parameters, the discriminator is updated so that the discriminator generates target data domain correlation parameters. The correlation between data domains represented by the target data domain correlation parameters is consistent with or tends to be consistent with the correlation between data domains required by the model to be trained. The generator is invoked to generate synthetic data based on the target data domain correlation parameters.

3. The method according to claim 1, characterized in that, The process of calling the generator and discriminator to perform data processing iterations until the data processing iterations meet the iteration cutoff condition includes: The generator is invoked to generate synthetic data based on the data domain correlation parameters corresponding to the original business data, and the synthetic data is mixed with the acquired business data and input into the discriminator. The discriminator is invoked to generate the corresponding data domain correlation parameters based on the input data, and the data domain correlation parameters are transmitted to the generator; The generator is invoked to generate synthetic data based on the transmitted data domain correlation parameters, and the synthetic data is mixed with the previously generated synthetic data and input into the discriminator. The discriminator is invoked to generate the corresponding data domain correlation parameters based on the input data, until the data processing iteration meets the iteration cutoff condition.

4. The method according to claim 3, characterized in that, The step of calling the discriminator to generate the corresponding data domain correlation parameters based on the input data includes: The discriminator is invoked to split the input data into data items according to the data domain; The discriminator is invoked to randomly swap the data items, resulting in multiple combined data. The discriminator is invoked to score the combined data based on the data domain correlation parameters corresponding to the original business data, and a scoring result is obtained; If the scoring result does not meet the scoring cutoff condition, the discriminator is invoked to randomly swap the data items again, adding the combined data and scoring the combined data until the scoring result meets the scoring cutoff condition. Then, the discriminator is invoked to generate the corresponding data domain correlation result parameter based on the combined data.

5. The method according to claim 4, characterized in that, The scoring cutoff criteria include: The proportion of the combined data whose scoring results are greater than or equal to the target score is higher than or equal to a preset proportion threshold. And / or, The number of combined data has reached the target number.

6. The method according to claim 1, characterized in that, The iteration cutoff condition includes: the number of data processing iterations reaches the preset number of iterations.

7. The method according to claim 1, characterized in that, Also includes: If the data processing iteration meets the iteration deadline, analyze the target synthetic data to obtain at least one index parameter of the target synthetic data; Based on the aforementioned index parameters, the quality score of the target synthetic data is obtained; If the quality score is less than the preset standard score, the generator and discriminator are called again to perform data processing iterations to obtain newly generated target synthetic data until the quality score is greater than or equal to the preset standard score.

8. A business model training device, characterized in that, include: The preprocessing module is used to acquire raw business data, extract data domains from the raw business data, and obtain data domain correlation parameters corresponding to the raw business data based on the raw business data. The data domain correlation parameters are used to characterize the correlation between data domains. A data domain is a category item into which data is divided. The training module is used to call the generator and discriminator to perform data processing iterations until the data processing iterations meet the iteration deadline. The target synthetic data is used to train the model to be trained obtained from the central server to obtain the trained model. The generator generates synthetic data according to the data domain correlation parameters, and mixes the synthetic data with business data or with the previously generated synthetic data and inputs it into the discriminator. The discriminator generates the data domain correlation parameters according to the input data and transmits the data domain correlation parameters to the generator for iteration. The target synthetic data is obtained by the generator based on the data domain correlation parameters of the most recent data processing iteration. The upload module is used to upload the trained model to the central server so that the central server can generate a comprehensive model based on the trained model fed back from multiple parties. The training module is further configured to, when the integrated model fails to meet the expected standard conditions, use the integrated model issued by the central server as the model to be trained, call the generator and discriminator to collaboratively generate synthetic data, use the synthetic data to train the model to be trained, and call the upload module to upload the trained model to the central server, until the integrated model generated by the central server based on feedback from multiple parties meets the expected standard conditions, and then determine the integrated model as the business model.

9. A business model training device, characterized in that, include: Processor and memory storing computer program instructions; When the processor executes the computer program instructions, it implements the business model training method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the business model training method as described in any one of claims 1 to 7.

Images