A data contribution incentive method and device for supporting machine learning model training

Abstract

The present invention relates to the technical field of machine learning, and specifically relates to a data contribution incentive method and device for supporting the training of machine learning models. The present invention uses the training data provided by the data holding terminal to train the machine learning model, calculates the model performance of the machine learning model after training, and the model performance is used to characterize the quality of the machine learning model after training. Then, an incentive value is calculated based on the model performance, and the incentive value can reflect the contribution of the training data provided by the data holding terminal to the training of the machine learning model, that is, the incentive value reflects the quality of the training data provided by the data holding terminal or the size of the value of the training data. When there is a new model training requirement next time, the training data in the data holding terminal with a large incentive value can be preferentially considered, so as to obtain high-quality training data, and further better train the machine learning model.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular, to a data contribution incentive method and device for supporting the training of machine learning models. Background Art

[0002] The digital economy is a new type of economy. The key element of the digital economy is data resources, and the digital economy is the economic value generated by the circulation of data resources. A typical example in the digital economy is machine learning, and machine learning technologies include supervised learning, unsupervised learning, and semi-supervised learning. Machine learning has been widely applied in many fields such as biometric recognition and robotics. Machine learning models are the products of this technology. Trading machine learning models is an indirect form of data trading and one of the ways for data to indirectly generate value. It can not only overcome the constraints of data privacy and security, but also enhance the circulation of data and increase the value of data. This has given rise to the concept of a machine learning model market, that is, a large amount of data is required during the training of machine learning, and this large amount of data needs to be obtained from data holders.

[0003] The existing machine learning model market, as Figure 2 shown, consists of a model demander (the party that purchases the machine learning model obtained after training) and a machine learning market (the party responsible for training the machine learning model). The morphological functions of the existing machine learning model market platforms are relatively single, without providing the process of model trading and unable to meet the complex needs of model demanders. Specifically, the following deficiencies exist in the existing machine learning model trading market:

[0004] (1) From the perspective of model demanders, the types of models provided by the existing machine learning model market platforms are limited. When demanders put forward new demands, due to the lack of a training mechanism for new models, the demands of model demanders cannot be met in a timely manner.

[0005] (2) From the perspective of data holders, the existing machine learning model market platforms lack a data circulation mechanism among model demanders, data holders, and the machine learning market, and problems such as the data rights and interests of data holders not being guaranteed and model demanders having difficulty finding "high-quality" data will be encountered.

[0006] (3) From the perspective of market transactions, the existing machine learning model market platforms lack an effective incentive mechanism, resulting in blurred rights and interests of models and data, reducing the circulation and utilization rate of data. At the same time, the lack of a market operation mechanism will also bring problems such as the confirmation of rights, authorization, and protection of rights in the model trading market.

[0007] In summary, in the existing machine learning model market platform, it is difficult to obtain high-quality data for the training of machine learning models due to the lack of an incentive mechanism. Therefore, the existing technology still needs to be improved. Summary of the Invention

[0008] To solve the above technical problems, the present invention provides a data contribution incentive method and device for supporting the training of machine learning models, which solves the problem that it is difficult to obtain high-quality data for existing machine learning models due to the lack of an incentive mechanism.

[0009] To achieve the above object, the present invention adopts the following technical solutions:

[0010] In a first aspect, the present invention provides a data contribution incentive method for supporting the training of machine learning models, which includes:

[0011] Obtain training data sent by a data holding terminal;

[0012] Train the machine learning model based on the training data;

[0013] Calculate the model performance corresponding to the machine learning model after training;

[0014] According to the model performance, assign a corresponding incentive value to the data holding terminal where the training data is located, and the incentive value can be used as a basis for selecting the data holding terminal for the next training of the machine learning model.

[0015] In one implementation, the obtaining of the training data sent by the data holding terminal includes:

[0016] Send the data information required for training the machine learning model to the data holding terminal;

[0017] Obtain the training data and the label corresponding to the training data sent by the data holding terminal according to the data information;

[0018] Package the training data and the label corresponding to the training data in a set format;

[0019] Save the packaged training data and the label corresponding to the training data in a smart contract transaction pool, and the smart contract transaction pool is used to provide data for subsequent training of the machine learning model.

[0020] In one implementation, the training data is encrypted data, and the training of the machine learning model based on the training data includes:

[0021] Decrypt the training data to obtain the decrypted training data;

[0022] Obtain the label corresponding to the decrypted training data;

[0023] Delete the data that does not match the label from all the training data to obtain the processed training data;

[0024] Train the machine learning model based on the processed training data.

[0025] In one implementation, the training of the machine learning model based on the training data includes:

[0026] Obtain the target data in the training data from the intelligent contract transaction pool according to the label;

[0027] Extract the interface functions defined by the intelligent contract transaction pool;

[0028] Train the machine learning model based on the target data;

[0029] Update the parameters of the machine learning model through the interface functions during the process of training the machine learning model to complete the training of the machine learning model.

[0030] In one implementation, the calculation of the model performance corresponding to the trained machine learning model includes:

[0031] Obtain the verification input data and the verification output data corresponding to the verification input data;

[0032] Input the verification input data into the trained machine learning model to obtain the data output by the trained machine learning model;

[0033] Compare the verification output data with the data output by the trained machine learning model to obtain the model precision in the model performance.

[0034] In one implementation, the data incentive method further includes:

[0035] Collect the model parameters covered by the machine learning model before training and the model parameters covered by the trained machine learning model;

[0036] Obtain the information gain of the trained machine learning model relative to the machine learning model before training based on the model parameters covered by the machine learning model before training and the model parameters covered by the trained machine learning model.

[0037] In one implementation, endowing the data holding terminal where the training data is located with a corresponding incentive value according to the model performance, where the incentive value is used as one of the bases for selecting the data holding terminal in the next training of the machine learning model, includes:

[0038] Calculate the product of the model precision rate and the information gain in the model performance to obtain the product result.

[0039] Endow the data holding terminal where the training data is located with a corresponding incentive value according to the product result.

[0040] In one implementation, endowing the data holding terminal where the training data is located with a corresponding incentive value according to the product result includes:

[0041] Add the product result to a set constant to obtain the addition result.

[0042] Perform a power operation on the addition result as the exponent of a set base number, where the set base number is greater than one, to obtain the power operation result.

[0043] Endow the data holding terminal where the training data is located with a corresponding incentive value according to the power operation result.

[0044] In one implementation, the data incentive method further includes:

[0045] Perform an integration operation on the power operation result for the training data sent by the data holding terminal to obtain an integration result.

[0046] Endow the data holding terminal where the training data is located with a corresponding incentive value according to the integration result.

[0047] In one implementation, after endowing the data holding terminal where the training data is located with a corresponding incentive value according to the model performance, where the incentive value is used as the basis for selecting the data holding terminal in the next training of the machine learning model, it further includes:

[0048] Calculate the amount to be paid to the data holding terminal according to the incentive value.

[0049] Obtain the cost required for the platform for training the machine learning model.

[0050] Send the trained machine learning model to the model demand terminal.

[0051] Statistically obtain the market price of the trained machine learning model for the model demand terminal.

[0052] Add the amount, the cost, and the market price to obtain the price corresponding to the trained machine learning model.

[0053] In a second aspect, an embodiment of the present invention further provides a data contribution incentive device for supporting the training of a machine learning model. The device includes the following components:

[0054] A data acquisition module, configured to obtain training data sent by a data holding terminal;

[0055] A training module, configured to train the machine learning model according to the training data;

[0056] A performance calculation module, configured to calculate the model performance corresponding to the trained machine learning model;

[0057] An incentive module, configured to assign a corresponding incentive value to the data holding terminal where the training data is located according to the model performance, and the incentive value is used as a basis for selecting the data holding terminal for the next training of the machine learning model.

[0058] In a third aspect, an embodiment of the present invention further provides a terminal device. The terminal device includes a memory, a processor, and a data contribution incentive program for supporting the training of a machine learning model stored in the memory and executable on the processor. When the processor executes the data contribution incentive program for supporting the training of a machine learning model, the steps of the above-mentioned data contribution incentive method for supporting the training of a machine learning model are implemented.

[0059] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium. A data contribution incentive program for supporting the training of a machine learning model is stored on the computer-readable storage medium. When the data contribution incentive program for supporting the training of a machine learning model is executed by a processor, the steps of the above-mentioned data contribution incentive method for supporting the training of a machine learning model are implemented.

[0060] Beneficial effects: The present invention uses the training data provided by the data holding terminal to train the machine learning model, calculates the model performance of the machine learning model after training, and the model performance is used to characterize the quality of the machine learning model after training. Then, an incentive value is calculated based on the model performance. The incentive value can reflect the contribution of the training data provided by the data holding terminal to the training of the machine learning model, that is, the incentive value reflects the quality of the training data provided by the data holding terminal or the magnitude of the value of the training data, and thus can be used as a basis for rewarding the data provider. This can also encourage each data provider to provide high-quality data for model training. In addition, when the next training of a new machine learning model is required, the training data in the data holding terminal with a large incentive value can be preferentially considered, so as to obtain high-quality training data, and then the machine learning model can be better trained. Description of the Drawings

[0061] Figure 1 is the overall flowchart of the present invention;

[0062] Figure 2 is the machine learning model market in the background art;

[0063] Figure 3 is the machine learning model trading market platform of the present invention;

[0064] Figure 4 is the machine learning model initialization program code diagram;

[0065] Figure 5 is the schematic diagram of the relationship between the incentive value μ, the model performance, and the information gain;

[0066] Figure 6 is the interaction schematic diagram among the data provider, the model requester, and the machine learning model trading market;

[0067] Figure 7 is the internal structure principle block diagram of the terminal device provided by the embodiment of the present invention. Detailed Embodiments

[0068] The following combines the embodiments and the drawings of the specification to clearly and completely describe the technical solutions in the present invention. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0069] The digital economy is a new type of economy. The key element of the digital economy is data resources, and the digital economy is the economic value generated by the circulation of data resources. A typical example in the digital economy is machine learning, which includes supervised learning, unsupervised learning, and semi-supervised learning. Machine learning has been widely applied in many fields such as biometric recognition and robotics. Machine learning models are the products of this technology. Trading machine learning models is an indirect form of data trading and one of the ways for data to indirectly generate value. It can not only overcome the constraints of data privacy and security but also enhance the circulation of data and increase the value of data. This has given rise to the concept of the machine learning model market, that is, a large amount of data is needed during the training of machine learning, and this large amount of data needs to be purchased from data holders.

[0070] The existing machine learning model market, as Figure 2 shown, consists of model demanders (the party that purchases the machine learning model obtained after training) and the machine learning market (the party responsible for training the machine learning model). The form and function of the existing machine learning model market are relatively single, without providing the process of model trading and unable to meet the complex needs of model demanders. Specifically, the following deficiencies exist in the existing machine learning model trading market:

[0071] (1) From the perspective of model demanders, the types of models provided by the existing machine learning model market are limited. When demanders put forward new demands, due to the lack of a training mechanism for new models, the demands of model demanders cannot be met in a timely manner.

[0072] (2) From the perspective of data holders, the existing machine learning model market lacks a data circulation mechanism among model demanders, data holders, and the machine learning model market, and will encounter problems such as the data rights and interests of data holders not being guaranteed and model demanders having difficulty finding "high-quality" data.

[0073] (3) From the perspective of market transactions, the existing machine learning model market lacks an effective incentive mechanism, resulting in blurred rights and interests of models and data, reducing the circulation and utilization rate of data. At the same time, the lack of a market operation mechanism will also bring problems such as the confirmation of rights, authorization, and protection of rights in the model trading market.

[0074] In summary, in the existing machine learning model market, it is difficult to obtain high-quality data for machine learning model training due to the lack of an incentive mechanism. Therefore, the existing technology still needs to be improved.

[0075] To solve the above technical problems, the present invention provides a method and apparatus for incentivizing data contribution for machine learning model training, which solves the problem that it is difficult to obtain high-quality data in the existing machine learning model trading market due to the lack of a data contribution incentive mechanism. Specifically, in implementation, first, training data sent by a data holding terminal is obtained; then, the machine learning model is trained based on the training data; after that, the model performance corresponding to the trained machine learning model is calculated; finally, according to the model performance, an incentive value is given to the data holding terminal where the training data is located. In this embodiment, when it is necessary to train a new machine learning model next time, the training data in the data holding terminal with a large incentive value can be preferably selected, so as to obtain high-quality training data, and then the machine learning model can be better trained.

[0076] For example, there are two data holding terminals A and B. The machine learning model is continuously trained using the training data provided by data holding terminal A. After the training is completed, the performance of the trained machine learning model is evaluated, and an incentive value a is calculated based on this performance. Similarly, the machine learning model is continuously trained using the training data provided by data holding terminal B, and an incentive value b is obtained. When the incentive value b is greater than the incentive value a, it indicates that the data quality owned by data holding terminal B is higher than the data quality owned by data holding terminal A. When it is necessary to provide data for the training of a new machine learning model, data holding terminal B can be given priority, so as to better complete the training of the new machine learning model.

[0077] Exemplary method

[0078] The data contribution incentive method for supporting machine learning model training in this embodiment can be applied to a terminal device, and the terminal device can be a terminal product with a video playback function, such as a television, a computer, etc. In this embodiment, as Figure 1 shown, the data contribution incentive method for supporting machine learning model training specifically includes the following steps:

[0079] S100, obtain the training data sent by the data holding terminal.

[0080] The data incentive method in this embodiment is applied to a data trading market platform composed of a data provider (data holding terminal), a machine learning model trading market platform, and a model demander (model demand terminal) as Figure 3 shown.

[0081] The machine learning model trading market platform in this embodiment is for facilitating the mutual trading between model demanders lacking data for training models and data holders having data but not knowing how to generate value from it. Through the machine learning model trading market platform, the data held by data holders can be circulated to reflect the intrinsic value of the data they possess, thereby attracting more model demanders to purchase the data of data holders.

[0082] As Figure 3 shown, in this embodiment, the model demander first sends a model demand to the machine learning model trading market platform (i.e., what kind of machine learning model the model demander needs to obtain). The model demand includes content such as model task objectives, model structure requirements, verification means, and incentive functions. After receiving the model demand, the machine learning model trading market platform parses the model demand to obtain the type of training data required for training the machine learning model, and then the machine learning model trading market platform sends this type to the data holder, and the data holder will send the corresponding training data to the machine learning model trading market platform, that is, the machine learning model trading market platform obtains the training data provided by the data holder. At the same time, after receiving the data request sent by the model demander, the machine learning model trading market platform will also write relevant smart contract code according to the task description. The rules in the contract include start and end times, termination conditions for machine learning model training, trading functions related to funds, and model update rule functions. Using the program as Figure 4 shown, the model trading market platform deploys the smart contract through a contract deployment tool and obtains a contract address Addr = Deploy(Contract), where Contract contains the contract content, Deploy is the contract deployment interface function provided by the blockchain platform, and Addr is the contract address returned after successfully deploying the contract. The contract address is the unique identifier for the contract to exist on the blockchain in the form of a transaction. After successful deployment, the contract address will be returned to the user who deploys the contract. Other users can obtain the contract address from the user who deploys the contract or can also obtain the address by scanning the blockchain system.

[0083] Based on the above principle, step S100 includes the following steps S101, S102, S103, and S104:

[0084] S101, Send the data information required for training the machine learning model to the data holding terminal.

[0085] The model requester lacks the necessary training data to train the corresponding machine learning model, so it submits a model purchase request to the machine learning model trading market platform. At the same time, the model requester provides the necessary task description (i.e., model requirements) to the machine learning model trading market platform. The task description includes, but is not limited to, information such as the model type, model initialization parameters, validation dataset (data used to validate the trained machine learning model), incentive amount (if the trained machine learning model needed can be obtained by using the training data provided by the data holder, the model requester will give a certain reward amount to the data holder), and data requirements. After receiving the model requirements, the machine learning model trading market platform parses the model requirements to obtain the data information corresponding to the training data required for training the model requested by the model requester.

[0086] In this embodiment, after the machine learning model trading market platform receives the task description from the model requester, it parses the task description to obtain the data information, and then sends the data information to the data holder (data holding terminal).

[0087] S102, Obtain the training data and the labels corresponding to the training data sent by the data holding terminal according to the data information.

[0088] After the data holder in step S101 receives the data information, the data holder will send the corresponding training data and the labels corresponding to the training data to the machine learning model trading market platform according to the data information.

[0089] S103, Package the training data and the labels corresponding to the training data in a set format.

[0090] S104, Save the packaged training data and the labels corresponding to the training data in the smart contract trading pool, and the smart contract trading pool is used to provide data for subsequent training of the machine learning model.

[0091] After the machine learning model trading market platform receives the training data and the labels corresponding to the training data, it will package the two in a set format, and the set format is pre-saved in the smart contract (the so-called smart contract is a special piece of code that contains the function of saving the set data format). The set format in this embodiment is, for example, <data, label>. For example, if the data is the data used to represent the characteristics of the animal cat, then the corresponding label is cat.

[0092] The specific processes of step S103 and step S104 are as follows:

[0093] If the data holder has training data that conforms to the data information, the training data can be uploaded to the machine learning model trading market platform for trading. This process can be described as follows: First, the data holder submits a deposit to the machine learning model trading market platform and provides data resources, i.e., Addr.Deposit(user, coins), where user and coins are the user identifier of the data contributor and the deposit respectively, and Deposit is the interface function for submitting the deposit defined in the contract. When the deposit is successfully paid, the model market will package it into the corresponding contract trading format according to the data requirement format in the contract (such as <data, label>) and send it to the trading pool where the smart contract is located by calling Addr.Receiver_data(user, data), where user and data are the user identifier of the data holder and the training data to be submitted respectively, and Receiver_data is the interface function for submitting data defined in the contract. To ensure the security and privacy of the data, the data uploaded by the data holder is encrypted data. The encryption algorithm can adopt symmetric or asymmetric encryption algorithms according to security needs, such as AES, SM2, SM4.

[0094] S200. Train the machine learning model based on the training data.

[0095] Figure 3 After the machine learning model trading market platform in receives the training data sent by the data holder, it does not directly train the machine learning model with the training data. Instead, it first processes the training data to obtain the preprocessed training data and then trains the machine learning model. Step S200 includes the following steps S201, S202, S203, and S204:

[0096] S201. Decrypt the training data to obtain the decrypted training data.

[0097] The reason for decrypting the training data is that the data holder encrypted the data before uploading it to ensure data security.

[0098] S202. Obtain the label corresponding to the decrypted training data.

[0099] S203. Delete the data that does not match the label from all the training data to obtain the processed training data.

[0100] Step S202 and step S203 constitute a verification mechanism, which is used to verify whether the data uploaded by the data holder meets the requirements. In this embodiment, the data maintainer of the machine learning model trading market platform first decrypts the data data (the data uploaded by the data holder), then searches for the rules defined in the contract from the contract address Addr in the transaction, and checks the uploaded data according to the rules. For example, it is confirmed whether the data and the label are consistent. If the rules are met, the data is submitted to the blockchain using the corresponding consensus method. If not, the transaction information (the data uploaded by the data holder) is deleted. If the data maintainer of the market platform discovers that the data holder maliciously disrupts the market order (such as false pictures and other behaviors), the deposit submitted by the user will be withheld in the contract as a penalty mechanism (the penalty mechanism also includes not giving priority to the punished data holder when conducting model training next time). If the data holder normally exits the market transaction, the deposit will be refunded.

[0101] The processed training data is placed in the blockchain and stored in the smart contract trading pool on the blockchain. When step S204 starts to train the machine learning model, the required training data is fetched from the blockchain.

[0102] S204, training the machine learning model based on the processed training data.

[0103] The training data in step S204 is sourced from the blockchain. Step S204 includes the following steps S2041, S2042, S2043, and S2044:

[0104] S2041, obtaining the target data in the training data from the smart contract trading pool according to the label.

[0105] S2042, extracting the interface function defined by the smart contract trading pool.

[0106] S2043, training the machine learning model based on the target data.

[0107] S2044, updating the parameters of the machine learning model through the interface function during the process of training the machine learning model to complete the training of the machine learning model.

[0108] The detailed processes of steps S2041 to S2044 are as follows: After the data is packaged into a transaction and submitted to the smart contract, the model training engine obtains new data through Addr.Retrieval_data(data_index), where data_index is the identifier of the data, and Retrieval_data is the interface function for data retrieval defined in the contract. Then, it calls the interface for model training, updates the model with the new data to obtain new model parameters, and calls the corresponding interface to update the model parameters Addr.Model_update(new_model) in the contract, where new_model is the new model parameter and Model_update is the interface function for model update defined in the contract.

[0109] Step 200 of this embodiment may also only include steps S2041, S2042, S2043, and S2044, that is, it is not necessary to verify the data uploaded by the data holder, but directly save it in the smart contract transaction pool of the blockchain, which can improve the speed of data upload.

[0110] S300, calculate the model performance corresponding to the machine learning model after training.

[0111] This embodiment can use the accuracy rate, precision rate, and recall rate of the machine learning model after training to represent the model performance. When using the precision rate to represent the model performance, step S300 includes the following steps S301, S302, and S303:

[0112] S301, obtain the verification input data and the verification output data corresponding to the verification input data.

[0113] The verification input data and the verification output data corresponding to the verification input data constitute the verification dataset. The verification dataset is also uploaded by the data holder to the machine learning model trading market platform. However, the verification dataset is only saved in the isolated server of the platform and can only access the verification service and cannot read and write the verification dataset, which can prevent the leakage of the verification dataset and thus ensure the security of the data.

[0114] S302, input the verification input data into the machine learning model after training to obtain the data output by the machine learning model after training.

[0115] S303, compare the verification output data with the data output by the machine learning model after training to obtain the model precision rate in the model performance.

[0116] The verification output data is known sample data. The closer the data output by the machine learning model is to the verification output data, the higher the precision rate of the model.

[0117] In this embodiment, while calculating the performance of the model, the information gain of the machine learning model after training relative to before training is also calculated, and based on the model performance and the information gain, the incentive to be given to the data holder for the data uploaded by the data holder is calculated. Calculating the information gain includes the following steps S304 and S305:

[0118] S304, collect the model parameters covered by the machine learning model before training and the model parameters covered by the machine learning model after training.

[0119] S305, based on the model parameters covered by the machine learning model before training and the model parameters covered by the machine learning model after training, obtain the information gain of the machine learning model after training relative to the machine learning model before training.

[0120] There are numerous model parameters involved in the machine learning model. Before and after the training of the machine learning model, these model parameters will change, thus generating information gain. The information gain is to calculate the probability distribution of the model parameters after training and the probability distribution of the model parameters after training, and the difference between these two probability distributions.

[0121] S400, based on the model performance, assign a corresponding incentive value to the data holding terminal where the training data is located, and the incentive value is used as the basis for selecting the data holding terminal for the next training of the machine learning model.

[0122] This embodiment comprehensively considers the model performance and the information gain to assign a corresponding incentive value to the data holder. Step S400 includes the following steps S401, S402, S403, and S404:

[0123] S401, calculate the product of the model precision in the model performance and the information gain to obtain the product result.

[0124] S402, add the product result to a set constant to obtain the addition result.

[0125] S403, perform a power operation on the addition result as the exponent of a set base number, where the set base number is greater than one, to obtain the power operation result.

[0126] S404, based on the power operation result, assign a corresponding incentive value to the data holding terminal where the training data is located.

[0127] Steps S401 to S404 obtain the power operation result f(x i ):

[0128]

[0129] In the formula, A(x i ) is the model performance corresponding to the model after training with the training data x uploaded by the data holder, E(x i ) is the information gain, c is a set constant, and p is a set base number. The result of the power operation f(x i ) is directly used as the incentive value for the data holding terminal. i ) is directly used as the incentive value for the data holding terminal.

[0130] The above steps S401 to S404 calculate the incentive value using the training data x uploaded by the data holder once (such as the data uploaded for the i-th time). In this embodiment, the result of the power operation f(x) corresponding to any training data x uploaded by the data holder can also be calculated, and then the integral of the result of the power operation f(x) is used to obtain the incentive value μ. The steps for calculating μ in this embodiment include S405 and S406: i The above steps S401 to S404 calculate the incentive value using the training data x uploaded by the data holder once (such as the data uploaded for the i-th time). In this embodiment, the result of the power operation f(x) corresponding to any training data x uploaded by the data holder can also be calculated, and then the integral of the result of the power operation f(x) is used to obtain the incentive value μ. The steps for calculating μ in this embodiment include S405 and S406:

[0131] S405, perform an integration operation on the result of the power operation for the training data sent by the data holding terminal to obtain an integration result.

[0132] S406, according to the integration result, assign a corresponding incentive value to the data holding terminal where the training data is located.

[0133] This embodiment calculates f(x) and μ using the following formula:

[0134] f(x) = p A(x)E(x)+c , p > 1

[0135] μ = ∫f(x)dx

[0136] As Figure 5 shown, the value of μ shows an increasing trend as the model is trained, Figure 5 and the incentive function is f(x).

[0137] The model requester, data holder, and machine learning model trading market in this embodiment complete the training of the machine learning model using the flow chart as Figure 6 shown, and deliver the trained machine learning model to the model requester. At the same time, the incentive value is also saved in the smart contract. The specific process of saving includes: after obtaining new model parameters, through the data trading incentive mechanism, according to predetermined rules (such as the accuracy improvement of the new and old models and the information gain of the model), calculate the corresponding rewards. The platform packages the improvement of the data on the model and the corresponding rewards and sends them to the smart contract.

[0138] The smart contract is Addr.Reward(sender, reward), which completes this transaction. Here, sender and reward are respectively the user identifier of the data holder who should receive the reward and the corresponding reward amount, and Reward is the interface function for distributing rewards defined in the contract.

[0139] Figure 6 As can be seen, in the process of training the machine learning model in this embodiment, a model training engine, a data verification engine, and a data trading incentive mechanism are involved. The following will explain them separately:

[0140] Data verification engine: This engine aims to verify the availability of the data uploaded by the data holder, and it can prevent low-quality data from flowing into the transaction and affecting the final model quality. The data verification engine in this embodiment mainly allows the data holder to initiate a data upload request to the trading platform according to the requirements and submit a deposit. The data holder uploads data in the format of <sample, label> to the platform. In the model trading market platform, the data maintainer verifies the uploaded data. If it meets the data correctness, it is submitted to the smart contract using the corresponding consensus method. If it does not meet the requirements, the data trading information is deleted and the user data is returned. If a data holder maliciously disrupts the market trading rules, the market will deduct the deposit of the holder.

[0141] Model training engine: This engine aims to execute the model training task and perform the model training of machine learning through the existing computing power foundation of the market platform. When the data meets the requirements, the model market platform calls the model training engine to learn new data. For example, the trading platform proposed in this patent can train the model for common classification models. The classification model learns from the training data set to establish a mapping f = x → Y from the input space x to the output space Y.

[0142] Data trading incentive mechanism: This mechanism aims to execute tasks including but not limited to data contribution calculation, model pricing, etc.

[0143] By calculating μ through steps S100 to S400, this embodiment also sets a price for the machine learning model after training according to the size of μ, including the following steps S501, S502, S503, S504, S505:

[0144] S501, calculate the amount to be paid to the data holding terminal according to the incentive value.

[0145] The incentive value in this embodiment is μ. The larger μ is, the larger the amount that the model purchaser needs to pay to the data holder, and vice versa.

[0146] S502, obtain the cost required by the platform for training the machine learning model.

[0147] The platform for training the machine learning model, i.e., Figure 3 the machine learning model trading market platform in [], the cost is the amount that the model demander needs to pay to the machine learning model trading market platform. The cost of the platform will be calculated based on the smooth cost of model training and the labor cost.

[0148] S503, send the trained machine learning model to the model demand terminal.

[0149] S504, count the market price of the trained machine learning model faced by the model demand terminal.

[0150] The model demander will evaluate the market price of the trained machine learning model. This market price can be the amount that the model demander hopes to make a profit after deducting the amount in step S501 and the cost in step S502.

[0151] S505, add the amount, the cost, and the market price to obtain the price corresponding to the trained machine learning model.

[0152] In summary, the present invention uses the training data provided by the data holding terminal to train the machine learning model, calculates the model performance of the trained machine learning model, and the model performance is used to characterize the quality of the trained machine learning model. Then, an incentive value is calculated according to the model performance, and the incentive value can reflect the contribution of the training data provided by the data holding terminal to training the machine learning model, that is, the incentive value reflects the quality of the training data provided by the data holding terminal or the size of the value of the training data. When it is necessary to train a new machine learning model next time, the training data in the data holding terminal with a large incentive value can be preferably selected, so as to obtain high-quality training data, and then the machine learning model can be better trained.

[0153] In addition, the present invention proposes an effective strategy to ensure data availability. Through the smart contract technology, multi-party verification is realized, ensuring the data availability for model update.

[0154] The model training engine proposed by the present invention can meet common machine learning models, including but not limited to supervised models, unsupervised models, or semi-supervised models, etc.

[0155] When the data trading incentive mechanism of the present invention receives a model requirement from the model requester, it requires the model requester to upload a model validation set or validation requirements. The market side also ensures the confidentiality of the validation set to prevent users from sending data resources for the validation set, which may cause chaos in the market rules. Based on smart contracts, the data trading incentive mechanism is deployed to verify data availability, ensuring the effectiveness of model training. The incentive criteria are based on the performance improvement of the model and the information gain of the model, and are verified by multiple parties through the machine learning model trading platform to provide reasonable incentives for data holders.

[0156] Exemplary device

[0157] This embodiment also provides a data contribution incentive device for supporting machine learning model training. The device includes the following components:

[0158] A data acquisition module, configured to obtain training data sent by a data holding terminal;

[0159] A training module, configured to train the machine learning model according to the training data;

[0160] A performance calculation module, configured to calculate the model performance corresponding to the machine learning model after training;

[0161] An incentive module, configured to assign a corresponding incentive value to the data holding terminal where the training data is located according to the model performance, and the incentive value is used as one of the bases for selecting the data holding terminal for the next training of the machine learning model.

[0162] Based on the above embodiments, the present invention further provides a terminal device, and its principle block diagram can be as Figure 7 shown. The terminal device includes a processor, a memory, a network interface, a display screen, and a temperature sensor connected through a system bus. Among them, the processor of the terminal device is used to provide computing and control capabilities. The memory of the terminal device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the terminal device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements a data contribution incentive method for supporting machine learning model training. The display screen of the terminal device can be a liquid crystal display screen or an electronic ink display screen, and the temperature sensor of the terminal device is pre-set inside the terminal device to detect the operating temperature of the internal device.

[0163] Those skilled in the art can understand, Figure 7The principle block diagram shown only shows the block diagram of some structures related to the solution of the present invention, and does not constitute a limitation on the terminal device to which the solution of the present invention is applied. The specific terminal device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0164] In one embodiment, a terminal device is provided. The terminal device includes a memory, a processor, and a data contribution incentive program stored in the memory and operable on the processor to support machine learning model training. When the processor executes the data contribution incentive program for supporting machine learning model training, the following operation instructions are implemented:

[0165] Obtain the training data sent by the data holding terminal;

[0166] Train the machine learning model based on the training data;

[0167] Calculate the model performance corresponding to the machine learning model after training;

[0168] According to the model performance, assign a corresponding incentive value to the data holding terminal where the training data is located. The incentive value is used as one of the bases for selecting the data holding terminal for the next training of the machine learning model.

[0169] Those of ordinary skill in the art can understand that all or part of the processes in the above method embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above method embodiments. Among them, any reference to the memory, storage, database, or other media used in the embodiments provided by the present invention can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0170] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data contribution incentive method for supporting the training of machine learning models, characterized in that, it includes: The model demand side submits a model purchase demand to the machine learning model trading market platform, and the machine learning model trading market platform obtains the training data sent by the data holding terminal; Training the machine learning model based on the training data; Calculating the model performance corresponding to the machine learning model after training; According to the model performance, assigning a corresponding incentive value to the data holding terminal where the training data is located, and the incentive value can be used as one of the bases for selecting the data holding terminal for the next training of the machine learning model; Collecting the model parameters covered by the machine learning model before training and the model parameters covered by the machine learning model after training; Based on the model parameters covered by the machine learning model before training and the model parameters covered by the machine learning model after training, obtaining the information gain of the machine learning model after training relative to the machine learning model before training; Calculating the product of the model precision rate in the model performance and the information gain to obtain the product result; According to the product result, assigning a corresponding incentive value to the data holding terminal where the training data is located.

2. The data contribution incentive method for supporting the training of machine learning models according to claim 1, characterized in that, The obtaining of the training data sent by the data holding terminal includes: Sending the data information required for training the machine learning model to the data holding terminal; Obtaining the training data sent by the data holding terminal according to the data information and the label corresponding to the training data; Packing the training data and the label corresponding to the training data in a set format; Storing the packed training data and the label corresponding to the training data in the smart contract trading pool, and the smart contract trading pool is used to provide data for the subsequent training of the machine learning model.

3. The data contribution incentive method for supporting the training of machine learning models according to claim 1, characterized in that, The training data is encrypted data, and the training of the machine learning model based on the training data includes: Decrypting the training data to obtain the decrypted training data; Obtaining the label corresponding to the decrypted training data; Deleting the data that is not consistent with the label from all the training data to obtain the processed training data; Training the machine learning model based on the processed training data.

4. The data contribution incentive method for supporting the training of machine learning models according to claim 2, characterized in that, The training of the machine learning model based on the training data includes: Obtaining the target data in the training data from the smart contract trading pool according to the label; Extracting the interface function defined by the smart contract trading pool; Training the machine learning model based on the target data. During the process of training the machine learning model, update the parameters of the machine learning model through the interface function to complete the training of the machine learning model.

5. The data contribution incentive method for supporting the training of a machine learning model according to claim 1, wherein, calculating the model performance corresponding to the machine learning model after training includes: obtaining verification input data and verification output data corresponding to the verification input data; inputting the verification input data into the machine learning model after training to obtain the data output by the machine learning model after training; comparing the verification output data with the data output by the machine learning model after training to obtain the model precision rate in the model performance.

6. The data contribution incentive method for supporting the training of a machine learning model according to claim 1, wherein, assigning a corresponding incentive value to the data holding terminal where the training data is located according to the product result includes: adding the product result to a set constant to obtain an addition result; performing a power operation on the addition result as the exponent of a set base number, where the set base number is greater than one; assigning a corresponding incentive value to the data holding terminal where the training data is located according to the power operation result.

7. The data contribution incentive method for supporting the training of a machine learning model according to claim 6, wherein, the data incentive method further includes: performing an integration operation on the power operation result for the training data sent by the data holding terminal to obtain an integration result; assigning a corresponding incentive value to the data holding terminal where the training data is located according to the integration result.

8. The data contribution incentive method for supporting the training of a machine learning model according to claim 1, wherein, assigning a corresponding incentive value to the data holding terminal where the training data is located according to the model performance, and the incentive value is used as a basis for selecting the data holding terminal for the next training of the machine learning model. After that, it further includes: calculating the amount to be paid to the data holding terminal according to the incentive value; obtaining the cost required for the platform for training the machine learning model; sending the machine learning model after training to the model demand terminal; statistical market price of the machine learning model after training facing the model demand terminal; adding the amount, the cost, and the market price to obtain the price corresponding to the machine learning model after training.

9. A data contribution incentive device for supporting the training of a machine learning model, wherein, the device includes the following components: a data acquisition module for obtaining training data sent by a data holding terminal; a training module for training the machine learning model according to the training data; a performance calculation module for calculating the model performance corresponding to the machine learning model after training; An incentive module, configured to assign a corresponding incentive value to the data holding terminal where the training data is located according to the model performance, and the incentive value is used as one of the bases for selecting the data holding terminal for the next training of the machine learning model.

10. A terminal device, characterized in that the terminal device includes a memory, a processor, and a data contribution incentive program for supporting the training of a machine learning model, which is stored in the memory and can run on the processor. When the processor executes the data contribution incentive program for supporting the training of the machine learning model, the steps of the data contribution incentive method for supporting the training of the machine learning model according to any one of claims 1-8 are implemented.

11. A computer-readable storage medium, characterized in that the computer-readable storage medium stores a data contribution incentive program for supporting the training of a machine learning model. When the data contribution incentive program for supporting the training of the machine learning model is executed by a processor, the steps of the data contribution incentive method for supporting the training of the machine learning model according to any one of claims 1-8 are implemented.

Images