WO2022160578A1

WO2022160578A1 - State transition core optimization-based data processing method, apparatus and device, and medium

Info

Publication number: WO2022160578A1
Application number: PCT/CN2021/101998
Authority: WO
Inventors: 姜迪
Original assignee: 深圳前海微众银行股份有限公司
Priority date: 2021-01-27
Filing date: 2021-06-24
Publication date: 2022-08-04
Also published as: CN113158223A

Abstract

The present application discloses a state transition core optimization-based data processing method, apparatus and device, and a medium. Said method comprises: in the process of a first participant training local model parameters each time, dynamically determining a state sampling algorithm of preset local sample data according to resource attribute information of the first participant, to obtain identification state information of the preset local sample data, so as to determine combined state information of all preset local model parameters of the first participant; determining, according to the combined state information, target model parameters to be federalized; and performing federation training with each second participant on the basis of said target model parameters, so as to obtain a preset prediction model of the first participant.

Description

Data processing method, device, device and medium based on state transition kernel optimization

This application claims the priority of the Chinese patent application filed on January 27, 2021, the application number is 202110115051.9, and the title is "Data Processing Method, Apparatus, Equipment and Medium Based on State Transfer Kernel Optimization", which is hereby incorporated in its entirety. Reference.

technical field

The present application relates to the field of artificial intelligence technology of financial technology (Fintech), and in particular, to a data processing method, apparatus, device and medium based on state transition kernel optimization.

Background technique

With the continuous development of financial technology, especially Internet technology finance, more and more technologies are applied in the financial field, but the financial industry also puts forward higher requirements for technology, such as the financial industry's data processing based on state transfer kernel optimization There are also higher requirements.

At present, in the process of training models through machine learning, participants usually directly exchange data with other participants, and direct data exchange with other participants will violate the privacy of users and cause security risks. When modeling, the sample data often have different recognition states. For example, in the speech recognition process (speech recognition process is to recognize data frames as states, combine states as factors, and combine factors into words), data frames can be recognized. It is A state, B state, C state, etc. Among them, A state, B state, C state and other types have different recognition probabilities. In the process of determining the recognition probabilities of different states of sample data, related technologies often use a fixed method. The identification probability of different states of the sample data is determined in a fixed way, which leads to the problem of poor resource adaptation in the model training process.

technical problem

The main purpose of this application is to provide a data processing method, device, equipment and medium based on state transition kernel optimization, which aims to solve the problem of determining the identification probability of different states of sample data in a fixed manner in the related art, so that the resources in the model training process are suitable. The technical problems of poor compatibility and easy violation of user privacy.

technical solutions

In order to achieve the above object, the present application provides a data processing method based on state transition core optimization, which is applied to a first participant, and the first participant and the second participant are connected by federated communication, and the state transition core optimization is based on the state transition core. The data processing methods include:

In the process of each training of local model parameters by the first participant, the state sampling algorithm of the preset local sample data is dynamically determined according to the resource attribute information of the first participant, so as to obtain the identification state information of the preset local sample data, to determine the combined state information of all preset local model parameters of the first participant;

According to the combined state information, determine the target model parameters to be federated;

Based on the target model parameters to be federated, federation training is performed with each second participant to obtain a preset prediction model of the first participant.

The present application also provides a data processing apparatus based on state transition core optimization, which is applied to a first participant, and the first participant and the second participant are connected by federated communication, and the data processing apparatus based on state transition core optimization include:

The first determination module is configured to dynamically determine the state sampling algorithm of the preset local sample data according to the resource attribute information of the first participant in each process of training the local model parameters by the first participant, so as to obtain the preset local model. Identification status information of the sample data to determine the combined status information of all preset local model parameters of the first participant;

a second determination module, configured to determine the target model parameters to be federated according to the combined state information;

The federation module is configured to perform federated training with each second participant based on the target model parameters to be federated to obtain a preset prediction model of the first participant.

The present application also provides a data processing device optimized based on a state transition core, the data processing device optimized based on the state transition core is an entity device, and the data processing device optimized based on the state transition core includes: a memory, a processor, and a storage device. The program of the state transition core-optimized data processing method on the memory and executable on the processor, when the program of the state transition core-optimized data processing method is executed by the processor, can realize the following The steps of the above-mentioned data processing method based on state transition kernel optimization.

The present application also provides a medium on which a program for implementing the above-mentioned data processing method based on state transition kernel optimization is stored, and when the program of the data processing method based on state transition kernel optimization is executed by a processor, the above-mentioned program is implemented Steps of a data processing method based on state transition kernel optimization.

The present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the above-mentioned data processing method based on state transition kernel optimization.

beneficial effect

The present application provides a data processing method, device, equipment and medium based on state transition kernel optimization. In the related art, different participants directly exchange data, and determine the identification probabilities of different states of sample data in a fixed manner, so that the model Compared with the poor applicability of resources and violation of user privacy in the training process, the present application dynamically determines the state of the preset local sample data according to the resource attribute information of the first participant in the process of training the local model parameters by the first participant. a sampling algorithm to obtain identification state information of preset local sample data to determine combined state information of all preset local model parameters of the first participant; according to the combined state information, determine the target model parameters to be federated; based on the The target model parameters to be federated are federated with each second participant to obtain a preset prediction model of the first participant. In this application, since the first participant performs federated training with each of the second participants, the direct data interaction between different participants is avoided to cause user privacy and security risks. In addition, in this application, in the first participant In the process of training local model parameters each time, the state sampling algorithm of the preset local sample data is dynamically determined according to the resource attribute information of the first participant, so as to obtain the identification state information of the preset local sample data, so as to determine the first participant. The combined state information of all the preset local model parameters of the party, that is, the identification state information of dynamically determining the identification state of the local sample data based on the resource attribute information, so as to determine the combined state information of all the preset local model parameters of the first party, Instead of determining the identification state information of different states of the sample data, such as the identification probability, in a fixed way, it improves the resource adaptability in the model training process, and solves the problem of determining the identification probability of different states of the sample data in a fixed way in the related art. The technical problem of poor adaptability of resources in the model training process and easy violation of user privacy.

Description of drawings

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

In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the accompanying drawings required for describing the embodiments or related technologies will be briefly introduced below. Obviously, for those skilled in the art, On the premise of no creative labor, other drawings can also be obtained from these drawings.

1 is a schematic flowchart of a first embodiment of a data processing method based on state transition kernel optimization of the present application;

2 is a data processing method based on state transition kernel optimization of the present application, in the process of each training of local model parameters by a first participant, dynamically determining the state of preset local sample data according to resource attribute information of the first participant A sampling algorithm to obtain the identification state information of the preset local sample data, and a schematic flow chart of the refinement steps of the steps of determining the combined state information of all preset local model parameters of the first participant;

FIG. 3 is a schematic diagram of the device structure of the hardware operating environment involved in the solution of the embodiment of the present application.

The realization, functional features and advantages of the present application will be further described with reference to the accompanying drawings in conjunction with the embodiments.

Embodiments of the present invention

It should be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application.

An embodiment of the present application provides a data processing method based on state transition core optimization. In the first embodiment of the data processing method based on state transition core optimization of the present application, referring to FIG. 1 , it is applied to a first participant, and the first A participant is connected to the second participant for federated communication, and the data processing method based on state transition kernel optimization includes:

Step S10, in the process of training the local model parameters by the first participant each time, dynamically determine the state sampling algorithm of the preset local sample data according to the resource attribute information of the first participant, so as to obtain the identification of the preset local sample data state information to determine the combined state information of all preset local model parameters of the first participant;

Step S20, determining the target model parameters to be federated according to the combined state information;

Step S30, based on the target model parameters to be federated, perform federation training with each second participant to obtain a preset prediction model of the first participant.

Specific steps are as follows:

In this embodiment, it should be noted that the data processing method based on state transition core optimization can be applied to a data processing system based on state transition core optimization (especially, applied to the No. a participant), the data processing system based on state transition core optimization is subordinate to the data processing device based on state transition core optimization, and for the data processing system based on state transition core optimization, a second participant may also be built in, or To communicate with the second participant, it should be noted that the first participant and the second participant (which may both belong to the data processing system optimized based on the state transfer core) can directly perform a federated communication connection. In addition, the A federated communication connection may also be indirectly performed between the first participant and the second participant through a third participant.

It should be noted that before the first participant and the second participant are federated, the first participant needs to train its own model parameters locally. Specifically, for example, the first participant iteratively trains 500 times locally. After (training model parameters), conduct federated communication with the second party to obtain the aggregation parameters, then replace the local model parameters with the aggregation parameters as replacement update model parameters, and continue the next round of iterative training based on the replacement update model parameters , until the desired model is finally obtained.

In this embodiment, the data processing method based on state transition kernel optimization is applied in the process of the first participant training its own model parameters. In this process, it needs to be dynamically determined according to the resource attribute information of the first participant. Preset the state sampling algorithm of the local sample data, and then obtain the identification state information of the preset local sample data according to the state sampling algorithm, wherein the resource attribute information includes information such as computing resources, storage resources and transmission resources, and the preset local sample data The state sampling algorithms include sampling algorithms with replacement, sampling without replacement, federated Montepross-Hattings sampling, and optimized federated Montepross-Hattings sampling, among which, each state sampling algorithm can be pre-existing The first participant is local, or each state sampling algorithm is temporarily invoked or generated. According to the state sampling algorithm, the identification state information of the preset local sample data is obtained. Specifically, according to the state sampling algorithm, the preset local sample data is obtained. The recognition probability of each recognition state, and then according to the state sampling algorithm, the values of the model parameters corresponding to each recognition state of the preset local sample data are obtained. It should be noted that no matter what sampling algorithm is used, the recognition probability of each recognition state is consistent. , or within the preset error range, but different sampling algorithms, resource consumption and sampling rates are different, specifically, for example, in the process of locally training local model parameters by the first participant, a certain sample data exists State-B state-C state, through different sampling algorithms, it is possible that state A accounts for 70%, state B accounts for 20%, and state C accounts for 10%, but through the federal Montepros-Haiting Sampling algorithm, it is possible to obtain 70% of the resources in A state is M1 memory consumption, and through the optimized federated Montepros-Hattings sampling algorithm, it is possible to obtain 70% of the resources in A state is M2 memory consumption, while Through the sampling with replacement, it is possible to obtain the resource that the A state accounts for 70% of the M3 memory consumption, of which the M2 memory consumption is the least.

It should be noted that the sample data is data composed of multiple sample features, and the state sampling algorithm for dynamically determining the preset local sample data according to the resource attribute information of the first participant includes: according to the resource attribute information of the first participant The attribute information dynamically determines the state sampling algorithm of different sample features in the preset local sample data, that is, for the corresponding sample features, the corresponding pre-stored state sampling algorithms are different, and then the recognition probability of different states corresponding to the sample features is obtained. Therefore, the probability of the overall sample data corresponding to the output data can be obtained by multiplying the recognition probabilities of different sample data, and in order to obtain the resource consumption of the probability of the output data, the recognition probability of the different states of each sample feature can be obtained first. Calculate the consumed computing power resources, etc., and then add the corresponding computing power resources to obtain the required overall resources. In the related art, the fixed state sampling algorithm is used to determine the identification probability of the state of the sample feature in the corresponding sample data, and the fixed state sampling algorithm is used to determine the identification probability of the state of the sample feature in the corresponding sample data, which is difficult to take into account the resources during model training. Configuration, specifically, for example, if the resource configuration of the first participant is small, the fixed state sampling algorithm consumes a lot, which will cause the server to crash, making it difficult to meet the resource limitation requirements during model training. If the resource configuration of the first participant is long, The fixed state sampling algorithm consumes less resources, which makes it difficult to meet the speed requirement during model training. In this embodiment, the balance or adaptation of resource consumption attributes and speed during model training can be achieved.

It should be noted that, due to the existence of different sample features in the preset local sample data, each sample feature corresponds to a different state sampling algorithm, and different sample features of the preset local sample data are obtained through each sample feature corresponding to a different state sampling algorithm. After the identification state information of the preset local sample data is obtained, the identification state information of the preset local sample data is obtained, and further, the combined state information of all preset local model parameters of the first participant is determined, that is, the identification state information of the preset local sample data implicitly Corresponding to the combined state information of the preset local model parameters, and further, the combined state information of all the preset local model parameters of the first participant can be obtained. Specifically, for example, it is assumed that the sample features have three states: A state, B state, and C state , A state, B state, and C state, the recognition probabilities corresponding to the three states are 70%, 20%, and 10% respectively. When the sample feature is recognized as A state, the corresponding recognition model parameters such as weight can be 0.7.

Wherein, referring to FIG. 2 , in the process of training the local model parameters by the first participant each time, the state sampling algorithm of the preset local sample data is dynamically determined according to the resource attribute information of the first participant, so as to obtain the preset state sampling algorithm. The steps of identifying state information of local sample data to determine the combined state information of all preset local model parameters of the first participant include:

Step S11, in each process of training the local model parameters by the first participant, determining the upper limit of memory consumption according to the resource attribute information of the first participant;

In this embodiment, each time the first participant trains local model parameters, the upper limit of memory consumption is determined according to the resource attribute information of the first participant, that is, the memory capacity of the server of the first participant is read. Determine the memory consumption limit.

Step S12, dynamically determine the state sampling algorithm of the preset local sample data according to the upper limit of memory consumption and the preset sampling consumption calculation rule, so as to obtain the identification state information of the preset local sample data, so as to determine each preset sample data of the first participant. The state information of the local model parameters is set to determine the combined state information of all preset local model parameters of the first participant.

According to the upper limit of memory consumption, such as 500G, and the preset sampling consumption calculation rules, such as the consumption calculation rules of each sampling time and the consumption calculation rules of different state types, the state sampling algorithm of the preset local sample data is dynamically determined, so as to obtain the preset local sample data. The identification state information of the sample data to determine the state information of each preset local model parameter of the first participant to determine the combined state information of all preset local model parameters of the first participant. It should be noted that different state sampling algorithms The resource consumption of obtaining the identification state information of the preset local sample data is different. There is a first preset association relationship between different state sampling algorithms and resource consumption, or there is a preset second association between different state sampling algorithms, state types, etc. and resource consumption. relation.

Wherein, the state sampling algorithm of the preset local sample data is dynamically determined according to the upper limit of memory consumption and the preset sampling consumption calculation rule, so as to obtain the identification state information of the preset local sample data, so as to determine each The steps of presetting state information of local model parameters to determine the combined state information of all preset local model parameters of the first participant include:

Step a1, respectively determining the sub-memory consumption upper limit of each preset local model parameter;

The sub-memory consumption upper limit of each preset local model parameter is determined respectively, and the method for determining the sub-memory consumption upper limit of each preset local model parameter includes:

Method 1: Determine the upper limit of the corresponding sub-memory consumption according to the type of each preset local model parameter;

Method 2: Determine the upper limit of the corresponding sub-memory consumption according to the weight of each preset local model parameter.

Step a2, determine the type and quantity of each preset local model parameter, according to the sub-memory consumption upper limit, the preset sampling consumption calculation rule and the state type and quantity, determine the preset local sample data by traversal. State sampling algorithm;

After determining the upper limit of sub-memory consumption, according to the upper limit of sub-memory consumption, the preset sampling consumption calculation rule and the state type and quantity, the state sampling algorithm of the preset local sample data is determined by traversal method, that is, from the pre-existing state sampling algorithm. The state sampling algorithms of the preset local sample data are correspondingly determined among the state sampling algorithms locally of the first participant.

Specifically, for example, if the sub-memory consumption upper limit is 100 consumption metering values, according to the preset sampling consumption calculation rule, the state type and quantity, traverse each state sampling algorithm (in order to save resources, the traversal process only calculates The actual sampling operation is not performed), and the upper limit of memory consumption corresponding to the specific process of each state sampling algorithm is obtained: 200 consumption metering value, 300 consumption metering value, 150 consumption metering value, 90 consumption metering value, then select 90 consumption metering value. State sampling algorithm for sampling.

Step a3, determine the minimum state transition route of each preset local model parameter corresponding to the state sampling algorithm;

Determining the minimum state transition route of each preset local model parameter corresponding to the state sampling algorithm, specifically, the minimum state transition route may refer to: determining the adopted route of the minimum number of state combinations, specifically, for example, there is A state, B state and D state, in the sampling process, A state and B state can be used as a group, D state as a group, the minimum number of state combinations is 2 groups, if there are A state, B state and D state, In the sampling process, the A state, the B state and the D state can be regarded as different groups, and the minimum number of state combinations is 3 groups. It should be noted that the consumption in the sampling process is different for different groups. Therefore, it is necessary to Under the condition that the resource consumption limit is satisfied, the minimum state transition route of each preset local model parameter corresponding to the state sampling algorithm is determined.

In step a4, according to the state sampling algorithm and the minimum state transition route, the identification state of the preset local sample data is obtained, so as to determine the combined state information of all preset local model parameters of the first participant.

According to the state sampling algorithm and each minimum state transition route, the identification state of the preset local sample data is combined to determine the combined state information of all preset local model parameters of the first participant.

In this embodiment, the target model parameters to be federated are determined according to the combined state information. For example, it is assumed that the sample features have three states: A state, B state, and C state, A state, B state, and C state. The corresponding recognition probabilities are 70%, 20%, and 10%, respectively. When the sample feature is recognized as state A, the corresponding recognition model parameters such as weight can be 0.7. After obtaining the model parameters, based on the target model parameters to be federated, perform federation training with each second participant to obtain a preset prediction model of the first participant.

The present application provides a data processing method, device, equipment and medium based on state transition kernel optimization. In the related art, different participants directly exchange data, and determine the identification probabilities of different states of sample data in a fixed manner, so that the model Compared with the poor applicability of resources and violation of user privacy in the training process, the present application dynamically determines the state of the preset local sample data according to the resource attribute information of the first participant in the process of training the local model parameters by the first participant. a sampling algorithm to obtain identification state information of preset local sample data to determine combined state information of all preset local model parameters of the first participant; according to the combined state information, determine the target model parameters to be federated; based on the The target model parameters to be federated are federated with each second participant to obtain a preset prediction model of the first participant. In this application, since the first participant performs federated training with each of the second participants, the direct data interaction between different participants is avoided to cause user privacy and security risks. In addition, in this application, in the first participant In the process of training local model parameters each time, the state sampling algorithm of the preset local sample data is dynamically determined according to the resource attribute information of the first participant, so as to obtain the identification state information of the preset local sample data, so as to determine the first participant. The combined state information of all preset local model parameters of the party, that is, the identification state of the local sample data is dynamically determined based on the resource attribute information, so as to determine the combined state information of all the preset local model parameters of the first party, rather than fixed The recognition probability of different states of the sample data is determined by the method, so the adaptability of resources in the model training process is improved, and the recognition probability of different states of the sample data is determined in a fixed way in the related art, resulting in poor resource adaptability in the model training process. , and technical problems that are prone to infringe on user privacy.

Further, based on the first embodiment in this application, another embodiment is provided. In this embodiment, the step of respectively determining the sub-memory consumption upper limit of each preset local model parameter includes:

Step A1, determining the degree of influence of each preset local model parameter on the model training result;

Ways to determine the degree of influence of each preset local model parameter on the model training result include:

Determine the weight of each preset local model parameter to determine the degree of influence on the model training result, or determine the size of the influence factor of each preset local model parameter to determine the degree of influence on the model training result.

Step A2: Determine the sub-memory consumption upper limit of each preset local model parameter according to the influence degree.

According to the influence degree, the upper limit of sub-memory consumption of each preset local model parameter is determined, wherein, if the degree of influence is large, it is determined that the upper limit of sub-memory consumption of each preset local model parameter is high, wherein the degree of influence can be determined by the size of the influence factor .

In this embodiment, the influence degree of each preset local model parameter on the model training result is determined; according to the influence degree, the sub-memory consumption upper limit of each preset local model parameter is determined. In this embodiment, the sub-memory consumption upper limit of each preset local model parameter is accurately determined.

Further, based on the first embodiment in this application, another embodiment is provided. In this embodiment, specifically, in the process of dynamically determining the state sampling algorithm of the preset local sample data, the upper limit of memory consumption is also determined according to the memory consumption limit. And the preset sampling consumption calculation rule determines the intermediate sampling parameters that need to be saved.

In this embodiment, when memory resources are abundant, memory resources can be used for efficiency. Specifically, the sampling intermediate parameters that need to be saved are determined according to the memory consumption upper limit and the preset sampling consumption calculation rule. For example, since the sample data includes various Sample feature Q1 feature, Q2 feature, Q3 feature, each sample feature has different states such as A state, B state, C state of Q1 feature, D state of Q2 feature, E state, F state of Q3 feature, G state and H state, in the process of obtaining combined state information based on each sample feature corresponding to each state, such as Q1 feature-A state, Q2 feature D state and Q3 feature F state combination, or Q1 feature-B state, Q2 feature D state and The Q3 feature F state combination, etc., in the related art, are randomly combined, but in this embodiment, in order to improve the combination efficiency, an orderly combination is performed. Specifically, during the combination process, the Q1 feature-A can be saved. state (occupying a certain amount of memory, whether it can be determined according to the upper limit of memory consumption and the preset sampling consumption calculation rule), and then, first combine the Q1 feature-A state and each state of the Q2 feature, and each state of the Q3 feature, and then combine the Q1 feature-A state and each state of the Q2 feature. The feature-B state is combined with each state of the Q2 feature and each state of the Q3 feature, thereby improving the efficiency of obtaining the output data. Specifically, in this embodiment, an alias table (AliasTable) may also be set for different features, and then states are combined based on each alias table to improve efficiency.

Further, based on the first embodiment and the second embodiment in the present application, the first model parameter to be federated is used to perform federation training with each second participant to obtain the preset of the first participant. The steps of the predictive model include:

Step B1, based on the first model parameters to be federated, by executing a preset federation process, aggregate with the second model parameters to be federated of each second participant to obtain aggregated parameters, which are treated based on the aggregated parameters. The first model parameters of the federation are replaced and updated to obtain the replaced and updated model parameters of the first participant;

In this embodiment, based on the first model parameters to be federated, by executing a preset federation process, it is directly aggregated with the second model parameters to be federated of each second participant, so as to obtain aggregated parameters, which are based on the predetermined federation process. The aggregation parameters are replaced and updated to the first model parameters of the federation to obtain the replaced and updated model parameters of the first participant.

Step B2, continue to dynamically determine the state sampling algorithm of the replaced and updated model parameters, so as to continue to determine other model parameters of the first participant to be federated, and continue to perform iterative training until the preset training completion condition is reached, Get a preset predictive model.

As with the above-mentioned state sampling algorithm for determining model parameters, continue to dynamically determine the state sampling algorithm for the replaced and updated model parameters, so as to continue to determine other model parameters of the first participant to be federated in the next round to replace the updated model. parameters, and continue to perform iterative training until a preset training completion condition is reached, such as the preset loss function convergence, and a preset prediction model is obtained.

the first participant is connected to the second participant for federal communication through a third party;

Based on the first model parameters to be federated, by executing a preset federation process, aggregate with the second model parameters to be federated of each second participant to obtain aggregated parameters, so as to treat federation based on the aggregated parameters The steps of replacing and updating the first model parameters of the first participant, and obtaining the replaced and updated model parameters of the first participant, include:

Step C1, encrypting the first model parameters to be federated and sent to a third party for the third party to use based on the first model parameters to be federated and the received data of each second participant to be federated. The second model parameters are aggregated to obtain aggregated parameters;

The first model parameters to be federated are encrypted and sent to a third party to avoid leakage of model parameters for the third party to use based on the first model parameters to be federated and the received data of each second participant. The second model parameters to be federated are aggregated to obtain aggregated parameters, and the first model parameters to be federated are encrypted and sent to a third party for the third party to receive based on the first model parameters to be federated. The obtained second model parameters of each second participant to be federated are aggregated to obtain aggregated parameters.

Step C2: Receive the aggregation parameters encrypted and sent by the third party, replace and update the first model parameters of the federation based on the aggregation parameters, and obtain the replaced and updated model parameters of the first participant.

Receive the aggregation parameter encrypted and sent by the third party, and replace and update the first model parameter of the federation based on the aggregation parameter to obtain the replaced and updated model parameter of the first participant.

In this embodiment, the preset prediction model is accurately obtained through the federated model.

Referring to FIG. 3 , FIG. 3 is a schematic diagram of the device structure of the hardware operating environment involved in the solution of the embodiment of the present application.

As shown in FIG. 3 , the data processing device based on state transition core optimization may include: a processor 1001 , such as a CPU, a memory 1005 , and a communication bus 1002 . Among them, the communication bus 1002 is used to realize the connection communication between the processor 1001 and the memory 1005 . The memory 1005 may be a high-speed RAM memory or a non-volatile memory (non-volatile memory). memory), such as disk storage. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001 .

Optionally, the data processing device based on state transition kernel optimization may also include a rectangular user interface, a network interface, a camera, an RF (Radio Frequency, radio frequency) circuits, sensors, audio circuits, WiFi modules, etc. The rectangular user interface may include a display screen (Display) and an input sub-module such as a keyboard (Keyboard), and the optional rectangular user interface may also include a standard wired interface and a wireless interface. Optional network interfaces may include standard wired interfaces and wireless interfaces (such as WI-FI interfaces).

Those skilled in the art can understand that the structure of the data processing device optimized based on the state transition core shown in FIG. 3 does not constitute a limitation on the data processing device optimized based on the state transition core, and may include more or less than shown components, or a combination of certain components, or a different arrangement of components.

As shown in FIG. 3 , the memory 1005 as a medium may include an operating system, a network communication module, and a data processing program optimized based on the state transition kernel. The operating system is a program that manages and controls the hardware and software resources of the data processing device optimized based on the state transition core, and supports the operation of the data processing program optimized based on the state transition core and other software and/or programs. The network communication module is used to realize the communication between various components in the memory 1005, as well as the communication with other hardware and software in the data processing system optimized based on the state transition core.

In the data processing device based on state transition core optimization shown in FIG. 3 , the processor 1001 is configured to execute the data processing program based on state transition core optimization stored in the memory 1005 to implement any one of the above state transition core based data processing programs Steps for an optimized data processing method.

The specific implementation manner of the data processing device optimized based on the state transition core of the present application is basically the same as the above-mentioned embodiments of the data processing method based on the optimization of the state transition core, and will not be repeated here.

The present application also provides a data processing apparatus based on state transition core optimization, which is applied to a first participant, and the first participant is connected to a second participant for federated communication, and the data processing apparatus based on state transition core optimization include:

Optionally, the first determining module includes:

a first determining unit, configured to determine the upper limit of memory consumption according to the resource attribute information of the first participant during each process of training the local model parameters by the first participant;

The second determination unit is configured to dynamically determine the state sampling algorithm of the preset local sample data according to the upper limit of memory consumption and the preset sampling consumption calculation rule, so as to obtain the identification state information of the preset local sample data, so as to determine the first participation state information of each preset local model parameter of the first party to determine the combined state information of all preset local model parameters of the first participant.

Optionally, the second determining unit includes:

a first determining subunit, configured to respectively determine the sub-memory consumption upper limit of each preset local model parameter;

The second determination subunit is used to determine the type and quantity of each preset local model parameter, and according to the sub-memory consumption upper limit, the preset sampling consumption calculation rule, and the state type and quantity, determine the preset by traversal. Set the state sampling algorithm of local sample data;

a third determination subunit, configured to determine the minimum state transition route of each preset local model parameter corresponding to the state sampling algorithm;

The fourth determination subunit is configured to obtain the identification state of the preset local sample data according to the state sampling algorithm and the minimum state transition route, so as to determine the combined state information of all preset local model parameters of the first participant.

Optionally, the first determination subunit is used to implement:

Determine the degree of influence of each preset local model parameter on the model training result;

According to the influence degree, the upper limit of sub-memory consumption of each preset local model parameter is determined.

Optionally, in the process of dynamically determining the state sampling algorithm of the preset local sample data, the intermediate sampling parameters that need to be saved are also determined according to the upper limit of memory consumption and the preset sampling consumption calculation rule.

Optionally, the federation module includes:

an aggregation unit, configured to perform aggregation with the second model parameters to be federated of each second participant based on the first model parameters to be federated, by executing a preset federation process, so as to obtain aggregation parameters to be based on the aggregation The parameters are replaced and updated by the first model parameters of the federation to obtain the replaced and updated model parameters of the first participant;

The third determination unit is configured to continue to dynamically determine the state sampling algorithm of the replaced and updated model parameters, so as to continue to determine other model parameters of the first participant to be federated, and to continuously perform iterative training until the preset value is reached. After the training is completed, the preset prediction model is obtained.

Optionally, the first participant is connected with the second participant through a third party for federal communication;

The third determining unit includes:

The sending unit is configured to encrypt and send the first model parameters to be federated to a third party, so that the third party can use the first model parameters to be federated and the received data from each second participant to be sent to a third party. The second model parameters of the federation are aggregated to obtain aggregated parameters;

A receiving unit, configured to receive the aggregation parameters encrypted and sent by the third party, and to replace and update the first model parameters of the federation based on the aggregation parameters to obtain the replaced and updated model parameters of the first participant.

The specific implementations of the data processing apparatus based on state transition kernel optimization of the present application are basically the same as the above-mentioned embodiments of the data processing method based on state transition kernel optimization, and will not be repeated here.

An embodiment of the present application provides a medium, and the medium stores one or more programs, and the one or more programs can also be executed by one or more processors to implement any one of the above The steps of a data processing method based on state transition kernel optimization.

The specific implementation manner of the medium of the present application is basically the same as the above-mentioned embodiments of the data processing method based on the optimization of the state transition kernel, and details are not repeated here.

The specific implementation manner of the computer program product of the present application is basically the same as the above-mentioned embodiments of the data processing method based on the optimization of the state transition kernel, and will not be repeated here.

It should be noted that, herein, the terms "comprising", "comprising" or any other variation thereof are intended to encompass non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements, It also includes other elements not expressly listed or inherent to such a process, method, article or apparatus. Without further limitation, an element qualified by the phrase "comprising a..." does not preclude the presence of additional identical elements in a process, method, article or apparatus that includes the element.

The above-mentioned serial numbers of the embodiments of the present invention are only for description, and do not represent the advantages or disadvantages of the embodiments.

From the description of the above embodiments, those skilled in the art can clearly understand that the method of the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is better implementation. Based on such understanding, the technical solutions of the present invention can be embodied in the form of software products in essence or the parts that make contributions to the prior art, and the computer software products are stored in a medium (such as ROM/RAM, magnetic disk, optical disk) ), including several instructions to enable a terminal device (which may be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in the various embodiments of the present invention.

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

Claims

A data processing method based on state transition kernel optimization, wherein, applied to a first participant, the first participant is connected with a second participant for federated communication, and the data processing method based on state transition kernel optimization includes:

In the process of each training of local model parameters by the first participant, the state sampling algorithm of the preset local sample data is dynamically determined according to the resource attribute information of the first participant, so as to obtain the identification state information of the preset local sample data, to determine the combined state information of all preset local model parameters of the first participant;

According to the combined state information, determine the target model parameters to be federated;

Based on the target model parameters to be federated, federation training is performed with each second participant to obtain a preset prediction model of the first participant.
The data processing method based on state transition kernel optimization according to claim 1, wherein, in the process of each training of local model parameters by the first participant, the predetermined prediction is dynamically determined according to the resource attribute information of the first participant. The steps of setting the state sampling algorithm of the local sample data to obtain the identification state information of the preset local sample data to determine the combined state information of all preset local model parameters of the first participant include:

In each process of training the local model parameters by the first participant, determining the upper limit of memory consumption according to the resource attribute information of the first participant;

According to the upper limit of memory consumption and the preset sampling consumption calculation rule, the state sampling algorithm of the preset local sample data is dynamically determined, so as to obtain the identification state information of the preset local sample data, so as to determine each preset local model of the first participant The state information of the parameters is used to determine the combined state information of all preset local model parameters of the first participant.
The data processing method based on state transition core optimization according to claim 2, wherein the state sampling algorithm of the preset local sample data is dynamically determined according to the upper limit of memory consumption and the preset sampling consumption calculation rule, so as to obtain the preset sampling algorithm. The steps of setting the identification state information of the local sample data to determine the state information of each preset local model parameter of the first participant to determine the combined state information of all the preset local model parameters of the first participant include:

Determine the sub-memory consumption upper limit of each preset local model parameter separately;

Determine the type and quantity of each preset local model parameter, and determine the state sampling algorithm of the preset local sample data by traversing according to the sub-memory consumption upper limit, the preset sampling consumption calculation rule, and the state type and quantity ;

Determine the minimum state transition route of each preset local model parameter corresponding to the state sampling algorithm;

According to the state sampling algorithm and the minimum state transition route, the identification state of the preset local sample data is obtained, so as to determine the combined state information of all preset local model parameters of the first participant.
The data processing method based on state transition kernel optimization according to claim 3, wherein the step of respectively determining the sub-memory consumption upper limit of each preset local model parameter comprises:

Determine the degree of influence of each preset local model parameter on the model training result;

According to the influence degree, the upper limit of sub-memory consumption of each preset local model parameter is determined.
The data processing method based on state transition kernel optimization according to claim 4, wherein the manner of determining the degree of influence of each preset local model parameter on the model training result comprises:

Determine the weight of each preset local model parameter to determine the degree of influence on the model training result, or determine the size of the influence factor of each preset local model parameter to determine the degree of influence on the model training result.
The data processing method based on state transition kernel optimization according to claim 2, wherein, in the process of dynamically determining the state sampling algorithm of the preset local sample data, it is further determined according to the upper limit of memory consumption and the calculation rule of preset sampling consumption Sample intermediate parameters that need to be saved.
The data processing method based on state transition kernel optimization according to claim 1, wherein the federation training is performed with each second participant based on the first model parameters to be federated to obtain the data of the first participant. The steps to preset a predictive model include:

Based on the first model parameters to be federated, by executing a preset federation process, aggregate with the second model parameters to be federated of each second participant to obtain aggregation parameters, so as to treat the federated first model parameters based on the aggregation parameters A model parameter is replaced and updated to obtain the replaced and updated model parameter of the first participant;

Continue to dynamically determine the state sampling algorithm of the replaced and updated model parameters to continue to determine other model parameters of the first participant to be federated, and continue to perform iterative training until the preset training completion condition is reached, and a preset prediction model.
The data processing method based on state transition kernel optimization according to claim 7, wherein the first participant is connected with the second participant through a third party for federated communication;

Based on the first model parameters to be federated, by executing a preset federation process, aggregate with the second model parameters to be federated of each second participant to obtain aggregated parameters, so as to treat federation based on the aggregated parameters The steps of replacing and updating the first model parameters of the first participant, and obtaining the replaced and updated model parameters of the first participant, include:

encrypting the first model parameters to be federated and sent to a third party for the third party to use the first model parameters to be federated and the received second models of each second participant to be federated The parameters are aggregated to obtain the aggregated parameters;

Receive the aggregation parameter encrypted and sent by the third party, and replace and update the first model parameter of the federation based on the aggregation parameter to obtain the replaced and updated model parameter of the first participant.
The data processing method based on state transition core optimization according to any one of claims 1 to 8, wherein the resource attribute information includes computing resources, storage resources and transmission resources, and the state sampling of the preset local sample data The algorithms include replacement sampling algorithm, non-replacement sampling algorithm, federated Montepross-Hattings sampling algorithm, and optimized federated Montepross-Hattings sampling algorithm.
The data processing method based on state transition kernel optimization according to any one of claims 1 to 8, wherein the state sampling algorithm for dynamically determining preset local sample data according to the resource attribute information of the first participant comprises:

The state sampling algorithm for different sample features in the preset local sample data is dynamically determined according to the resource attribute information of the first participant.
A data processing apparatus based on state transition core optimization, wherein, applied to a first participant, the first participant is connected with a second participant for federated communication, and the data processing apparatus based on state transition core optimization includes:

The first determination module is configured to dynamically determine the state sampling algorithm of the preset local sample data according to the resource attribute information of the first participant in each process of training the local model parameters by the first participant, so as to obtain the preset local model. Identification status information of the sample data to determine the combined status information of all preset local model parameters of the first participant;

a second determination module, configured to determine the target model parameters to be federated according to the combined state information;

The federation module is configured to perform federated training with each second participant based on the target model parameters to be federated to obtain a preset prediction model of the first participant.
A data processing device based on state transition core optimization, wherein the data processing device based on state transition core optimization includes: a memory, a processor, and data processing stored on the memory for implementing the state transition core optimization method procedure,

The memory is used for storing a program for realizing the data processing method optimized based on the state transition kernel;

The processor is configured to execute a program for implementing the data processing method optimized based on the state transition core, so as to realize the following steps:

In each process of training local model parameters by the first participant, the state sampling algorithm of the preset local sample data is dynamically determined according to the resource attribute information of the first participant, so as to obtain the identification state information of the preset local sample data, to determine the combined state information of all preset local model parameters of the first participant;

According to the combined state information, determine the target model parameters to be federated;

Based on the target model parameters to be federated, federation training is performed with each second participant to obtain a preset prediction model of the first participant.
The data processing device based on state transition core optimization according to claim 12, wherein the processor is configured to execute a program for implementing the data processing method based on state transition core optimization, so as to realize the following steps:

In each process of training the local model parameters by the first participant, determining the upper limit of memory consumption according to the resource attribute information of the first participant;

According to the upper limit of memory consumption and the preset sampling consumption calculation rule, the state sampling algorithm of the preset local sample data is dynamically determined, so as to obtain the identification state information of the preset local sample data, so as to determine each preset local model of the first participant The state information of the parameters is used to determine the combined state information of all preset local model parameters of the first participant.
The data processing device based on state transition core optimization according to claim 13, wherein the processor is configured to execute a program for implementing the data processing method based on state transition core optimization, so as to realize the following steps:

Determine the sub-memory consumption upper limit of each preset local model parameter separately;

Determine the type and quantity of each preset local model parameter, and determine the state sampling algorithm of the preset local sample data by traversing according to the sub-memory consumption upper limit, the preset sampling consumption calculation rule, and the state type and quantity ;

Determine the minimum state transition route of each preset local model parameter corresponding to the state sampling algorithm;

According to the state sampling algorithm and the minimum state transition route, the identification state of the preset local sample data is obtained, so as to determine the combined state information of all preset local model parameters of the first participant.
The data processing device based on state transition core optimization according to claim 14, wherein the processor is configured to execute a program for implementing the data processing method based on state transition core optimization, so as to realize the following steps:

Determine the degree of influence of each preset local model parameter on the model training result;

According to the influence degree, the upper limit of sub-memory consumption of each preset local model parameter is determined.
The data processing device based on state transition core optimization according to claim 13, wherein, in the process of dynamically determining the state sampling algorithm of the preset local sample data, it is further determined according to the upper limit of memory consumption and the calculation rule of preset sampling consumption Sample intermediate parameters that need to be saved.
The data processing device based on state transition core optimization according to claim 12, wherein the processor is configured to execute a program for implementing the data processing method based on state transition core optimization, so as to realize the following steps:

Based on the first model parameters to be federated, by executing a preset federation process, aggregate with the second model parameters to be federated of each second participant to obtain aggregation parameters, so as to treat the federated first model parameters based on the aggregation parameters A model parameter is replaced and updated to obtain the replaced and updated model parameter of the first participant;

Continue to dynamically determine the state sampling algorithm of the replaced and updated model parameters to continue to determine other model parameters of the first participant to be federated, and continue to perform iterative training until the preset training completion condition is reached, and a preset prediction model.
The data processing device based on state transition core optimization of claim 17, wherein the first participant is connected to the second participant in federated communication through a third party;

Based on the first model parameters to be federated, by executing a preset federation process, aggregate with the second model parameters to be federated of each second participant to obtain aggregated parameters, so as to treat federation based on the aggregated parameters The steps of replacing and updating the first model parameters of the first participant, and obtaining the replaced and updated model parameters of the first participant, include:

encrypting the first model parameters to be federated and sent to a third party for the third party to use the first model parameters to be federated and the received second models of each second participant to be federated The parameters are aggregated to obtain the aggregated parameters;

Receive the aggregation parameter encrypted and sent by the third party, and replace and update the first model parameter of the federation based on the aggregation parameter to obtain the replaced and updated model parameter of the first participant.
A medium, wherein a program for realizing a data processing method optimized based on a state transition kernel is stored on the medium, and the program for realizing a data processing method optimized based on a state transition kernel is executed by a processor to realize the method as claimed in claim 1 to Steps of the data processing method based on state transition kernel optimization described in any one of 10.
A computer program product comprising a computer program, wherein the computer program, when executed by a processor, implements the method of any one of claims 1 to 10.