CN115935189A - Method, device, system and equipment for training trans-city federal migration model - Google Patents

Abstract

The invention discloses a method, a device, a system and equipment for training a trans-city federal migration model, which relate to the technical field of intelligent cities, and the method comprises the following steps: determining prediction index data and prediction city labels based on an initial cross-city federal migration model and training index data in an obtained training index data set; adjusting model parameters of the initial cross-city federal migration model based on standard index data corresponding to the prediction index data and standard city labels corresponding to the prediction city labels to obtain a trained target cross-city federal migration model; and sending the model parameters of the target cross-city federal migration model to the target city client, so that the target city client builds a standard cross-city federal migration model based on the received model parameters, and determines a target city label based on the standard cross-city federal migration model. The embodiment of the invention solves the problem that the target city cannot finish the city portrait task due to the missing of index data.

Description

Training method, device, system and equipment for cross-city federated migration model

technical field

The embodiments of the present invention relate to the technical field of smart cities, and in particular to a training method, device, system and equipment for a cross-city federation migration model.

Background technique

The combination of big data and urban development is an important development direction of cities in the future. The city portrait task refers to predicting city labels (such as population data, traffic data, etc.) Such as commercial popularity index, consumption index, etc.) to guide the future development direction of the city.

Training neural network models is an effective way to predict city labels. However, due to the different development levels and directions of different cities, some cities may have a small amount of indicator data, making them unsuitable for training neural network models. . To solve this problem, the existing cross-city migration method is to train the cross-city federated migration model in the source city with rich data resources, and migrate the cross-city federated migration model to the target city with scarce data resources to assist the target city. Complete the city portrait mission.

In the process of realizing the present invention, it is found that there are at least the following technical problems in the prior art:

In actual urban scenarios, not only will there be missing city labels in the target city, but some types of indicator data may also be missing. Therefore, the cross-city federated migration model migrating to the target city cannot complete the city portrait task in the absence of both city labels and missing indicator data.

Contents of the invention

Embodiments of the present invention provide a training method, device, system, and equipment for a cross-city federated migration model to solve the problem that the target city cannot complete the city portrait task due to lack of index data, and broaden the application scenarios of the cross-city migration method .

According to an embodiment of the present invention, a training method for a cross-city federation migration model is provided, which is applied to a source city client, and the method includes:

In response to obtaining the training indicator data set, based on the initial cross-city federation transfer model and the training indicator data in the training indicator data set, determine the predictive indicator data and the predicted city label;

Based on the standard index data corresponding to the predictive index data and the standard city label corresponding to the predicted city label, adjust the model parameters of the initial cross-city federated migration model to obtain a trained target cross-city federated migration model ;

Send the model parameters of the target cross-city federation migration model to the target city client, so that the target city client builds a standard cross-city federation migration model based on the received model parameters, and migrates across cities based on the standard The model determines the target city labels.

According to another embodiment of the present invention, a training method of a cross-city federation migration model is provided, which is applied to a target city client, and the method includes:

Receive the model parameters of the trained target cross-city federation migration model sent by the source city client, and build a standard cross-city federation migration model based on the model parameters;

Input the test indicator data into the standard cross-city federation migration model to obtain the output target city label;

Wherein, the target cross-city federated migration model is that the source city client determines the predictive index data and the predicted city label based on the initial cross-city federated migration model and the training index data in the training index data set, and based on the correlation with the predictive index The standard index data corresponding to the data and the standard city label corresponding to the predicted city label are obtained by performing a training operation on the initial cross-city federation transfer model.

According to another embodiment of the present invention, a training device for a cross-city federation migration model is provided, which is applied to a source city client, and the device includes:

A predictive city label determination module, configured to determine predictive index data and predictive city labels based on the initial cross-city federation transfer model and the training index data in the training index data set in response to obtaining the training index data set;

A target inter-city federated migration model determination module, configured to adjust the model parameters of the initial inter-city federated migration model based on the standard indicator data corresponding to the predicted indicator data and the standard city label corresponding to the predicted city label , to obtain the trained target inter-city federation migration model;

The model parameter sending module is used to send the model parameters of the target cross-city federation migration model to the target city client, so that the target city client builds a standard cross-city federation migration model based on the received model parameters, and based on The standard cross-city federated migration model determines the destination city label.

According to another embodiment of the present invention, a training device for a cross-city federation migration model is provided, which is applied to a target city client, and the device includes:

A standard cross-city federated migration model building module, used to receive the model parameters of the trained target cross-city federated migration model sent by the source city client, and build a standard cross-city federated migration model based on the model parameters;

The target city label determination module is used to input the test indicator data into the standard cross-city federation migration model to obtain the output target city label;

Wherein, the target cross-city federated migration model is that the source city client determines the predictive index data and the predicted city label based on the initial cross-city federated migration model and the training index data in the training index data set, and based on the correlation with the predictive index The standard index data corresponding to the data and the standard city label corresponding to the predicted city label are obtained by performing a training operation on the initial cross-city federation transfer model.

According to another embodiment of the present invention, a training system for cross-city federation migration model is provided, the system includes: a source city client and a target city client;

Wherein, the source city client is configured to determine predictive index data and predicted city labels based on the initial cross-city federation transfer model and the training index data in the training index data set in response to obtaining the training index data set, and based on The standard index data corresponding to the predictive index data and the standard city label corresponding to the predicted city label are adjusted to the model parameters of the initial cross-city federated migration model to obtain a trained target cross-city federated migration model, and send the model parameters of the target cross-city federation migration model to the target city client;

The target city client is configured to construct a standard cross-city federated migration model based on the received model parameters, and determine a target city label based on the standard cross-city federated migration model.

According to another embodiment of the present invention, a terminal device is provided, and the terminal device includes:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein,

The memory stores a computer program that can be executed by the at least one processor, and the computer program is executed by the at least one processor, so that the at least one processor can execute the method described in any embodiment of the present invention. A training method for federated transfer models across cities.

According to another embodiment of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement any of the embodiments of the present invention when executed. A training method for cross-city federated transfer models.

In the technical solution of the embodiment of the present invention, by responding to the acquisition of the training index data set, based on the initial cross-city federation transfer model and the training index data in the training index data set, the prediction index data and the prediction city label are determined, based on the corresponding prediction index data The standard index data and the standard city labels corresponding to the predicted city labels are used to adjust the model parameters of the initial cross-city federated migration model to obtain the trained target cross-city federated migration model, so that the trained target cross-city federated migration model can be simultaneously With the ability to predict index data and predict city labels, it solves the problem that the target city cannot complete the city portrait task due to lack of index data, and broadens the application scenarios of the cross-city migration method.

It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of the present invention, nor is it intended to limit the scope of the present invention. Other features of the present invention will be easily understood from the following description.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings that need to be used in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those skilled in the art, other drawings can also be obtained based on these drawings without creative effort.

Fig. 1 is the flowchart of the training method of a kind of cross-city federation migration model provided by one embodiment of the present invention;

Fig. 2 is a relationship diagram between different index data of different cities provided by an embodiment of the present invention;

Fig. 3 is the flowchart of the concrete example of the training method of a kind of cross-city federation migration model provided by one embodiment of the present invention;

FIG. 4 is a flow chart of another training method for a cross-city federation migration model provided by an embodiment of the present invention;

FIG. 5 is a flowchart of a training method for a cross-city federation migration model provided by an embodiment of the present invention;

FIG. 6 is a schematic structural diagram of a training device for a cross-city federation migration model provided by an embodiment of the present invention;

FIG. 7 is a schematic structural diagram of a training device for a cross-city federation migration model provided by an embodiment of the present invention;

FIG. 8 is a schematic structural diagram of a training system for a cross-city federation migration model provided by an embodiment of the present invention;

FIG. 9 is a system schematic diagram of a training system for a cross-city federation migration model provided by an embodiment of the present invention;

Fig. 10 is a schematic structural diagram of a terminal device provided by an embodiment of the present invention.

Detailed ways

In order to enable those skilled in the art to better understand the solutions of the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only It is an embodiment of a part of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts shall fall within the protection scope of the present invention.

It should be noted that the terms "first" and "second" in the description and claims of the present invention and the above drawings are used to distinguish similar objects, but not necessarily used to describe a specific sequence or sequence. It is to be understood that the data so used are interchangeable under appropriate circumstances such that the embodiments of the invention described herein can be practiced in sequences other than those illustrated or described herein. Furthermore, the terms "comprising" and "having", as well as any variations thereof, are intended to cover a non-exclusive inclusion, for example, a process, method, system, product or device comprising a sequence of steps or elements is not necessarily limited to the expressly listed instead, may include other steps or elements not explicitly listed or inherent to the process, method, product or apparatus.

Fig. 1 is a flow chart of a training method for a cross-city federated migration model provided by an embodiment of the present invention. This embodiment is applicable to the case of performing a city portrait task for a target city with missing index data. The training device of the city federation migration model can be implemented in the form of hardware and/or software, and the training device of the cross-city federation migration model can be configured in the source city client. Wherein, the source city client may represent a client corresponding to a city with abundant index data resources. As shown in Figure 1, the method includes:

S110. In response to obtaining the training index data set, determine predictive index data and predicted city labels based on the initial cross-city federation transfer model and the training index data in the training index data set.

Wherein, for example, the training index data can be used to represent the data generated in the development process of the source city, for example, the training index data can be population data, traffic data, tax data and so on.

Wherein, for example, the predicted city labels may be consumption index, traffic convenience degree, air quality index and so on. Among them, the consumption index can provide a reference basis for subsequent business district planning, the degree of traffic convenience can provide a reference basis for subsequent traffic scheduling, and the air quality index can provide a reference basis for the formulation and implementation of subsequent environmental governance plans.

In one embodiment, specifically, the training indicator data includes point-of-interest area data, road network data, and population data, wherein the point-of-interest area data includes the number of points of interest, the total number of points of interest, and the total number of points of interest corresponding to at least one type of point of interest. At least one of the entropy values of the points of interest, the road network data includes the number of roads corresponding to at least one road type, the population data includes the number of working population and/or the number of living population, correspondingly, the predictive index data is the number of consumption population, and the predicted City labels are categories for commercial popularity.

Among them, specifically, POI area data can be used to describe the functions or attributes of different areas in the city, especially, the number and type of POI (Point of Interest, POI) can reflect the popularity of an area. Among them, for example, the types of POIs include but are not limited to catering services, scenic spots, public facilities, shopping, transportation facilities services, financial insurance services, science, education and cultural services, business residences, life services, sports services, healthcare services, government Institutions and social groups, accommodation services and green parks, etc., there is no limitation on the types of data points here.

In one embodiment, specifically, the point-of-interest area data includes the number of points of interest d ^pf corresponding to each type of point of interest, the total number of points of interest d ^pn and the entropy value of points of interest d ^pe , wherein the point-of-interest entropy value d ^pe satisfies formula:

表示第i中兴趣点类型对应的兴趣点数量。其中，兴趣点熵值d^pe可反映一个区域的功能多样性。in,

Indicates the number of interest points corresponding to the type of interest point in i. Among them, the interest point entropy value d ^pe can reflect the functional diversity of a region.

Among them, concretely, the road network data can reflect the traffic convenience of the city. Wherein, for example, the road types include but not limited to expressways, expressways, living streets, forestry roads, competition roads, etc., and the number of roads corresponding to at least one road type can be represented by d ^rf .

Specifically, the population data may reflect the distribution of different population types in different regions of the city. Exemplarily, the population types include but are not limited to working population, resident population, and consumption population. In one embodiment, specifically, the population data includes the number of working population d ^wp and the number of resident population d ^rp .

In one embodiment, specifically, based on the initial cross-city federation migration model and the training indicator data in the training indicator dataset, determining the predictive indicator data and the predicted city label includes: inputting the training indicator data in the training indicator dataset into the initial In the initial data generation module and initial label generation module in the cross-city federated migration model; through the initial data generation module, based on the input training indicator data, output prediction indicator data; through the initial label generation module, based on the input training indicator data and the initial The predictive indicator data output by the data generation module outputs the predicted city label.

FIG. 2 is a relationship diagram between different index data of different cities provided by an embodiment of the present invention. Specifically, Figure 2 reflects the relationship between the normalized POI entropy value and the normalized consumer population for four different cities (city A, city B, city C, and city D respectively). . From Figure 2, it can be concluded that the relationship knowledge between different index data is similar among different cities. Due to the different development levels or development directions of different cities, there are usually relatively large differences between the same type of indicator data produced by different cities. However, there are similarities in relational knowledge between different types of index data in different cities (as shown in Figure 2). Therefore, training the initial data generation module in the initial cross-city federation transfer model based on the training index data of the source city can make the initial The data generation module learns the relationship knowledge between different index data, which can ensure the trainability of the initial data generation module and the accuracy of the output prediction index data.

In one embodiment, specifically, the initial data generation module includes a feature extractor, a data regressor, and a domain classifier. Correspondingly, based on the input training indicator data, the output prediction indicator data includes: inputting the training indicator data to the feature In the extractor, the output training indicator features are obtained; through the data regressor, based on the training indicator features output by the feature extractor, output prediction indicator data, and through the domain classifier, based on the training indicator features output by the feature extractor, output the predicted city classification results.

Specifically, the feature extractor can learn to extract the relationship features between the training index data and the prediction index data. Exemplarily, the feature extractor may contain two fully connected layers. The network architecture of the feature extractor is not limited here.

In one embodiment, specifically, the predictive indicator data is the consumption population. Specifically, the predictive index data output by the data regressor is the predicted consumption population. Wherein, for example, the data regressor may include two fully connected layers and one regression layer. The network architecture of the feature extractor is not limited here.

Specifically, only using the feature extractor and the data regressor may cause domain drift. To solve this problem, the initial data generation module in this embodiment includes a domain classifier. Wherein, for example, the domain classifier may include two fully connected layers and one classification layer. The network architecture of the domain classifier is not limited here.

Specifically, the predicted city classification result can be used to represent which source city client the training feature data comes from.

Among them, specific and accurate category predictions of commercial popularity can provide strong data support for the formulation and implementation of commercial district planning schemes. Exemplarily, the number of tags corresponding to the category of commercial popularity may be 5, which are respectively very high, high, medium, low and very low.

S120. Based on the standard index data corresponding to the predictive index data and the standard city labels corresponding to the predicted city labels, adjust the model parameters of the initial inter-city federated transfer model to obtain a trained target inter-city federated transfer model.

In this embodiment, specifically, the training index data set contains training index data corresponding to at least two source city clients respectively, and correspondingly, based on the standard index data corresponding to the predictive index data and the standard index data corresponding to the predicted city label City label, adjust the model parameters of the initial cross-city federated migration model to obtain the trained target cross-city federated migration model, including: for each iteration process of the initial cross-city federated migration model, based on the standards corresponding to the predictive index data The index data and the standard city labels corresponding to the predicted city labels determine the target loss function; based on the target loss function, determine the current model parameters corresponding to the initial cross-city federation migration model of the current iteration until the target loss function converges. The initial cross-city federated migration model is used as the target cross-city federated migration model after training.

Among them, specifically, the training index data corresponding to at least two source city clients are obtained, and at least two training index data are respectively input into the initial cross-city federation transfer model to obtain the output predictive index data and predicted city labels, based on The standard indicator data corresponding to the predicted indicator data and the standard city label corresponding to the predicted city label determine the target loss function, and based on the target loss function, determine the current model parameters corresponding to the initial cross-city federated migration model of the current iteration until the target loss When the function converges, the initial cross-city federated migration model of the current iteration is used as the target cross-city federated migration model after training. Wherein, at least two source city clients include the current source city client.

It should be noted that the acquisition, storage, and application of the training index data of the source cities involved in the embodiments of the present invention all comply with relevant laws and regulations.

S130. Send the model parameters of the target cross-city federation migration model to the target city client, so that the target city client builds a standard cross-city federation migration model based on the received model parameters, and determines the target city based on the standard cross-city federation migration model Label.

Specifically, the target city client can build a standard cross-city federated migration model based on the received model parameters and the model architecture of the initial cross-city federated migration model, and input the test indicator data corresponding to the training indicator data into the standard cross-city federated migration model. In the federated migration model, the standard data generation module in the standard cross-city federated migration model outputs missing indicator data based on the input test index data, and the standard label module in the standard cross-city federated migration model generates modules based on the input test index data and standard data Output the missing indicator data, and output the target city label. Wherein, the test index data may be data of the same index type as the training index data existing in the target city, for example, the test index data includes POI data, road network data and population data of the target city.

FIG. 3 is a flow chart of a specific example of a training method for a cross-city federated migration model provided by an embodiment of the present invention. Specifically, multiple source cities (city A and city B in FIG. 3 ) contain abundant index data resources, for example, the index data resources can be used to represent the index data of different regions in the city. The initial cross-city federation migration model is trained based on the training indicator data of multiple source cities. For example, the training indicator data can be tax data, monitoring data, and trajectory data. Specifically, the initial cross-city federated migration model includes an initial data generation module and an initial label generation module, wherein the initial data generation module can be used to learn the relationship knowledge between indicator data and generate predictive indicator data based on the learned relationship knowledge, the initial The label generation module generates predicted city labels based on the training indicator data and the prediction indicator data. Exemplarily, the predictive index data may be population data, and the predicted city label may be a category of commercial popularity, specifically, the predicted city label is very high, high, medium, low or very low. After the trained target cross-city federated migration model is obtained, the target cross-city federated migration model is migrated to the target city (city C in Figure 3) to obtain a standard cross-city federated migration model. City C inputs its own tax data, monitoring data and trajectory data into the standard cross-city federal migration model, and the standard data generation module in the standard cross-city federal migration model outputs the predicted population based on the input tax data, monitoring data and trajectory data Data, the standard label generation module in the standard cross-city federated migration model outputs the category of commercial popularity of city C based on the input tax data, monitoring data and trajectory data and predicted population data.

Specifically, the standard data generation module in the standard cross-city federated migration model includes a trained feature extractor and data regressor.

In the technical solution of this embodiment, by responding to the acquisition of the training index data set, based on the initial cross-city federation transfer model and the training index data in the training index data set, the prediction index data and the prediction city label are determined, based on the corresponding prediction index data The standard index data and the standard city labels corresponding to the predicted city labels are used to adjust the model parameters of the initial cross-city federated migration model to obtain the trained target cross-city federated migration model, so that the trained target cross-city federated migration model has both The ability to predict index data and predict city labels solves the problem that the target city cannot complete the city portrait task due to lack of index data, and broadens the application scenarios of the cross-city migration method.

Fig. 4 is a flow chart of another cross-city federation migration model training method provided by an embodiment of the present invention. The standard city label corresponding to the label, the model parameters of the initial cross-city federation migration model are adjusted, and the technical characteristics of the trained target cross-city federation migration model are further refined. As shown in Figure 4, the method includes:

S210. In response to obtaining the training indicator data set, determine predictive indicator data and predicted city labels based on the initial cross-city federation transfer model and the training indicator data in the training indicator data set.

In this embodiment, the training index data set includes the training index data corresponding to the current source city client.

S220. For each iterative process of the initial inter-city federated migration model, determine a target loss function based on the standard index data corresponding to the predictive index data and the standard city label corresponding to the predicted city label.

In one embodiment, specifically, the target loss function includes a first target loss function corresponding to the initial data generation module and a second target loss function corresponding to the initial label generation module, correspondingly, based on the standard corresponding to the predictive index data The index data and the standard city labels corresponding to the predicted city labels determine the target loss function, including: constructing the first loss function based on the predicted index data and the standard index data, and constructing the second loss function based on the predicted city classification results and the standard city classification results. Loss function; based on the first loss function and the second loss function, determine the first target loss function corresponding to the initial data generation module; based on the predicted city label and the standard city label, construct the second target loss function corresponding to the initial label generation module .

Among them, for example, the least square error algorithm can be used to construct the first loss function based on the predictive index data and the standard index data. Specifically, the first loss function L _DR satisfies the formula:

与第i个训练指标数据对应的预测指标数据，N表示训练指标数据的训练样本量。Among them, a _i represents the standard index data corresponding to the i-th training index data,

The predictive index data corresponding to the i-th training index data, N represents the training sample size of the training index data.

Among them, for example, the cross-entropy minimization algorithm can be used to construct the second loss function based on the predicted city classification results and the standard city classification results. Specifically, the second loss function L _DC satisfies the formula:

表示第j个源城市客户端对应的第i个训练指标数据所属的预测城市分类结果，|C|表示源城市客户端的数量。Among them, d _ij represents the standard city classification result to which the i-th training indicator data corresponding to the j-th source city client belongs,

Indicates the predicted city classification result to which the i-th training indicator data corresponding to the j-th source city client belongs, and |C| represents the number of source city clients.

On the basis of the above embodiments, specifically, the initial data generation module further includes a gradient inversion layer, the gradient inversion layer is set between the feature extractor and the domain classifier, and the gradient inversion layer is used in the initial data generation module In the backpropagation process, the function gradient of the second loss function input into the gradient inversion layer is multiplied by a preset negative number, and correspondingly, the coefficient of the second loss function in the first target loss function is a preset negative number.

Among them, specifically, the first objective loss function L ₁ satisfies the formula:

L ₁ =L _DR -λL _DC

The advantage of this setting is that since the training index features corresponding to the training index data of different cities are very different, the feature extractor should learn the common feature representation between the training index data of different cities, rather than distinguishing feature representation. For this reason, the training objective of the feature extractor should maximize the second loss function L _DC . The purpose of the feature inversion layer (GRL) is to have no effect on the training index features during the forward propagation of the initial data generation module, but in the process of backpropagation, the feature inversion layer will take the second loss function The function gradient is multiplied by a preset negative number. In this embodiment, by adopting a supervised confrontation strategy, the feature extractor can learn to extract common feature representations between training index data of different cities during the training process.

Wherein, for example, the minimum cross-entropy algorithm can be used to determine the second target loss function based on the predicted city label and the standard city label. Specifically, the second target loss function L ₂ satisfies the formula:

表示第j种预测城市标签对应的第i个训练指标数据对应的预测城市标签，|K|表示预测城市标签对应的标签数量。Among them, y _ij represents the standard city label corresponding to the i-th training index data corresponding to the j-th predicted city label,

Indicates the predicted city label corresponding to the i-th training index data corresponding to the j-th predicted city label, and |K| indicates the number of labels corresponding to the predicted city label.

S230. Based on the target loss function, determine the current model parameters corresponding to the initial cross-city federated migration model of the current iteration, and send the current model parameters to the central server, so that the central server can send the current model parameters based on the received at least two source city clients. , determine the aggregated model parameters, and send the aggregated model parameters to the clients of each source city respectively.

In this embodiment, at least two source city clients include the current source city client.

Specifically, each source city client trains the initial cross-city federation transfer model based on local training index data, and sends the current model parameters updated for each iteration to the central processing unit.

Wherein, for example, the central processing unit may determine the aggregated model parameters based on at least two current model parameters by using an averaging or weighted average aggregation algorithm.

S240. Use the aggregated model parameters received from the central server as the model parameters of the initial cross-city federated migration model of the current iteration, until the target loss function converges, use the initial cross-city federated migration model of the current iteration as the target cross-city federation model after training. Urban Federation Migration Model.

S250. Send the model parameters of the target cross-city federation migration model to the target city client, so that the target city client builds a standard cross-city federation migration model based on the received model parameters, and determines the target city based on the standard cross-city federation migration model Label.

In one embodiment, specifically, the model parameters of the target cross-city federation migration model are sent to the central server, so that the central server receives the model parameters of the target cross-city federation migration model respectively sent by at least two source city clients parameters, determine the target model parameters, and send the target model parameters to the target city client.

Wherein, for example, the central processor may use an average or a weighted average aggregation algorithm to determine the target model parameters based on the model parameters of the target cross-city federation migration models respectively sent by at least two source city clients.

Constrained by law or reality, the training index data corresponding to the clients in each source city may involve privacy protection issues and cannot be disclosed to the public. Therefore, there may be situations where it is not possible to train the initial inter-city federated transfer model using centralized modeling. In the technical solution of this embodiment, by using multiple source city clients to train their respective initial cross-city federation migration models at the same time, for each source city client, the initial cross-city federation migration model in the current source city client In each iteration, the target loss function is determined based on the standard indicator data corresponding to the predicted indicator data and the standard city label corresponding to the predicted city label. Model parameters, and send the current model parameters to the central server, and receive the aggregated model parameters sent by the central server after performing an aggregation operation on each current model parameter, and use the aggregated model parameters as the model parameters of the initial cross-city federation migration model of the current iteration, Until the target loss function converges, the initial cross-city federated migration model of the current iteration is used as the target cross-city federated migration model after training, which solves the privacy protection problem that exists when multiple source city clients interact with training index data, and satisfies It meets the demand for privacy protection of data in the process of cross-city migration.

Fig. 5 is a flow chart of a training method for a cross-city federated migration model provided by an embodiment of the present invention. This embodiment is applicable to the case of performing a city portrait task on a target city with missing index data. The training device of the city federation migration model can be implemented by the training device of the cross-city federation migration model. The training device of the cross-city federation migration model can be implemented in the form of hardware and/or software. The training device of the cross-city federation migration model can be configured in the target city client. Wherein, the target city client may represent a client corresponding to a city with scarce index data resources. As shown in Figure 5, the method includes:

S310. Receive the model parameters of the trained target cross-city federation migration model sent by the source city client, and build a standard cross-city federation migration model based on the model parameters.

In this embodiment, the target cross-city federated migration model is determined by the source city client based on the initial cross-city federated migration model and the training index data in the training index data set to determine the predictive index data and predicted city labels, and based on the corresponding prediction index data The standard indicator data and the standard city labels corresponding to the predicted city labels are obtained by performing the training operation on the initial cross-city federated transfer model.

Wherein, for example, the training index data can be used to represent the data generated in the development process of the source city, for example, the training index data can be population data, traffic data, tax data and so on.

In one embodiment, specifically, the training indicator data includes point-of-interest area data, road network data, and population data, wherein the point-of-interest area data includes the number of points of interest, the total number of points of interest, and the total number of points of interest corresponding to at least one type of point of interest. At least one of the entropy values of the points of interest, the road network data includes the number of roads corresponding to at least one road type, the population data includes the number of working population and/or the number of living population, correspondingly, the predictive index data is the number of consumption population, and the predicted City labels are categories for commercial popularity.

In one embodiment, specifically, the initial cross-city federated migration model includes an initial data generation module and an initial label generation module, wherein the initial data generation module is used to output the prediction index based on the training index data in the input training index data set The data, the initial label generation module, is used to output the predicted city label based on the training indicator data in the input training indicator data set and the prediction indicator data output by the initial data generation module.

In one embodiment, specifically, the model parameters include the data model parameters corresponding to the target data generation module in the target cross-city federation migration model and the label model parameters corresponding to the target label generation module. Correspondingly, based on the model parameters, the standard cross The city federation migration model, including: based on the data model parameters, constructing the standard data generation module in the standard cross-city federation migration model, and inputting the reference index data into the standard data generation module to obtain the output missing index data; based on the label model parameters , build a reference label generation module, and train the reference label generation module based on the reference index data and missing index data, and obtain the standard label generation module in the standard cross-city federation transfer model.

其中，缺失指标数据

满足如下关系：In one embodiment, specifically, it is assumed that the training index data includes the number of points of interest d ^pf corresponding to the types of points of interest of the source city client, the total number of points of interest d ^pn , the entropy value of points of interest d ^pe , and the number of points of interest corresponding to each type of road. The number of roads d ^rf , the number of working population d ^wp , and the number of resident populations d ^rp . Correspondingly, the reference index data includes the number of points of interest b ^pf , the total number of points of interest b ^pn , and the number of interest points corresponding to each type of point of interest in the target city client. The point entropy value b ^pe , the number of roads b ^rf corresponding to each road type, the number of working population b ^wp and the number of resident population b ^rp , input the reference index data X _t into the standard data generation module in the standard cross-city federal migration model , to get the missing indicator data for the output

Among them, the missing indicator data

Satisfy the following relationship:

Among them, Θ _RK represents the data model parameters.

输入到构建的参考标签生成模块中，得到输出的目标城市的预测城市标签

基于目标城市的预测城市标签和目标城市的标准城市标签，对参考标签生成模块进行训练，得到标准跨城市联邦迁移模型中的标准标签生成模块。其中，目标城市的预测城市标签

满足如下关系：Specifically, refer to the index data X _t and the missing index data

Input to the constructed reference label generation module to get the predicted city label of the output target city

Based on the predicted city labels of the target city and the standard city labels of the target city, the reference label generation module is trained to obtain the standard label generation module in the standard cross-city federated transfer model. Among them, the predicted city label of the target city

Satisfy the following relationship:

Among them, Θ _task represents the label model parameters.

Wherein, for example, the minimum cross-entropy algorithm can be used to construct the loss function corresponding to the reference label generation module. Specifically, the loss function L _t satisfies the formula:

表示第j种预测城市标签对应的第i个参考指标数据对应的标准城市标签，

表示第j种预测城市标签对应的第i个参考指标数据对应的预测城市标签，|K|表示预测城市标签对应的标签数量，M表示参考指标数据的数据量。in,

Indicates the standard city label corresponding to the i-th reference index data corresponding to the j-th predicted city label,

Indicates the predicted city label corresponding to the i-th reference indicator data corresponding to the j-th predicted city label, |K| indicates the number of labels corresponding to the predicted city label, and M indicates the data volume of the reference indicator data.

The advantage of this setting is that in the cross-city migration scenario, due to the lack of indicator data in the target city, and the city labels still have differences between different cities. Therefore, the standard data generation module constructed by the data model parameters can be directly frozen in the target city to generate the missing index data of the target city. Then, based on a small amount of standard label data in the target city, the reference label generation module constructed based on the label model parameters is fine-tuned and trained to obtain the standard label generation module of the target city. This further improves the accuracy of the target city label output by the standard inter-city federated migration model corresponding to the target city.

S320. Input the test index data into the standard inter-city federation migration model to obtain the output target city label.

Input the test indicator data corresponding to the training indicator data into the standard cross-city federated migration model, the standard data generation module in the standard cross-city federated migration model outputs missing indicator data based on the input test indicator data, and the standard cross-city federated migration model The standard label module of , based on the input test indicator data and the missing indicator data output by the standard data generation module, outputs the target city label. Wherein, the test index data may be data of the same index type as the training index data existing in the target city, for example, the test index data includes POI data, road network data and population data of the target city.

In the technical solution of this embodiment, the model parameters of the target cross-city federation migration model sent by the source city client are received, and based on the model parameters, a standard cross-city federation migration model is constructed; the test index data is input into the standard cross-city federation migration model. In the federated migration model, the output target city labels are obtained, where the target cross-city federated migration model is the source city client determines the predictive indicator data and predicted city labels based on the initial cross-city federated migration model and the training indicator data in the training indicator data set , and based on the standard index data corresponding to the predictive index data and the standard city labels corresponding to the predicted city labels to perform training operations on the initial cross-city federation transfer model, which solves the problem that the target city cannot complete the city portrait task due to lack of index data problem, broadening the application scenarios of the cross-city migration method.

FIG. 6 is a schematic structural diagram of a training device for a cross-city federated migration model provided by an embodiment of the present invention, which can be configured in a source city client. As shown in FIG. 6 , the device includes: a predicted city label determination module 410 , a target cross-city federation migration model determination module 420 and a model parameter sending module 430 .

Among them, the predicted city label determination module 410 is used to determine the predicted indicator data and the predicted city label based on the initial cross-city federation transfer model and the training indicator data in the training indicator dataset in response to the acquisition of the training indicator data set;

The target cross-city federation migration model determination module 420 is used to adjust the model parameters of the initial cross-city federation migration model based on the standard index data corresponding to the predictive index data and the standard city label corresponding to the predicted city label, and obtain the trained Target cross-city federated migration model;

The model parameter sending module 430 is used to send the model parameters of the target cross-city federation migration model to the target city client, so that the target city client builds a standard cross-city federation migration model based on the received model parameters, and based on the standard cross-city federation A federated migration model determines target city labels.

In the technical solution of this embodiment, by responding to the acquisition of the training index data set, based on the initial cross-city federation transfer model and the training index data in the training index data set, the prediction index data and the prediction city label are determined, based on the corresponding prediction index data The standard index data and the standard city labels corresponding to the predicted city labels are used to adjust the model parameters of the initial cross-city federated migration model to obtain the trained target cross-city federated migration model, so that the trained target cross-city federated migration model has both The ability to predict index data and predict city labels solves the problem that the target city cannot complete the city portrait task due to lack of index data, and broadens the application scenarios of the cross-city migration method.

On the basis of the above-mentioned embodiments, specifically, the predicted city label determination module 410 includes:

The training indicator data input unit is used to input the training indicator data in the training indicator data set into the initial data generation module and the initial label generation module in the initial cross-city federation migration model;

The predictive index data output unit is used to output the predictive index data based on the input training index data through the initial data generation module;

The predicted city label output unit is used to output the predicted city label through the initial label generation module based on the input training indicator data and the prediction indicator data output by the initial data generation module.

On the basis of the above embodiments, specifically, the initial data generation module includes a feature extractor, a data regressor, a domain classifier, and a predictive index data output unit, specifically for:

Input the training index data into the feature extractor to obtain the output training index features;

Through the data regressor, based on the training indicator features output by the feature extractor, the predictive indicator data is output, and through the domain classifier, based on the training indicator features output by the feature extractor, the predicted city classification results are output.

On the basis of the above-mentioned embodiments, specifically, the training index data set contains the training index data corresponding to the current source city client, and the target cross-city federation migration model determination module 420 includes:

The target loss function determination unit is used to determine the target loss function based on the standard indicator data corresponding to the predictive indicator data and the standard city label corresponding to the predicted city label for each iteration process of the initial cross-city federated migration model;

The current model parameter sending unit is used to determine the current model parameters corresponding to the initial cross-city federation migration model of the current iteration based on the target loss function, and send the current model parameters to the central server, so that the central server is based on the received at least two The current model parameters sent by the source city clients respectively, determine the aggregation model parameters, and send the aggregation model parameters to each source city client respectively; wherein, at least two source city clients include the current source city client;

The first target cross-city federated migration model determination unit is used to use the received aggregation model parameters sent by the central server as the model parameters of the initial cross-city federated migration model of the current iteration, until the target loss function converges, the initial The cross-city federated migration model is used as the target cross-city federated migration model after training.

On the basis of the above embodiments, specifically, the model parameter sending module 430 is specifically used for:

Send the model parameters of the target cross-city federation migration model to the central server, so that the central server determines the target model parameters based on the received model parameters of the target cross-city federation migration model sent by at least two source city clients, and sends The target model parameters are sent to the target city client.

On the basis of the above-mentioned embodiments, specifically, the training index data set contains training index data corresponding to at least two source city clients respectively, and the target inter-city federation migration model determination module 420 includes:

The target loss function determination unit is used to determine the target loss function based on the standard indicator data corresponding to the predictive indicator data and the standard city label corresponding to the predicted city label for each iteration process of the initial cross-city federated migration model;

The second target cross-city federated migration model determination unit is used to determine the current model parameters corresponding to the initial cross-city federated migration model of the current iteration based on the target loss function, until the target loss function converges, the initial cross-city federated migration of the current iteration The model is used as the target cross-city federation migration model after training.

On the basis of the above embodiments, specifically, the target loss function includes a first target loss function corresponding to the initial data generation module and a second target loss function corresponding to the initial label generation module, and the target loss function determination unit is specifically used for :

Construct a first loss function based on the predictive index data and standard index data, and construct a second loss function based on the predicted city classification results and standard city classification results;

Based on the first loss function and the second loss function, determine a first target loss function corresponding to the initial data generation module;

Based on the predicted city labels and standard city labels, a second objective loss function corresponding to the initial label generation module is constructed.

On the basis of the above embodiments, specifically, the initial data generation module further includes a gradient inversion layer, the gradient inversion layer is set between the feature extractor and the domain classifier, and the gradient inversion layer is used in the initial data generation module In the backpropagation process, the function gradient of the second loss function input into the gradient inversion layer is multiplied by a preset negative number, and correspondingly, the coefficient of the second loss function in the first target loss function is a preset negative number.

On the basis of the above-mentioned embodiments, specifically, the training index data includes point-of-interest area data, road network data, and population data, wherein the point-of-interest area data includes the number of points of interest, the total number of points of interest corresponding to at least one type of point of interest, respectively. At least one of quantity and interest point entropy value, the road network data includes the number of roads corresponding to at least one road type, the population data includes the number of working population and/or the number of living population, and correspondingly, the predictive index data is the number of consumption population , predicting the city label as a category of commercial popularity.

FIG. 7 is a schematic structural diagram of a training device for a cross-city federation transfer model provided by an embodiment of the present invention, which can be configured in a target city client. As shown in FIG. 7 , the device includes: a standard cross-city federation migration model building module 510 and a target city label determination module 520 .

Wherein, the standard cross-city federated migration model building module 510 is used to receive the model parameters of the trained target cross-city federated migration model sent by the source city client, and build a standard cross-city federated migration model based on the model parameters;

The target city label determination module 520 is used to input the test indicator data into the standard cross-city federation migration model to obtain the output target city label;

Among them, the target cross-city federated migration model is based on the initial cross-city federated migration model and the training index data in the training index data set by the source city client to determine the predictive index data and predicted city labels, and based on the standard index data corresponding to the predictive index data And the standard city labels corresponding to the predicted city labels are obtained by performing the training operation on the initial cross-city federated transfer model.

In the technical solution of this embodiment, the model parameters of the target cross-city federation migration model sent by the source city client are received, and based on the model parameters, a standard cross-city federation migration model is constructed; the test index data is input into the standard cross-city federation migration model. In the federated migration model, the output target city labels are obtained, where the target cross-city federated migration model is the source city client determines the predictive indicator data and predicted city labels based on the initial cross-city federated migration model and the training indicator data in the training indicator data set , and based on the standard index data corresponding to the predictive index data and the standard city labels corresponding to the predicted city labels to perform training operations on the initial cross-city federated migration model, which solves the problem that the target city cannot complete the city portrait task due to lack of index data problem, broadening the application scenarios of the cross-city migration method.

On the basis of the above embodiments, specifically, the model parameters include the data model parameters corresponding to the target data generation module in the target cross-city federation migration model and the label model parameters corresponding to the target label generation module, and the standard cross-city federation migration model construction module 510, specifically for:

Based on the data model parameters, construct the standard data generation module in the standard cross-city federation migration model, and input the reference index data into the standard data generation module to obtain the output missing index data;

Based on the parameters of the label model, a reference label generation module is constructed, and based on the reference index data and missing index data, the reference label generation module is trained to obtain the standard label generation module in the standard cross-city federation transfer model.

The training device for the cross-city federated migration model provided by the above-mentioned embodiments of the present invention can execute the training method for the cross-city federated migration model provided by the above-mentioned embodiments of the present invention, and has corresponding functional modules and beneficial effects for executing the method.

Fig. 8 is a schematic structural diagram of a training system for a cross-city federated migration model provided by an embodiment of the present invention. Serve.

As shown in FIG. 8 , the training system 600 of the cross-city federated migration model includes: a source city client 610 and a target city client 620; wherein, the source city client 610 is used to respond to the acquisition of the training index data set, based on The initial cross-city federation migration model and the training index data in the training index data set determine the predictive index data and predicted city labels, and based on the standard index data corresponding to the predictive index data and the standard city labels corresponding to the predicted city labels, the initial inter-city The model parameters of the city federation migration model are adjusted to obtain the trained target cross-city federation migration model, and the model parameters of the target cross-city federation migration model are sent to the target city client 620; the target city client 620 is used to receive Based on the obtained model parameters, a standard cross-city federation migration model is constructed, and based on the standard cross-city federation migration model, the target city label is determined.

Wherein, for example, the training index data can be used to represent the data generated in the development process of the source city, for example, the training index data can be population data, traffic data, tax data and so on.

Wherein, for example, the predicted city labels may be consumption index, traffic convenience degree, air quality index and so on. Among them, the consumption index can provide a reference basis for subsequent business district planning, the degree of traffic convenience can provide a reference basis for subsequent traffic scheduling, and the air quality index can provide a reference basis for the formulation and implementation of subsequent environmental governance plans.

In one embodiment, specifically, the source city client 610 is specifically configured to: respectively input the training indicator data in the training indicator data set into the initial data generation module and the initial label generation module in the initial cross-city federation migration model; Through the initial data generation module, output prediction index data based on the input training index data; through the initial label generation module, based on the input training index data and the prediction index data output by the initial data generation module, output the predicted city label.

In one embodiment, specifically, the source city client 610 is specifically configured to: respectively input the training indicator data in the training indicator data set into the initial data generation module and the initial label generation module in the initial cross-city federation migration model; Through the initial data generation module, output prediction index data based on the input training index data; through the initial label generation module, based on the input training index data and the prediction index data output by the initial data generation module, output the predicted city label.

In one embodiment, specifically, the training index data set contains the training index data corresponding to the current source city client 610, and correspondingly, the source city client 610 is specifically used for: each iteration of the initial cross-city federation migration model The process is to determine the target loss function based on the standard indicator data corresponding to the predicted indicator data and the standard city label corresponding to the predicted city label; based on the target loss function, determine the current model parameters corresponding to the initial cross-city federation migration model of the current iteration, and Sending the current model parameters to the central server, so that the central server determines the aggregation model parameters based on the received current model parameters sent by at least two source city clients 610, and sends the aggregation model parameters to the source city clients respectively 610; where at least two source city clients 610 include the current source city client 610; the received aggregation model parameters sent by the central server are used as the model parameters of the initial cross-city federation migration model of the current iteration until the target loss function When converging, the initial cross-city federated migration model of the current iteration is used as the target cross-city federated migration model after training.

In another embodiment, specifically, the number of source city clients 610 is at least two. Correspondingly, the training system 600 of the cross-city federation transfer model further includes a central server, and the central server is used to receive at least two The current model parameters sent by the source city client 610 respectively, determine the aggregation model parameters, and send the aggregation model parameters to each source city client 610 respectively; the source city client 610 is specifically used for: for the initial cross-city federation migration model In each iteration process, based on the standard indicator data corresponding to the predicted indicator data and the standard city label corresponding to the predicted city label, the target loss function is determined, and based on the target loss function, the current iteration corresponding to the initial cross-city federation migration model is determined. model parameters, and send the current model parameters to the central server; use the received aggregated model parameters as the model parameters of the current iteration's initial cross-city federation migration model, until the target loss function converges, the current iteration's initial cross-city federation migration model The model is used as the target cross-city federation migration model after training.

In one embodiment, specifically, the model parameters include the data model parameters corresponding to the target data generation module in the target cross-city federation migration model and the label model parameters corresponding to the target label generation module. The target city client 620 is specifically used for: Based on the data model parameters, construct the standard data generation module in the standard cross-city federation migration model, and input the reference index data into the standard data generation module to obtain the output missing index data; based on the label model parameters, construct the reference label generation module, And based on the reference index data and missing index data, the reference label generation module is trained to obtain the standard label generation module in the standard cross-city federation transfer model.

FIG. 9 is a system schematic diagram of a training system for a cross-city federation migration model provided by an embodiment of the present invention. Specifically, the central processor sends the model architecture of the initial cross-city federation migration model to the source city-1 client, source city-2 client... The model parameters of the initial cross-city federated migration model are initialized. The source city-1 client, the source city-2 client...the source city-J client performs local training on the initial cross-city federation transfer model in the source city client 610 based on the local training index data respectively. For each source city client 610, during each iteration of the initial cross-city federation migration model, an initial cross-city federation is determined based on the standard indicator data corresponding to the predicted indicator data and the standard city label corresponding to the predicted city label. The current model parameters of the migration model are sent to the central processor, and the central processor performs an aggregation operation on the current model parameters sent by the client 610 of each source city to obtain a global model, and the aggregated model parameters in the global model They are respectively sent to the source city-1 client, source city-2 client...source city-J client, so that each source city client 610 continues training based on the received aggregation model parameters.

Among them, the initial cross-city federation migration model includes an initial data generation module and an initial label generation module, wherein the initial data generation module includes a feature extractor, a data regressor, a domain classifier and a feature inversion layer, based on the corresponding The first loss function L _DR and the second loss function L _DC corresponding to the domain classifier adjust the model parameters of the initial data generation module. Among them, through the initial label generation module, based on the training index data and the prediction index data output by the initial data generation module, the predicted city label is output, and the second target loss function L ₂ constructed based on the predicted city label and the standard city label generates the initial label The model parameters of the module are adjusted.

Wherein, specifically, the central server determines the target model parameters based on the received model parameters of the target inter-city federation migration model sent by at least two source city clients 610 , and sends the target model parameters to the target city client 620 . The target city client 620 freezes the standard data generation module built based on the data model parameters, and fine-tunes and trains the reference tag generation module built based on the tag model parameters with its corresponding standard city tags based on the reference index data to obtain a standard tag generation module.

In the technical solution of this embodiment, through the source city client in the training system of the cross-city federated migration model, in response to the acquisition of the training index data set, based on the initial cross-city federated migration model and the training index data in the training index data set, determine Prediction index data and prediction city labels, based on the standard index data corresponding to the prediction index data and the standard city labels corresponding to the prediction city labels, adjust the model parameters of the initial cross-city federation migration model to obtain the trained target cross-city federation The migration model enables the trained target cross-city federation migration model to have the ability to predict index data and city labels at the same time, which solves the problem that the target city cannot complete the city portrait task due to lack of index data, and broadens the cross-city migration method. Application scenarios.

Fig. 10 is a schematic structural diagram of a terminal device provided by an embodiment of the present invention. Terminal device 10 is intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other suitable computers. Terminal devices may also represent various forms of mobile devices, such as personal digital processing, cellular phones, smartphones, wearable devices (eg, helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are by way of example only, and are not intended to limit implementations of the inventions described and/or claimed herein.

In this embodiment, the terminal device 10 is a source city client or a target city client.

As shown in FIG. 10 , the terminal device 10 includes at least one processor 11, and a memory communicatively connected with the at least one processor 11, such as a read-only memory (ROM) 12, a random access memory (RAM) 13, etc., wherein the memory stores There is a computer program executable by at least one processor, and the processor 11 can operate according to a computer program stored in a read-only memory (ROM) 12 or loaded from a storage unit 18 into a random access memory (RAM) 13. Various appropriate actions and processes are performed. In the RAM 13, various programs and data necessary for the operation of the terminal device 10 can also be stored. The processor 11, ROM 12, and RAM 13 are connected to each other through a bus 14. An input/output (I/O) interface 15 is also connected to the bus 14 .

Multiple components in the terminal device 10 are connected to the I/O interface 15, including: an input unit 16, such as a keyboard, a mouse, etc.; an output unit 17, such as various types of displays, speakers, etc.; a storage unit 18, such as a magnetic disk, an optical disk etc.; and a communication unit 19, such as a network card, a modem, a wireless communication transceiver, and the like. The communication unit 19 allows the terminal device 10 to exchange information/data with other devices through a computer network such as the Internet and/or various telecommunication networks.

Processor 11 may be various general and/or special purpose processing components having processing and computing capabilities. Some examples of processor 11 include, but are not limited to, central processing units (CPUs), graphics processing units (GPUs), various dedicated artificial intelligence (AI) computing chips, various processors that run machine learning model algorithms, digital signal processing processor (DSP), and any suitable processor, controller, microcontroller, etc. The processor 11 executes the various methods and processes described above, for example, the training method of the cross-city federation migration model.

In some embodiments, the method for training a federated migration model across cities can be implemented as a computer program tangibly embodied in a computer-readable storage medium, such as the storage unit 18 . In some embodiments, part or all of the computer program can be loaded and/or installed on the terminal device 10 via the ROM 12 and/or the communication unit 19. When the computer program is loaded into the RAM 13 and executed by the processor 11, one or more steps of the training method of the cross-city federation migration model described above can be performed. Alternatively, in other embodiments, the processor 11 may be configured in any other appropriate manner (for example, by means of firmware) to execute a method for training a cross-city federation migration model.

Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips Implemented in a system of systems (SOC), load programmable logic device (CPLD), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may include being implemented in one or more computer programs executable and/or interpreted on a programmable system including at least one programmable processor, the programmable processor Can be special-purpose or general-purpose programmable processor, can receive data and instruction from storage system, at least one input device, and at least one output device, and transmit data and instruction to this storage system, this at least one input device, and this at least one output device an output device.

The computer program for implementing the training method of the cross-city federated migration model of the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus, so that the computer program causes the functions/operations specified in the flowcharts and/or block diagrams to be implemented when executed by the processor. A computer program may execute entirely on the machine, partly on the machine, as a stand-alone software package partly on the machine and partly on a remote machine or entirely on the remote machine or server.

An embodiment of the present invention also provides a computer-readable storage medium, the computer-readable storage medium stores computer instructions, and the computer instructions are used to make a processor execute a training method for a cross-city federation migration model, the method comprising:

In response to obtaining the training indicator data set, based on the initial cross-city federation transfer model and the training indicator data in the training indicator data set, determine the predictive indicator data and the predicted city label;

Based on the standard index data corresponding to the predicted index data and the standard city label corresponding to the predicted city label, the model parameters of the initial cross-city federation migration model are adjusted to obtain the trained target cross-city federation migration model;

Send the model parameters of the target cross-city federation migration model to the target city client, so that the target city client builds a standard cross-city federation migration model based on the received model parameters, and determines the target city label based on the standard cross-city federation migration model.

Alternatively, the computer instructions are used to cause the processor to execute another cross-city federation transfer model training method, the method comprising:

Receive the model parameters of the trained target cross-city federation migration model sent by the source city client, and build a standard cross-city federation migration model based on the model parameters;

Input the test indicator data into the standard cross-city federation migration model to get the output target city label;

Among them, the target cross-city federated migration model is based on the initial cross-city federated migration model and the training index data in the training index data set by the source city client to determine the predictive index data and predicted city labels, and based on the standard index data corresponding to the predictive index data And the standard city labels corresponding to the predicted city labels are obtained by performing the training operation on the initial cross-city federated transfer model.

In the context of the present invention, a computer readable storage medium may be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus or device. A computer readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. Alternatively, a computer readable storage medium may be a machine readable signal medium. More specific examples of machine-readable storage media would include one or more wire-based electrical connections, portable computer discs, hard drives, random access memory (RAM), read only memory (ROM), erasable programmable read only memory (EPROM or flash memory), optical fiber, compact disk read only memory (CD-ROM), optical storage, magnetic storage, or any suitable combination of the foregoing.

To provide for interaction with the user, the systems and techniques described herein can be implemented on a terminal device having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) display device for displaying information to the user. monitor); and a keyboard and a pointing device (for example, a mouse or a trackball), through which the user can provide input to the terminal device. Other kinds of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and can be in any form (including Acoustic input, speech input or, tactile input) to receive input from the user.

The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., as a a user computer having a graphical user interface or web browser through which a user can interact with embodiments of the systems and techniques described herein), or including such backend components, middleware components, Or any combination of front-end components in a computing system. The components of the system can be interconnected by any form or medium of digital data communication, eg, a communication network. Examples of communication networks include: local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

A computing system can include clients and servers. Clients and servers are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by computer programs running on the respective computers and having a client-server relationship to each other. The server can be a cloud server, also known as a cloud computing server or a cloud host. It is a host product in the cloud computing service system to solve the problems of difficult management and weak business expansion in traditional physical hosts and VPS services. defect.

It should be understood that steps may be reordered, added or deleted using the various forms of flow shown above. For example, each step described in the present invention may be executed in parallel, sequentially, or in a different order, as long as the desired result of the technical solution of the present invention can be achieved, there is no limitation herein.

The above specific implementation methods do not constitute a limitation to the protection scope of the present invention. It should be apparent to those skilled in the art that various modifications, combinations, sub-combinations and substitutions may be made depending on design requirements and other factors. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims (17)

1. A training method of a cross-city federal migration model is applied to a source city client and comprises the following steps:

in response to the acquisition of a training index data set, determining prediction index data and prediction city labels based on an initial cross-city federal migration model and the training index data in the training index data set;

adjusting model parameters of the initial cross-city federal migration model based on standard index data corresponding to the prediction index data and standard city labels corresponding to the prediction city labels to obtain a trained target cross-city federal migration model;

and sending the model parameters of the target cross-city federal migration model to a target city client, so that the target city client builds a standard cross-city federal migration model based on the received model parameters, and determines a target city label based on the standard cross-city federal migration model.

2. The method of claim 1, wherein determining predictive index data and predictive city labels based on an initial cross-city federated migration model and training index data in the training index dataset comprises:

respectively inputting the training index data in the training index data set into an initial data generation module and an initial label generation module in an initial cross-city federal migration model;

outputting, by the initial data generation module, prediction index data based on the input training index data;

and outputting a predicted city label based on the input training index data and the predicted index data output by the initial data generation module through the initial label generation module.

3. The method of claim 2, wherein the initial data generation module comprises a feature extractor, a data regressor and a domain classifier, and wherein outputting prediction index data based on the input training index data comprises:

inputting the training index data into the feature extractor to obtain an output training index feature;

outputting, by the data regressor, predictive index data based on the training index features output by the feature extractor, and outputting, by the domain classifier, a predictive city classification result based on the training index features output by the feature extractor.

4. The method according to claim 3, wherein the training index dataset includes training index data corresponding to a current source city client, and accordingly, the adjusting the model parameters of the initial cross-city federation migration model based on standard index data corresponding to the prediction index data and standard city labels corresponding to the prediction city labels to obtain a trained target cross-city federation migration model includes:

determining a target loss function based on standard index data corresponding to the prediction index data and a standard city label corresponding to the prediction city label for each iteration process of the initial cross-city federal migration model;

determining current model parameters corresponding to an initial cross-city federal migration model of current iteration based on the target loss function, and sending the current model parameters to a central server, so that the central server determines aggregation model parameters based on the received current model parameters respectively sent by at least two source city clients, and sends the aggregation model parameters to each source city client; wherein the at least two source city clients comprise the current source city client;

and taking the received aggregation model parameters sent by the central server as model parameters of the initial cross-city federal migration model of the current iteration until the target loss function is converged, and taking the initial cross-city federal migration model of the current iteration as a trained target cross-city federal migration model.

5. The method of claim 4, wherein sending model parameters of the target cross-city federated migration model to a target city client comprises:

sending the model parameters of the target cross-city federal migration model to the central server, so that the central server determines target model parameters based on the received model parameters of the target cross-city federal migration model, which are sent by the at least two source city clients respectively, and sends the target model parameters to the target city clients.

6. The method according to claim 3, wherein the training index dataset includes training index data corresponding to at least two source city clients, and accordingly, the adjusting the model parameters of the initial cross-city federal migration model based on the standard index data corresponding to the prediction index data and the standard city labels corresponding to the prediction city labels to obtain a trained target cross-city federal migration model includes:

determining a target loss function based on standard index data corresponding to the prediction index data and a standard city label corresponding to the prediction city label for each iteration process of the initial cross-city federal migration model;

and determining current model parameters corresponding to the initial cross-city federal migration model of the current iteration based on the target loss function, and taking the initial cross-city federal migration model of the current iteration as a trained target cross-city federal migration model until the target loss function is converged.

7. The method according to claim 4 or 6, wherein the objective loss function comprises a first objective loss function corresponding to the initial data generation module and a second objective loss function corresponding to the initial label generation module, and wherein determining the objective loss function based on the standard index data corresponding to the prediction index data and the standard city label corresponding to the prediction city label comprises:

constructing a first loss function based on the prediction index data and the standard index data, and constructing a second loss function based on the prediction city classification result and the standard city classification result;

determining a first target loss function corresponding to the initial data generation module based on the first loss function and the second loss function;

and constructing a second target loss function corresponding to the initial label generation module based on the predicted city label and the standard city label.

8. The method according to claim 7, wherein the initial data generation module further comprises a gradient inversion layer disposed between the feature extractor and the domain classifier, the gradient inversion layer is configured to multiply a function gradient of the second loss function input into the gradient inversion layer by a preset negative number during back propagation of the initial data generation module, and accordingly, a coefficient of the second loss function in the first target loss function is the preset negative number.

9. The method of claim 1, wherein the training index data comprises point of interest area data, road network data and population data, wherein the point of interest area data comprises at least one of a number of points of interest, a total number of points of interest and a point of interest entropy value corresponding to at least one type of point of interest, wherein the road network data comprises a number of roads corresponding to at least one type of roads, wherein the population data comprises a number of working populations and/or a number of residential populations, and wherein the prediction index data is a number of consuming populations and wherein the prediction city label is a category of business popularity.

10. A training method of a cross-city federal migration model is applied to a target city client, and comprises the following steps:

receiving model parameters of a trained target cross-city federal migration model sent by a source city client, and constructing a standard cross-city federal migration model based on the model parameters;

inputting the test index data into the standard trans-city federal migration model to obtain an output target city label;

the target cross-city federal migration model is obtained by the source city client determining prediction index data and prediction city labels based on an initial cross-city federal migration model and training index data in a training index data set, and executing training operation on the initial cross-city federal migration model based on standard index data corresponding to the prediction index data and standard city labels corresponding to the prediction city labels.

11. The method according to claim 10, wherein the model parameters include data model parameters corresponding to a target data generation module and tag model parameters corresponding to a target tag generation module in the target cross-city federal migration model, and accordingly, the constructing a standard cross-city federal migration model based on the model parameters includes:

constructing a standard data generation module in a standard cross-city federal migration model based on the data model parameters, and inputting reference index data into the standard data generation module to obtain output missing index data;

and constructing a reference label generation module based on the label model parameters, and training the reference label generation module based on the reference index data and the missing index data to obtain a standard label generation module in a standard trans-city federal migration model.

12. A training device of a cross-city federal migration model is applied to a source city client, and comprises:

the prediction city label determining module is used for responding to the obtained training index data set, and determining prediction index data and prediction city labels based on an initial trans-city federal migration model and the training index data in the training index data set;

a target cross-city federal migration model determination module, configured to adjust model parameters of the initial cross-city federal migration model based on standard index data corresponding to the prediction index data and standard city labels corresponding to the prediction city labels, so as to obtain a trained target cross-city federal migration model;

and the model parameter sending module is used for sending the model parameters of the target cross-city federal migration model to the target city client, so that the target city client builds a standard cross-city federal migration model based on the received model parameters, and determines a target city label based on the standard cross-city federal migration model.

13. The utility model provides a training device of cross city federal migration model which is characterized in that, is applied to target city client, includes:

the standard trans-city federal migration model building module is used for receiving model parameters of a trained target trans-city federal migration model sent by a source city client, and building a standard trans-city federal migration model based on the model parameters;

the target city label determining module is used for inputting the test index data into the standard trans-city federal migration model to obtain an output target city label;

the target cross-city federal migration model is obtained by the source city client determining prediction index data and prediction city labels based on an initial cross-city federal migration model and training index data in a training index data set, and executing training operation on the initial cross-city federal migration model based on standard index data corresponding to the prediction index data and standard city labels corresponding to the prediction city labels.

14. A training system for a cross-city federal migration model, comprising: a source city client and a target city client;

the source city client is used for responding to the obtained training index data set, determining prediction index data and prediction city labels based on an initial cross-city federal migration model and training index data in the training index data set, adjusting model parameters of the initial cross-city federal migration model based on standard index data corresponding to the prediction index data and standard city labels corresponding to the prediction city labels to obtain a trained target cross-city federal migration model, and sending the model parameters of the target cross-city federal migration model to a target city client;

the target city client is used for constructing a standard cross-city federal migration model based on the received model parameters and determining a target city label based on the standard cross-city federal migration model.

15. The system according to claim 14, wherein the number of the source city clients is at least two, and accordingly, the system further comprises a central server, and the central server is configured to determine an aggregation model parameter based on the received current model parameters respectively sent by the at least two source city clients, and send the aggregation model parameter to each of the source city clients;

the source city client is specifically configured to: aiming at each iteration process of the initial cross-city federal migration model, determining a target loss function based on standard index data corresponding to the prediction index data and a standard city label corresponding to the prediction city label, determining current model parameters corresponding to the initial cross-city federal migration model of the current iteration based on the target loss function, and sending the current model parameters to a central server;

and taking the received aggregation model parameters as model parameters of the initial cross-city federal migration model of the current iteration until the target loss function is converged, and taking the initial cross-city federal migration model of the current iteration as a trained target cross-city federal migration model.

16. A terminal device, characterized in that the terminal device comprises:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein,

the memory stores a computer program executable by the at least one processor to enable the at least one processor to perform the method of training a cross-city federal migration model as defined in any of claims 1-9 or the method of training a cross-city federal migration model as defined in any of claims 10-11.

17. A computer-readable storage medium storing computer instructions for causing a processor to implement the method for training across-city federal migration model of any of claims 1-9 or the method for training across-city federal migration model of any of claims 10-11 when executed.

Images