Data processing methods, devices and storage media

Abstract

This invention provides a data processing method, apparatus, and storage medium. The method includes: acquiring data to be processed; processing the data to be processed using a decision tree model to obtain a processing result; wherein each tree model in the decision tree model is trained using intermediate data corresponding to multiple sample data calculated in groups and multiple remote sample data sent by a second server. Since each tree model in this solution requires intermediate data sent by the second server (i.e., the partner) for joint training with multiple sample data calculated in groups, it does not require the second server's base data, thus preventing data leakage from the other party. Furthermore, because multiple sample data are calculated in groups, the computational resources consumed are less. Therefore, this solution can effectively prevent data leakage from the partner and improve computational efficiency.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a data processing method, apparatus and storage medium. Background Technology

[0002] In a multi-party collaborative data processing model, we need to process the data using a trained model. Current modeling solutions require centralizing the data in one location, loading it all into memory at once, and then training the model. Therefore, partner data will inevitably be leaked to our organization. Furthermore, the solution of loading data into memory all at once for training is limited by computing resources, resulting in reduced computational efficiency. Therefore, the current solution presents the following technical problems: easy leakage of partner data and low computational efficiency. Summary of the Invention

[0003] The data processing method, apparatus, and storage medium provided in this invention can effectively prevent data leakage from partners and improve computing efficiency.

[0004] The technical solution of this invention is implemented as follows: This invention provides a data processing method, including: Obtain the data to be processed; The data to be processed is processed using a decision tree model to obtain the processing result; wherein, Each tree model in the decision tree model is trained using intermediate data corresponding to multiple sample data calculated in groups and multiple remote sample data sent by the second server.

[0005] In the above scheme, before processing the data to be processed using a decision tree model to obtain the processing result of the data to be processed, the method further includes: Based on the gradient vectors calculated from the multiple business parameters of the multiple sample data, the split point vectors of multiple features corresponding to the multiple sample data after feature grouping are determined, as well as the off-site split point vectors of multiple off-site features corresponding to the multiple off-site sample data; each sample data includes: the corresponding business parameters and the feature values ​​of multiple features; each off-site sample data includes: the off-site feature values ​​corresponding to multiple off-site features. By combining the split point vector and the off-site split point vector to determine the optimal split point threshold, iterative splitting of the multiple sample data or the multiple off-site sample data is performed until the first predetermined condition is met, and the current tree model is obtained. The current tree model is used to process the multiple sample data to obtain multiple residuals, and then the next tree model is trained based on the multiple residuals until the training stops when the second predetermined condition is met, resulting in multiple tree models. The multiple tree models are then combined to form the decision tree model.

[0006] In the above scheme, determining the split point vectors of multiple features corresponding to the multiple sample data after feature grouping, and the off-site split point vectors of multiple off-site features corresponding to the multiple off-site sample data, based on the gradient vectors calculated from the multiple business parameters of the multiple sample data, includes: Based on the gradient vector calculated from the multiple business parameters of the multiple sample data, the split point vectors corresponding to the multiple features after feature grouping are determined; The gradient vector is sent to the second server, and the second server receives multiple remote gradient histograms and their corresponding remote thresholds based on the gradient vector, thereby calculating the remote split point vector; wherein... The plurality of off-site gradient histograms are determined by the second server based on the plurality of off-site feature values ​​of the plurality of off-site sample data; the plurality of sample data and the plurality of off-site sample data include the same identification information; the plurality of off-site gradient histograms correspond to the plurality of off-site features; the intermediate data includes: the plurality of off-site gradient histograms and their corresponding plurality of off-site thresholds.

[0007] In the above scheme, before determining the split point vectors corresponding to the multiple features after feature grouping based on the gradient vectors calculated from the multiple business parameters of the multiple sample data, the method further includes: Multiple sample data are acquired, and the multiple features contained in the multiple sample data are grouped to obtain multiple feature groups; each feature group includes multiple feature values ​​of at least one feature corresponding to the multiple sample data. The step of determining the split point vectors corresponding to multiple features after feature grouping by calculating multiple gradient vectors based on multiple business parameters of the multiple sample data includes: Based on the gradient vector, at least one optimal threshold corresponding to at least one feature in each feature group is calculated in the order of the feature groups, thereby obtaining multiple optimal thresholds corresponding to the multiple features; The multiple optimal thresholds are sorted to form the split point vector.

[0008] In the above scheme, the step of calculating at least one optimal threshold corresponding to at least one feature in each feature group according to the gradient vector and in the order of the feature groups includes: Multiple thresholds are obtained for the multiple feature values ​​of each feature. Multiple histograms corresponding to each feature are constructed according to the multiple thresholds in the order of the feature group. The multiple feature values ​​corresponding to each feature are divided into the corresponding histograms to obtain multiple intermediate gradient histograms corresponding to the multiple features. Multiple gradients of the multiple sample data are calculated based on the gradient vector, and the multiple gradients are respectively accumulated in the intermediate gradient histograms within the corresponding numerical range to obtain multiple gradient histograms; The current threshold is determined from the plurality of thresholds, and the current information gain is calculated; If the current information gain is less than the previous information gain, then the threshold is iterated according to the new threshold corresponding to each gradient histogram until the current information gain is greater than the previous current information gain, and then the optimal threshold corresponding to each gradient histogram is obtained, thereby obtaining the at least one optimal threshold.

[0009] In the above scheme, the step of sending the gradient vector to the second server, receiving multiple remote gradient histograms and their corresponding remote thresholds from the second server based on the gradient vector, and then calculating the remote split point vector includes: The gradient vector is encrypted and sent to the second server. The second server then receives multiple remote gradient histograms and their corresponding remote thresholds based on the gradient vector. The current threshold is determined from the plurality of remote thresholds, and the current remote information gain is calculated; If the current off-site information gain is smaller than the previous off-site information gain, then the off-site threshold is iterated according to the new off-site threshold corresponding to each off-site gradient histogram until the current off-site information gain is larger than the previous off-site information gain, and the optimal off-site threshold corresponding to each off-site gradient histogram is obtained, thereby obtaining multiple optimal off-site thresholds for the multiple location features. The multiple optimal off-site thresholds are sorted to form the off-site split point vector.

[0010] In the above scheme, the step of combining the optimal split point threshold determined by the split point vector and the off-site split point vector to iteratively split the multiple sample data or the multiple off-site sample data until a first predetermined condition is met, thereby obtaining the current tree model, includes: Among the multiple current information gains corresponding to multiple optimal thresholds included in the split point vector, the maximum information gain is determined from among the multiple off-site information gains corresponding to multiple optimal off-site thresholds included in the off-site split point vector. The optimal threshold or optimal off-site value corresponding to the maximum information gain is determined as the optimal split point threshold. If the optimal split point threshold belongs to the split point vector, then the first identification information and the multiple features corresponding to the multiple first feature values ​​less than the optimal split point threshold are divided into left nodes, and the multiple second identification information and the multiple features corresponding to the multiple second feature values ​​greater than the optimal split point threshold are divided into right nodes. The iterative splitting of the multiple sample data is performed sequentially until a predetermined number of splits is reached, and then the current tree model is obtained. Wherein, the multiple first feature values ​​and the multiple second feature values ​​belong to the feature values ​​corresponding to the features of the optimal split point threshold; the multiple first identification information are the identification information corresponding to the multiple first sample data corresponding to the multiple first feature values; and the multiple second identification information are the identification information corresponding to the multiple second sample data corresponding to the multiple second feature values.

[0011] In the above scheme, after determining the optimal threshold or optimal off-site threshold corresponding to the maximum information gain, and the optimal split point threshold, the method further includes: If the optimal split point threshold belongs to the off-site split point vector, then the optimal split point threshold is sent to the second server, and the second server is received multiple third identifiers and multiple fourth identifiers based on the off-site feature values ​​corresponding to the optimal split point threshold; wherein, The plurality of third identification information are a plurality of first off-site feature values ​​that are less than the optimal split point threshold, and the identification information corresponding to the plurality of first off-site sample data; the plurality of fourth identification information are a plurality of second off-site feature values ​​that are greater than the optimal split point threshold, and the identification information corresponding to the plurality of second off-site sample data. The multiple third identifiers and the multiple remote features are divided into left nodes, and the multiple fourth identifiers and the multiple remote features are divided into right nodes. The multiple remote sample data are iteratively split in sequence until a predetermined number of splits is reached, and then the current tree model is obtained.

[0012] In the above scheme, the process of obtaining multiple residuals from the multiple sample data using the current tree model, and then training the next tree model based on the multiple residuals, continues until the training stops when a second predetermined condition is met, resulting in multiple tree models, including: The multiple sample data are input into the current tree model to obtain the multiple residuals; The current gradient vector is calculated based on the multiple residuals. Then, the next tree model is determined using the current gradient vector. The process continues until the number of tree models is greater than a predetermined number or the multiple residuals are less than a residual threshold, thus obtaining the multiple tree models.

[0013] This invention also provides a data processing method, characterized in that it includes: Receive the gradient vector sent by the first server; Intermediate data is determined based on the gradient vector combined with feature-grouped data from multiple geographically dispersed samples. The intermediate data is sent to the first server, which then trains multiple tree models based on the intermediate data and multiple sample data calculated in groups. The multiple tree models are then combined to form a decision tree model.

[0014] In the above scheme, before determining the intermediate data based on the gradient vector and multiple cross-regional sample data grouped by feature grouping, the method further includes: Multiple off-site sample data are acquired, and the multiple off-site features contained in the multiple off-site sample data are grouped to obtain multiple off-site feature groups; each off-site feature group includes: multiple off-site feature values ​​of at least one off-site feature corresponding to the multiple off-site sample data. The process of determining intermediate data based on the gradient vector combined with feature-grouped multiple samples from different locations includes: Multiple remote thresholds are obtained for the multiple remote feature values ​​of each remote feature. Based on the multiple remote thresholds in the order of the remote feature group, multiple remote histograms corresponding to each remote feature are constructed. The multiple remote feature values ​​corresponding to each remote feature are divided into the corresponding remote histograms to obtain multiple intermediate remote gradient histograms corresponding to the multiple remote features. Multiple gradients of the multiple remote sample data are calculated based on the gradient vector, and the multiple remote gradients are accumulated in the intermediate remote gradient histograms within the corresponding numerical range to obtain multiple remote gradient histograms. The intermediate data is formed by combining the multiple remote gradient histograms with the multiple remote thresholds.

[0015] In the above scheme, after determining the intermediate data based on the gradient vector and multiple cross-regional sample data grouped by feature grouping, the method further includes: Receive the optimal split point threshold sent by the first server; Determine multiple third identifiers for multiple first off-site sample data corresponding to multiple first off-site feature values ​​that are less than the optimal split point threshold; Determine multiple fourth identifiers for multiple second off-site sample data corresponding to multiple second off-site feature values ​​that are greater than the optimal split point threshold; The plurality of third identification information and the plurality of fourth identification information are sent to the first server.

[0016] This invention also provides a data processing apparatus, comprising: The data acquisition unit is used to acquire data to be processed. The processing unit is used to process the data to be processed using a decision tree model to obtain the processing result of the data to be processed; wherein, Each tree model in the decision tree model is trained using intermediate data corresponding to multiple sample data calculated in groups and multiple remote sample data sent by the second server.

[0017] This invention also provides a data processing apparatus, comprising: A receiving unit is used to receive gradient vectors sent by the first server; The determining unit is used to determine intermediate data based on the gradient vector combined with multiple cross-regional sample data after feature grouping; The sending unit is used to send the intermediate data to the first server, so that the first server can train multiple tree models based on the intermediate data and multiple sample data calculated in groups, and then combine the multiple tree models to form a decision tree model.

[0018] This invention also provides a data processing apparatus, including a first memory and a first processor. The first memory stores a computer program that can run on the first processor. When the first processor executes the program, it implements the steps in the method on the first server side.

[0019] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a first processor, implements the steps of the method on the first server side.

[0020] This invention also provides a data processing apparatus, including a second memory and a second processor. The second memory stores a computer program that can run on the second processor. When the second processor executes the program, it implements the steps in the method on the second server side.

[0021] This invention also provides a computer-readable storage medium storing a computer program thereon, characterized in that the computer program, when executed by a second processor, implements the steps of the method on the second server side.

[0022] In this embodiment of the invention, data to be processed is acquired; a decision tree model is used to process the data to obtain the processing result; wherein, each tree model in the decision tree model is trained using intermediate data corresponding to multiple sample data calculated in groups and multiple remote sample data sent by a second server. Since each tree model in this solution requires intermediate data sent by the second server (i.e., the partner) and is jointly trained with multiple sample data calculated in groups, the basic data of the second server is not required, thus preventing the other party's data from being leaked. Moreover, since multiple sample data are calculated in groups, the computational resources occupied are less. Therefore, this solution can effectively prevent the partner's data from being leaked and improve computational efficiency. Attached Figure Description

[0023] Figure 1 A schematic flowchart of an optional data processing method provided in an embodiment of the present invention; Figure 2 A schematic flowchart of an optional data processing method provided in an embodiment of the present invention; Figure 3 A schematic flowchart of an optional data processing method provided in an embodiment of the present invention; Figure 4 A schematic flowchart of an optional data processing method provided in an embodiment of the present invention; Figure 5 A schematic flowchart of an optional data processing method provided in an embodiment of the present invention; Figure 6 A schematic flowchart of an optional data processing method provided in an embodiment of the present invention; Figure 7 A schematic flowchart of an optional data processing method provided in an embodiment of the present invention; Figure 8 A schematic flowchart of an optional data processing method provided in an embodiment of the present invention; Figure 9 A schematic flowchart of an optional data processing method provided in an embodiment of the present invention; Figure 10 A schematic flowchart of an optional data processing method provided in an embodiment of the present invention; Figure 11 A schematic flowchart of an optional data processing method provided in an embodiment of the present invention; Figure 12 An interactive schematic diagram of the data processing method provided in an embodiment of the present invention; Figure 13 Schematic diagram of the data processing apparatus provided in the embodiments of the present invention Figure 1 ; Figure 14 A hardware entity illustration of the data processing device provided in the embodiments of the present invention. Figure 1 ; Figure 15 Schematic diagram of the data processing apparatus provided in the embodiments of the present invention Figure 2 ; Figure 16 A hardware entity illustration of the data processing device provided in the embodiments of the present invention. Figure 2 . Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The described embodiments should not be regarded as limitations on the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0026] If similar descriptions such as "first / second" appear in the invention document, the following explanation shall be added: In the following description, the terms "first / second / third" are used only to distinguish similar objects and do not represent a specific order of objects. It is understood that "first / second / third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of the invention described herein can be implemented in an order other than that illustrated or described herein.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit the invention.

[0028] Figure 1 This is an optional flowchart illustrating a data processing method provided in an embodiment of the present invention, which will be combined with... Figure 1 The steps shown are explained.

[0029] S101. Obtain the data to be processed.

[0030] In this embodiment of the invention, the first server acquires the data to be processed.

[0031] In this embodiment of the invention, the first server can obtain data to be processed from multiple clients through pre-established communication lines with multiple clients.

[0032] In this embodiment of the invention, the first server can obtain the data to be processed from the corresponding database.

[0033] The data to be processed can be multiple attribute data corresponding to a target. The target can include any one of users, stores, and items. For example, the data to be processed can be multiple attribute data of a user. These attribute data can include information such as the user's transaction habits, gender, age, education level, and location. The data to be processed can also be attribute data of a store. The data to be processed can also be attribute data of a specific item.

[0034] In this embodiment of the invention, the server can obtain the data to be processed through a third-party storage device. Specifically, the target object of the server uploads the data to be processed to the server via the server's interface.

[0035] In this embodiment of the invention, the server can obtain data to be processed from other servers through a pre-established communication line with other servers.

[0036] The first server can be the server of the organizer in the collaborative data processing project.

[0037] S102. Use the decision tree model to process the data to be processed and obtain the processing result of the data to be processed; wherein, each tree model in the decision tree model is trained by grouping and calculating multiple sample data and intermediate data corresponding to multiple remote sample data sent by the second server.

[0038] In this embodiment of the invention, the first server uses a decision tree model to process the data to be processed and obtain the processing result of the data to be processed; wherein, each tree model in the decision tree model is trained by grouping and calculating multiple sample data and intermediate data corresponding to multiple remote sample data sent by the second server.

[0039] Among them, the decision tree model is used to predict the probability of performing a certain business operation on the data to be processed. The decision tree model is formed by combining multiple tree models.

[0040] The second server can be the server of a partner in the collaborative data processing project.

[0041] In this embodiment of the invention, the identification information of multiple sample data and corresponding multiple remote sample data is the same. That is, each sample data and its corresponding remote sample data correspond to the same target. Furthermore, the types of the multiple sample data, the multiple remote sample data, and the targets corresponding to the data to be processed are the same.

[0042] In this embodiment of the invention, a first server can group multiple features corresponding to multiple sample data to obtain multiple feature groups. Each feature group includes multiple feature values ​​corresponding to the features of the multiple sample data. The first server calculates the gradient histogram corresponding to each feature group, and then calculates a split point vector based on a preset threshold for each gradient histogram. The first server sends multiple remote gradient histograms (intermediate data) corresponding to multiple remote features to a second server. The first server calculates remote split point vectors for multiple preset remote thresholds corresponding to the multiple remote gradient histograms. The first server determines the optimal split point threshold from the split point vectors and the remote split point vectors, and then iteratively splits multiple sample data or multiple remote sample data according to the optimal split point threshold until a first predetermined condition is met, resulting in the current tree model. The first server then uses the current tree model to process multiple residuals of the multiple sample data, and then trains the next tree model based on the multiple residuals until a second predetermined condition is met, resulting in multiple tree models. These multiple tree models are then combined to form a decision tree model.

[0043] In this embodiment of the invention, data to be processed is acquired; a decision tree model is used to process the data to obtain the processing result; wherein, each tree model in the decision tree model is trained using intermediate data corresponding to multiple sample data calculated in groups and multiple remote sample data sent by a second server. Since each tree model in this solution requires intermediate data sent by the second server (i.e., the partner) and is jointly trained with multiple sample data calculated in groups, the basic data of the second server is not required, thus preventing the other party's data from being leaked. Moreover, since multiple sample data are calculated in groups, the computational resources occupied are less. Therefore, this solution can effectively prevent the partner's data from being leaked and improve computational efficiency.

[0044] In some embodiments, see Figure 2 , Figure 2 This is an optional flowchart illustrating the data processing method provided in an embodiment of the present invention. Figure 1 The implementation of S102 is preceded by S103 to S105, which will be explained in conjunction with each step.

[0045] S103. Based on the gradient vectors calculated from the multiple business parameters of multiple sample data, determine the split point vectors of multiple features corresponding to the multiple sample data after feature grouping, and the off-site split point vectors of multiple off-site features corresponding to the multiple off-site sample data.

[0046] In this embodiment of the invention, the first server calculates gradient vectors based on multiple business parameters of multiple sample data, and determines the split point vectors of multiple features corresponding to the multiple sample data after feature grouping, as well as the off-site split point vectors of multiple off-site features corresponding to the multiple off-site sample data. Each sample data includes: corresponding business parameters and feature values ​​of multiple features; each off-site sample data includes: off-site feature values ​​corresponding to multiple off-site features.

[0047] In this embodiment of the invention, the first server calculates gradient vectors based on multiple business parameters of multiple sample data, and determines the split point vectors corresponding to multiple features after feature grouping. The first server sends the gradient vectors to the second server, and receives multiple remote gradient histograms and their corresponding remote thresholds fed back by the second server based on the gradient vectors, and then calculates the remote split point vectors.

[0048] S104. Combine the split point vector and the optimal split point threshold determined by the split point vector and the off-site split point vector, and perform iterative splitting of multiple sample data or multiple off-site sample data until the first predetermined condition is met, and then stop to obtain the current tree model.

[0049] In this embodiment of the invention, the first server combines the split point vector and the remote split point vector to determine the optimal split point threshold, and performs iterative splitting of multiple sample data or multiple remote sample data until the first predetermined condition is met, thereby obtaining the current tree model.

[0050] In this embodiment of the invention, the first server determines the split point threshold with the largest information gain from the split point vector and the remote split point vector. If the optimal split point threshold belongs to the split point vector, iterative splitting of multiple sample data is performed until a predetermined number of splits is reached, resulting in the current tree model. If the optimal split point threshold belongs to the remote split point vector, the optimal split point threshold is sent to the second server. The server receives relevant data from multiple remote sample data and performs iterative splitting of multiple remote sample data until a predetermined number of splits is reached, resulting in the current tree model.

[0051] For example, the predetermined number of splits can be 5 or 6. No specific limitation is made in this embodiment of the invention.

[0052] S105. Use the current tree model to process multiple residuals of multiple sample data, and then train the next tree model based on the multiple residuals until the second predetermined condition is met, and the training stops, resulting in multiple tree models. Combine the multiple tree models to form a decision tree model.

[0053] In this embodiment of the invention, the first server uses the current tree model to process multiple residuals of multiple sample data, and then trains the next tree model based on the multiple residuals until the training stops when the second predetermined condition is met, thus obtaining multiple tree models. The obtained multiple tree models are then combined to form a decision tree model.

[0054] In this embodiment of the invention, the first server inputs multiple sample data into the current tree model, obtaining multiple residuals corresponding to the multiple sample data. The first server then uses the multiple residuals to calculate a new gradient vector corresponding to the multiple samples, and combines the new gradient vector to calculate a new split point vector and a remote split point vector. The first server then determines a new optimal split point threshold, and iteratively splits the multiple sample data or multiple remote sample data according to the optimal split point threshold to obtain the next tree model. This process continues until the number of tree models reaches the number threshold or the multiple residuals are less than the residual threshold, resulting in multiple tree models. The first server then combines the multiple tree models according to different weights to obtain a decision tree model.

[0055] Each tree model in the decision tree model of this solution requires the second server to send multiple remote gradient histograms and their corresponding remote thresholds, which are then jointly trained with multiple sample data calculated in groups. This eliminates the need for the second server's base data, ensuring that the other party's data is not leaked. Moreover, since multiple sample data are calculated in groups, the computational resources required are less. Therefore, this solution can effectively prevent data leakage from the partner and improve computational efficiency.

[0056] In some embodiments, see Figure 3 , Figure 3 This is an optional flowchart illustrating the data processing method provided in an embodiment of the present invention. Figure 2 The shown S103 can be implemented through S106 to S107, which will be explained in conjunction with each step.

[0057] S106. Based on the gradient vectors calculated from multiple business parameters of multiple sample data, determine the split point vectors corresponding to multiple features after feature grouping.

[0058] In this embodiment of the invention, the first server determines the split point vectors corresponding to multiple features after feature grouping by calculating the gradient vectors based on multiple business parameters of multiple sample data. Each sample data includes: identification information, business parameters, and feature values ​​of multiple features.

[0059] In this embodiment of the invention, the first server first acquires multiple sample data. The first server calculates gradient vectors based on multiple business parameters. The first server groups the feature values ​​within multiple features corresponding to the multiple sample data to obtain multiple feature groups. The first server acquires multiple thresholds for multiple feature values ​​of each feature, and constructs multiple histograms corresponding to each feature according to the order of the feature groups and multiple thresholds. The multiple feature values ​​corresponding to each feature are then assigned to the corresponding histograms to obtain multiple intermediate gradient histograms corresponding to the multiple features. Multiple gradients of the multiple sample data are calculated based on the gradient vectors, and the multiple gradients are accumulated in the intermediate gradient histograms within the corresponding numerical ranges to obtain multiple gradient histograms.

[0060] The first server calculates the gradient vector based on multiple business parameters, which means that the first server calculates the first and second derivatives of multiple business parameters.

[0061] For example, multiple sample data may include ten samples. Each sample data may include corresponding identification information, business parameters, and feature values ​​of ten features. The feature categories of each sample data are the same. The first server can divide the ten features of the multiple sample data into two groups. Each group includes multiple feature values ​​of five features corresponding to the multiple sample data. In this embodiment of the invention, the first server may also use other grouping methods, which are not limited in this embodiment.

[0062] S107. Send the gradient vector to the second server, receive multiple remote gradient histograms and their corresponding remote thresholds from the second server based on the gradient vector, and then calculate the remote split point vector.

[0063] In this embodiment of the invention, the first server sends the gradient vector to the second server, and receives multiple remote gradient histograms and their corresponding remote thresholds from the second server based on the gradient vector, thereby calculating the remote split point vector. The multiple remote gradient histograms are determined by the second server based on multiple remote feature values ​​of multiple remote sample data; the multiple sample data and the multiple remote sample data include the same identification information; the multiple remote gradient histograms correspond to multiple remote features; the intermediate data includes: multiple remote gradient histograms and their corresponding remote thresholds.

[0064] In this embodiment of the invention, after the first server calculates the gradient vector, it encrypts the gradient vector and sends it to the second server. At this time, the second server has already grouped the multiple remote sample data into feature groups. The second server obtains multiple remote thresholds for multiple remote feature values ​​for each remote feature, and constructs multiple remote histograms corresponding to each remote feature according to the order of the remote feature groups and multiple remote thresholds. It then divides the multiple remote feature values ​​corresponding to each remote feature into the corresponding remote histograms, obtaining multiple intermediate remote gradient histograms corresponding to the multiple remote features. The second server calculates multiple remote gradients for the multiple remote sample data based on the gradient vector, and accumulates the multiple remote gradients in the intermediate remote gradient histograms within the corresponding numerical ranges, obtaining multiple remote gradient histograms. The second server sends the multiple remote gradient histograms to the first server. The first server, combining the multiple remote thresholds corresponding to the multiple remote gradient histograms, calculates the remote split point vector using a method similar to that used for the multiple gradient histograms.

[0065] In some embodiments, see Figure 4 , Figure 4 This is an optional flowchart illustrating the data processing method provided in an embodiment of the present invention. Figure 3 The shown S106 can be implemented through S108 to S110, which will be explained in conjunction with each step.

[0066] S108. Obtain multiple sample data, and group the multiple features contained in the multiple sample data to obtain multiple feature groups.

[0067] In this embodiment of the invention, a first server acquires multiple sample data sets, and groups the multiple features contained in each sample data set into multiple feature groups. Each feature group includes multiple feature values ​​of at least one feature corresponding to the multiple sample data sets.

[0068] In this embodiment of the invention, multiple sample data include multiple features of the same number and type. Each sample data includes feature values ​​corresponding to the multiple features. The first server divides the multiple features into multiple feature groups. The first server adds the feature values ​​corresponding to the multiple features from the multiple sample data into the corresponding features.

[0069] For example, the multiple sample data includes ten sample data points, each containing five features: transaction habit information, gender information, age information, education information, and location information. Each sample data point has a corresponding feature value for each of the five features. The first server divides the transaction habit information, gender information, and age information into a first group, and the first server divides the education information and location information into a second group. The first server then adds the feature values ​​of the ten sample data points corresponding to the transaction habit information, gender information, and age information to the corresponding features, obtaining the final first feature group. The first server then adds the feature values ​​of the ten sample data points corresponding to the education information and location information to the corresponding features, obtaining the final second feature group. In this embodiment of the invention, the first server may also use other grouping methods, which are not limited in this embodiment.

[0070] S109. Based on the gradient vector, calculate at least one optimal threshold for at least one feature in each feature group according to the order of the feature groups, and then obtain multiple optimal thresholds for multiple features.

[0071] In this embodiment of the invention, the first server calculates at least one optimal threshold corresponding to at least one feature in each feature group according to the gradient vector and the order of the feature groups, thereby obtaining multiple optimal thresholds corresponding to multiple features.

[0072] In this embodiment of the invention, a first server obtains multiple thresholds for each of the multiple features. The first server constructs intermediate gradient histograms for the multiple features according to the order of the feature groups. The first server calculates multiple gradients for the multiple sample data based on the gradient vectors, and adds these gradients to the intermediate gradient histograms corresponding to the numerical ranges, resulting in multiple gradient histograms. The first server calculates the current information gain based on the current threshold of each gradient histogram. If the current information gain is smaller than the previous information gain, the current threshold is iterated until the current information gain is greater than the previous information gain. At least one optimal threshold for at least one feature is obtained, and thus multiple optimal thresholds for multiple features are obtained.

[0073] S110. Sort the multiple optimal thresholds to form a split point vector.

[0074] In this embodiment of the invention, the first server sorts multiple optimal thresholds to form a split point vector.

[0075] In this embodiment of the invention, the first server sorts multiple optimal thresholds according to the order of multiple features to form a split point vector.

[0076] In this embodiment of the invention, the first server calculates at least one optimal threshold corresponding to at least one feature in each feature group according to the order of the feature groups, thereby obtaining multiple optimal thresholds corresponding to multiple features, and finally forming a split point vector. Since the first server performs calculations on the data in groups, the occupation of computing resources is reduced, thereby improving the computing efficiency.

[0077] In some embodiments, see Figure 5 , Figure 5 This is an optional flowchart illustrating the data processing method provided in an embodiment of the present invention. Figure 4 S109 shown can be implemented through S111 to S114, which will be explained in conjunction with each step.

[0078] S111. Obtain multiple thresholds for multiple feature values ​​of each feature. Based on the multiple thresholds and in the order of the feature group, construct multiple histograms corresponding to each feature. Divide the multiple feature values ​​corresponding to each feature into the corresponding histograms to obtain multiple intermediate gradient histograms corresponding to multiple features.

[0079] In this embodiment of the invention, the first server obtains multiple thresholds for multiple feature values ​​of each feature, constructs multiple histograms corresponding to each feature according to the order of the feature group and multiple thresholds, and divides the multiple feature values ​​corresponding to each feature into the corresponding histograms to obtain multiple intermediate gradient histograms corresponding to multiple features.

[0080] In this embodiment of the invention, the staff sends multiple threshold values ​​corresponding to each feature to the first server via a human-computer interaction device. The first server constructs multiple histograms corresponding to each feature based on the multiple threshold values, following the order of the feature group. The first server then assigns the multiple feature values ​​corresponding to each feature to the corresponding histograms, obtaining multiple intermediate gradient histograms for the multiple features.

[0081] For example, in this embodiment of the invention, taking age as an example, an intermediate gradient histogram corresponding to the age feature is constructed. The age feature includes multiple feature values: 20, 22, 23, 25, 26, 29, 26, 26, 31, 35, 38, 39, 35, 40, 42, 44, 45, and 43. The first server obtains multiple thresholds including: 0, 25, 30, 40, and 45. The first server constructs histograms for values ​​from 0 to 25, 25 to 30, 30 to 40, and 40 to 45 using the multiple thresholds 0, 25, 30, 40, and 45. The first server divides 20, 22, 23, and 25 into histograms from 0 to 25; the first server divides 26, 29, 26, and 26 into histograms from 25 to 30; the first server divides 31, 35, 38, 39, and 35 into histograms from 30 to 40; and the first server divides 40, 42, 44, 45, and 43 into histograms from 40 to 45, thus obtaining the intermediate gradient histograms of age characteristics.

[0082] S112. Calculate multiple gradients for multiple sample data based on the gradient vector, and accumulate the multiple gradients in the intermediate gradient histograms within the corresponding numerical ranges to obtain multiple gradient histograms.

[0083] In this embodiment of the invention, the first server calculates multiple gradients of multiple sample data based on the gradient vector, and accumulates the multiple gradients in the intermediate gradient histograms within the corresponding numerical range to obtain multiple gradient histograms.

[0084] For example, the gradient of the first sample data calculated by the first server based on the gradient vector is 26. The first server detects that the gradient 26 belongs to the feature value range corresponding to the age feature. Then, the first server adds the gradient 26 to the histogram of the middle gradient histogram of the age feature from 25 to 30, thereby obtaining the gradient histogram of the age feature. The first server can then calculate multiple gradient histograms corresponding to multiple features in the same way.

[0085] S113. Determine the current threshold from among multiple thresholds and calculate the current information gain.

[0086] In this embodiment of the invention, the first server determines the current threshold from multiple thresholds and calculates the current information gain.

[0087] In this embodiment of the invention, the first server can use the LightGBM algorithm to calculate the current information gain in combination with the current threshold.

[0088] S114. If the current information gain is less than the previous information gain, then according to the new threshold corresponding to each gradient histogram, the threshold is iterated until the current information gain is greater than the previous current information gain, and the optimal threshold corresponding to each gradient histogram is obtained, thus obtaining at least one optimal threshold, and finally obtaining multiple optimal thresholds.

[0089] In this embodiment of the invention, if the current information gain is less than the previous information gain, the threshold is iterated according to the new threshold corresponding to each gradient histogram until the current information gain is greater than the previous current information gain. Then, the optimal threshold corresponding to each gradient histogram is obtained, and at least one optimal threshold is obtained, and finally multiple optimal thresholds are obtained.

[0090] In this embodiment of the invention, the first server calculates multiple optimal thresholds corresponding to multiple gradient histograms, thereby accurately determining the split point vector and providing a data foundation for the splitting of the tree model.

[0091] In some embodiments, see Figure 6 , Figure 6 This is an optional flowchart illustrating the data processing method provided in an embodiment of the present invention. Figure 3 The shown S107 can be implemented through S115 to S118, which will be explained in conjunction with each step.

[0092] S115. Encrypt the gradient vector and send it to the second server. Receive multiple remote gradient histograms and their corresponding remote thresholds from the second server based on the gradient vector.

[0093] In this embodiment of the invention, the first server encrypts the gradient vector and sends it to the second server, and receives multiple remote gradient histograms and their corresponding remote thresholds from the second server based on the gradient vector.

[0094] S116. Determine the current threshold from multiple remote thresholds and calculate the current remote information gain.

[0095] In this embodiment of the invention, the first server determines the current threshold from multiple remote thresholds and calculates the current remote information gain.

[0096] S117. If the current off-site information gain is smaller than the previous off-site information gain, then based on the new off-site threshold corresponding to each off-site gradient histogram, the off-site threshold is iterated until the current off-site information gain is greater than the previous off-site information gain, and the optimal off-site threshold corresponding to each off-site gradient histogram is obtained, thereby obtaining multiple optimal off-site thresholds for multiple off-site features.

[0097] In this embodiment of the invention, if the current off-site information gain calculated by the first server is less than the previous off-site information gain, the off-site threshold is iterated according to the new off-site threshold corresponding to each off-site gradient histogram until the current off-site information gain is greater than the previous off-site information gain, and then the optimal off-site threshold corresponding to each off-site gradient histogram is obtained, thereby obtaining multiple optimal off-site thresholds for multiple off-site features.

[0098] S118. Sort multiple optimal off-site thresholds to form an off-site split point vector.

[0099] In this embodiment of the invention, the first server sorts multiple optimal remote thresholds to form a remote split point vector.

[0100] In this embodiment of the invention, the first server sorts multiple optimal remote thresholds according to the order of multiple remote features to obtain a remote split point vector.

[0101] In this embodiment of the invention, the first server receives multiple remote gradient histograms and their corresponding remote thresholds fed back by the second server based on the gradient vector, thereby avoiding the first server receiving basic data sent by the second server to train the tree model. Therefore, this solution can effectively prevent data leakage from partners.

[0102] In some embodiments, see Figure 7 , Figure 7 This is an optional flowchart illustrating the data processing method provided in an embodiment of the present invention. Figure 3 The steps S104 to S105 shown can be implemented by S119 to S123, which will be explained in conjunction with each step.

[0103] S119. Among the multiple current information gains corresponding to multiple optimal thresholds included in the split point vector, determine the maximum information gain among the multiple off-site information gains corresponding to multiple optimal off-site thresholds included in the off-site split point vector.

[0104] In this embodiment of the invention, the first server determines the maximum information gain among the multiple current information gains corresponding to multiple optimal thresholds included in the split point vector, and among the multiple remote information gains corresponding to multiple optimal remote thresholds included in the remote split point vector.

[0105] In this embodiment of the invention, since multiple optimal thresholds for the split point vector correspond to multiple current information gains, and multiple optimal remote thresholds for the remote split point vector correspond to multiple remote information gains, the first server can determine the maximum information gain among multiple current information gains and multiple remote information gains.

[0106] Information gain represents the degree to which information uncertainty is reduced under a given condition.

[0107] S120. Determine the optimal threshold or optimal off-site threshold corresponding to the maximum information gain as the optimal split point threshold.

[0108] In this embodiment of the invention, the first server determines the optimal threshold or optimal remote threshold corresponding to the maximum information gain as the optimal split point threshold.

[0109] S121. If the optimal split point threshold belongs to the split point vector, then the first identification information and multiple features corresponding to multiple first feature values ​​less than the optimal split point threshold are divided into left nodes, and the multiple second identification information and multiple features corresponding to multiple second feature values ​​greater than the optimal split point threshold are divided into right nodes. The iterative splitting of multiple sample data is performed sequentially until the predetermined number of splits is reached, and then the current tree model is obtained.

[0110] In this embodiment of the invention, if the optimal split point threshold belongs to the split point vector, the first server divides the first identification information and multiple features corresponding to multiple first feature values ​​that are less than the optimal split point threshold into left nodes, and divides the multiple second identification information and multiple features corresponding to multiple second feature values ​​that are greater than the optimal split point threshold into right nodes. The server then performs iterative splitting of multiple sample data until the predetermined number of splits is reached, and then stops to obtain the current tree model.

[0111] Among them, multiple first feature values ​​and multiple second feature values ​​belong to the feature values ​​corresponding to the optimal split point threshold. Multiple first identification information refers to the identification information corresponding to multiple first sample data corresponding to multiple first feature values; multiple second identification information refers to the identification information corresponding to multiple second sample data corresponding to multiple second feature values.

[0112] For example, the optimal split point threshold is 30. The feature corresponding to the optimal split point threshold is the age feature. The first server identifies multiple first feature values ​​less than 30 in the age feature. The first server divides the multiple first identifiers and features of the multiple first sample data corresponding to the multiple first feature values ​​into left nodes. The first server identifies multiple second feature values ​​greater than 30 in the age feature. The first server divides the multiple second identifiers and features of the multiple second sample data corresponding to the multiple second feature values ​​into right nodes. Subsequently, the first server splits the sample data in the left and right nodes again until a predetermined number of splits is reached, resulting in the current tree model.

[0113] S122. Input multiple sample data into the current tree model to obtain multiple residuals.

[0114] In this embodiment of the invention, the first server inputs multiple sample data into the current tree model to obtain multiple residuals.

[0115] S123. Calculate the current gradient vector based on multiple residuals, and then use the current gradient vector to determine the next tree model. Stop when the number of tree models is greater than the predetermined number or multiple residuals are less than the residual threshold, thus obtaining multiple tree models.

[0116] In this embodiment of the invention, the first server calculates the current gradient vector based on multiple residuals, and then uses the current gradient vector to determine the current tree model. The process continues until the number of tree models is greater than a predetermined number or multiple residuals are less than a residual threshold, thus obtaining multiple tree models.

[0117] In some embodiments, see Figure 8 , Figure 8 This is an optional flowchart illustrating the data processing method provided in an embodiment of the present invention. Figure 7 The shown S121 can be implemented through S124 to S125, which will be explained in conjunction with each step.

[0118] S124. If the optimal split point threshold belongs to the off-site split point vector, then the optimal split point threshold is sent to the second server, and the second server is received multiple third identification information and multiple fourth identification information based on the off-site feature value corresponding to the optimal split point threshold.

[0119] In this embodiment of the invention, if the optimal split point threshold belongs to the off-site split point vector, the first server sends the optimal split point threshold to the second server and receives multiple third identification information and multiple fourth identification information from the second server based on the off-site feature value corresponding to the optimal split point threshold.

[0120] S125. Divide multiple third-level identifiers and multiple remote features into left nodes, and multiple fourth-level identifiers and multiple remote features into right nodes. Iterate and split multiple remote sample data in sequence until the predetermined number of splits is reached, and then stop to obtain the current tree model.

[0121] In this embodiment of the invention, the first server divides multiple third identification information and multiple remote features into left nodes, and multiple fourth identification information and multiple remote features into right nodes, and performs iterative splitting of multiple remote sample data in sequence until a predetermined number of splits is reached, thus obtaining the current tree model.

[0122] In this embodiment of the invention, the first server divides multiple third identification information and multiple remote features into left nodes, and multiple fourth identification information and multiple remote features into right nodes. Then, iteratively splits the remote sample data in the left nodes and the remote sample data in the right nodes until a predetermined number of splits is reached, and then stops to obtain the current tree model.

[0123] In this embodiment of the invention, when the optimal split point threshold is in the remote threshold vector, the first server sends the optimal split point threshold to the second server and receives multiple third identification information and multiple fourth identification information sent by the second server, and then performs iterative splitting of multiple remote sample data. Since the iterative splitting of the tree model in this scheme does not require the basic data of the second server, it effectively prevents the leakage of the partner's data.

[0124] In some embodiments, see Figure 9 , Figure 9 This is an optional flowchart illustrating the data processing method provided in an embodiment of the present invention, which will be explained in conjunction with each step.

[0125] S401, master homomorphic encryption.

[0126] S402, Sample set.

[0127] S403, master.

[0128] S404, Off-site sample set.

[0129] In this embodiment of the invention, the first server preprocesses the sample features, dividing them into multiple groups. It can even group a single feature into its own group, loading the corresponding feature column values ​​of all samples into memory and participating in parameter calculation. For example, all sample corresponding feature columns are divided into groups f1 to fn+1, where n is a positive integer greater than or equal to 1. Essentially, this method divides the number of samples a process or thread needs to process into multiple parts, significantly reducing single-machine resource consumption. For instance, for a set of 100 feature samples, compared to a distributed scheme dividing them into 10 feature groups in the traditional mode, the traditional mode consumes 10 times more resources per machine than the distributed mode, while the distributed mode significantly reduces hardware requirements.

[0130] The feature data partitioning is done by the master process, and the feature column names for the corresponding groups are assigned when workers request them. This is equivalent to vertically splitting the workload of the training task on one side.

[0131] Based on the vertical splitting task, the master node on the first server side calculates the first and second derivatives of the sample, and at the same time uses a homomorphic encryption algorithm to encrypt the data before passing it to the master node on the second server side.

[0132] Feature column dimension calculation: Each worker node can obtain the derivative vector of the sample from the master node and calculate the gradient histogram of the assigned feature column accordingly. Then, the information gain and Gini coefficient of each feature at each value are calculated, thereby calculating the optimal split point of this feature. The worker performs calculations on all assigned features, forming a vector of optimal split points for each feature, and passes it to the master. The specific calculation method is that the first server calculates the information gain based on each threshold in the gradient histogram. If the gain obtained by a certain threshold is better than the previously recorded value, the optimal threshold of that feature is updated. Finally, the optimal thresholds of all features are found, forming a vector similar to: [optimal split point value of feature 1 to optimal split point value of feature n].

[0133] Parallel computation of tree node dimensions: Because each feature is independent, a feature column dimension computation strategy can be used for parallel feature computation. Different nodes in the same layer are also independent of each other, so a parallel strategy can be used to compute the tasks under all nodes in the same layer at the same time to achieve the purpose of accelerating computation.

[0134] The first server sends the derivative histogram calculated based on its own features to the second server. The first server can then calculate the optimal split point for the features on the second server. Having already obtained the optimal split point vectors for all features on its side, the first server will ultimately find the optimal feature split point for the current iteration. The first server then passes the optimal split point to the second server (master). The optimal split point vector on the second server is calculated using the same method as on the first server, except that the first server performs the calculation instead, providing only the gradient histogram of the features. The first server calculates the optimal split point vector for the second server based on the gradient histogram provided by the second server and merges it with its own data to find the globally optimal feature split point. This feature could be from either the second or first server. With the optimal split point, samples on nodes can be split according to the actual feature values, causing the tree to grow—the tree model generation process. Subsequent splits are performed on the newly generated nodes (repeating the above steps to find the optimal feature split point). When the number of leaf nodes reaches the configured value, the entire tree is complete. The next iteration is performed using the same algorithm to obtain a new tree, and finally all the trees are combined to form a complete model.

[0135] Each split of the tree node will form a new leaf node. The final shape of the decision tree is determined by the master node through merging the parameter information from the first and second servers and through multiple rounds of calculation.

[0136] In some embodiments, see Figure 10 , Figure 10 This is an optional flowchart illustrating the data processing method provided in an embodiment of the present invention, which will be explained in conjunction with each step.

[0137] S201, Receive the gradient vector sent by the first server.

[0138] In this embodiment of the invention, the second server receives the gradient vector sent by the first server.

[0139] S202. Determine intermediate data based on multiple sample data from different locations after combining gradient vectors with feature grouping.

[0140] In this embodiment of the invention, the second server determines intermediate data based on multiple remote sample data grouped by combining gradient vectors with features.

[0141] S203. The intermediate data is sent to the first server, which then trains multiple tree models based on the intermediate data and the multiple sample data calculated in groups, and then combines the multiple tree models to form a decision tree model.

[0142] In this embodiment of the invention, the second server sends intermediate data to the first server, which then trains multiple tree models based on the intermediate data and multiple sample data calculated in groups, and combines the multiple tree models to form a decision tree model.

[0143] Since each tree model in the decision tree model of this solution requires intermediate data sent by a second server (i.e., the partner) to be jointly trained with multiple sample data calculated in groups, the basic data of the second server is not required, thus preventing the other party's data from being leaked. Moreover, since multiple sample data are calculated in groups, the computational resources occupied are less. Therefore, this solution can effectively prevent the partner's data from being leaked and improve computational efficiency.

[0144] In some embodiments, see Figure 11 , Figure 11 This is an optional flowchart illustrating the data processing method provided in an embodiment of the present invention. Figure 10 S204 is also included before S201 shown, and will be explained in conjunction with each step.

[0145] S204. Obtain multiple remote sample data, and group the multiple remote features contained in the multiple remote sample data to obtain multiple remote feature groups; each remote feature group includes: multiple remote feature values ​​of at least one remote feature corresponding to the multiple remote sample data.

[0146] In this embodiment of the invention, a second server acquires multiple remote sample data sets, and groups the multiple remote features contained in each of the multiple remote sample data sets to obtain multiple remote feature groups. Each remote feature group includes multiple remote feature values ​​of at least one remote feature corresponding to the multiple remote sample data sets.

[0147] In some embodiments, see Figure 11 , Figure 11 This is an optional flowchart illustrating the data processing method provided in an embodiment of the present invention. Figure 10 The shown S202 can be implemented through S205 to S207, which will be explained in conjunction with each step.

[0148] S205. Obtain multiple remote thresholds for multiple remote feature values ​​of each remote feature. Construct multiple remote histograms corresponding to each remote feature according to the order of the remote feature group and multiple remote thresholds. Divide the multiple remote feature values ​​corresponding to each remote feature into the corresponding remote histograms to obtain multiple intermediate remote gradient histograms corresponding to multiple remote features.

[0149] In this embodiment of the invention, the second server obtains multiple remote thresholds for multiple remote feature values ​​of each remote feature, constructs multiple remote histograms corresponding to each remote feature according to the order of the remote feature group and multiple remote thresholds, and divides the multiple remote feature values ​​corresponding to each remote feature into the corresponding remote histograms to obtain multiple intermediate remote gradient histograms corresponding to multiple remote features.

[0150] S206. Calculate multiple gradients of multiple sample data from different locations based on the gradient vector, and accumulate the multiple gradients in the intermediate gradient histograms of the corresponding numerical ranges to obtain multiple gradient histograms of different locations.

[0151] In this embodiment of the invention, the second server calculates multiple gradients of multiple remote sample data based on the gradient vector, and accumulates the multiple remote gradients in the intermediate remote gradient histograms within the corresponding numerical range to obtain multiple remote gradient histograms.

[0152] S207. Combine multiple remote gradient histograms with multiple remote thresholds to form intermediate data.

[0153] In this embodiment of the invention, the second server combines multiple remote gradient histograms with multiple remote thresholds to form intermediate data.

[0154] In some embodiments, see Figure 11 , Figure 11 This is an optional flowchart illustrating the data processing method provided in an embodiment of the present invention. Figure 10The implementation of S203 is followed by S208 to S211, which will be explained in conjunction with each step.

[0155] S208, Receive the optimal split point threshold sent by the first server.

[0156] In this embodiment of the invention, the second server receives the optimal split point threshold sent by the first server.

[0157] S209. Determine multiple third identifier information of multiple first off-site sample data corresponding to multiple first off-site feature values ​​that are less than the optimal split point threshold.

[0158] In this embodiment of the invention, the second server determines multiple third identifiers of multiple first off-site sample data corresponding to multiple first off-site feature values ​​that are less than the optimal split point threshold.

[0159] S210. Determine multiple fourth identifiers for multiple second off-site sample data corresponding to multiple second off-site feature values ​​that are greater than the optimal split point threshold.

[0160] In this embodiment of the invention, the second server determines multiple fourth identification information of multiple second off-site sample data corresponding to multiple second off-site feature values ​​that are greater than the optimal split point threshold.

[0161] S211. Send multiple third identification information and multiple fourth identification information to the first server.

[0162] In this embodiment of the invention, the second server sends multiple third identification information and multiple fourth identification information to the first server.

[0163] In some embodiments, see Figure 12 , Figure 12 This is an interactive schematic diagram of the data processing method provided in the embodiments of the present invention, which will be explained in conjunction with each step.

[0164] S301. The first server calculates the gradient vector based on multiple business parameters of multiple sample data, and determines the split point vectors corresponding to multiple features after feature grouping.

[0165] The detailed implementation of step S301 is the same as that of S103, and will not be repeated here.

[0166] S302. The first server sends the gradient vector to the second server, and receives multiple remote gradient histograms and their corresponding remote thresholds from the second server based on the gradient vector, and then calculates the remote split point vector.

[0167] The detailed implementation of step S302 is the same as that of S104, and will not be repeated here.

[0168] S303. The first server determines the optimal split point threshold by combining the split point vector and the remote split point vector, and performs iterative splitting of multiple sample data or multiple remote sample data until the first predetermined condition is met, and then stops to obtain the current tree model.

[0169] The detailed implementation of step S303 is the same as that of S105, and will not be repeated here.

[0170] S304. The first server uses the current tree model to process multiple residuals of multiple sample data, and then trains the next tree model based on the multiple residuals until the second predetermined condition is met, and the training stops, resulting in multiple tree models. The multiple tree models are then combined to form a decision tree model.

[0171] The detailed implementation of step S304 is the same as that of S106, and will not be repeated here.

[0172] In some embodiments, see Figure 13 , Figure 13 Schematic diagram of the data processing apparatus provided in the embodiments of the present invention Figure 1 .

[0173] This invention also provides a data processing device 800, including a data acquisition unit 803 and a processing unit 804.

[0174] Data acquisition unit 803 is used to acquire data to be processed; Processing unit 804 is used to process the data to be processed using a decision tree model to obtain the processing result of the data to be processed; wherein, Each tree model in the decision tree model is trained using intermediate data corresponding to multiple sample data calculated in groups and multiple remote sample data sent by the second server.

[0175] In this embodiment of the invention, the data processing device 800 is further configured to determine, based on the gradient vectors calculated from the multiple business parameters of the multiple sample data, the split point vectors of multiple features corresponding to the multiple sample data after feature grouping, and the off-site split point vectors of multiple off-site features corresponding to the multiple off-site sample data; each sample data includes: corresponding business parameters and feature values ​​of multiple features; each off-site sample data includes: off-site feature values ​​corresponding to multiple off-site features; combining the split point vectors and the off-site split point vectors to determine the optimal split point threshold, iteratively splitting the multiple sample data or the multiple off-site sample data until a first predetermined condition is met, thus obtaining the current tree model; using the current tree model to process and obtain multiple residuals of the multiple sample data, and then training the next tree model based on the multiple residuals until a second predetermined condition is met, thus obtaining multiple tree models; and combining the obtained multiple tree models to form the decision tree model.

[0176] In this embodiment of the invention, the data processing device 800 is further configured to determine the split point vectors corresponding to multiple features after feature grouping based on the gradient vectors calculated from the gradient vectors of multiple business parameters of the multiple sample data; each sample data includes: identification information, business parameters, and feature values ​​of multiple features; the gradient vector is sent to the second server, and the second server receives multiple remote gradient histograms and their corresponding multiple remote thresholds fed back by the gradient vector, thereby calculating the remote split point vector; the multiple remote gradient histograms are determined by the second server based on multiple remote feature values ​​of the multiple remote sample data; the multiple sample data and the multiple remote sample data include the same identification information; the multiple remote gradient histograms correspond to multiple remote features; the intermediate data includes: the multiple remote gradient histograms and their corresponding multiple remote thresholds; In this embodiment of the invention, the data acquisition unit 803 in the data processing device 800 is further configured to acquire multiple sample data, group the multiple features contained in the multiple sample data into multiple feature groups; each feature group includes multiple feature values ​​of at least one feature corresponding to the multiple sample data; according to the gradient vector, calculate at least one optimal threshold corresponding to the at least one feature in each feature group in the order of the feature groups, thereby obtaining multiple optimal thresholds corresponding to the multiple features; sort the multiple optimal thresholds to form the split point vector.

[0177] In this embodiment of the invention, the processing unit 804 in the data processing device 800 is further configured to obtain multiple thresholds for the multiple feature values ​​of each feature, construct multiple histograms corresponding to each feature according to the multiple thresholds in the order of the feature group, divide the multiple feature values ​​corresponding to each feature into the corresponding histograms, and obtain multiple intermediate gradient histograms corresponding to the multiple features; calculate multiple gradients of the multiple sample data according to the gradient vector, and accumulate the multiple gradients in the intermediate gradient histograms within the corresponding numerical ranges to obtain multiple gradient histograms; determine the current threshold among the multiple thresholds and calculate the current information gain; if the current information gain is smaller than the previous information gain, iterate the threshold according to the new threshold corresponding to each gradient histogram until the current information gain is larger than the previous information gain, obtain the optimal threshold corresponding to each gradient histogram, and thus obtain the at least one optimal threshold.

[0178] In this embodiment of the invention, the processing unit 804 in the data processing device 800 is further configured to encrypt and send the gradient vector to the second server, receive the second server to determine the current threshold from the multiple remote gradient histograms and their corresponding multiple remote thresholds based on the gradient vector feedback, and calculate the current remote information gain; if the current remote information gain is less than the previous remote information gain, then the remote threshold is iterated based on the new remote threshold corresponding to each remote gradient histogram until the current remote information gain is greater than the previous current remote information gain, thus obtaining the optimal remote threshold corresponding to each remote gradient histogram, and thereby obtaining multiple optimal remote thresholds for the multiple remote features; the multiple optimal remote thresholds are sorted to form the remote split point vector.

[0179] In this embodiment of the invention, the processing unit 804 in the data processing device 800 is further configured to determine the maximum information gain among the multiple current information gains corresponding to multiple optimal thresholds included in the split point vector, and among the multiple off-site information gains corresponding to multiple optimal off-site thresholds included in the off-site split point vector; determine the optimal threshold or optimal off-site threshold corresponding to the maximum information gain as the optimal split point threshold; if the optimal split point threshold belongs to the split point vector, then divide the first identification information and the multiple features corresponding to multiple first feature values ​​less than the optimal split point threshold into left nodes, divide the multiple second identification information and the multiple features corresponding to multiple second feature values ​​greater than the optimal split point threshold into right nodes, and sequentially perform iterative splitting of the multiple sample data until a predetermined number of splits is reached, thereby obtaining the current tree model; the multiple first feature values ​​and the multiple second feature values ​​belong to the feature values ​​of the corresponding features of the optimal split point threshold; the multiple first identification information are the identification information corresponding to the multiple first sample data corresponding to the multiple first feature values; the multiple second identification information are the identification information corresponding to the multiple second sample data corresponding to the multiple second feature values.

[0180] In this embodiment of the invention, the processing unit 804 in the data processing device 800 is further configured to: if the optimal split point threshold belongs to the off-site split point vector, send the optimal split point threshold to the second server; receive multiple third identification information and multiple fourth identification information from the second server based on the off-site feature values ​​corresponding to the off-site features of the optimal split point threshold; the multiple third identification information are identification information corresponding to multiple first off-site sample data corresponding to multiple first off-site feature values ​​less than the optimal split point threshold; the multiple fourth identification information are identification information corresponding to multiple second off-site sample data corresponding to multiple second off-site feature values ​​greater than the optimal split point threshold; divide the multiple third identification information and the multiple off-site features into left nodes, divide the multiple fourth identification information and the multiple off-site features into right nodes, and perform iterative splitting of the multiple off-site sample data sequentially until a predetermined number of splits is reached, thereby obtaining the current tree model.

[0181] In this embodiment of the invention, the data to be processed is acquired by the data acquisition unit 803; the data to be processed is processed by the processing unit 804 using a decision tree model to obtain the processing result; wherein, each tree model in the decision tree model is trained by multiple sample data calculated in groups and intermediate data corresponding to multiple remote sample data sent by the second server. Since each tree model in the decision tree model of this solution requires intermediate data sent by the second server (i.e., the partner) and is jointly trained with multiple sample data calculated in groups, the basic data of the second server is not required, thus preventing the leakage of the other party's data. Moreover, since multiple sample data are calculated in groups, the computational resources occupied are less. Therefore, this solution can effectively prevent the leakage of partner's data and improve computational efficiency.

[0182] It should be noted that, in the embodiments of the present invention, if the above-described data processing method is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of the present invention, or the part that contributes to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a data processing device (which may be a personal computer, etc.) to execute all or part of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a magnetic disk, or an optical disk. Thus, the embodiments of the present invention are not limited to any specific hardware and software combination.

[0183] Correspondingly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method.

[0184] Correspondingly, this embodiment of the invention provides a data processing device 800, including a first memory 802 and a first processor 801. The first memory 802 stores a computer program that can run on the first processor 801. When the first processor 801 executes the program, it implements the steps in the above method.

[0185] It should be noted that the descriptions of the storage medium and device embodiments above are similar to those of the method embodiments above, and have similar beneficial effects. For technical details not disclosed in the storage medium and device embodiments of the present invention, please refer to the descriptions of the method embodiments of the present invention for understanding.

[0186] It should be noted that, Figure 14 A hardware entity illustration of the data processing device provided in the embodiments of the present invention. Figure 1 ,like Figure 14 As shown, the hardware entity of the data processing device 800 includes: a first processor 801 and a first memory 802, wherein; The first processor 801 typically controls the overall operation of the data processing device 800.

[0187] The first memory 802 is configured to store instructions and applications executable by the processor 801, and can also cache data to be processed or already processed by the first processor 801 and the various modules in the data processing device 800 (e.g., image data, audio data, voice communication data and video communication data), which can be implemented by flash memory or random access memory (RAM).

[0188] In some embodiments, see Figure 15 , Figure 15 Schematic diagram of the data processing apparatus provided in the embodiments of the present invention Figure 2 .

[0189] This invention also provides a data processing device 900, including: a receiving unit 903, a determining unit 904, and a sending unit 905.

[0190] The receiving unit 903 is used to receive the gradient vector sent by the first server; The determining unit 904 is used to determine intermediate data based on the gradient vector combined with the feature grouped data of multiple samples from different locations. The sending unit 905 is used to send the intermediate data to the first server, so that the first server can train multiple tree models based on the intermediate data and multiple sample data calculated in groups, and then combine the multiple tree models to form a decision tree model.

[0191] In this embodiment of the invention, the receiving unit 903 in the data processing device 900 is used to acquire multiple remote sample data, group the multiple remote features contained in the multiple remote sample data into multiple remote feature groups; each remote feature group includes: multiple remote feature values ​​of at least one remote feature corresponding to the multiple remote sample data; the determining unit 904 is used to acquire multiple remote thresholds of the multiple remote feature values ​​of each remote feature, construct multiple remote histograms corresponding to each remote feature according to the order of the remote feature groups and the multiple remote thresholds, divide the multiple remote feature values ​​corresponding to each remote feature into the corresponding remote histograms, and obtain multiple intermediate remote gradient histograms corresponding to the multiple remote features; calculate multiple remote gradients of the multiple remote sample data according to the gradient vector, accumulate the multiple remote gradients into the intermediate remote gradient histograms within the corresponding numerical range, and obtain multiple remote gradient histograms; combine the multiple remote gradient histograms with the multiple remote thresholds to form the intermediate data.

[0192] In this embodiment of the invention, the receiving unit 903 in the data processing device 900 is used to receive the optimal split point threshold sent by the first server; the determining unit 904 is used to determine multiple third identification information of multiple first off-site sample data corresponding to multiple first off-site feature values ​​that are less than the optimal split point threshold; and to determine multiple fourth identification information of multiple second off-site sample data corresponding to multiple second off-site feature values ​​that are greater than the optimal split point threshold; the sending unit 905 is used to send the multiple third identification information and the multiple fourth identification information to the first server.

[0193] In this embodiment of the invention, a receiving unit 903 receives a gradient vector sent by a first server; a determining unit 904 determines intermediate data based on the gradient vector and multiple remote sample data grouped by features; and a sending unit 905 sends the intermediate data to the first server, which then trains multiple tree models based on the intermediate data and the multiple sample data calculated in groups, and combines the multiple tree models to form a decision tree model. Since each tree model in this solution requires intermediate data sent by a second server (i.e., a partner) for joint training with the multiple sample data calculated in groups, the solution does not require the second server's basic data, thus preventing data leakage. Furthermore, because the multiple sample data are calculated in groups, the computational resources required are less. Therefore, this solution can effectively prevent data leakage from partners and improve computational efficiency.

[0194] Correspondingly, this embodiment of the invention provides a data processing device 900, including a second memory 902 and a second processor 901. The second memory 902 stores a computer program that can run on the second processor 901. When the second processor 901 executes the program, it implements the steps in the above-described method.

[0195] It should be noted that the descriptions of the storage medium and device embodiments above are similar to those of the method embodiments above, and have similar beneficial effects. For technical details not disclosed in the storage medium and device embodiments of the present invention, please refer to the descriptions of the method embodiments of the present invention for understanding.

[0196] It should be noted that, Figure 16 A hardware entity illustration of the data processing device provided in the embodiments of the present invention. Figure 2 ,like Figure 16 As shown, the hardware entity of the data processing device 900 includes: a second processor 901 and a second memory 902, wherein; The second processor 901 typically controls the overall operation of the data processing device 900.

[0197] The second memory 902 is configured to store instructions and applications executable by the processor 901, and can also cache data to be processed or already processed by the second processor 901 and the various modules in the data processing device 900 (e.g., image data, audio data, voice communication data and video communication data), which can be implemented by flash memory or random access memory (RAM).

[0198] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of the invention. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of the invention, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the invention. The sequence numbers of the above-described embodiments of the invention are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0199] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0200] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the apparatus or units can be electrical, mechanical, or other forms.

[0201] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0202] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0203] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0204] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0205] The above description is merely an embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A data processing method, characterized in that, include: Based on the gradient vectors calculated from multiple business parameters of multiple sample data, the split point vectors corresponding to multiple features after feature grouping are determined. The gradient vector is sent to the second server, and the second server receives multiple remote gradient histograms and their corresponding remote thresholds based on the gradient vector, thereby calculating the remote split point vector; the intermediate data includes: the multiple remote gradient histograms and their corresponding remote thresholds. The optimal splitting point threshold is determined by combining the splitting point vector and the off-site splitting point vector. Iterative splitting of the multiple sample data or multiple off-site sample data is performed until the first predetermined condition is met, and the current tree model is obtained. The current tree model is used to process the multiple sample data to obtain multiple residuals, and then the next tree model is trained based on the multiple residuals until the training stops when the second predetermined condition is met, resulting in multiple tree models. The multiple tree models are then combined to form a decision tree model. Obtain data to be processed from other servers; wherein, the data to be processed consists of multiple attribute data of the corresponding target; The data to be processed is processed using a decision tree model to obtain the processing result.

2. The data processing method according to claim 1, characterized in that, The multiple remote gradient histograms are determined by the second server based on multiple remote feature values ​​of the multiple remote sample data; the multiple sample data and the multiple remote sample data include the same identification information; the multiple remote gradient histograms correspond to multiple remote features.

3. The data processing method according to claim 1, characterized in that, Before determining the split point vectors corresponding to the multiple features after feature grouping by using the gradient vectors calculated based on the multiple business parameters of the multiple sample data, the method further includes: Multiple sample data are acquired, and the multiple features contained in the multiple sample data are grouped to obtain multiple feature groups; each feature group includes multiple feature values ​​of at least one feature corresponding to the multiple sample data. The step of determining the split point vectors corresponding to multiple features after feature grouping by calculating the gradient vector based on multiple business parameters of the multiple sample data includes: Based on the gradient vector, at least one optimal threshold corresponding to at least one feature in each feature group is calculated in the order of the feature groups, thereby obtaining multiple optimal thresholds corresponding to the multiple features; The multiple optimal thresholds are sorted to form the split point vector.

4. The data processing method according to claim 3, characterized in that, The step of calculating at least one optimal threshold corresponding to at least one feature in each feature group according to the gradient vector and in the order of the feature groups includes: Multiple thresholds for the multiple feature values ​​of each feature are obtained. Based on the multiple thresholds in the order of the feature group, multiple histograms corresponding to each feature are constructed. The multiple feature values ​​corresponding to each feature are divided into the corresponding histograms to obtain multiple intermediate gradient histograms corresponding to the multiple features. Multiple gradients of the multiple sample data are calculated based on the gradient vector, and the multiple gradients are respectively accumulated in the intermediate gradient histograms within the corresponding numerical range to obtain multiple gradient histograms; The current threshold is determined from the plurality of thresholds, and the current information gain is calculated; If the current information gain is less than the previous information gain, then the threshold is iterated according to the new threshold corresponding to the gradient histogram until the current information gain is greater than the previous current information gain, and then the optimal threshold corresponding to the gradient histogram is obtained, thereby obtaining the at least one optimal threshold.

5. The data processing method according to claim 1, characterized in that, The step of sending the gradient vector to the second server, receiving multiple off-site gradient histograms and their corresponding off-site thresholds from the second server based on the gradient vector, and then calculating the off-site split point vector includes: The gradient vector is encrypted and sent to the second server, and the second server is received multiple remote gradient histograms and their corresponding multiple remote thresholds based on the gradient vector. The current threshold is determined from the plurality of remote thresholds, and the current remote information gain is calculated; If the current off-site information gain is smaller than the previous off-site information gain, then the off-site threshold is iterated according to the new off-site threshold corresponding to the off-site gradient histogram until the current off-site information gain is larger than the previous off-site information gain, and the optimal off-site threshold corresponding to the off-site gradient histogram is obtained, thereby obtaining multiple optimal off-site thresholds for multiple off-site features. The multiple optimal off-site thresholds are sorted to form the off-site split point vector.

6. The data processing method according to claim 1, characterized in that, The optimal split point threshold determined by combining the split point vector and the off-site split point vector is used to iteratively split the multiple sample data or the multiple off-site sample data until a first predetermined condition is met, thereby obtaining the current tree model, including: Among the multiple current information gains corresponding to multiple optimal thresholds included in the split point vector, the maximum information gain is determined from among the multiple off-site information gains corresponding to multiple optimal off-site thresholds included in the off-site split point vector. The optimal threshold or optimal off-site threshold corresponding to the maximum information gain is determined as the optimal split point threshold; If the optimal split point threshold belongs to the split point vector, then the first identification information and the multiple features corresponding to the multiple first feature values ​​that are less than the optimal split point threshold are divided into left nodes, and the multiple second identification information and the multiple features corresponding to the multiple second feature values ​​that are greater than the optimal split point threshold are divided into right nodes. The multiple sample data are iteratively split until the predetermined number of splits is reached, and then the current tree model is obtained. Wherein, the plurality of first feature values ​​and the plurality of second feature values ​​belong to the feature values ​​of the corresponding features of the optimal split point threshold; the first identification information is the identification information corresponding to the plurality of first sample data corresponding to the plurality of first feature values; the plurality of second identification information is the identification information corresponding to the plurality of second sample data corresponding to the plurality of second feature values.

7. The data processing method according to claim 6, characterized in that, After determining the optimal threshold or optimal off-site threshold corresponding to the maximum information gain, and the optimal split point threshold, the method further includes: If the optimal split point threshold belongs to the off-site split point vector, then the optimal split point threshold is sent to the second server, and the second server is received multiple third identifiers and multiple fourth identifiers based on the off-site feature values ​​corresponding to the optimal split point threshold; wherein, The plurality of third identification information are a plurality of first off-site feature values ​​that are less than the optimal split point threshold, and the identification information corresponding to the plurality of first off-site sample data; the plurality of fourth identification information are a plurality of second off-site feature values ​​that are greater than the optimal split point threshold, and the identification information corresponding to the plurality of second off-site sample data. The multiple third identifiers and multiple remote features are divided into left nodes, and the multiple fourth identifiers and multiple remote features are divided into right nodes. The multiple remote sample data are iteratively split in sequence until a predetermined number of splits is reached, and then the current tree model is obtained.

8. The data processing method according to claim 1, characterized in that, The process involves using the current tree model to obtain multiple residuals from the multiple sample data, then training the next tree model based on these residuals, until training stops when a second predetermined condition is met, resulting in multiple tree models, including: The multiple sample data are input into the current tree model to obtain the multiple residuals; The current gradient vector is calculated based on the multiple residuals. Then, the next tree model is determined using the current gradient vector. The process continues until the number of tree models is greater than a predetermined number or the multiple residuals are less than a residual threshold, thus obtaining the multiple tree models.

9. A data processing method, characterized in that, include: Receive the gradient vector sent by the first server; Multiple remote thresholds are obtained for multiple remote feature values ​​of each remote feature. Based on the multiple remote thresholds in the order of the remote feature group, multiple remote histograms corresponding to each remote feature are constructed. The multiple remote feature values ​​corresponding to each remote feature are divided into the corresponding remote histograms to obtain multiple intermediate remote gradient histograms corresponding to the multiple remote features. Multiple gradients of multiple remote sample data are calculated based on the gradient vector. The multiple gradients of multiple remote samples are then accumulated in the intermediate gradient histograms of the corresponding numerical ranges to obtain multiple gradient histograms of multiple remote samples. The multiple geographical gradient histograms and the multiple geographical thresholds are combined to form intermediate data; The intermediate data is sent to the first server, which then trains multiple tree models based on the intermediate data and the multiple sample data calculated in groups. The multiple tree models are then combined to form a decision tree model. The types of the multiple sample data, the multiple remote sample data and the target data to be processed are the same.

10. The data processing method according to claim 9, characterized in that, Before determining intermediate data based on the gradient vector and multiple cross-regional sample data grouped by feature grouping, the method further includes: Multiple off-site sample data are acquired, and the multiple off-site features contained in the multiple off-site sample data are grouped to obtain multiple off-site feature groups; each off-site feature group includes multiple off-site feature values ​​of at least one off-site feature corresponding to the multiple off-site sample data.

11. The data processing method according to claim 9, characterized in that, After determining the intermediate data based on the gradient vector and the multiple cross-regional sample data grouped by feature grouping, the method further includes: Receive the optimal split point threshold sent by the first server; Determine multiple third identifiers for multiple first off-site sample data corresponding to multiple first off-site feature values ​​that are less than the optimal split point threshold; Determine multiple fourth identifiers for multiple second off-site sample data corresponding to multiple second off-site feature values ​​that are greater than the optimal split point threshold; The plurality of third identification information and the plurality of fourth identification information are sent to the first server.

12. A data processing apparatus, characterized in that, include: The processing unit is used to determine the split point vectors corresponding to multiple features after feature grouping based on the gradient vectors calculated from multiple business parameters of multiple sample data. The processing unit is used to send the gradient vector to the second server, receive multiple remote gradient histograms and their corresponding multiple remote thresholds fed back by the second server based on the gradient vector, and then calculate the remote split point vector; the intermediate data includes: the multiple remote gradient histograms and their corresponding multiple remote thresholds. The processing unit is used to combine the split point vector and the optimal split point threshold determined by the off-site split point vector to perform iterative splitting of the multiple sample data or multiple off-site sample data until the first predetermined condition is met, and then stop to obtain the current tree model. The processing unit is used to process the current tree model to obtain multiple residuals of the multiple sample data, and then train the next tree model based on the multiple residuals until the training stops when the second predetermined condition is met, thereby obtaining multiple tree models, and combining the multiple tree models to form a decision tree model. A data acquisition unit is used to acquire data to be processed from other servers; wherein the data to be processed is multiple attribute data of a corresponding target; The processing unit is used to process the data to be processed using a decision tree model to obtain the processing result of the data to be processed.

13. A data processing apparatus, characterized in that, include: A receiving unit is used to receive gradient vectors sent by the first server; The determining unit is used to obtain multiple remote thresholds for multiple remote feature values ​​of each remote feature, construct multiple remote histograms corresponding to each remote feature according to the order of the remote feature group and the multiple remote thresholds, and divide the multiple remote feature values ​​corresponding to each remote feature into the corresponding remote histograms to obtain multiple intermediate remote gradient histograms corresponding to the multiple remote features. The determining unit is used to calculate multiple gradients of multiple remote sample data based on the gradient vector, and to accumulate the multiple remote gradients in the intermediate remote gradient histograms within the corresponding numerical range to obtain multiple remote gradient histograms. A determining unit is used to combine the plurality of off-site gradient histograms with the plurality of off-site thresholds to form intermediate data; The sending unit is used to send the intermediate data to the first server, so that the first server can train multiple tree models based on the intermediate data and multiple sample data calculated in groups, and then combine the multiple tree models to form a decision tree model; wherein, the multiple sample data, the multiple remote sample data and the target corresponding to the data to be processed are of the same type.

14. A data processing apparatus, characterized in that, It includes a first memory and a first processor, the first memory storing a computer program that can run on the first processor, and the first processor executing the program implementing the steps of the method according to any one of claims 1 to 8.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the first processor, it implements the steps of the method according to any one of claims 1 to 8.

16. A data processing apparatus, characterized in that, It includes a second memory and a second processor, the second memory storing a computer program that can run on the second processor, the second processor executing the program to implement the steps of the method according to any one of claims 9 to 11.

17. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the second processor, it implements the steps of the method according to any one of claims 9 to 11.

Images