Method and device for determining user classification labels

Abstract

The invention discloses a method and device for determining a user classification label, and the method comprises the steps: determining a first virtual label and a second virtual label based on a predicted label and a real label of a sample user in the ith round, and dividing the first virtual label and the second virtual label into a first virtual sub-label and a second virtual sub-label respectively; sending the first virtual sub-tag and the second virtual sub-tag to each second participant; receiving a first grouping label value and a second grouping label value sent by each second participant; respectively summarizing the first grouping label value and the second grouping label value into a first characteristic value and a second characteristic value according to the second dimension characteristic of each second participant; obtaining a decision tree corresponding to the sample user according to the first feature value and the second feature value; the decision tree is used for determining classification labels of non-sample users. By adopting the method, the communication overhead and the calculation complexity can be reduced on the premise of security calculation of a plurality of participants.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and in particular to a method and device for determining user classification labels. Background Art

[0002] In the field of secure computing, there are many similar scenarios where multiple participants need to be united to perform classification or other tasks while meeting the privacy protection requirement. For example, each participant has one or more characteristics of the user, and the participants do not want their own user characteristics to be known by the other party, but each participant needs to combine the user characteristics of other participants to jointly determine the user's classification label.

[0003] The traditional method is that each participant encrypts its own user features and determines the user's classification label based on the encrypted user features. However, this method involves encryption and decryption processes, resulting in high computational complexity and high communication overhead. Summary of the invention

[0004] The present application provides a method and device for determining user classification labels, which are used to reduce communication overhead and computational complexity under the premise of secure computing by multiple participants.

[0005] In a first aspect, an embodiment of the present application provides a method for determining a user classification label, the method being applied to a first participant having a classification label of a sample user, the method being executable by a device for determining a user classification label, the method comprising: for at least one sample user, based on the predicted label of the sample user in the i-th round and the real label of the sample user, determining a first virtual label of the sample user and a second virtual label of the sample user, and splitting the first virtual label of the sample user and the second virtual label of the sample user into a first virtual sub-label corresponding to each second participant and a second virtual sub-label corresponding to each second participant, respectively; each second participant having a second dimensional feature of the sample user and setting a feature value of the sample user under each second dimensional feature to a group number according to a self-defined grouping method; sending to each second participant a message in the at least one sample user's first virtual label and a second virtual sub-label corresponding to each second participant's first virtual label; The method comprises the following steps: receiving a first virtual sub-label and a second virtual sub-label corresponding to each sample user; receiving a first grouping label value and a second grouping label value for each second dimensional feature sent by each second participant, wherein the first grouping label value is a summary value of the first virtual sub-labels of the at least one sample user under the same group number, and the second grouping label value is a summary value of the second virtual sub-labels of the at least one sample user under the same group number; for the second dimensional feature of each second participant, summarizing the first grouping label value and the second grouping label value of each second participant under the second dimensional feature into the first feature value and the second feature value of the second dimensional feature under each group number respectively; obtaining a decision tree corresponding to the at least one sample user according to the first feature value and the second feature value of each second dimensional feature under each group number; and the decision tree is used to determine the classification label of the non-sample user.

[0006] By adopting the above method, two virtual labels of sample users are calculated based on the real labels and predicted labels of sample users, and then each virtual label is split to form virtual sub-labels, so that each second participant calculates the first feature value and the second feature value under each dimension feature based on the virtual sub-label and the feature grouping number. Based on the first feature and the second feature, the decision tree corresponding to the sample user is obtained. In this way, the first participant protects the real label of the user, and the second participant protects the user feature information by grouping the second dimension features. The entire training process only splits the label, which reduces the computational complexity and communication overhead compared to the traditional method of training decision trees based on encryption and decryption.

[0007] In a possible implementation, before obtaining the decision tree corresponding to the at least one sample user, it also includes: for each first dimensional feature of the first participant, setting the characteristic value of the sample user under each first dimensional feature as the corresponding group number according to the grouping method set by the first participant; determining the first grouping label value and the second grouping label value for each first dimensional feature; obtaining the decision tree corresponding to the at least one sample user according to the first characteristic value and the second characteristic value of each second dimensional feature under each group number, including: obtaining the decision tree corresponding to the at least one sample user according to the first grouping label value and the second grouping label value of each first dimensional feature, and the first characteristic value and the second characteristic value of each second dimensional feature under each group number.

[0008] By adopting the above method, when the first participant also has part of the characteristic information of the sample user, the decision tree can also be trained in combination with the user characteristic information of the first participant.

[0009] In a possible implementation, a decision tree corresponding to the at least one sample user is obtained, including: determining multiple candidate split nodes of the decision tree according to each group number of each first dimensional feature and each group number of each second dimensional feature; for each candidate split node, obtaining a first sub-gain based on a first eigenvalue and a second eigenvalue of a group number belonging to a first side of the candidate split node, obtaining a second sub-gain based on a first eigenvalue and a second eigenvalue of a group number belonging to a second side of the candidate split node, and obtaining a third sub-gain based on the first eigenvalue and the second eigenvalue of all group numbers of the candidate split node; determining a gain value of the candidate split node based on the first sub-gain, the second sub-gain, and the third sub-gain; determining a candidate split node with the best gain value as a split node; until a decision tree is obtained.

[0010] In a possible implementation, the at least one sample user is a sample user sampled in the i-th round; after obtaining the decision tree, it also includes: determining the leaf node corresponding to the at least one sample user in the decision tree and the weight of the leaf node; based on the weight of the leaf node of the decision tree, determining the predicted label of the sample user sampled in the i+1th round under the decision tree, and updating the decision tree through the sample user sampled in the i+1th round.

[0011] By using the above method, the predicted labels corresponding to the leaf nodes can be continuously updated to make the predicted labels of the sample users closer to the real labels of the sample users. In this way, when the predicted labels of the sample users are close to the real labels, it can be judged that the decision tree training is completed.

[0012] In a possible implementation, the sum of the first virtual sub-tags corresponding to each second participant is equal to the first virtual tag of the sample user; the sum of the second virtual sub-tags corresponding to each second participant is equal to the second virtual tag of the sample user; for the second participant, the first grouping tag value is the sum of the first virtual sub-tags of the at least one sample user under the same group number; for the first participant, the first grouping tag value is the sum of the first virtual tags of the at least one sample user under the same group number; for the second participant, the second grouping tag value is the sum of the second virtual sub-tags of the at least one sample user under the same group number; for the first participant, the second grouping tag value is the sum of the second virtual tags of the at least one sample user under the same group number.

[0013] In a possible implementation, when there is only one second participant, a third participant who does not need to provide the second dimensional feature is set as another second participant.

[0014] In a second aspect, an embodiment of the present application provides a device for determining a user classification label, the device comprising: a processing module for determining, for at least one sample user, a first virtual label and a second virtual label of the sample user based on the predicted label of the sample user in the i-th round and the real label of the sample user, and splitting the first virtual label of the sample user and the second virtual label of the sample user into a first virtual sub-label corresponding to each second participant and a second virtual sub-label corresponding to each second participant, respectively; each second participant has a second dimensional feature of the sample user and sets the feature value of the sample user under each second dimensional feature as a group number according to a self-defined grouping method; a sending module for sending the first virtual sub-label and the second virtual sub-label corresponding to each of the at least one sample user to each second participant, respectively; a receiving module A block is used to receive the first grouping label value and the second grouping label value for each second dimensional feature sent by each second participant, wherein the first grouping label value is the summary value of the first virtual sub-label of the at least one sample user under the same group number, and the second grouping label value is the summary value of the second virtual sub-label of the at least one sample user under the same group number; a summarizing module is used to summarize the first grouping label value and the second grouping label value of each second participant under the second dimensional feature into the first feature value and the second feature value of the second dimensional feature under each group number respectively for the second dimensional feature of each second participant; the processing module is also used to obtain a decision tree corresponding to the at least one sample user according to the first feature value and the second feature value of each second dimensional feature under each group number; the decision tree is used to determine the classification label of the non-sample user.

[0015] In a possible implementation, the processing module is also used to, for each first dimensional feature of the first participant, set the characteristic value of the sample user under each first dimensional feature as a corresponding group number according to the grouping method set by the first participant; determine the first grouping label value and the second grouping label value for each first dimensional feature; and obtain a decision tree corresponding to the at least one sample user based on the first grouping label value and the second grouping label value of each first dimensional feature and the first characteristic value and the second characteristic value of each second dimensional feature under each group number.

[0016] In a possible implementation, the processing module is specifically used to determine multiple candidate split nodes of the decision tree according to each group number of each first dimensional feature and each group number of each second dimensional feature; for each candidate split node, obtain a first sub-gain based on the first eigenvalue and the second eigenvalue of the group number belonging to the first side of the candidate split node, obtain a second sub-gain based on the first eigenvalue and the second eigenvalue of the group number belonging to the second side of the candidate split node, and obtain a third sub-gain based on the first eigenvalue and the second eigenvalue of all group numbers of the candidate split node; determine the gain value of the candidate split node based on the first sub-gain, the second sub-gain and the third sub-gain; determine the candidate split node with the best gain value as the split node; until a decision tree is obtained.

[0017] In a possible implementation, the at least one sample user is a sample user sampled in the i-th round; the processing module is also used to determine the leaf node corresponding to the at least one sample user in the decision tree and the weight of the leaf node; based on the weight of the leaf node of the decision tree, determine the predicted label of the sample user sampled in the i+1th round under the decision tree, and update the decision tree through the sample user sampled in the i+1th round.

[0018] In a possible implementation, the sum of the first virtual sub-tags corresponding to each second participant is equal to the first virtual tag of the sample user; the sum of the second virtual sub-tags corresponding to each second participant is equal to the second virtual tag of the sample user; for the second participant, the first grouping tag value is the sum of the first virtual sub-tags of the at least one sample user under the same group number; for the first participant, the first grouping tag value is the sum of the first virtual tags of the at least one sample user under the same group number; for the second participant, the second grouping tag value is the sum of the second virtual sub-tags of the at least one sample user under the same group number; for the first participant, the second grouping tag value is the sum of the second virtual tags of the at least one sample user under the same group number.

[0019] In a possible implementation, when there is only one second participant, a third participant who does not need to provide the second dimensional feature is set as another second participant.

[0020] In a third aspect, an embodiment of the present application further provides a device for determining a user classification label, the device comprising a memory and a processor, the memory comprising a computer program or instructions, the processor being used to call the computer program or instructions stored in the memory to execute a method as in any possible implementation of the first aspect.

[0021] In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium, wherein instructions are stored in the computer-readable storage medium. When a computer reads and executes the instructions, the computer executes a method in any possible implementation of the first aspect.

[0022] In a fifth aspect, the present invention provides a computer program product, wherein instructions are stored in the computer program product. When a computer reads and executes the instructions, the computer executes the method in any possible implementation of the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0024] Figure 1 A flowchart corresponding to a method for determining a user classification label proposed in an embodiment of the present application;

[0025] Figure 2 A histogram corresponding to a first eigenvalue based on age provided in an embodiment of the present application;

[0026] Figure 3 A schematic diagram of a decision tree provided in an embodiment of the present application;

[0027] Figure 4 A schematic diagram of internal modules of an apparatus 4000 for determining a user classification label provided in an embodiment of the present application;

[0028] Figure 5 A schematic diagram of the structure of a device 5000 for determining a user classification label provided in an embodiment of the present application. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0030] In the financial field, it is usually necessary to determine the user's classification label, such as whether the user is on the blacklist or the user's credit score. In order to more accurately determine the user's classification label, it is usually necessary to combine the user characteristics of multiple participants. Each participant does not want the user characteristics known to other participants. Therefore, multiple participants need to jointly determine the user's classification label under the premise of secure computing.

[0031] The participants are generally banks, insurance institutions, credit reporting agencies or other institutions that have user characteristic information. These participants can be divided into two categories, one is the institution that has the user's real label, and the other is the institution that has user characteristic information.

[0032] In a specific scenario, the first participant, such as a credit reporting agency, has the real labels of some users. The real labels can be whether the user belongs to the blacklist or the whitelist. The second participant, such as an insurance agency, and the third participant, such as a bank, have more characteristic information of users. The real label can also be other labels, such as whether the user is an ordinary user or a high-quality user, and whether the user's Sesame Score is a high-quality score, a medium score, or a low score. This application does not limit the type of user labels. The number of some users and more users is also not limited in this application. Some users can be 100,000 users, and more users can be 1 million users, including some users. The first participant provides the real labels y of 100,000 users and some characteristic information x of 100,000 users; the second participant and the third participant provide some other characteristic information x of the 100,000 users. Through the real labels y and characteristic information x of these 100,000 users, a classification model is trained, and the classification labels of the remaining 900,000 users are determined according to the trained classification model.

[0033] For the above mentioned scenario, the traditional method is that the three participants encrypt their own user information so that other participants do not know their own user information, and train the classification model based on the encrypted user information. However, this method generally faces two problems: high computational complexity and high communication overhead due to the encryption and decryption process.

[0034] Based on this, an embodiment of the present application proposes a method for determining user classification labels, which is used to reduce communication overhead and computational complexity under the premise of secure computing by multiple participants.

[0035] Before introducing this application, let me explain two professional terms: Secure Multi-Party Computation (MPC) is a subfield of cryptography, whose goal is for all parties to jointly calculate a given function while maintaining the privacy of these inputs. While ensuring that the participants obtain the correct calculation results, they cannot obtain any information other than the calculation results. The cryptography in the model protects the privacy of each other's participants.

[0036] Light Gradient Boosting Machine (LightGBM) is an efficient and scalable machine learning algorithm. The algorithm significantly improves the speed of model training and the accuracy of prediction through a series of optimization techniques, such as histogram gradient boosting, growth strategy, gradient-based one-side sampling algorithm (GOSS) and exclusive feature bundling algorithm (EFB).

[0037] The LightGBM algorithm is a commonly used classification model. The embodiment of the present application proposes a new LightGBM training scheme based on secret sharing, which splits the LightGBM algorithm tasks and uses secret sharing encryption calculations for key steps. This can not only protect the privacy of each participant, but also achieve efficient model training and prediction.

[0038] Figure 1 A flowchart corresponding to a method for determining a user classification label proposed in an embodiment of the present application is shown. The method is applied to a first participant having a sample user classification label. The method can be executed by a device for determining a user classification label, such as Figure 1 As shown, the process includes the following steps:

[0039] Step 101, the device for determining user classification labels determines, for at least one sample user, a first virtual label and a second virtual label of the sample user based on the predicted label of the sample user in the i-th round and the real label of the sample user, and splits the first virtual label of the sample user and the second virtual label of the sample user into a first virtual sub-label corresponding to each second participant and a second virtual sub-label corresponding to each second participant, respectively.

[0040] The device for determining the user classification label may be a device corresponding to the first participant having the user's real label, which may be a computer, a handheld computer or other intelligent device. Each participant determines multiple users based on the common users, and the multiple users include sample users and other users. The sample users are some of the above users, and the multiple users are more of the above users.

[0041] For at least one sample user, based on the predicted label and true label of the sample user in the i-th round, determine the first virtual label and the second virtual label of the sample user, where i is a positive integer greater than or equal to 1. If the sample user belongs to the blacklist, the true label can be defined as 0, and if the sample user belongs to the whitelist, the true label can be defined as 1. The predicted labels of the sample users in the first round are set according to the staff and can be uniformly set to 0.5. The first virtual label and the second virtual label of the sample user can be calculated according to the true label and the predicted label of the sample user using a specific calculation method, and are calculated according to the first-order gradient and the second-order gradient formula in the embodiment of the present application. Exemplarily, the calculation formulas for the first-order gradient gi and the second-order gradient hi are gi=y_pred–y, hi=y_pred*(1–y_pred), where y_pred is the predicted label of the sample user, y is the true label of the sample user, that is, gi is the first virtual label of the sample user, and hi is the second virtual label of the sample user.

[0042] And the first virtual label and the second virtual label of the sample user are respectively split into the first virtual sub-label corresponding to each second participant and the second virtual sub-label corresponding to each second participant. Exemplarily, the first virtual label is split into the first virtual sub-label corresponding to each second participant, and the second virtual label is split into the second virtual sub-label corresponding to each second participant. The splitting method can be a certain splitting method set, such as the product of the split first virtual sub-label is the first virtual label; the result of the split first virtual sub-label after the set operation is the first virtual label; the second virtual sub-label is also split in the same way. The embodiment of the present application does not limit the specific splitting method. In the embodiment of the present application, the sum of the first virtual sub-label corresponding to each second participant is equal to the first virtual label of the sample user; the sum of the second virtual sub-label corresponding to each second participant is equal to the second virtual label of the sample user. Split gi into the first virtual sub-label corresponding to each second participant, so that the sum of the split first virtual sub-label is gi. For example, if gi is 5, it can be split into 1 and 4, or it can be split into -995 and 1000. Hi is also split in the same way. It should be noted that when the number of participants is greater than three, the number of first virtual sub-tags and the number of second virtual sub-tags may be greater than two.

[0043] Exemplarily, the serial number of at least one sample user is 12345, and the first virtual label gi, the second virtual label hi, the first virtual sub-label (gi0 and gi1) and the second virtual sub-label (hi0 and hi1) of the sample users with serial numbers 1 to 5 are shown in Table 1, where i represents the serial number of the sample user.

[0044] Table 1

[0045] gi = y_pred - y gi0 gi1 hi=y_pred*(1-y_pred) hi0 hi1 0.5 0 0.5 0.25 0 0.25 -0.5 -0.1 -0.4 0.25 0.1 0.15 0.5 0 0.5 0.25 0.2 0.05 -0.5 0 -0.5 0.25 0.3 -0.05 0.5 1000 -999.5 0.25 0.25 0

[0046] Each second participant has the second dimension characteristics of the sample users and sets the characteristic values ​​of the sample users under each second dimension characteristic as the group number according to the self-defined grouping method. Exemplarily, the second dimension characteristics can be user characteristics possessed by the second participant, such as age and annual income. According to the self-defined grouping method, the characteristic values ​​are grouped according to the value of the characteristic. Taking age as an example, the characteristic value of 0-18 years old can be set as group number 0, the characteristic value of 19-30 years old can be set as group number 1, the characteristic value of 31-40 years old can be set as group number 2, the characteristic value of 41-50 years old can be set as group number 3, and the characteristic value of 51-60 years old can be set as group number 4. Specifically, if the second participant is a bank, the bank has the age information of the user, the serial numbers of the five sample users are 12345, and the ages of the five sample users are 17, 25, 35, 46, and 27 respectively, then the characteristic values ​​of the five sample users under the dimension characteristic of age are 01231.

[0047] Furthermore, the first second participant has the second dimension characteristics of the sample users, and the second participant has other second dimension characteristics of the sample users. For example, the first second participant has the user's age characteristics, and the second participant has the sample user's annual income characteristics. Since the second participant sets the characteristic values ​​of the sample users under each second dimension characteristic as the grouping number according to the self-defined grouping method, the second participant also sets the characteristic values ​​of the sample users under the annual income as the grouping number according to the self-defined grouping method. According to the self-defined grouping method, the characteristic values ​​of the 5 sample users under the dimension characteristic of annual income are 01234, and different groups are set according to the specific values ​​of the annual income. The above-mentioned setting of different groups according to the specific age is similar, and will not be elaborated here.

[0048] The first participant sends the characteristic value of 01231 of the 5 sample users under the dimension feature of age to the second participant, and the second participant sends the characteristic value of 01234 of the 5 sample users under the dimension feature of annual income to the first participant. Then the first second participant and the second participant both have the age characteristic value and annual income characteristic value of the sample users, and the five sample users correspond to the 5 rows of information in the above Table 1 in order.

[0049] Each second participant may have multiple second dimensional features. For example, a bank not only has the user's age information, but also the user's educational information. Each participant may group all the second dimensional features of the user that it has, or may group a part of the second dimensional features that it has. The grouped second dimensional features are the features that participate in the method of determining the user classification label. Generally speaking, the more second dimensional features that participate, the more accurate the user's classification label is. The embodiment of the present application is explained by taking each participant providing one second dimensional feature as an example. The processing method for multiple second dimensional features is the same, and no further elaboration is made here. When the number of second participants is one, a second participant who does not need to provide a second dimensional feature is set as another second participant. The number of second participants may also be two or more. When the number of second participants is greater than two, multiple second participants only need to perform similar processing.

[0050] Step 102: The device for determining the user classification label sends the first virtual sub-label and the second virtual sub-label corresponding to at least one sample user to each second participant respectively.

[0051] After determining the first virtual sub-label and the second virtual sub-label, the first virtual sub-label and the second virtual sub-label corresponding to each of the at least one sample user are sent to each second participant. In the embodiment of the present application, a first participant and two second participants are used as an example for explanation. Exemplarily, the data of the gi0 column and the hi0 column in the above Table 1 are sent to the first second participant, and the data of the gi1 column and the hi1 column in the above Table 1 are sent to the second second participant. Optionally, the gi0 of the five sample users can be formed into a vector glist0[0,-0.1,0,0,1000], and the hi0 of the five sample users can be formed into a vector hlist0[0,0.1,0.2,0.3,0.25] and sent to the first second participant; similarly, glist1[0.5,-0.4,0.5,0.5,-999.5] and hlist1[0.25,0.15,0.05,-0.05,0] are sent to the second second participant.

[0052] Step 103: The device for determining the user classification label receives the first grouping label value and the second grouping label value for each second dimensional feature sent by each second participant.

[0053] The first participant receives the first grouping label value and the second grouping label value for each second dimensional feature sent by the first second participant and the second second participant. Among them, the first grouping label value is the summary value of the first virtual sub-labels of the sample users under the same group number, and the second grouping label value is the summary value of the second virtual sub-labels of the sample users under the same group number. The summary value can be the sum of the virtual sub-labels, or it can be other calculation methods for the virtual sub-labels. In the embodiment of the present application, the first grouping label value is the sum of the first virtual sub-labels of the sample users under the same group number, and the second grouping label value is the sum of the second virtual sub-labels of the sample users under the same group number.

[0054] Table 2 shows the first group label value and the second group label value obtained by summing the first virtual sub-label and the second virtual sub-label under the same group number for the first and second participants according to the age and annual income characteristics. Specifically, since the characteristic value of the five sample users under the dimension feature of age is 01231, there are four group numbers corresponding to age, 0123, of which group number 1 is the second sample user and the fifth sample user. For group number 0 under age, the corresponding first virtual sub-label is 0; for group number 1 under age, the corresponding first virtual sub-label sum is -0.1+1000=999.9; for group number 2 under age, the corresponding first virtual sub-label is 0; group number 3 is also 0, that is, according to the group number 0123 corresponding to age, the first virtual sub-label is summed to obtain 0, 999.9, 0, 0; since the group number of the annual income feature of the five sample users is 01234, there is no group corresponding to two sample users. Therefore, according to the annual income feature, the sum of the first virtual sub-label is 0, -0.1, 0, 0, 1000, and the first grouping label value can be expressed in the form of a vector as [0, 999.9, 0, 0] and [0, -0.1, 0, 0, 1000]. According to the group number 0123 corresponding to the age, the sum of the second virtual sub-label is 0, 0.35, 0.2, 0.3; according to the annual income feature, the sum of the second virtual sub-label is 0, 0.1, 0.2, 0.3, 0.25. The second grouping label value can be expressed in the form of a vector as [0, 0.35, 0.2, 0.3] and [0, 0.1, 0.2, 0.3, 0.25].

[0055] Table 2

[0056] age gi0 sum hi0 sum Annual income gi0 sum hi0 sum 0 0 0 0 0 0 1 999.9 0.35 1 -0.1 0.1 2 0 0.2 2 0 0.2 3 0 0.3 3 0 0.3 4 1000 0.25

[0057] Similarly, Table 3 shows the first grouping label value and the second grouping label value obtained by summing the first virtual sub-label and the second virtual sub-label under the same grouping number according to the age and annual income characteristics of the second second participant. The first grouping label value can be expressed in the form of a vector as [0.5, -0.4, 0.5, -0.5, -999.5] and [0.5, -999.9, 0.5, -0.5]; the second grouping label value can be expressed in the form of a vector as [0.25, 0.15, 0.05, -0.05, 0] and [0.25, 0.15, 0.05, -0.05]. The specific calculation method is the same as above, so no further details are given here.

[0058] Table 3

[0059] Annual income gi1 sum hi1 sum age gi1 sum hi1 sum 0 0.5 0.25 0 0.5 0.25 1 -0.4 0.15 1 -999.9 0.15 2 0.5 0.05 2 0.5 0.05 3 -0.5 -0.05 3 -0.5 -0.05 4 -999.5 0

[0060] Step 104, the device for determining the user classification label summarizes the first grouping label value and the second grouping label value of each second participant under the second dimensional feature into the first characteristic value and the second characteristic value of the second dimensional feature under each group number, respectively.

[0061] Exemplarily, for the second dimensional feature of each second participant, the first grouping label value is summarized as the first eigenvalue of the second dimensional feature under each group; and the second grouping label value is summarized as the second eigenvalue of the second dimensional feature under each group. The summing method may be summation or other calculation methods, which is not specifically limited in the present application. In the embodiment of the present application, the summation is performed. Specifically, taking age as an example, the first grouping label values ​​[0,999.9,0,0] and [0.5,-999.9,0.5,-0.5] are summed to obtain a first eigenvalue of [0.5, 0, 0.5, -0.5], and the second grouping label values ​​[0,0.35,0.2,0.3] and [0.25,0.15,0.05,-0.05] are summed to obtain a second eigenvalue of [0.25,0.5,0.25,0.25]; taking annual income as an example, the first grouping label values ​​[0,999.9,0,0] and [0.5,-999.9,0.5,-0.5] are summed to obtain a second eigenvalue of [0.25,0.5,0.25,0.25]. The sum of the label values ​​[0,-0.1,0,0,1000] and [0.5,-0.4,0.5,-0.5,-999.5] gives the first eigenvalue [0.5,-0.5,0.5,-0.5,0.5], and the sum of the second group label values ​​[0,0.1,0.2,0.3,0.25] and [0.25,0.15,0.05,-0.05,0] gives the second eigenvalue [0.25,0.25,0.25,0.25,0.25]. Table 4 shows the first and second eigenvalues.

[0062] Table 4

[0063]

[0064]

[0065] Figure 2 A histogram corresponding to a first characteristic value based on age provided in an embodiment of the present application, from Figure 2 It can be seen that the horizontal axis is the group number of age, and the vertical axis is the first eigenvalue corresponding to the group. The histogram corresponding to the second eigenvalue based on age and the histogram corresponding to the first eigenvalue and the second eigenvalue based on annual income are not described in detail here.

[0066] Step 105, the device for determining the user classification label obtains a decision tree corresponding to at least one sample user according to the first feature value and the second feature value of each second dimensional feature under each group number, and the decision tree is used to determine the classification label of the non-sample user.

[0067] According to the first eigenvalue and the second eigenvalue of each second dimension under each group, the decision tree corresponding to the sample user is determined. The root node of the decision tree corresponds to the five sample users with serial numbers 1, 2, 3, 4, and 5 processed above. The split point is determined according to the characteristics of each second dimension until it can no longer be split, and the final decision tree is obtained to determine the classification label of the non-sample user. Specifically, it can be seen from Table 4 that there are 4 groups for age characteristics and 5 groups for annual income characteristics. Therefore, the candidate split nodes of the sample users corresponding to the root node can be selected from 3+4=7, with the age group number less than 1 being divided into one group, and the age group number greater than or equal to 1 being divided into 1 group; the age group number less than 2 being divided into one group, and the age group number greater than or equal to 2 being divided into one group, and so on, which will not be described in detail here. After the split node is determined, for example, the age group number less than 2 is used as the split node, the sample users corresponding to the left subtree include users with serial numbers 1, 2, and 5, and the samples corresponding to the right subtree include users with serial numbers 3 and 4. Continue to split the left subtree and the right subtree until they can no longer be split.

[0068] Optionally, before obtaining the decision tree corresponding to at least one sample user, it may also include: for each first dimensional feature of the first participant, setting the characteristic value of the sample user under each first dimensional feature as the corresponding group number according to the grouping method set by the first participant, determining the first grouping label value and the second grouping label value for each first dimensional feature, and obtaining the decision tree corresponding to at least one sample user according to the first characteristic value and the second characteristic value of each second dimensional feature under each group number, including: obtaining the decision tree corresponding to at least one sample user according to the first grouping label value and the second grouping label value of each first dimensional feature and the first characteristic value and the second characteristic value of each second dimensional feature under each group number.

[0069] Specifically, in addition to providing the user's real label, the first participant may also have the user's characteristic information, such as the first participant also has the user's gender characteristics, which are first-dimensional characteristics. The number of first-dimensional characteristics may also be multiple, and one first-dimensional characteristic is used as an example in the embodiment of the present application. The gender grouping method is set to 0 for male and 1 for female. If the genders of users with serial numbers 1, 2, 3, 4, and 5 are male, female, female, male, and female, respectively, the characteristic values ​​of users with serial numbers 1, 2, 3, 4, and 5 under the gender characteristics are 01101.

[0070] Determine the first grouping label value and the second grouping label value for gender characteristics. For the first participant, the first grouping label value is the sum of the first virtual labels of the above sample users under the same group number, and the second grouping label value is the sum of the second virtual labels of the above sample users under the same group number. As can be seen from the above, the first virtual label is gi and the second virtual label is hi. Therefore, since the first virtual label of the user with serial number 12345 is 0.5, -0.5, 0.5, -0.5, 0.5; the second virtual label of the user with serial number 12345 is 0.25, 0.25, 0.25, 0.25, 0.25, therefore, the result of summing the first virtual label and the second virtual label according to the gender group number 0 and 1 is shown in Table 5.

[0071] Table 5

[0072] gender gi sum hi sum 0 0 0.5 1 0.5 0.75

[0073] Furthermore, multiple candidate split nodes of the decision tree are determined according to the group numbers of the first dimension features and the group numbers of the second dimension features. Specifically, the sample users corresponding to the root node are still the users with serial numbers 1, 2, 3, 4, and 5. It can be seen from Tables 5 and 4 that due to the addition of gender group numbers 0 and 1, one more candidate split node is added, which is whether the gender is 0, that is, the decision tree has a total of 8 candidate split nodes.

[0074] For each candidate split node, a first sub-gain is obtained based on the first eigenvalue and the second eigenvalue of the group number belonging to the first side of the candidate split node, a second sub-gain is obtained based on the first eigenvalue and the second eigenvalue of the group number belonging to the second side of the candidate split node, and a third sub-gain is obtained based on the first eigenvalue and the second eigenvalue of all the group numbers of the candidate split node; based on the first sub-gain, the second sub-gain and the third sub-gain, a gain value of the candidate split node is determined.

[0075] Exemplarily, the first side and the second side are the two sides of the decision tree. In the embodiment of the present application, the first side can be understood as the left side of the decision tree, and the second side can be understood as the right side of the decision tree. Taking the candidate split node as the age group number less than 2 as an example, the group numbers on the first side are 0 and 1, and the group numbers on the second side are 2 and 3. It can be seen from Table 4 above that on the first side, the first eigenvalue of group number 0 is 0.5, and the first eigenvalue of group number 1 is 0; the second eigenvalue of group number 0 is 0.25, and the second eigenvalue of group number 1 is 0.5; on the second side, the first eigenvalue of group number 2 is 0.5, and the first eigenvalue of group number 3 is -0.5; the second eigenvalue of group number 2 is 0.25, and the second eigenvalue of group number 3 is 0.25. A first sub-gain is obtained based on the first eigenvalue and the second eigenvalue of group numbers 0 and 1; a second sub-gain is obtained based on the first eigenvalue and the second eigenvalue of group numbers 2 and 3; a third sub-gain is obtained based on the first eigenvalue and the second eigenvalue of group numbers 0, 1, 2, and 3, and a gain value of the candidate split node is determined according to the first sub-gain, the second sub-gain, and the third sub-gain according to a preset calculation method.

[0076] Optionally, a preset calculation method for calculating a gain value of a candidate segmentation node provided in an embodiment of the present application is as shown in Formula 1:

[0077]

[0078] in, is the first sub-gain, is the second sub-gain, is the third sub-gain, λ is a constant, gl is the sum of the first eigenvalues ​​of the first side grouping numbers, which is 0.5+0=0.5; hl is the sum of the second eigenvalues ​​of the first side grouping numbers, which is 0.25+0.5=0.75; gr is the sum of the first eigenvalues ​​of the second side grouping numbers, which is 0.5-0.5=0; hr is the sum of the second eigenvalues ​​of the second side grouping numbers, which is 0.25+0.25=0.5; g is the sum of gl and gr, and h is the sum of hl and hr.

[0079] The candidate split node with the best gain value is determined as the split node until a decision tree is obtained. Exemplarily, the gain value of each of the above 8 candidate split nodes is calculated to determine whether the maximum gain value is greater than 0. If it is greater than 0, the candidate split node with the largest gain value is determined as the split node. Figure 3 A schematic diagram of a decision tree provided in an embodiment of the present application, from Figure 3It can be seen that determining whether the age group is less than 2 is a split node. According to this split node, users can be divided into two groups. Users 1, 2, and 5 are on the side with an age group less than 2, and users 3 and 4 are on the other side with an age group greater than or equal to 2. Based on the split decision tree, the split is continued. For the grouping of users 1, 2, and 5, the split is continued according to gender grouping, age grouping, and income grouping. The same is true for users 3 and 4. Figure 3 It can be seen that for users 1, 2, and 5, if the split node is determined to be 0 for gender grouping, users 1, 2, and 5 can be divided into user 1 and user 2 and 5. For users 3 and 4, if the gain values ​​of the corresponding candidate split nodes are not greater than 0, the split is not continued. It should be noted that when the sample space changes, the eigenvalues ​​of the corresponding groups may also change. It is necessary to recalculate the first eigenvalue and the second eigenvalue corresponding to each group under each feature according to the new sample space.

[0080] Optionally, the sample users are sample users of the i-th round of sampling, and the sampling method is random or preset, which is not specifically limited in the embodiment of the present application. After obtaining the decision tree, it also includes:

[0081] Determine the leaf node corresponding to at least one sample user in the decision tree and the weight of the leaf node. Based on the weight of the leaf node of the decision tree, determine the predicted label of the sample user in the i+1th round of sampling under the decision tree, and update the decision tree through the sample user in the i+1th round of sampling. Specifically, after obtaining the decision tree, determine the weight of the leaf node. The weight calculation formula of each leaf node is shown in Formula 2 and Formula 3:

[0082]

[0083] Among them, Formula 2 is the weight calculation formula when the leaf node is the left node, Formula 3 is the weight calculation formula when the leaf node is the right node, lr and λ are constants, and the calculation methods of gl, gr, hl, and hr are the same as above, so we will not go into details here. According to the weight of each leaf node, the predicted label of the leaf node is determined, and the calculation formula of the predicted label is shown in Formula 4:

[0084]

[0085] After the i-th decision tree is determined based on the sample users in the i-th round of sampling, the above 100,000 sample users are brought into the i-th decision tree, and the predicted labels of the 100,000 sample users are updated according to the predicted labels corresponding to the leaf nodes. For example, the predicted label of the first sample user in the i-1th round is 0.3. Based on the feature information of the sample user and the i-th corresponding decision tree, the leaf node corresponding to the sample user is found, and the predicted label of the sample user is determined. If the predicted label is 0.28, the predicted label of the sample user is updated.

[0086] Based on the updated 10w sample users, a part of users are sampled as sample users for the i+1th round, and the i+1th decision tree is determined based on the sample users for the i+1th round. Then, based on the predicted labels corresponding to the leaf nodes of the i+1th decision tree, the predicted labels of the 10w sample users are updated until the difference between the predicted labels of the 10w sample users and the true labels is less than the set threshold, indicating that the decision tree training is successful.

[0087] According to the trained decision tree and the features of 90w users to be tested, the leaf nodes corresponding to the 90w users to be tested are determined, and the predicted labels corresponding to the leaf nodes are the predicted labels of the users to be tested. In this way, a more accurate decision tree is determined based on multiple rounds of training of the decision tree, and the predicted labels of the users to be tested are determined based on the trained decision tree. In the embodiment of the present application, the predicted label is a black sample user or a white sample user. Further, whether the user has a loan limit can be determined based on whether the user is a black sample user or a white sample user.

[0088] Optionally, according to the decision tree training method provided in the embodiment of the present application, other predicted labels of the user can also be determined, such as determining the user's Sesame Credit score, determining the user's credit score, etc.

[0089] Figure 4 Schematic diagram of the internal modules of a device 4000 for determining a user classification label provided in an embodiment of the present application. Figure 4 As shown, the device may include a processing module 401, a sending module 402, a receiving module 403, and a summarizing module 404. Optionally, it also includes a storage module, which is used to store computer instructions or programs. The processing module 401 can call the computer instructions or programs stored in the storage module.

[0090] Among them, the processing module 401 is used to determine, for at least one sample user, the first virtual label and the second virtual label of the sample user based on the predicted label of the sample user in the i-th round and the real label of the sample user, and split the first virtual label of the sample user and the second virtual label of the sample user into a first virtual sub-label corresponding to each second participant and a second virtual sub-label corresponding to each second participant respectively; each second participant has the second dimensional characteristics of the sample user and sets the characteristic value of the sample user under each second dimensional characteristic as the group number according to the self-defined grouping method; the sending module 402 is used to send the first virtual sub-label and the second virtual sub-label corresponding to each of the at least one sample user to each second participant respectively; the receiving module 403 is used to receive the first virtual sub-label and the second virtual sub-label sent by each second participant. The first grouping label value and the second grouping label value for each second dimensional feature are sent, wherein the first grouping label value is the summary value of the first virtual sub-label of the at least one sample user under the same group number, and the second grouping label value is the summary value of the second virtual sub-label of the at least one sample user under the same group number; the summarizing module 404 is used to summarize the first grouping label value and the second grouping label value of each second participant under the second dimensional feature into the first feature value and the second feature value of the second dimensional feature under each group number respectively for the second dimensional feature of each second participant; the processing module 401 is also used to obtain the decision tree corresponding to the at least one sample user according to the first feature value and the second feature value of each second dimensional feature under each group number; the decision tree is used to determine the classification label of the non-sample user.

[0091] In a possible implementation, the processing module 401 is also used to, for each first dimensional feature of the first participant, set the characteristic value of the sample user under each first dimensional feature as a corresponding group number according to the grouping method set by the first participant; determine the first grouping label value and the second grouping label value for each first dimensional feature; and obtain a decision tree corresponding to the at least one sample user based on the first grouping label value and the second grouping label value of each first dimensional feature and the first characteristic value and the second characteristic value of each second dimensional feature under each group number.

[0092] In a possible implementation, the processing module 401 is specifically used to determine multiple candidate split nodes of the decision tree according to each group number of each first dimensional feature and each group number of each second dimensional feature; for each candidate split node, obtain a first sub-gain based on the first eigenvalue and the second eigenvalue of the group number belonging to the first side of the candidate split node, obtain a second sub-gain based on the first eigenvalue and the second eigenvalue of the group number belonging to the second side of the candidate split node, and obtain a third sub-gain based on the first eigenvalue and the second eigenvalue of all group numbers of the candidate split node; determine the gain value of the candidate split node based on the first sub-gain, the second sub-gain and the third sub-gain; determine the candidate split node with the best gain value as the split node; until a decision tree is obtained.

[0093] In a possible implementation, the at least one sample user is a sample user sampled in the i-th round; the processing module 401 is also used to determine the leaf node corresponding to the at least one sample user in the decision tree and the weight of the leaf node; based on the weight of the leaf node of the decision tree, determine the predicted label of the sample user sampled in the i+1th round under the decision tree, and update the decision tree through the sample user sampled in the i+1th round.

[0094] In a possible implementation, the sum of the first virtual sub-tags corresponding to each second participant is equal to the first virtual tag of the sample user; the sum of the second virtual sub-tags corresponding to each second participant is equal to the second virtual tag of the sample user; for the second participant, the first grouping tag value is the sum of the first virtual sub-tags of the at least one sample user under the same group number; for the first participant, the first grouping tag value is the sum of the first virtual tags of the at least one sample user under the same group number; for the second participant, the second grouping tag value is the sum of the second virtual sub-tags of the at least one sample user under the same group number; for the first participant, the second grouping tag value is the sum of the second virtual tags of the at least one sample user under the same group number.

[0095] In a possible implementation, when there is only one second participant, a third participant who does not need to provide the second dimensional feature is set as another second participant.

[0096] Figure 5 FIG. 5 is a schematic diagram of a device 5000 for determining a user classification label provided in an embodiment of the present application. Figure 5 As shown, it includes at least one processor 501 and a memory 502 connected to the at least one processor 501. The specific connection medium between the processor 501 and the memory 502 is not limited in the embodiment of the present application. Figure 5For example, the processor 501 and the memory 502 are connected via a bus. The bus can be divided into an address bus, a data bus, a control bus, etc.

[0097] In the embodiment of the present application, the memory 502 stores instructions that can be executed by at least one processor 501. The at least one processor 501 can implement the steps of the above-mentioned method for determining the user classification label by executing the instructions stored in the memory 502.

[0098] The processor 501 is the control center of the computer device, and can use various interfaces and lines to connect various parts of the computer device, and perform resource settings by running or executing instructions stored in the memory 502 and calling data stored in the memory 502. Optionally, the processor 501 may include one or more processing units, and the processor 501 may integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface, and application programs, and the modem processor mainly processes wireless communications. It is understandable that the above-mentioned modem processor may not be integrated into the processor 501. In some embodiments, the processor 501 and the memory 502 may be implemented on the same chip, and in some embodiments, they may also be implemented separately on independent chips.

[0099] The processor 501 may be a general-purpose processor, such as a central processing unit (CPU), a digital signal processor, an application-specific integrated circuit (ASIC), a field programmable gate array or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component, and may implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of the present application. A general-purpose processor may be a microprocessor or any conventional processor, etc. The steps of the method disclosed in the embodiments of the present application may be directly embodied as being executed by a hardware processor, or may be executed by a combination of hardware and software modules in the processor.

[0100] The memory 502 is a non-volatile computer-readable storage medium that can be used to store non-volatile software programs, non-volatile computer executable programs and modules. The memory 502 may include at least one type of storage medium, such as a flash memory, a hard disk, a multimedia card, a card-type memory, a random access memory (Random Access Memory, RAM), a static random access memory (Static Random Access Memory, SRAM), a programmable read-only memory (Programmable Read Only Memory, PROM), a read-only memory (Read Only Memory, ROM), an electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, EEPROM), a magnetic memory, a disk, an optical disk, etc. The memory 502 is any other medium that can be used to carry or store a desired program code in the form of an instruction or data structure and can be accessed by a computer, but is not limited thereto. The memory 502 in the embodiment of the present application can also be a circuit or any other device that can realize a storage function, for storing program instructions and / or data.

[0101] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0102] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0103] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0104] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0105] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for determining a user classification label, characterized in that: The method is applied to a first participant having a classification label of a sample user, and the method comprises: For at least one sample user, based on the predicted label of the sample user in the i-th round and the real label of the sample user, determine the first virtual label of the sample user and the second virtual label of the sample user, and split the first virtual label of the sample user and the second virtual label of the sample user into a first virtual sub-label corresponding to each second participant and a second virtual sub-label corresponding to each second participant, respectively; each second participant has the second dimensional feature of the sample user and sets the feature value of the sample user under each second dimensional feature as the grouping number according to the self-defined grouping method; Sending to each second participant respectively the first virtual sub-label and the second virtual sub-label respectively corresponding to the at least one sample user; Receive a first grouping label value and a second grouping label value for each second dimension feature sent by each second participant, wherein the first grouping label value is a summary value of first virtual sub-labels of the at least one sample user under the same group number, and the second grouping label value is a summary value of second virtual sub-labels of the at least one sample user under the same group number; For the second dimensional feature of each second participant, the first grouping label value and the second grouping label value of each second participant under the second dimensional feature are respectively summarized into the first characteristic value and the second characteristic value of the second dimensional feature under each group number; A decision tree corresponding to the at least one sample user is obtained according to the first eigenvalue and the second eigenvalue of each second dimensional feature under each group number; the decision tree is used to determine the classification label of the non-sample user.

2. The method according to claim 1, characterized in that Before obtaining the decision tree corresponding to the at least one sample user, the method further includes: For each first dimensional feature of the first participant, set the feature value of the sample user under each first dimensional feature as a corresponding grouping number according to the grouping method set by the first participant; Determine a first grouping label value and a second grouping label value for each first dimension feature; Obtaining a decision tree corresponding to the at least one sample user according to the first eigenvalue and the second eigenvalue of each second dimensional feature under each group number, including: A decision tree corresponding to the at least one sample user is obtained according to the first grouping label value and the second grouping label value of each first dimensional feature and the first feature value and the second feature value of each second dimensional feature under each group number.

3. The method according to any one of claims 1 to 2, characterized in that Obtaining a decision tree corresponding to the at least one sample user includes: Determine a plurality of candidate split nodes of the decision tree according to each grouping number of each first dimensional feature and each grouping number of each second dimensional feature; For each candidate segmentation node, a first sub-gain is obtained based on a first eigenvalue and a second eigenvalue of a group number belonging to a first side of the candidate segmentation node, a second sub-gain is obtained based on a first eigenvalue and a second eigenvalue of a group number belonging to a second side of the candidate segmentation node, and a third sub-gain is obtained based on the first eigenvalue and the second eigenvalue of all group numbers of the candidate segmentation node; a gain value of the candidate segmentation node is determined based on the first sub-gain, the second sub-gain and the third sub-gain; The candidate split node with the best gain value is determined as the split node until a decision tree is obtained.

4. The method according to claim 3, characterized in that The at least one sample user is a sample user in the i-th round of sampling; After getting the decision tree, it also includes: Determine a leaf node corresponding to the at least one sample user in the decision tree and a weight of the leaf node; Based on the weights of the leaf nodes of the decision tree, the predicted labels of the sample users sampled in the i+1th round under the decision tree are determined, and the decision tree is updated through the sample users sampled in the i+1th round.

5. The method according to claim 3, characterized in that: The sum of the first virtual sub-labels corresponding to each second participant is equal to the first virtual label of the sample user; The sum of the second virtual sub-labels corresponding to each second participant is equal to the second virtual label of the sample user; For the second participant, the first group label value is the sum of the first virtual sub-labels of the at least one sample user under the same group number; For the first participant, the first group label value is the sum of the first virtual labels of the at least one sample user under the same group number; For the second participant, the second group label value is the sum of the second virtual sub-labels of the at least one sample user under the same group number; For the first participant, the second group label value is the sum of the second virtual labels of the at least one sample user under the same group number.

6. The method according to claim 3, characterized in that When there is only one second participant, a third participant who does not need to provide second dimensional features is set as another second participant.

7. A device for determining a user classification label, characterized in that: A processing module is used to determine, for at least one sample user, a first virtual label and a second virtual label of the sample user based on the predicted label of the sample user in the i-th round and the real label of the sample user, and to split the first virtual label of the sample user and the second virtual label of the sample user into a first virtual sub-label corresponding to each second participant and a second virtual sub-label corresponding to each second participant, respectively; each second participant has a second dimensional feature of the sample user and sets a feature value of the sample user under each second dimensional feature as a grouping number according to a self-defined grouping method; A sending module, configured to send the first virtual sub-label and the second virtual sub-label respectively corresponding to the at least one sample user to each second participant; A receiving module, configured to receive a first grouping label value and a second grouping label value for each second dimensional feature sent by each second participant, wherein the first grouping label value is a summary value of first virtual sub-labels of the at least one sample user under the same group number, and the second grouping label value is a summary value of second virtual sub-labels of the at least one sample user under the same group number; a summarizing module, for summarizing, for each second participant's second dimensional feature, the first grouping label value and the second grouping label value of each second participant under the second dimensional feature into the first characteristic value and the second characteristic value of the second dimensional feature under each group number; The processing module is also used to obtain a decision tree corresponding to the at least one sample user according to the first feature value and the second feature value of each second dimensional feature under each group number; the decision tree is used to determine the classification label of the non-sample user.

8. A device for determining a user classification label, characterized in that: include: Memory, used to store computer programs or instructions; A processor, configured to call a computer program or instruction stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when a computer reads and executes the instructions, the computer executes the method according to any one of claims 1 to 6.

10. A computer program product, characterized in that The computer program product stores instructions, and when a computer reads and executes the instructions, the computer executes the method according to any one of claims 1 to 6.

Images