CN112910953A

CN112910953A - Business data pushing method and device and server

Info

Publication number: CN112910953A
Application number: CN202110047198.9A
Authority: CN
Inventors: 陈珊珊
Original assignee: Industrial and Commercial Bank of China Ltd ICBC
Current assignee: Industrial and Commercial Bank of China Ltd ICBC
Priority date: 2021-01-14
Filing date: 2021-01-14
Publication date: 2021-06-04
Anticipated expiration: 2041-01-14
Also published as: CN112910953B

Abstract

The specification provides a service data pushing method, a service data pushing device and a service data pushing server. Based on the method, in the field of artificial intelligence, when a server pushes business data to a user object, the guest group type of the user object is determined, and then a prediction sub-model which is matched with the guest group type of the user object in a preset target user prediction model which is established in advance according to sample data of two parties based on horizontal federal learning is called to process the feature data of the user object, so that whether the user object belongs to a target user is determined; under the condition that the user object is determined to be a target user, processing feature data of the user object by calling a preset user portrait prediction model which is established in advance according to sample data of two parties based on longitudinal federal learning to obtain a portrait label of the user object; and then, according to the portrait label, generating and according to a matched target pushing rule, accurately pushing appropriate target service data to the user object in a targeted manner, and improving the pushing success rate.

Description

Business data pushing method and device and server

Technical Field

The specification belongs to the technical field of artificial intelligence, and particularly relates to a service data pushing method, a service data pushing device and a service server.

Background

In many service data pushing scenarios (for example, a recommendation scenario of a financial service, etc.), it is often difficult to accurately predict whether a user object to be pushed with service data is a potential target user who will receive the service data, and it is also difficult to accurately predict specific service data suitable for the user object. The technical problems of inaccurate pushing, poor business data pushing effect and the like often exist when business data are pushed to a user object based on the existing method.

In view of the above problems, no effective solution has been proposed.

Disclosure of Invention

The present specification provides a method, an apparatus, and a server for pushing service data, so as to accurately identify whether a user object is a target user object, and further determine and accurately push appropriate target service data to the user object according to a demand tendency of the user object when it is determined that the user object is the target user, thereby obtaining a better pushing effect and improving a pushing success rate.

The present specification provides a method for pushing service data, including:

acquiring identification information of a user object and characteristic data of the user object; wherein the feature data at least comprises non-traffic class feature data;

determining the guest group type of the user object according to the identification information of the user object;

calling a prediction sub-model matched with the guest group type of the user object in a preset target user prediction model to process the characteristic data of the user object so as to determine whether the user object is a target user; the preset target user prediction model is established in advance according to first sample data in a first data party and second sample data in a second data party on the basis of transverse federal learning;

under the condition that the user object is determined to be a target user, calling a preset user portrait prediction model to process feature data of the user object so as to obtain a portrait label of the user object; the preset user portrait prediction model is established in advance according to first sample data in a first data party and second sample data in a second data party on the basis of longitudinal federal learning;

generating a target pushing rule matched with the user object according to the portrait label of the user object;

and pushing corresponding target business data to the user object according to the target pushing rule.

In one embodiment, the guest group type includes: an administrative group and a non-administrative group.

In one embodiment, the preset target user prediction model comprises a first predictor model and a second predictor model; the first predictor model is matched with an administrative customer group, and the second predictor model is matched with a non-administrative customer group.

In one embodiment, the preset target user prediction model is established as follows:

the method comprises the steps that a first server responds to a first training request about a preset target user prediction model, and performs sample data fusion in cooperation with a second server according to a protocol rule based on horizontal federal learning so as to determine a first positive sample user, a first negative sample user and a first pseudo negative sample user in an output customer group sample user and a second positive sample user and a second mixed negative sample user in a non-output customer group sample user; the first server is deployed at one side of a first data party, the held first sample data at least comprises non-business characteristic data of a full amount of sample users, and the full amount of sample users comprise user management customer group sample users and non-user management customer group sample users; the second server is a server deployed at one side of a second data party, and second sample data held by the second server at least comprises service class characteristic data of a full positive sample user;

acquiring non-service characteristic data of a first positive sample user and non-service characteristic data of a first negative sample user from the first sample data to serve as first grouping sample data; acquiring non-service characteristic data of a second positive sample user and non-service characteristic data of a second mixed negative sample user from the first sample data to serve as second grouping sample data;

training by using the first grouping sample data to obtain a first predictor model; training to obtain a second predictor model by utilizing the second grouping of sample data;

and combining the first prediction submodel and the second prediction submodel to obtain a preset target user prediction model.

In one embodiment, training to obtain a first predictor model using the first packet of sample data includes:

extracting a plurality of data from the non-business characteristic data of a first positive sample user in the first grouping sample data, and marking the data as first positive sample training data;

extracting a plurality of data from the non-business class characteristic data of a first negative sample user in the first grouping sample data, and marking the data as first negative sample training data;

and training to obtain a first predictor model by using the first positive sample training data and the first negative sample training data.

In one embodiment, training a second predictor model using the second packet of sample data comprises:

extracting a plurality of data from the business class characteristic data of a second positive sample user in the second grouping sample data, and marking the data as second positive sample training data;

randomly extracting a plurality of data from the business class characteristic data of a second mixed negative sample user in the second grouped sample data, and marking the data as second negative sample training data;

and training to obtain a second predictor model by using the second positive sample training data and the second negative sample training data.

In one embodiment, the preset user portrait prediction model is built as follows:

the first server responds to a second training request about a preset user portrait prediction model, and cooperates with the second server to perform sample data fusion according to a protocol rule based on longitudinal federal learning to obtain third training data carrying a preset portrait label; wherein the third training data comprises non-traffic class feature data of the positive sample user; the preset portrait label is determined according to the business class characteristic data of the user who just samples;

and training to obtain a preset user portrait prediction model by using the third training data carrying the preset portrait label.

In one embodiment, the predetermined user portrait prediction model includes a plurality of portrait prediction submodels, wherein different portrait prediction submodels are used to predict different portrait tags.

In one embodiment, the target business data includes insurance business products and/or insurance business services that fit the user object; correspondingly, the target pushing rule comprises an insurance business marketing scheme.

In one embodiment, the non-traffic class characteristic data comprises at least one of: monthly income, monthly consumption amount, deposit balance; the service class characteristic data comprises at least one of: gift identification, premium, and premium.

In one embodiment, the portrait label includes at least one of: payment period tendency label, premium tendency label, and premium tendency label.

The present specification further provides a method for training a model, which is applied to a first server deployed on a first data side, and includes:

responding to a first training request about a preset target user prediction model, and performing sample data fusion in cooperation with a second server according to a protocol rule based on horizontal federal learning so as to determine a first positive sample user, a first negative sample user and a first pseudo negative sample user in an out-of-customer group sample user, and a second positive sample user and a second mixed negative sample user in a non-out-of-customer group sample user; the first sample data held by the first server at least comprises non-business class characteristic data of full-scale sample users, wherein the full-scale sample users comprise customer group sample users and non-customer group sample users; the second server is a server deployed at one side of a second data party, and second sample data held by the second server at least comprises service class characteristic data of a full positive sample user;

responding to a second training request about a preset user portrait prediction model, and performing sample data fusion by cooperating with a second server according to a protocol rule based on longitudinal federal learning to obtain third training data carrying a preset portrait label; wherein the third training data comprises non-traffic class feature data of the positive sample user; the preset portrait label is determined according to the business class characteristic data of the user who just samples; the first sample data held by the first server at least comprises non-business class characteristic data of a full amount of sample users; the second server is a server deployed at one side of a second data party, and second sample data held by the second server at least comprises service class characteristic data of a full positive sample user;

This specification also provides a service data pushing device, including:

the acquisition module is used for acquiring the identification information of the user object and the characteristic data of the user object; wherein the feature data at least comprises non-traffic class feature data;

the determining module is used for determining the guest group type of the user object according to the identification information of the user object;

the first calling module is used for calling a prediction sub-model matched with the user object passenger group type in a preset target user prediction model to process the characteristic data of the user object so as to determine whether the user object is a target user; the preset target user prediction model is established in advance according to first sample data in a first data party and second sample data in a second data party on the basis of transverse federal learning;

the second calling module is used for calling a preset user portrait prediction model to process the characteristic data of the user object under the condition that the user object is determined to be the target user so as to obtain a portrait label of the user object; the preset user portrait prediction model is established in advance according to first sample data in a first data party and second sample data in a second data party on the basis of longitudinal federal learning;

the generation module is used for generating a target pushing rule matched with the user object according to the portrait label of the user object;

and the pushing module is used for pushing corresponding target business data to the user object according to the target pushing rule.

The present specification also provides a server comprising a processor and a memory for storing processor-executable instructions, the processor implementing the relevant steps of the training method of the model and/or the pushing method of the business data when executing the instructions.

According to the pushing method, the pushing device and the pushing server for the business data, when the server pushes the business data to the user object, the client type of the user object can be determined firstly, and then a prediction sub-model which is matched with the client type of the user object in a preset target user prediction model established in advance based on horizontal federal learning according to first sample data in a first data party and second sample data in a second data party can be called to process the feature data of the user object, so that whether the user object belongs to a potential target user which can receive the pushed business data or not can be accurately identified; under the condition that the user object is determined to be the target user, further calling a preset user portrait prediction model established in advance according to the first sample data and the second sample data based on longitudinal federal learning to process the feature data of the user object so as to obtain a portrait label of the user object; further, according to the portrait label, the requirement tendency and the personal preference of the user object can be analyzed, and a target pushing rule matched with the user object is generated; and according to the target pushing rule, the appropriate target business data is pushed to the user object accurately in a targeted manner, so that a better pushing effect can be obtained, the pushing success rate is improved, and the accurate pushing of the business data of the user object is realized.

Drawings

In order to more clearly illustrate the embodiments of the present specification, the drawings needed to be used in the embodiments will be briefly described below, and the drawings in the following description are only some of the embodiments described in the present specification, and it is obvious to those skilled in the art that other drawings can be obtained according to the drawings without any creative effort.

Fig. 1 is a schematic diagram of an embodiment of a structural composition of a data processing system to which a pushing method of service data provided by an embodiment of the present specification is applied;

fig. 2 is a flowchart illustrating a pushing method of service data according to an embodiment of the present disclosure;

fig. 3 is a schematic diagram illustrating an embodiment of a pushing method for service data provided by an embodiment of the present specification, in a scenario example;

FIG. 4 is a schematic flow chart diagram of a model training method provided by one embodiment of the present description;

FIG. 5 is a schematic flow chart diagram of a model training method provided by one embodiment of the present description;

fig. 6 is a flowchart illustrating a pushing method of service data according to an embodiment of the present disclosure;

FIG. 7 is a schematic diagram of a server according to an embodiment of the present disclosure;

fig. 8 is a schematic structural composition diagram of a service data pushing device provided in an embodiment of the present specification;

FIG. 9 is a schematic structural component diagram of a model training apparatus provided in an embodiment of the present disclosure;

fig. 10 is a schematic diagram of an embodiment of a pushing method for service data provided by an embodiment of the present specification, in an example scenario;

fig. 11 is a schematic diagram of an embodiment of a pushing method for service data provided by an embodiment of the present specification, in an example scenario.

Detailed Description

In order to make those skilled in the art better understand the technical solutions in the present specification, the technical solutions in the embodiments of the present specification will be clearly and completely described below with reference to the drawings in the embodiments of the present specification, and it is obvious that the described embodiments are only a part of the embodiments of the present specification, and not all of the embodiments. All other embodiments obtained by a person skilled in the art based on the embodiments in the present specification without any inventive step should fall within the scope of protection of the present specification.

In consideration of the existing pushing method based on the business data, under the condition of data isolation, sample data in a single data party is often used for training a prediction model for predicting a potential target user which can receive the pushed business data, then the prediction model is used for predicting a user object to be pushed with the business data so as to determine whether the user object is a potential target user, and then the business data is pushed to the user object according to the prediction result.

Firstly, the sample data used by the method in the process of training the prediction model is often only the sample data in a single data party, and the sample data used is relatively limited, so that the model precision of the prediction model obtained by training is relatively poor.

Secondly, the positive samples in the commonly used sample data are often relatively definite, but the negative samples may have a lot of uncertainties, which further affects the model accuracy.

For example, in the model training process of most prediction models, a user object which historically and finally receives service data in sample data is often marked as a positive sample user; and directly marking the user objects which are remained in the sample data and do not receive the service data except the positive sample user as the negative sample user.

In practice, the situation of a user object that does not eventually accept business data is relatively complex. There are many user objects that have been pushed business data before, but do not accept the pushed business data eventually, and such user objects do belong to negative sample users (which may also be called pure negative sample users).

However, there are also many user objects that do not actually accept the business data because they have not been pushed before. Some of such user objects may accept the pushed service data if the user object has been previously pushed with the service data. Thus, such a user object is not a pure negative sample user, but a pseudo-negative sample user.

Based on the existing method, when a model is trained, the pseudo negative sample user is often marked as a negative sample user for use, so that the distribution deviation of the negative sample data is inevitably introduced in the model training process, further the model training is influenced, the obtained prediction model has errors and poor model precision, and the accuracy of the target user predicted by the prediction model in the subsequent use is influenced.

In addition, based on the existing method, only whether the user object is the target user is predicted, and personalized requirements and preferences of the user object on the business data are not considered and analyzed, so that the acceptance of the pushed business data by the user object is low.

For the root cause of the above problems, the present specification first considers that the first sample data in the first data party and the second sample data in the second data party can be comprehensively utilized through the protocol rules based on horizontal federal learning to expand the sample data for model training.

When specifically utilizing the first sample data in the first data party and the second sample data in the second data party, further finding out: different data parties have different data channels and data isolation, so that the different data parties often have great difference in marking sample data and even have the problem of disordered data labels; further, it is difficult to directly collaborate with sample data held by different data parties, subject to data isolation.

For example, the head office, although not exclusively responsible for pushing business data, may hold non-business class feature data for a full number of sample users. And the subsidiary company which is specially responsible for business data pushing generally performs business data pushing aiming at the existing sample users of the main company. When business data is pushed specifically, a main company often pushes the business data of a subsidiary company only through a shared channel shared with the subsidiary company. On the subsidiary side, besides using the shared channel, the business data can be pushed through the own channel which the main company does not have. For example, on the head office side, the record data based on the common channel for sample user a is marked as negative sample user. On the subsidiary side, the user a is marked as a positive sample for the same sample based on the recorded data of the own channel.

In some cases, the subsidiary company records only the business class feature data of the user who has successfully pushed the data (corresponding to the positive sample user), and does not record the relevant data of the user who has failed to push the data (corresponding to the negative sample user).

Therefore, the head office often holds the non-business class feature data (which can be recorded as the first sample data) of the full sample users, and the subsidiary often holds the business class feature data (which can be recorded as the second sample data) of the full positive sample users (i.e. users whose business data push is successful).

Further, under the constraint of the relevant regulations, the head office and the subsidiary companies generally cannot directly transfer the data held by each other, and it is difficult to directly use the data held by both parties in combination to train the model.

In view of the above, the present specification further combines specific features in a data processing scenario and the correlation between the first sample data and the second sample data, before specific implementation, for the influence of the false negative sample on model training, the first server deployed in the first data party and holding the first sample data, and the second server deployed in the second data party and holding the second sample data may distinguish a customer group of a manager and a customer group of a non-manager in sample users, perform data fusion through cooperation according to a protocol rule based on horizontal federal learning, and obtain a preset target user prediction model with higher precision, where the preset target user prediction model includes a first prediction submodel corresponding to the customer group of the manager and a second prediction submodel corresponding to the customer group of the non-manager, by comprehensively using data training held by both parties.

Further, the first server and the second server are used for carrying out data fusion through cooperation based on the prediction result of a preset target user prediction model and according to a protocol rule based on longitudinal federal learning, and a preset user portrait prediction model capable of finely depicting the personalized demand tendency of a user object is obtained through comprehensive utilization of data held by the two servers.

In this way, in specific implementation, after accessing the feature data of the user object of the service data to be pushed, the server may determine the guest group type of the user object, and may further invoke a prediction sub-model in the preset target user prediction model, which is matched with the guest group type of the user object, to process the feature data of the user object, so as to accurately identify whether the user object belongs to a potential target user who will accept the pushed service data. And under the condition that the user object is determined to be the target user, further calling a preset user portrait prediction model to process the characteristic data of the user object so as to obtain a portrait label of the user object. And further, a target pushing rule matched with the user object can be generated according to the portrait label, and appropriate target business data can be pushed to the user object in a targeted mode according to the target pushing rule. Therefore, a better pushing effect can be obtained, the pushing success rate is improved, and the accurate pushing of the service data of the user object is realized.

The embodiment of the present specification provides a method for pushing service data, which may be specifically applied to a data processing system including a first server and a second server. As can be seen in fig. 1. The first server and the second server can be connected in a wired or wireless mode to perform corresponding data interaction.

In this embodiment, the first server may be specifically understood as a server deployed on a side of a first data party (or referred to as a first data domain, for example, a head office). The first server may hold first sample data in a first data party. The first sample data may be specifically understood as sample data obtained from the first data party. Each piece of sample data in the first sample data may specifically include two pieces of data, namely, identification information of the user object (for example, an identity ID of the user a, a mobile phone number of the user a, a name of the user a, and the like) and non-business feature data (for example, occupation of the user a, monthly income of the user a, number of default times of the user a, and the like) corresponding to the identification information and not directly related to business data. The non-traffic class feature data may further comprise a plurality of different non-traffic class features.

In this embodiment, the second server may be specifically understood as a server deployed on the side of the second data party (or referred to as the second data domain, e.g., a subsidiary company). The second server may hold second sample data in the second data party. The second sample data may be sample data obtained from a second data party. Each sample data in the second sample data may specifically include two parts of data, i.e., identification information of the user object and service class feature data directly related to the service data corresponding to the identification information. The service class feature data may further comprise a plurality of different service class features.

The identification information of the user object in the second sample data is included in the identification information of the user object in the first sample data.

In addition, because there is data isolation between the first data party and the second data party, the first server cannot directly transmit the held first sample data to the second server. Accordingly, the second server cannot directly transmit the held second sample data to the first server.

In this embodiment, before the specific implementation (modeling stage), the first server and the second server may cooperate to perform data fusion according to a protocol rule based on horizontal federal learning, and determine a full positive sample user from the full sample users according to first sample data and second sample data respectively held by the first server and the second server; and separating to obtain first type training data (corresponding to a first positive sample user and a first negative sample user in the customer management group) with definite negative samples and no pseudo negative samples and second type training data (corresponding to a second positive sample user and a second mixed negative sample user in the customer non-management group) with pseudo negative samples by distinguishing the customer management group and the customer non-management group; respectively training by utilizing the first type of training data to obtain a first prediction submodel aiming at the customer group, and training by utilizing the second type of training data to obtain a second prediction submodel aiming at the non-customer group; and combining the first prediction submodel and the second prediction submodel to obtain a preset target user prediction model comprising the two prediction submodels, thereby reducing the influence of negative sample data distribution deviation introduced by using a pseudo negative sample on the model precision in the model training process and obtaining the preset target user prediction model with higher precision and better effect.

Further, the first server and the second server can perform data fusion through cooperation according to a protocol rule based on longitudinal federal learning, non-business feature data in the first sample data are used as feature data of a positive sample user when a training model is used, and business feature data in the second sample data are used for determining a corresponding preset portrait label to obtain third training data carrying the preset portrait label; and then, training by using the third training data carrying the preset portrait label to obtain a preset user portrait prediction model which can finely predict and depict the personalized demand tendency and personal preference of the user object to the service data and has a good effect.

According to the mode, the first server and the second server can obtain the preset target user prediction model and the preset user portrait prediction model which are high in precision and good in effect through cooperation training.

The first server and/or the second server may hold and call the two prediction models obtained by the training, or may provide the two prediction models to a third server of a third party according to a cooperation protocol for use.

In specific implementation (a pushing stage), a server (which may be a first server, a second server, or a third server) accessing a data object of service data to be pushed may determine a guest group type of a user object according to identification information of the user object; the feature data of the user object can be processed by calling a prediction sub-model matched with the guest group type of the user object in a preset target user prediction model, so that whether the user object belongs to a potential target user capable of receiving the pushed business data or not can be accurately identified; under the condition that the user object is determined to be the target user, a preset user portrait prediction model is further called to process the feature data of the user object so as to obtain a portrait label of the user object; and further, a target pushing rule matched with the user object can be generated according to the portrait label, and appropriate target business data can be pushed to the user object in a targeted mode according to the target pushing rule.

Therefore, the user object is more willing to accept the pushed service data, a better pushing effect is obtained, the pushing success rate is improved, and the accurate pushing of the service data of the user object is realized.

In this embodiment, the first server and the second server may specifically include a background server that is applied to a service data processing platform side and is capable of implementing functions such as data transmission and data processing. Specifically, the first server and the second server may be, for example, an electronic device having data operation, storage functions and network interaction functions. Alternatively, the first server and the second server may also be software programs running in the electronic device and providing support for data processing, storage and network interaction. In this embodiment, the number of servers included in the first server and the second server is not specifically limited. The first server and the second server may be specifically one server, or several servers, or a server cluster formed by several servers.

Referring to fig. 2, an embodiment of the present disclosure provides a method for pushing service data. When the method is implemented, the following contents may be included.

S201: acquiring identification information of a user object and characteristic data of the user object; wherein the feature data at least comprises non-traffic class feature data.

S202: and determining the guest group type of the user object according to the identification information of the user object.

S203: calling a prediction sub-model matched with the guest group type of the user object in a preset target user prediction model to process the characteristic data of the user object so as to determine whether the user object is a target user; the preset target user prediction model is established in advance according to first sample data in a first data party and second sample data in a second data party on the basis of horizontal federal learning.

S204: under the condition that the user object is determined to be a target user, calling a preset user portrait prediction model to process feature data of the user object so as to obtain a portrait label of the user object; the preset user portrait prediction model is established in advance according to first sample data in a first data party and second sample data in a second data party based on longitudinal federal learning.

S205: and generating a target pushing rule matched with the user object according to the portrait label of the user object.

S206: and pushing corresponding target business data to the user object according to the target pushing rule.

Through the embodiment, the server can accurately identify and determine whether the user object of the business data to be pushed is a target user of potential meeting business data or not based on the preset target user prediction model established by horizontal federal learning union according to sample data held by multiple parties respectively in advance; under the condition that the user object is determined to be the target user, further utilizing a preset user portrait prediction model established in advance according to multi-party held sample data respectively and based on longitudinal federal learning combination to finely determine portrait labels capable of reflecting individual demand tendency and personal preference of the user object; and further, according to the portrait label, the appropriate target business data can be pushed to the user object accurately, so that the user object is more willing to accept the pushed business data, and the success rate of pushing is improved.

In one embodiment, the method may be applied specifically on the server side. The server may be a first server deployed on a side of a first data party, a second server deployed on a side of a second data party, or a third server deployed on a side of a third party different from the first data party and the second data party.

The first data party may be a data party holding the first sample data. The first sample data may specifically include non-traffic class feature data of a full number of sample users. The second data party may specifically be a data party holding second sample data. The first sample data may specifically include service class feature data of a full number of positive sample users. The full positive sample users are a subset of full sample users. The third data party may be specifically different from the first data party and the second data party, and is not a third party holding sample data.

In an embodiment, the service class feature data may be specifically understood as feature data directly associated with service data to be pushed, and the non-service class feature data may be specifically understood as feature data indirectly associated with the service data to be pushed. The service data may specifically include: insurance business products, insurance business services, financing business products, financing business services, and the like. Of course, the above listed service data is only a schematic illustration. In specific implementation, the service data may also include other types of service data according to specific application scenarios and processing requirements. The present specification is not limited to these.

In an embodiment, the service class feature data and the non-service class feature data may specifically include different types of features respectively according to different service data, the first data party and the second data party.

The following description will specifically take the business data as an insurance business product, the first data party as a bank head office, and the second data party as an insurance business sub-company subordinate to the bank head office.

In this embodiment, the non-service feature data may specifically include one or more of the following features: monthly income, monthly consumption amount, inventory amount, occupation, etc. Correspondingly, the service class feature data may specifically include one or more of the following listed features: gift identification, historical premium, historical purchased risk category, and the like.

In this embodiment, the first server deployed on the first data side may hold non-business class feature data of a full number of sample users as the first sample data. The first data party is not mainly responsible for pushing the business data, but assists the second data party in pushing the business data through a shared channel (for example, bank insurance and the like) shared by the second data party. Therefore, the positive sample users (i.e., users that receive the pushed traffic data) that are marked out of the total number of sample users in the first sample data held by the first data party are often not complete.

The second server deployed on the second data side is mainly responsible for pushing the service data. Moreover, due to the association relationship between the second data party and the first data party, the second data party often mainly pushes related service data for the full number of sample users held by the first data party, and records the sample user (marked as a positive sample user) that is successfully pushed and the service class feature data of the positive sample user that is successfully pushed.

In addition, the second data party pushes the business data by using a shared channel shared by the first data party and also pushes the business data by using a self-owned channel (such as a self-selling channel of an insurance company) which is not possessed by other first data parties, and records the positive sample user and the business class feature data of the positive sample user. For such positive sample users, the first data party is not known or marked on the side of the first data party due to data isolation between the first data party and the second data party.

Thus, the second server may hold the traffic class feature data of the full number of positive sample users as the second sample data. And the above-mentioned full positive sample users are often also a subset of full sample users.

In an embodiment, when preparing to push service data to a user object, the server may first obtain identification information of the user object and feature data of the user object.

The identification information may specifically include identification information such as a name, an identity ID, a mobile phone number, and an account name for indicating the user object. The feature data of the user object may specifically be non-business feature data of the user object. Of course, the feature data of the user object may also be a combination of non-service class feature data and service class feature data of the user object.

In an embodiment, the server may determine, according to the identification information of the user object, a type of the guest group to which the user object belongs by data query, for example, by querying a preset guest group list.

In one embodiment, the guest group type may specifically include: an administrative group and a non-administrative group.

In this embodiment, the user object is firstly classified into the guest group types according to the guest group of the user and the guest group of the non-user, so that a prediction sub-model matched with the guest group type in a preset target user prediction model can be called subsequently, and whether the user object is the target user or not is accurately determined.

In an embodiment, the customer group for management may be specifically understood as a customer group for which a special person is responsible for service and maintenance. For the customer management group, the server often holds more comprehensive characteristic data and more complete history push records.

The non-management customer group can be specifically understood as a customer group without special people for service and maintenance. For this non-administrative customer base, the feature data held by the server is often incomplete, and the history push records held are also often incomplete.

Accordingly, the full number of sample users in the first sample data held by the first server may also include: the first data party provides a management customer group for which the special person is responsible for service and maintenance, and a non-management customer group for which the special person is not responsible for service and maintenance. For the management customer group, the first server holds a complete history pushing record based on a common channel; for non-administrative customer groups, the historical push records based on common channels held by the first server are complete. In addition, the first server does not hold history push records based on other channels (for example, the own channel of the second data party) except the shared channel for the administrative customer group and the non-administrative customer group.

In one embodiment, when the server is implemented, a matched prediction sub-model can be determined from prediction sub-models contained in a preset target user prediction model according to the guest group type of the user object; and then, the matched predictor model can be called to process the characteristic data of the user object, and whether the user object is a target user or not is determined according to a processing result output by the model.

In one embodiment, the target user may be specifically understood as a potential customer group user with a relatively high probability of accepting the pushed service data.

In an embodiment, the preset target user prediction model may be specifically understood as a prediction model that is obtained by the first server and the second server cooperatively in advance by using the first sample data and the second sample data held by the first server and the second server, based on horizontal federal learning training, and that is capable of predicting a probability value that a user object belongs to a target user based on feature data of the user object. The training mode of the preset target user prediction model will be further described later.

In one embodiment, in particular, the guest group types include: under the condition of the administrative customer group and the non-administrative customer group, the preset target user prediction model specifically may include two different prediction submodels, namely a first prediction submodel and a second prediction submodel. The first predictor model is matched with the customer management group, and the second predictor model is matched with the non-customer management group.

In this embodiment, the first predictor model may be specifically obtained by training in advance with non-business feature data (which may be referred to as first-class training data) of the customer group, and can relatively accurately determine whether the user object is a predictor model of the target user for the customer group. The second predictor model may be obtained by training in advance using non-service type feature data (which may be referred to as second type training data) of the non-managed customer group, and may be a predictor model that can more accurately determine whether the user object is a target user for the non-managed customer group.

By establishing and utilizing the preset target user prediction model comprising the first prediction sub-model corresponding to the management customer group and the second prediction sub-model corresponding to the non-management customer group, whether the user object is a potential target user for receiving the service data can be determined more finely.

In one embodiment, in the case that the guest group type of the user object is determined to be an administrator guest group, a first prediction submodel in a preset target user prediction model may be determined to be a matched prediction submodel; the first predictor model may then be invoked to process the feature data of the user object to determine whether the user object is a target user.

Under the condition that the user object is determined to be a non-user-management user group, a second prediction sub-model in a preset target user prediction model can be determined to be a matched prediction sub-model; a second predictor model may then be invoked to process the feature data for the user object to determine whether the user object is a target user.

In an embodiment, in a case that the user object is determined to be the target user by using the preset target user prediction model, further, the preset user portrait prediction model may be invoked to process the feature data of the user object, so as to finely determine the business portrait label of the user object according to the model output.

In an embodiment, the preset user portrait prediction model may be specifically understood as a prediction model that the first server and the second server cooperate in advance to use the first sample data and the second sample data held by the first server and the second server to predict a portrait label, which is matched with the user object and relates to business data, based on feature data of the user object, based on vertical federal learning training. The training method of the predetermined user portrait prediction model will be further described.

In one embodiment, the portrait tag may be specifically understood as tag data reflecting a requirement or preference of a user object for an attribute dimension of the service data that is prone to be accepted.

In an embodiment, in the case that the service data is insurance service product or insurance service, and the application scenario is insurance service marketing scenario, the portrait label may specifically include one or more of the following listed portrait labels: payment terms propensity label, premium propensity label, and the like. Of course, the above-mentioned image labels are only illustrative.

In specific implementation, the portrait label may also include other types of portrait labels according to specific service data and specific application scenarios. For example, in the case where the business data is a financial product and the application scenario is a marketing scenario of the financial product, the portrait label may be an annual profit margin tendency label, a fixed holding duration tendency label, a financial risk level tendency label, or the like.

In one embodiment, the predetermined user portrait prediction model may specifically include a plurality of portrait prediction submodels. Wherein each of the plurality of image predictor models corresponds to one image label.

Correspondingly, a plurality of portrait prediction submodels contained in a preset portrait prediction model of the user can be called to respectively process the characteristic data of the user object, a plurality of different portrait labels corresponding to the user object are obtained, and further the personalized demand tendency or preference of the user object can be more finely and comprehensively depicted by utilizing the different portrait labels.

In one embodiment, matching target push rules can be customized for a user object in a targeted manner according to portrait labels of the user object; and further, according to the target pushing rule, appropriate target business data can be pushed to the user object.

In an embodiment, the target pushing rule may specifically be a construction manner of business data matched with the user object and applicable to the user object, a marketing scheme of the business data, or a recommendation policy of the business data. The target push rule may also be rule data of other contents corresponding to different service data and application scenarios. The present specification is not limited to these.

In one embodiment, specifically, the server may determine a target push rule suitable for the user object in a targeted manner according to the portrait label of the user object; further, target business data with high matching degree with the user object and high user object receiving probability can be configured according to a target pushing rule; and then, according to a target pushing rule, selecting a pushing mode which is preferred by the user object and can be reached in time, and pushing the target service data to the user object.

When the user object receives the target service data pushed according to the target pushing rule, relatively, the user object has a higher probability of being willing to accept the target service data. Therefore, the pushing success rate can be effectively improved, and a better pushing effect is obtained.

As can be seen from the above, based on the method for pushing business data provided in this specification, when a server pushes business data to a user object, a guest group type of the user object may be determined first, and then a prediction sub-model that is matched with the guest group type of the user object in a preset target user prediction model established in advance based on horizontal federal learning according to first sample data in a first data party and second sample data in a second data party may be called to process feature data of the user object, so as to accurately identify whether the user object belongs to a potential target user that will accept the pushed business data; under the condition that the user object is determined to be the target user, further calling a preset user portrait prediction model established in advance according to the first sample data and the second sample data based on longitudinal federal learning to process the feature data of the user object so as to obtain a portrait label of the user object; and then, a target pushing rule matched with the user object can be generated according to the portrait label, and appropriate target business data can be pushed to the user object in a targeted manner according to the target pushing rule, so that a good pushing effect can be obtained, the pushing success rate is improved, and the accurate pushing of the business data of the user object is realized.

In an embodiment, the preset target user prediction model (which may be denoted as model a) invoked by the server may be specifically established by the first server and the second server in advance based on horizontal federal learning according to first sample data in the first data party and second sample data in the second data party respectively held by the first server and the second server.

In an embodiment, in implementation, for the first server side, the preset target user prediction model may be specifically trained in the following manner.

S1: the method comprises the steps that a first server responds to a first training request about a preset target user prediction model, and performs sample data fusion in cooperation with a second server according to a protocol rule based on horizontal federal learning so as to determine a first positive sample user, a first negative sample user and a first pseudo negative sample user in an output customer group sample user and a second positive sample user and a second mixed negative sample user in a non-output customer group sample user; the first server is deployed at one side of a first data party, the held first sample data at least comprises non-business characteristic data of a full amount of sample users, and the full amount of sample users comprise user management customer group sample users and non-user management customer group sample users; the second server is a server deployed at one side of a second data party, and second sample data held by the second server at least comprises service class characteristic data of a full positive sample user;

s2: acquiring non-service characteristic data of a first positive sample user and non-service characteristic data of a first negative sample user from the first sample data to serve as first grouping sample data; acquiring non-service characteristic data of a second positive sample user and non-service characteristic data of a second mixed negative sample user from the first sample data to serve as second grouping sample data;

s3: training by using the first grouping sample data to obtain a first predictor model (which can be marked as a model A1); training to obtain a second predictor model (which can be marked as a model A2) by using the second sample data;

s4: and combining the first prediction submodel and the second prediction submodel to obtain a preset target user prediction model.

Through the embodiment, the first server can cooperate with the second server through the horizontal federal learning framework according to the protocol rule based on horizontal federal learning, and on the premise that the sample data held by the first server cannot be out of range, the full positive sample users are determined and marked from the full sample users held by one side of the first server; dividing the full amount of sample users into user management group sample users and non-user management group sample users based on the type of the user group; and according to the characteristics of the user group to be managed and the user group to be unmanaged, the non-business characteristic data of the user group sample user in the held first sample data is used for training a first prediction sub-model corresponding to the user group to be managed, the non-business characteristic data of the user group sample to be unmanaged is used for training a second prediction sub-model corresponding to the user group to be unmanaged, and therefore a preset target user prediction model capable of identifying and determining potential target users more accurately can be obtained.

In an embodiment, the first training request may be specifically initiated by a first server, or may be initiated by a second server. The present specification is not limited to these.

In one embodiment, it is considered that in some cases, there may be data isolation between the first server and the second server, and the two parties cannot directly interact with each other to respectively hold the first sample data and the second sample data. Therefore, the horizontal federal learning framework can be utilized to perform safe sample data fusion according to the cooperation of the protocol rules based on the horizontal federal learning so as to break the limitation of data isolation and not to directly reveal the sample data held by the own party to the other party.

Through the data fusion, the first server can determine and mark a full number of positive sample users in the first sample data held by the own party (for example, preset marking information for indicating the positive sample users is set in the positive sample users); meanwhile, the first server can also divide the full sample data into two groups, namely a user sample of the user management client group and a user sample of the user non-user management client group.

Further, for the user of the user-account sample, the first server may first determine a positive sample user from the user-account sample users as the first positive sample user by retrieving the mark information of the user-account sample user. In addition, the first server also holds a complete shared channel-based historical push record for the user-customer group sample user. Therefore, the first server can screen the remaining user sample users of the user customer group, except the first positive sample user, from the user sample users of the user customer group according to the historical pushing records of the user sample users of the user customer group, and screen out the sample users who have previously pushed the service data but do not accept the pushed service data as the first negative sample user (i.e. pure negative sample user); meanwhile, the remaining sample users (i.e., sample users that have not previously pushed traffic data) are taken as the first pseudo-negative sample users.

For the non-user-managed customer group sample users, the first server may first determine a positive sample user from the non-user-managed customer group sample users as a second positive sample user by retrieving the tag information of the non-user-managed customer group sample users. Furthermore, the common channel-based historical push records held by the first server for the non-managed customer base sample users are incomplete. Therefore, for the non-user-managed group, the first server cannot clearly distinguish the pure negative sample users from the pseudo negative sample users as for the user-managed group. The first server may statistically determine remaining non-user-customer-group sample users of the non-user-customer-group sample users excluding the second positive sample user as second mixed negative sample users.

Through the embodiment, the first server and the second server can cooperate through a protocol rule based on horizontal federal learning to determine and identify a first positive sample user, a first negative sample user and a first pseudo negative sample user from sample users of a customer management customer group in a total number of sample users held by the first server; meanwhile, a second positive sample user and a second mixed negative sample user are determined and identified in the sample users of the non-user-managed group.

In one embodiment, for the user group, the first server may obtain, from the held first sample data, non-traffic class feature data of the first positive sample user and non-traffic class feature data of the first negative sample user as first packet sample data. The obtained first grouping sample data can not be doped with non-service characteristic data of a false negative sample user, and has no data offset of a negative sample and better training effect.

Meanwhile, for the non-administrative customer group, the first server may obtain, from the held first sample data, the service class feature data of the second positive sample user and the non-service class feature data of the second mixed sample user as second packet sample data.

Therefore, the first server can accurately split the first component sample data and the second component sample data which are used for training different prediction submodels aiming at different customer group types from the held first sample data.

In an embodiment, the training to obtain the first predictor model by using the first packet of sample data may include the following steps.

S1: extracting a plurality of data from the non-business characteristic data of a first positive sample user in the first grouping sample data, and marking the data as first positive sample training data;

s2: extracting a plurality of data from the non-business class characteristic data of a first negative sample user in the first grouping sample data, and marking the data as first negative sample training data;

s3: and training to obtain a first predictor model by using the first positive sample training data and the first negative sample training data.

By the embodiment, the second grouped sample data with the pseudo-negative sample is stripped in advance, so that the model error caused by the drift of the negative sample data introduced by using the pseudo-negative sample during model training can be effectively eliminated, and the first predictor model with higher precision and better effect for the customer group can be trained by utilizing the first positive sample training data and the first negative sample training data of the customer group.

In an embodiment, during the specific training, a Lookalike method may be adopted, and model training may be performed by using the first positive sample training data and the first negative sample training data to obtain a first predictor model meeting requirements.

The above Lookalike method may also be referred to as a similar population expansion method, and specifically may be a technology for finding more similar populations with potential relevance by using a certain algorithm evaluation model based on a seed user. Based on the Lookalike method, during specific implementation, various technologies such as collaborative filtering, node2vec and the like can be effectively and comprehensively applied to expand the users, and a prediction submodel with good effect for predicting the target users is obtained.

In an embodiment, the training to obtain the second predictor model by using the second packet of sample data may include the following steps.

S1: extracting a plurality of data from the business class characteristic data of a second positive sample user in the second grouping sample data, and marking the data as second positive sample training data;

s2: randomly extracting a plurality of data from the business class characteristic data of a second mixed negative sample user in the second grouped sample data, and marking the data as second negative sample training data;

s3: and training to obtain a second predictor model by using the second positive sample training data and the second negative sample training data.

By the embodiment, the error influence caused by the drift of the negative sample data introduced by using the pseudo negative sample during model training can be reduced to a certain extent, and the second predictor model for the non-user client group is obtained by training by using the second positive sample training data and the second negative sample training data of the non-user client group.

In this embodiment, the labeled first positive sample training data and the labeled first negative sample training data may be recorded as first type training data for training the first predictor model. And recording the marked second positive sample training data and the marked second negative sample training data as second type training data for training a second predictor model.

In an embodiment, during the specific training, a lookelike method may be adopted, and the model training is performed by using the second positive sample training data and the second mixed negative sample training data, so as to obtain a second predictor model meeting the requirement.

In one embodiment, in specific implementation, the first server may further obtain, by another way, historical service information associated with a second mixed negative sample user with respect to the second mixed negative sample user; and according to the associated historical service information, carrying out speculation to further screen out sample users which have higher probability to belong to pure negative sample users from the second mixed negative sample users as second prediction negative sample users. Correspondingly, the non-business class feature data of the second prediction negative sample user can be extracted in a targeted manner from the business class feature data of the second mixed negative sample user in the second grouping sample data, and the non-business class feature data serves as second negative sample training data to participate in training of the second prediction submodel. Therefore, the influence of negative sample data drifting during training is effectively reduced, and the prediction sub-model with higher precision and better effect for the non-user client group is obtained through training.

In an embodiment, when the second predictor model is specifically trained, a plurality of data may be respectively extracted from the first negative sample training data and the second mixed negative sample training data, so that the influence of the pseudo negative sample in the non-customer group is weakened by the pure negative sample in the customer group, and the combined second mixed negative sample data is obtained. Furthermore, the combined second mixed negative sample data can be used for participating in the training of the second predictor model so as to reduce the influence of the negative sample data drifting in the training process, and the predictor model with higher precision and better effect for the non-user client group is obtained.

In an embodiment, the preset user portrait prediction model (which may be denoted as model B) invoked by the server may be specifically established by the first server and the second server in advance based on the vertical federal learning according to first sample data in the first data party and second sample data in the second data party respectively held by the first server and the second server.

In one embodiment, the predetermined user profile prediction model may be trained on the first server side in the following manner.

S1: the first server responds to a second training request about a preset user portrait prediction model, and cooperates with the second server to perform sample data fusion according to a protocol rule based on longitudinal federal learning to obtain third training data carrying a preset portrait label; wherein the third training data comprises non-traffic class feature data of the positive sample user; the preset portrait label is determined according to the business class characteristic data of the user who just samples;

s2: and training to obtain a preset user portrait prediction model by using the third training data carrying the preset portrait label.

Through the embodiment, the first server can cooperate with the second server according to the protocol rule based on longitudinal federal learning, the second server determines the preset portrait label of the positive sample user by using the held business characteristic data, and then the portrait label is fused with the non-business characteristic data of the positive sample user held by the first server to obtain third training data of the preset portrait label carrying data knowledge containing business characteristic data; and then, a preset user portrait prediction model which can predict and determine portrait labels of users about business data more accurately is obtained through training by using the third training data.

In an embodiment, the second training request may be specifically initiated by the first server, or may be initiated by the second server. The present specification is not limited to these.

In one embodiment, in consideration of the possible data isolation between the first server and the second server in some scenarios, two parties cannot directly interact with each other to respectively hold the first sample data and the second sample data. Therefore, the longitudinal federal learning framework can be utilized to perform safe sample data fusion according to the cooperation of the protocol rules based on the longitudinal federal learning so as to break the limitation of data isolation and not directly reveal the data held by the own party to the other party.

Through the data fusion, the second server can determine preset portrait tags (such as payment deadline tendency tags, premium tendency tags and the like) of the positive sample user by using service class feature data in second sample data held by the own party; meanwhile, the first server can obtain the non-business characteristic data of the positive sample user from the held first sample data; and then the first server and the second server can combine the preset portrait label and the non-business characteristic data corresponding to the same positive sample user through cooperation to obtain the third training data carrying the preset portrait label.

In one embodiment, in the concrete integration, the first server and the second server can also combine three data, namely a preset portrait label corresponding to the same positive sample user, non-business characteristic data and business characteristic data, through cooperation to obtain relatively more complete and rich third training data carrying the preset portrait label, so that a more accurate preset user portrait prediction model can be trained.

In an embodiment, when the preset user portrait prediction model is specifically trained, the initial model may be learned and trained by using the third training data carrying the preset portrait label, so as to obtain the preset user portrait prediction model.

In one embodiment, the positive sample user training the preset user portrait prediction model may be a sample user that is determined in advance by using the preset target user prediction model and marked as a target user. Therefore, the trained preset target user prediction model and the preset user portrait prediction model have stronger associativity.

In one embodiment, the predetermined user portrait prediction model may specifically include one portrait prediction submodel, or may include a plurality of portrait prediction submodels, where different portrait prediction submodels are used to predict different portrait tags.

By the embodiment, a plurality of different portrait labels of the user object can be predicted simultaneously by using the preset user portrait prediction model; furthermore, according to the different portrait labels, the requirement tendency and the personal preference of the user object can be more finely depicted, so that a target pushing rule with higher matching degree and better effect can be generated for the user object.

In one embodiment, in the case that the preset user portrait prediction model includes a plurality of portrait prediction submodels, the second server and the first server may generate a plurality of different preset portrait labels according to different business class feature data of the held positive sample user during data fusion according to a protocol rule based on longitudinal federal learning. Thus, the third training data obtained through data fusion finally can contain multiple groups of third training data. Wherein, the third training data of the same group carries the same kind of preset portrait label.

Correspondingly, during training, a plurality of initial models can be trained by utilizing the plurality of groups of third training data respectively to obtain a plurality of portrait prediction submodels respectively corresponding to different portrait labels; and combining the plurality of portrait prediction submodels to obtain a preset user portrait prediction model.

In one embodiment, taking an insurance business marketing scenario as an example, the target business data may specifically include insurance business products and/or insurance business services; correspondingly, the target pushing rule may specifically include an insurance business marketing scheme.

In the insurance business marketing scene, whether the current user object to be marketed belongs to a potential target user who can accept the marketed insurance business can be determined by using a preset target user prediction model; under the condition that the current user object belongs to the target user, predicting a portrait label of the user object related to the insurance business by using a preset user portrait prediction model, and further determining the demand tendency and personal preference of the current user object for the insurance business according to the portrait label; further, an insurance business marketing scheme which is matched with the current user object and has higher acceptance of the current user object can be generated by combining the personalized demand tendency and the personal preference; and further, according to the insurance business marketing scheme, the appropriate marketing treatment of insurance business products and/or insurance business services can be accurately carried out on the current user object. Therefore, the marketing success rate can be effectively improved, meanwhile, the appropriate insurance business can be recommended to the user accurately according to the individual demand tendency and preference of the user, and the service experience of the user is improved.

In an embodiment, taking an insurance business marketing scenario as an example, the non-business class feature data may specifically include at least one of the following: monthly income, monthly consumption amount, deposit balance, etc.; the service class feature data may specifically include at least one of the following: gift identification, premium, etc. Of course, the above listed non-service class feature data and service class feature data are only schematic illustrations. In specific implementation, the non-service characteristic data and the service characteristic data may further include other types of data according to a specific application scenario and a processing requirement. The present specification is not limited to these.

By acquiring and utilizing the diversified business characteristic data and the non-business characteristic data, a preset target user prediction model and a preset user portrait prediction model which have higher precision and better effect and aim at an insurance business marketing scene can be trained.

In an embodiment, taking an insurance business marketing scenario as an example, the portrait label may specifically include at least one of the following: payment due trend labels, premium trend labels, and the like. Of course, the above-listed image labels are merely illustrative. In particular, the portrait tags may also include other types of tag data, depending on the particular application scenario and processing requirements. The present specification is not limited to these.

By utilizing the diversified portrait label, the individualized demand tendency and the individual preference of the user object about the insurance business can be more finely and comprehensively depicted in the insurance business marketing scene, so that the user object can be more accurately and properly marketed for the insurance business.

In one embodiment, as shown in fig. 3, after the preset target user prediction model and the preset user portrait prediction model are obtained through training in the above manner, the two models may be further divided into two layers to be combined together to obtain a combined processing model. The first layer model in the combined processing model is a preset target user prediction model A, and the second layer model is a preset user portrait prediction model B. The preset target user prediction model a may further include a first predictor model a1 and a second predictor model a2 corresponding to the tenant group and the non-tenant group, respectively. The predetermined user portrait prediction model B may further include a plurality of portrait label prediction submodels respectively corresponding to a plurality (e.g., n) of different portrait labels, which are respectively expressed as: a picture label predictor model B1, a picture label predictor model B2, and a picture label predictor model B3 … …, and a picture label predictor model Bn.

And only the characteristic data of the user object predicted as the target user by the first layer model flows into the second layer model through the first layer model so as to trigger the prediction of the portrait label of the user object.

In specific implementation, the feature data of the user object of the combined processing model is input, and after entering the first layer model, the feature data is shunted to a prediction sub-model (for example, the first prediction sub-model a1 or the second prediction sub-model a2) matched with the guest group type of the user object through a classifier to be processed so as to determine whether the user object belongs to the target user. After determining that the user object belongs to the target user, the feature data of the user object is flowed from the first layer model to the second layer model for the next processing.

After entering the second layer model, the feature data of the user object is processed by the image tag predictor models B1 and B2 … … Bn simultaneously flowing into the second layer model, so as to obtain n image tags corresponding to the user object.

Finally, n portrait tags determined as user objects of the target user may be output through the above-described combined processing model.

In an embodiment, in a specific implementation, the server may further push service data to the batch of user objects by using the combined processing model. Specifically, the server may combine the feature data input values of the batch of user objects with a first layer model in the processing model for processing. The first layer model can screen out a plurality of user objects belonging to the target user according to the characteristic data of the batch of user objects to obtain a target user list. And further, the characteristic data of the user object in the target user list can be transmitted to the second layer model for processing. The second layer model can determine the portrait labels of all the user objects on the target user list according to the characteristic data of the user objects on the target user list, and the portrait labels are output as a final model.

Furthermore, the server may generate a matching target push rule for each user object on the target user list according to the portrait label of each user object on the target user list. And then pushing appropriate target service data to a plurality of user objects on the target user list in batches according to the target pushing rule. Therefore, batch, efficient and accurate pushing of the service data can be realized.

As can be seen from the above, in the method for pushing service data provided in this specification, when pushing service data to a user object, a guest group type of the user object may be determined first, and then a prediction sub-model that is matched with the guest group type of the user object in a preset target user prediction model that is established in advance based on horizontal federal learning according to first sample data in a first data party and second sample data in a second data party may be called to process feature data of the user object, so as to accurately identify whether the user object belongs to a potential target user that will receive the pushed service data; under the condition that the user object is determined to be the target user, further calling a preset user portrait prediction model established in advance according to the first sample data and the second sample data based on longitudinal federal learning to process the feature data of the user object so as to obtain a portrait label of the user object; and then, a target pushing rule matched with the user object can be generated according to the portrait label, and appropriate target business data can be pushed to the user object in a targeted manner according to the target pushing rule, so that a good pushing effect can be obtained, the pushing success rate is improved, and the accurate pushing of the business data of the user object is realized. When a preset target user prediction model is trained, distinguishing an account management customer group and a non-account management customer group, enabling a first server holding first sample data in a first data party and a second server holding second sample data in a second data party to perform data fusion through cooperation according to a protocol rule based on horizontal federal learning, and separating to obtain first type training data with definite negative samples and no pseudo negative samples and second type training data with pseudo negative samples; and then respectively training a first prediction submodel only aiming at the customer group of the managed user by utilizing the first type of training data and training a second prediction submodel only aiming at the customer group of the unmanaged user by utilizing the second type of training data to obtain a preset target user prediction model comprising two prediction submodels, so that the influence of distribution deviation introduced by a pseudo negative sample on the model precision in the model training process is reduced, and the preset target user prediction model with higher precision and better effect is obtained. When a preset user portrait prediction model is trained, a first server holding first sample data in a first data party and a second server holding second sample data in a second data party are subjected to data fusion through cooperation according to a protocol rule based on longitudinal federal learning, non-business feature data in the first sample data are used as feature data of a positive sample user when the model is trained, and business feature data in the second sample data are used for determining a corresponding preset portrait label to obtain third training data carrying the preset portrait label; and then, the third training data carrying the preset portrait label is utilized to train to obtain a preset user portrait prediction model which can finely predict and depict the personalized demand tendency of the user object to the service data and has a good effect.

Referring to fig. 4, an embodiment of the present disclosure further provides a method for training a model to obtain a predetermined target user prediction model meeting requirements. The method can be applied to a first server deployed on the first data side. When implemented, the following may be included.

S401: responding to a first training request about a preset target user prediction model, and performing sample data fusion in cooperation with a second server according to a protocol rule based on horizontal federal learning so as to determine a first positive sample user, a first negative sample user and a first pseudo negative sample user in an out-of-customer group sample user, and a second positive sample user and a second mixed negative sample user in a non-out-of-customer group sample user; the first sample data held by the first server at least comprises non-business class characteristic data of full-scale sample users, wherein the full-scale sample users comprise customer group sample users and non-customer group sample users; the second server is a server deployed at one side of a second data party, and second sample data held by the second server at least comprises service class characteristic data of a full positive sample user;

s402: acquiring non-service characteristic data of a first positive sample user and non-service characteristic data of a first negative sample user from the first sample data to serve as first grouping sample data; acquiring non-service characteristic data of a second positive sample user and non-service characteristic data of a second mixed negative sample user from the first sample data to serve as second grouping sample data;

s403: training by using the first grouping sample data to obtain a first predictor model; training to obtain a second predictor model by utilizing the second grouping of sample data;

s404: and combining the first prediction submodel and the second prediction submodel to obtain a preset target user prediction model.

In this embodiment, when a preset target user prediction model is trained, by distinguishing and utilizing characteristics of sample data of a user-managed group and a user-unmanaged group, a first server holding first sample data in a first data party and a second server holding second sample data in a second data party perform data fusion through cooperation according to a protocol rule based on horizontal federal learning, so as to separate and obtain first type of training data (including first positive sample training data and first negative sample training data) with definite negative samples and without false negative samples, and second type of training data (including second positive sample training data and second negative sample training data) with false negative samples; and respectively training a first prediction submodel aiming at the customer group by utilizing the first type of training data and training a second prediction submodel aiming at the non-customer group by utilizing the second type of training data to obtain a preset target user prediction model comprising two prediction submodels, so that the influence of distribution deviation introduced by a pseudo negative sample on the model precision in the model training process is reduced, and the preset target user prediction model with higher precision and better effect is obtained.

Referring to fig. 5, an embodiment of the present disclosure further provides a model training method for training a predetermined user portrait prediction model meeting requirements. The method can be applied to a first server deployed on the first data side. When implemented, the following may be included.

S501: responding to a second training request about a preset user portrait prediction model, and performing sample data fusion by cooperating with a second server according to a protocol rule based on longitudinal federal learning to obtain third training data carrying a preset portrait label; wherein the third training data comprises non-traffic class feature data of the positive sample user; the preset portrait label is determined according to the business class characteristic data of the user who just samples; the first sample data held by the first server at least comprises non-business class characteristic data of a full amount of sample users; the second server is a server deployed at one side of a second data party, and second sample data held by the second server at least comprises service class characteristic data of a full positive sample user;

s502: and training to obtain a preset user portrait prediction model by using the third training data carrying the preset portrait label.

In this embodiment, when a preset user portrait prediction model is trained, a first server holding first sample data in a first data party and a second server holding second sample data in a second data party are subjected to data fusion through cooperation according to a protocol rule based on longitudinal federal learning, non-business class feature data in the first sample data are used as feature data of a positive sample user when the model is trained, and business class feature data in the second sample data are used for determining a corresponding preset portrait label, so that third training data carrying the preset portrait label are obtained; and then, training by using the third training data carrying the preset portrait label to obtain a preset user portrait prediction model which can finely predict and depict the personalized demand tendency and personal preference of the user object to the service data and has a good effect.

Referring to fig. 6, another method for pushing service data is further provided in the embodiments of the present disclosure. The method may be embodied as follows.

S601: acquiring identification information of a user object and characteristic data of the user object; wherein the feature data at least comprises non-traffic class feature data;

s602: determining the guest group type of the user object according to the identification information of the user object;

s603: calling a prediction sub-model matched with the guest group type of the user object in a preset target user prediction model to process the characteristic data of the user object so as to determine whether the user object is a target user; the preset target user prediction model is established in advance according to first sample data in a first data party and second sample data in a second data party on the basis of transverse federal learning;

s604: and under the condition that the user object is determined to be the target user, pushing corresponding target service data to the user object.

In this embodiment, by distinguishing the guest group type of the user object and using the matched predictor model in the preset target user prediction model to accurately determine whether the user object is a potential target user, the user object can be correspondingly pushed with service data, so that the success rate of pushing the service data can be improved.

Embodiments of the present specification further provide a server, including a processor and a memory for storing processor-executable instructions, where the processor, when implemented, may perform the following steps according to the instructions: acquiring identification information of a user object and characteristic data of the user object; wherein the feature data at least comprises non-traffic class feature data; determining the guest group type of the user object according to the identification information of the user object; calling a prediction sub-model matched with the guest group type of the user object in a preset target user prediction model to process the characteristic data of the user object so as to determine whether the user object is a target user; the preset target user prediction model is established in advance according to first sample data in a first data party and second sample data in a second data party on the basis of transverse federal learning; under the condition that the user object is determined to be a target user, calling a preset user portrait prediction model to process feature data of the user object so as to obtain a portrait label of the user object; the preset user portrait prediction model is established in advance according to first sample data in a first data party and second sample data in a second data party on the basis of longitudinal federal learning; generating a target pushing rule matched with the user object according to the portrait label of the user object; and pushing corresponding target business data to the user object according to the target pushing rule.

In order to complete the above instructions more accurately, referring to fig. 7, another specific server is provided in the embodiments of the present specification, where the server includes a network communication port 701, a processor 702, and a memory 703, and the above structures are connected by an internal cable, so that the structures may perform specific data interaction.

The network communication port 701 may be specifically configured to obtain identification information of a user object and feature data of the user object; wherein the feature data at least comprises non-traffic class feature data.

The processor 702 may be specifically configured to determine, according to the identification information of the user object, a guest group type of the user object; calling a prediction sub-model matched with the guest group type of the user object in a preset target user prediction model to process the characteristic data of the user object so as to determine whether the user object is a target user; the preset target user prediction model is established in advance according to first sample data in a first data party and second sample data in a second data party on the basis of transverse federal learning; under the condition that the user object is determined to be a target user, calling a preset user portrait prediction model to process feature data of the user object so as to obtain a portrait label of the user object; the preset user portrait prediction model is established in advance according to first sample data in a first data party and second sample data in a second data party on the basis of longitudinal federal learning; generating a target pushing rule matched with the user object according to the portrait label of the user object; and pushing corresponding target business data to the user object according to the target pushing rule.

The memory 703 may be specifically configured to store a corresponding instruction program.

In this embodiment, the network communication port 701 may be a virtual port that is bound to different communication protocols, so that different data can be sent or received. For example, the network communication port may be a port responsible for web data communication, a port responsible for FTP data communication, or a port responsible for mail data communication. In addition, the network communication port can also be a communication interface or a communication chip of an entity. For example, it may be a wireless mobile network communication chip, such as GSM, CDMA, etc.; it can also be a Wifi chip; it may also be a bluetooth chip.

In this embodiment, the processor 702 may be implemented in any suitable manner. For example, the processor may take the form of, for example, a microprocessor or processor and a computer-readable medium that stores computer-readable program code (e.g., software or firmware) executable by the (micro) processor, logic gates, switches, an Application Specific Integrated Circuit (ASIC), a programmable logic controller, an embedded microcontroller, and so forth. The description is not intended to be limiting.

In this embodiment, the memory 703 may include multiple layers, and in a digital system, the memory may be any memory as long as it can store binary data; in an integrated circuit, a circuit without a physical form and with a storage function is also called a memory, such as a RAM, a FIFO and the like; in the system, the storage device in physical form is also called a memory, such as a memory bank, a TF card and the like.

Embodiments of the present specification further provide a server, including a processor and a memory for storing processor-executable instructions, where the processor, when implemented, may perform the following steps according to the instructions: responding to a first training request about a preset target user prediction model, and performing sample data fusion in cooperation with a second server according to a protocol rule based on horizontal federal learning so as to determine a first positive sample user, a first negative sample user and a first pseudo negative sample user in an out-of-customer group sample user, and a second positive sample user and a second mixed negative sample user in a non-out-of-customer group sample user; the first sample data held by the first server at least comprises non-business class characteristic data of full-scale sample users, wherein the full-scale sample users comprise customer group sample users and non-customer group sample users; the second server is a server deployed at one side of a second data party, and second sample data held by the second server at least comprises service class characteristic data of a full positive sample user; acquiring non-service characteristic data of a first positive sample user and non-service characteristic data of a first negative sample user from the first sample data to serve as first grouping sample data; acquiring non-service characteristic data of a second positive sample user and non-service characteristic data of a second mixed negative sample user from the first sample data to serve as second grouping sample data; training by using the first grouping sample data to obtain a first predictor model; training to obtain a second predictor model by utilizing the second grouping of sample data; and combining the first prediction submodel and the second prediction submodel to obtain a preset target user prediction model.

Embodiments of the present specification further provide a server, including a processor and a memory for storing processor-executable instructions, where the processor, when implemented, may perform the following steps according to the instructions: responding to a second training request about a preset user portrait prediction model, and performing sample data fusion by cooperating with a second server according to a protocol rule based on longitudinal federal learning to obtain third training data carrying a preset portrait label; wherein the third training data comprises non-traffic class feature data of the positive sample user; the preset portrait label is determined according to the business class characteristic data of the user who just samples; the first sample data held by the first server at least comprises non-business class characteristic data of a full amount of sample users; the second server is a server deployed at one side of a second data party, and second sample data held by the second server at least comprises service class characteristic data of a full positive sample user; and training to obtain a preset user portrait prediction model by using the third training data carrying the preset portrait label.

Embodiments of the present specification further provide a server, including a processor and a memory for storing processor-executable instructions, where the processor, when implemented, may perform the following steps according to the instructions: acquiring identification information of a user object and characteristic data of the user object; wherein the feature data at least comprises non-traffic class feature data; determining the guest group type of the user object according to the identification information of the user object; calling a prediction sub-model matched with the guest group type of the user object in a preset target user prediction model to process the characteristic data of the user object so as to determine whether the user object is a target user; the preset target user prediction model is established in advance according to first sample data in a first data party and second sample data in a second data party on the basis of transverse federal learning; and under the condition that the user object is determined to be the target user, pushing corresponding target service data to the user object.

An embodiment of the present specification further provides a computer storage medium based on the above service data pushing method, where the computer storage medium stores computer program instructions, and when the computer program instructions are executed, the computer storage medium implements: acquiring identification information of a user object and characteristic data of the user object; wherein the feature data at least comprises non-traffic class feature data; determining the guest group type of the user object according to the identification information of the user object; calling a prediction sub-model matched with the guest group type of the user object in a preset target user prediction model to process the characteristic data of the user object so as to determine whether the user object is a target user; the preset target user prediction model is established in advance according to first sample data in a first data party and second sample data in a second data party on the basis of transverse federal learning; under the condition that the user object is determined to be a target user, calling a preset user portrait prediction model to process feature data of the user object so as to obtain a portrait label of the user object; the preset user portrait prediction model is established in advance according to first sample data in a first data party and second sample data in a second data party on the basis of longitudinal federal learning; generating a target pushing rule matched with the user object according to the portrait label of the user object; and pushing corresponding target business data to the user object according to the target pushing rule.

The present specification further provides a computer storage medium based on the training method of the model, where the computer storage medium stores computer program instructions, and when the computer program instructions are executed, the computer program instructions implement: responding to a first training request about a preset target user prediction model, and performing sample data fusion in cooperation with a second server according to a protocol rule based on horizontal federal learning so as to determine a first positive sample user, a first negative sample user and a first pseudo negative sample user in an out-of-customer group sample user, and a second positive sample user and a second mixed negative sample user in a non-out-of-customer group sample user; the first sample data held by the first server at least comprises non-business class characteristic data of full-scale sample users, wherein the full-scale sample users comprise customer group sample users and non-customer group sample users; the second server is a server deployed at one side of a second data party, and second sample data held by the second server at least comprises service class characteristic data of a full positive sample user; acquiring non-service characteristic data of a first positive sample user and non-service characteristic data of a first negative sample user from the first sample data to serve as first grouping sample data; acquiring non-service characteristic data of a second positive sample user and non-service characteristic data of a second mixed negative sample user from the first sample data to serve as second grouping sample data; training by using the first grouping sample data to obtain a first predictor model; training to obtain a second predictor model by utilizing the second grouping of sample data; and combining the first prediction submodel and the second prediction submodel to obtain a preset target user prediction model.

In this embodiment, the storage medium includes, but is not limited to, a Random Access Memory (RAM), a Read-Only Memory (ROM), a Cache (Cache), a Hard Disk Drive (HDD), or a Memory Card (Memory Card). The memory may be used to store computer program instructions. The network communication unit may be an interface for performing network connection communication, which is set in accordance with a standard prescribed by a communication protocol.

In this embodiment, the functions and effects specifically realized by the program instructions stored in the computer storage medium can be explained by comparing with other embodiments, and are not described herein again.

Referring to fig. 8, in a software level, an embodiment of the present specification further provides a service data pushing device, which may specifically include the following structural modules.

The obtaining module 801 may be specifically configured to obtain identification information of a user object and feature data of the user object; wherein the feature data at least comprises non-traffic class feature data;

the determining module 802 may be specifically configured to determine, according to the identification information of the user object, a guest group type of the user object;

the first invoking module 803 may be specifically configured to invoke a prediction sub-model, which is matched with the guest group type of the user object, in a preset target user prediction model to process feature data of the user object, so as to determine whether the user object is a target user; the preset target user prediction model is established in advance according to first sample data in a first data party and second sample data in a second data party on the basis of transverse federal learning;

a second calling module 804, configured to specifically call a preset user portrait prediction model to process feature data of the user object to obtain a portrait label of the user object when the user object is determined to be a target user; the preset user portrait prediction model is established in advance according to first sample data in a first data party and second sample data in a second data party on the basis of longitudinal federal learning;

a generating module 805, specifically configured to generate a target push rule matching the user object according to the portrait tag of the user object;

the pushing module 806 may be specifically configured to push corresponding target service data to the user object according to the target pushing rule.

It should be noted that, the units, devices, modules, etc. illustrated in the above embodiments may be implemented by a computer chip or an entity, or implemented by a product with certain functions. For convenience of description, the above devices are described as being divided into various modules by functions, and are described separately. It is to be understood that, in implementing the present specification, functions of each module may be implemented in one or more pieces of software and/or hardware, or a module that implements the same function may be implemented by a combination of a plurality of sub-modules or sub-units, or the like. The above-described embodiments of the apparatus are merely illustrative, and for example, the division of the units is only one logical division, and other divisions may be realized in practice, for example, a plurality of units or components may be combined or integrated into another system, or some features may be omitted, or not executed. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection through some interfaces, devices or units, and may be in an electrical, mechanical or other form.

As can be seen from the above, the service data pushing device provided in this specification can generate a target pushing rule matching with a user object in a targeted manner, and accurately push appropriate target service data to the user object according to the target pushing rule, so as to obtain a better pushing effect, improve a pushing success rate, and implement accurate pushing of service data of the user object.

Referring to fig. 9, on a software level, the embodiment of the present specification further provides a model training apparatus, which may specifically include the following structural modules.

The fusion module 901 may be specifically configured to respond to a first training request related to a preset target user prediction model, and perform sample data fusion in cooperation with a second server according to a protocol rule based on horizontal federal learning, so as to determine a first positive sample user, a first negative sample user, and a first pseudo negative sample user in an output customer group of sample users, and a second positive sample user and a second mixed negative sample user in a non-output customer group of sample users; the first sample data held by the first server at least comprises non-business class characteristic data of full-scale sample users, wherein the full-scale sample users comprise customer group sample users and non-customer group sample users; the second server is a server deployed at one side of a second data party, and second sample data held by the second server at least comprises service class characteristic data of a full positive sample user;

an obtaining module 902, which may be specifically configured to obtain, from the first sample data, non-service class feature data of a first positive sample user and non-service class feature data of a first negative sample user as first packet sample data; acquiring non-service characteristic data of a second positive sample user and non-service characteristic data of a second mixed negative sample user from the first sample data to serve as second grouping sample data;

a training module 903, which may be specifically configured to train to obtain a first predictor model by using the first packet of sample data; training to obtain a second predictor model by utilizing the second grouping of sample data;

the combining module 904 may be specifically configured to combine the first predictor model and the second predictor model to obtain a preset target user prediction model.

As can be seen from the above, the model training device provided based on the embodiments of the present specification can comprehensively utilize the first sample data in the first data party and the second sample data in the second data party, reduce the influence of the distribution offset introduced by the pseudo negative sample on the model precision in the model training process, and train to obtain the preset target user prediction model with higher precision and better effect.

In a specific scenario example, the above method provided in this specification may be applied to assist an insurance subsidiary under a bank in accurate marketing of insurance business. For a specific implementation process, the following can be referred to.

In the scenario example, the insurance precision marketing work is a kind of machine learning application scenario with important business complexity of the bank, and is used for assisting insurance agents and customer managers of the bank to accurately mine insurance potential purchasing customers (e.g., target users) and improve the marketing efficiency of insurance products.

At present, insurance products under a 'bank + insurance' (hereinafter referred to as 'bank insurance') mode have the following problems: 1. and (5) isolating data. An insurance subsidiary company is usually set in a large bank, but the insurance marketing data of the large bank is limited by a supervision policy, and the data of the insurance subsidiary company and the bank cannot be directly circulated, so that sample data for insurance marketing modeling is separated, and only modeling can be performed based on each data. 2. Existing models lack business interpretability. Insurance products have a complex business and great marketing difficulty. Insurance agents are required to explain insurance terms in detail, online purchasing proportions are small, and recommendation models lacking business interpretability are of limited help to insurance agents. 3. Negative examples are not defined exactly, and pure negative examples are few. In a marketing modeling scenario, generally, the number of pure negative examples (examples of marketing failures, for example, pure negative example users) is limited, negative examples are defined as unsound customer examples in most cases, but such examples include pseudo negative examples (for example, pseudo negative example users) which are not bought because they are not marketed, and the examples are shifted, so that the effect of modeling based on such negative examples is poor.

In this example, marketing modeling is considered to be possible by introducing both horizontal and vertical federal learning.

The federal learning refers to that each participant can perform joint modeling by means of data of other parties in the process of performing machine learning. All parties do not need to share original data, namely under the condition that the data are not local, data combined training is carried out, and a shared machine learning model is established. The federal learning generally includes three types, in the scenario example, two types of federal learning technologies, namely horizontal federal and vertical federal are mainly used, and a modeling scheme of double-layer federal learning is provided.

The principle of federal learning is to protect privacy and realize safe multi-party data combined model training by encrypting the intermediate result of an interactive mode model instead of an original data mode. Wherein, the encryption interaction: all information exchange in the federal learning process adopts an encryption mode, and only own data can be seen by participants. Data desensitization: the method is characterized in that original data are exchanged in the federal learning process, and lossless model training effect can be obtained under the condition that original information does not need to be exchanged through irreversible feature extraction and conversion (namely, the federal learning effect is consistent with the model training effect under the condition that data of two parties are completely visible).

The horizontal federal learning specifically means that under the condition that sample features of a data set are overlapped more and sample IDs are overlapped less, data are aligned according to feature dimensions by adopting the horizontal federal, and the data with the same sample features and not identical users are taken out for training.

The longitudinal federal learning specifically means that under the condition that sample IDs of a data set are overlapped more and sample features are overlapped less, the longitudinal federal is adopted to align the data with the sample IDs, and the part of the data with the same data IDs and not identical data features is taken out for training.

In the scene example, the problem of the insurance precision marketing modeling in the bank + insurance mode is solved by adopting a double-layer federal learning technology based on a Lookalike method. Enabling secure converged modeling and model interpretability of data by a bank (e.g., a first data party) and an insurance carrier (e.g., a second data party) of the bank for which data isolation exists due to regulatory policy restrictions. The client mining accuracy of the insurance marketing model is improved.

In specific implementation, firstly, the data of banks and insurance companies (for example, first sample data of a first data party and second sample data of a second data party) are subjected to safety fusion through a horizontal federal learning technology, and an insurance client mining model is trained to obtain an insurance hidden passenger list which is a first-layer federal model (for example, a preset target user prediction model). And training a plurality of insurance portrait models based on an insurance client marketing list output by the model to generate insurance portrait labels, and appointing a marketing strategy according to the label tendency, wherein the marketing strategy is a second federal model (for example, a preset user portrait prediction model).

When modeling is specifically trained, reference can be made to fig. 10. The process of the silver insurance accurate marketing system established based on double-layer federal learning comprises the following steps: 1, respectively carrying out data preparation work by a bank and an insurance subsidiary company; 2, implementing safety fusion on sample data of banks and insurance companies based on a Lookalike method and a horizontal federal learning technology and training a first-layer federal model, namely an insurance customer mining model; 3, generating an insurance purchase intention marketing list; 4, training a second layer of federal model based on a longitudinal federal training technology, wherein the second layer of federal model comprises a plurality of 'insurance portrait models'; and 5, generating a marketing strategy (for example, a target pushing rule) of the insurance agent, and then carrying out insurance marketing according to the marketing strategy (for example, pushing target business data according to the target pushing rule).

Specifically, step 1: the bank and the insurance subsidiary perform data preparation work separately.

Step 2: and implementing safety fusion on sample data of a head office and an insurance subsidiary company and training an insurance client mining model based on a horizontal federal learning technology adopting a Lookalike method. The model is divided into: bank administrator customer model a1 (e.g., a first predictor model), and bank non-administrator model a2 (e.g., a second predictor model).

In actual application, the insurance agent corresponding to the customer group can use the A1 model to predict the customer group. When marketing is carried out on non-administrative customer groups, prediction can be carried out by using an A2 model, and insurance marketing is carried out in a mode with lower marketing cost, such as WeChat mass-sending and intelligent voice robot telemarketing.

The management customer group of the bank has the service of the special person of the customer manager, the asset star level and the credit score are high, the marketing record is complete (for example, the historical pushing record) and the bank characteristic (for example, the non-business characteristic data) is comprehensive, but the number of the management customer group occupying the whole bank is small. The non-administrative customer group has no special service of a customer manager and no marketing record, the proportion of false negative samples in the samples is large, and the bank characteristics are lost more. Thus, the administrative and non-administrative customer groups can be distinguished for separate modeling. Specifically, models a1 and a2 may be established for customer groups and non-customer groups, respectively.

In building the a1 model, when sample selection is performed, positive samples can be made to purchase insurance customer base lists for the bank insurance + self-selling channel. Negative examples are random samples of the list of unsophisticated clients (including pure negative examples and pseudo-negative examples). The scene is modeled based on a horizontal federal learning framework, namely, the client information appears in two parties at the same time, and the characteristics are the same. In the scene, the client of the subsidiary company is the bank client, so the sample is selected as the client shared by the two organizations, and the selected sample is the characteristic of the bank side.

In the feature definition, the features on the bank side are used. Bank features are typically high-dimensional, comprehensive in asset and consumption features, but lack policy features (e.g., business-like feature data), such as important insurance portraits like premium, guarantee period, applicant, insured life, and gift identification. The banking features may include: assets, liabilities, deposit balances, insurance balances, academic calendars, educational backgrounds, professional information, bank transaction information, monthly average consumed amounts, large customer identification, customer star, number of marketing contacts, number of responses, insurance purchase risk flags, online shopping frequency, development wage identification, and the like. The insurance company has fewer characteristics compared with the bank, especially the characteristics of assets, transactions and wages are lost. The insurance features include: premium, term of coverage, history of insurances, applicant, insured person, gift certificate, etc.

When label definition is performed, a positive sample can be defined as an insurance client; negative examples are defined as no risk-purchasing customers. And because the total customer base number of the bank is very huge, the pseudo negative samples without insurance purchase are very huge without marketing, and it is unlikely that all hundreds of millions of samples are trained under the federal learning framework, so that all the negative samples are selected, and the pseudo negative samples are randomly sampled.

The A1 model is retrained to output an estimate to obtain a list of premium customers.

When the a2 model is constructed, the process is similar to that of the a1 model except that the corresponding customer group selected by the sample is a non-customer group, and the details are not repeated herein.

And step 3: an insurance agent marketing list is generated. And generating an insurance hidden passenger list (such as a target user list) for the insurance agent to use based on the model of the second step.

And 4, step 4: and training a plurality of insurance portrait models based on a longitudinal federal learning technology according to the client list generated in the third step. The insurance image label may include a payment deadline tendency, a premium tendency, and a premium tendency. The three tags may be determined according to characteristics of the possession of the insurance carrier, the bank side not having such characteristics because of regulatory policy restrictions. This is the second layer security image model B.

The definition of the longitudinal federation is that users of the mechanisms of the two parties are consistent, the characteristics are separated, and the modeling uses the common characteristics of the two parties. Specific sample selection and feature selection, as explained in detail below. The B model mines the client list to be marketed estimated according to the first layer insurance client mining models A1 and A2.

In this scenario example, three insurance portrait models (e.g., portrait label predictor models) are specifically trained, as distinguished from the first-level model, with the B model belonging to a regression model. Wherein the label definitions respectively include: payment period tendency B1, premium tendency B2 and premium tendency B3. The pre-estimated value of the model obtained by training is the specific payment age, the value of the premium and the like.

Taking the premium tendency model B2 as an example, the specific modeling scheme includes the following key contents: when the sample is selected, all positive samples are obtained from the insurance subsidiary company. Such a client may have both mechanism features, with the tag defined on the subsidiary side. In the feature definition, bank-side features are used. Because the model to be predicted contains non-subsidiary head office unskilled users, such customers do not have subsidiary features. So model B only selects bank features. When defining the label, the corresponding insurance amount of the sample is only defined at the insurance subsidiary company side, and the bank side does not have the label.

In specific use, the list of clients to be marketed output by the first-layer model can be used as a pre-estimation object. And processing the predicted objects by using a second-layer model to obtain a predicted value as a client's warranty suggestion output by the first-time federal model.

And 5: and generating a marketing strategy of the insurance agent. The insurance application tendency of each client is generated according to the models B1, B2 and B3, a refined marketing strategy can be formulated according to the insurance application tendency of the clients such as payment years, premium values and the like, and the professional degree of the insurance agent on the marketing of the clients is improved. For example, when an insurance agent markets a certain client, the client can be more accurately advised to buy specific premium value and premium value, and the interpretability of insurance products is effectively improved. If only the first layer model is available, the client receives the telephone of the insurance agent, only one suggestion is to buy insurance, but the client will immediately consult the categories of the insurance amount and the categories of the premium, in which case the second layer model can immediately assist the insurance agent to answer the client's question and even give the suggestion voluntarily.

In the above flow process, referring to fig. 11, after the data preparation work is completed, the data distribution owned by the bank and its insurance companies is relatively complex and chaotic. The bank customers are divided into an administrative customer group and a non-administrative customer group, the administrative customer group is provided with a corresponding bank customer manager, the asset star level is high, marketing records exist, a purely negative sample (marketing failure sample) can be provided, and the modeling effect based on the model is good. The assets of the non-managed customer group are low in rating, no marketing record is recorded, only purchasing records are recorded, and the proportion of false negative samples (no marketing and no purchasing) in the negative samples corresponding to the customer group is high. The modeling effect based on the model is poor.

Specifically, the sample distribution is characterized as follows: the insurance purchase channel is divided into a self-selling channel (e.g., self-owned channel) and a banking channel (e.g., shared channel). The insurance sub-company has a full insurance purchase positive sample (e.g., a full positive sample user), including self-marketing positive samples and silver insurance samples. The customers of the insurance company are generally the customers of the bank, but the bank is restricted by the regulatory policy to have no self-selling channel of the insurance company to be a positive sample. The part of the customers have negative labels at the bank side and positive label conflicts at the subsidiary side.

Specifically, referring to fig. 11, a guest group (i) represents a guest who purchases insurance of a subsidiary company in a head office, and the guest group (i) corresponds to a guest group (ii) of the subsidiary insurance company. Both the guest group i and the guest group ii are positive samples. The client group c represents the insurance purchasing client of the insurance company from the channel-selling company, and is a positive sample on the insurance company side. The guest group (r) corresponds to the guest group (r) at the head office, and the head office does not know the purchase condition of the guest group (r), and therefore, the guest group (r) is regarded as a negative sample at the head office side. Therefore, in the data preparation stage, both sides need to perform data security fusion, and then training modeling can be started.

Through the scene example, the method provided by the specification is applied, and the insurance precision marketing modeling under the bank + insurance mode based on the double-layer federal learning technology is realized. The bank and the insurance company which cannot be realized due to the limitation of the data supervision authority can break the data isolation to carry out the combined modeling. The insurance model building time insurance sample size and the characteristic dimension are expanded, the insurance marketing accuracy is effectively improved, the marketing cost of insurance agents is saved, the marketing success rate is improved, and the marketing strategy is optimized.

Although the present specification provides method steps as described in the examples or flowcharts, additional or fewer steps may be included based on conventional or non-inventive means. The order of steps recited in the embodiments is merely one manner of performing the steps in a multitude of orders and does not represent the only order of execution. When an apparatus or client product in practice executes, it may execute sequentially or in parallel (e.g., in a parallel processor or multithreaded processing environment, or even in a distributed data processing environment) according to the embodiments or methods shown in the figures. The terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, the presence of additional identical or equivalent elements in a process, method, article, or apparatus that comprises the recited elements is not excluded. The terms first, second, etc. are used to denote names, but not any particular order.

Those skilled in the art will also appreciate that, in addition to implementing the controller as pure computer readable program code, the same functionality can be implemented by logically programming method steps such that the controller is in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, embedded microcontrollers and the like. Such a controller may therefore be considered as a hardware component, and the means included therein for performing the various functions may also be considered as a structure within the hardware component. Or even means for performing the functions may be regarded as being both a software module for performing the method and a structure within a hardware component.

This description may be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, classes, etc. that perform particular tasks or implement particular abstract data types. The specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media including memory storage devices.

From the above description of the embodiments, it is clear to those skilled in the art that the present specification can be implemented by software plus necessary general hardware platform. With this understanding, the technical solutions in the present specification may be essentially embodied in the form of a software product, which may be stored in a storage medium, such as a ROM/RAM, a magnetic disk, an optical disk, etc., and includes several instructions for enabling a computer device (which may be a personal computer, a mobile terminal, a server, or a network device, etc.) to execute the methods described in the embodiments or some parts of the embodiments in the present specification.

The embodiments in the present specification are described in a progressive manner, and the same or similar parts among the embodiments are referred to each other, and each embodiment focuses on the differences from the other embodiments. The description is operational with numerous general purpose or special purpose computing system environments or configurations. For example: personal computers, server computers, hand-held or portable devices, tablet-type devices, multiprocessor systems, microprocessor-based systems, set top boxes, programmable electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments that include any of the above systems or devices, and the like.

While the specification has been described with examples, those skilled in the art will appreciate that there are numerous variations and permutations of the specification that do not depart from the spirit of the specification, and it is intended that the appended claims include such variations and modifications that do not depart from the spirit of the specification.

Claims

1. A method for pushing service data is characterized by comprising the following steps:

2. The method of claim 1, wherein the guest group type comprises: an administrative group and a non-administrative group.

3. The method of claim 2, wherein the pre-defined target user prediction model comprises a first predictor model and a second predictor model; the first predictor model is matched with an administrative customer group, and the second predictor model is matched with a non-administrative customer group.

4. The method of claim 3, wherein the predetermined target user prediction model is established as follows:

5. The method of claim 4, wherein training using the first packet of sample data to obtain a first predictor model comprises:

extracting a plurality of data from the non-business characteristic data of a first positive sample user in the first grouping sample data, and marking the data as first positive sample training data; extracting a plurality of data from the non-business class characteristic data of a first negative sample user in the first grouping sample data, and marking the data as first negative sample training data;

6. The method of claim 4, wherein using the second packet of sample data to train a second predictor model comprises:

extracting a plurality of data from the business class characteristic data of a second positive sample user in the second grouping sample data, and marking the data as second positive sample training data; randomly extracting a plurality of data from the business class characteristic data of a second mixed negative sample user in the second grouped sample data, and marking the data as second negative sample training data;

7. The method of claim 1, wherein the pre-defined user portrait prediction model is built as follows:

8. The method of claim 7, wherein the pre-defined user portrait prediction model comprises a plurality of portrait prediction submodels, wherein different portrait prediction submodels are used to predict different portrait tags.

9. The method according to claim 8, wherein the target business data comprises insurance business products and/or insurance business services suitable for the user object; correspondingly, the target pushing rule comprises an insurance business marketing scheme.

10. The method of claim 9, wherein the non-traffic class characteristic data comprises at least one of: monthly income, monthly consumption amount, deposit balance; the service class characteristic data comprises at least one of: gift identification, premium, and premium.

11. The method of claim 10, wherein the portrait label includes at least one of: payment period tendency label, premium tendency label, and premium tendency label.

12. A method for training a model, applied to a first server deployed on a first data side, includes:

13. A method for training a model, applied to a first server deployed on a first data side, includes:

14. A method for pushing service data is characterized by comprising the following steps:

and under the condition that the user object is determined to be the target user, pushing corresponding target service data to the user object.

15. A device for pushing service data, comprising:

16. A server comprising a processor and a memory for storing processor-executable instructions which, when executed by the processor, implement the steps of the method of any one of claims 1 to 11, 12, 13 or 14.