CN108205766A

CN108205766A - Information-pushing method, apparatus and system

Info

Publication number: CN108205766A
Application number: CN201611179080.7A
Authority: CN
Inventors: 谭译泽; 吴保华
Original assignee: Alibaba Group Holding Ltd
Current assignee: Alibaba Group Holding Ltd
Priority date: 2016-12-19
Filing date: 2016-12-19
Publication date: 2018-06-26

Abstract

The present invention provides a kind of information-pushing methods, apparatus and system, by obtaining positive sample crowd and negative sample crowd with the relevant essential information of target object according to input by user, learning training is carried out to positive sample crowd and negative sample crowd, to obtain the first prediction model, and positive sample crowd and negative sample crowd are modified using the transfer learning model of structure, learning training is carried out to revised positive sample crowd and negative sample crowd, to obtain the second prediction model, and potential crowd is predicted to obtain target group respectively using the first prediction model and the second prediction model, target object is pushed to target group.In the present embodiment, by having the machine learning module of supervision and transfer learning model, target group is got from potential crowd so that prediction result is more accurate, and pushes targeted advertisements to it.Due to prediction result, precisely then correspondingly target group is then the best crowd of advertisement delivery effect, and then advertiser can be helped to improve the effect that advertisement is launched.

Description

Information-pushing method, apparatus and system

Technical field

The invention belongs to data processing field more particularly to a kind of information-pushing method, apparatus and systems.

Background technology

Advertiser on the basis of mass data, can pass through engineering to the dispensing effect growing interest of advertisement at present It practises according to default rule, influence of the preset label to advertisement can be excavated, advertiser can be according to label to advertisement Part labels are chosen in influence, and potential crowd is obtained by the label chosen, and will obtain potential crowd and launch advertisement as advertiser Dispensing object, then send corresponding advertising information to launching object.

And since the truthful data in practical application can change, for example, buying the user of a certain commodity of a certain brand Group can change with the influence power and economic conditions of brand, and default rule can be difficult to capture this variation, it is difficult to Dispensing effect is made to maintain a preferable state, launching effect can decay quickly, and then lead to the dispensing effect of advertisement most Difference.

Invention content

The present invention provides a kind of information-pushing method, apparatus and system, and potential crowd's work is obtained for solving existing label For launch object, there are advertisement delivery effect it is poor the problem of.

To achieve these goals, the present invention provides a kind of information transmission system, including：Front end and data processing are put down Platform；

The front end is sent to the data processing for receiving the input by user and relevant essential information of target object Platform；

The data processing platform (DPP), for obtaining positive sample crowd and negative sample crowd according to the essential information, to institute It states positive and negative sample crowd and the negative sample crowd carries out learning training, to obtain the first prediction model, and utilize the migration of structure Learning model is modified the positive sample crowd and the negative sample crowd, to the revised positive sample crowd and institute It states negative sample crowd and carries out learning training, to obtain the second prediction model and utilize first prediction model and described the Two prediction models are respectively predicted to obtain target group the potential crowd, and the target is pushed to the target group Object.

To achieve these goals, the present invention provides a kind of interactive device, including：

For receiving the input by user and relevant essential information of target object, the essential information is sent to for front end Data processing platform (DPP) and the target group that the data processing platform (DPP) is obtained according to the essential information is received, and to described Target group pushes the target object.

To achieve these goals, the present invention provides a kind of information push-delivery apparatus, including：

Data processing platform (DPP), for obtaining positive sample crowd and negative sample crowd according to essential information relevant with target, Learning training is carried out to the positive and negative sample crowd and the negative sample crowd, to obtain the first prediction model, and utilizes structure Transfer learning model is modified the positive sample crowd and the negative sample crowd, to the revised positive sample crowd Learning training is carried out with the negative sample crowd, to obtain the second prediction model and utilize first prediction model and institute It states the second prediction model respectively to predict to obtain target group the potential crowd, to described in target group push Target object.

To achieve these goals, the present invention provides a kind of model prediction method, including：

Feature extraction is carried out to positive sample crowd and negative sample crowd, to obtain the fisrt feature information of sample；

The type parameter of the fisrt feature information and each sample is input in preset machine learning model and is carried out Training, to obtain the first prediction model；

The revised positive sample crowd and the negative sample crowd characteristic are extracted, to obtain the second feature of sample Information；

The type parameter of the second feature information and each sample is inputted in the machine learning model and is trained, To obtain the second prediction model.

To achieve these goals, the present invention provides a kind of model training apparatus, including：

Characteristic extracting module, for carrying out feature extraction to positive sample crowd and negative sample crowd, to obtain the of sample One characteristic information and the revised positive sample crowd and the negative sample crowd characteristic are extracted, to obtain sample Second feature information；

Training module, for the type parameter of the fisrt feature information and each sample to be input to preset engineering It practises and being trained in model, to obtain the first prediction model and join the second feature information and the type of each sample Number is inputted in the machine learning model and is trained, to obtain the second prediction model.

To achieve these goals, the present invention provides a kind of target group's acquisition methods, including：

According to input by user positive sample crowd and negative sample crowd are obtained with the relevant essential information of target object；

Learning training is carried out to the positive sample crowd and the negative sample crowd, to obtain the first prediction model；

The positive sample crowd and the negative sample crowd are modified using the transfer learning model of structure；

Learning training is carried out to the revised positive sample crowd and the negative sample crowd, to obtain the second prediction mould Type；

The potential crowd is predicted respectively using first prediction model and second prediction model to obtain To target group.

To achieve these goals, the present invention provides a kind of target group's acquisition device, including：

Acquisition module, for obtaining positive sample crowd according to the relevant essential information of input by user and target object and bearing Sample population；

Training module, for carrying out learning training to the positive sample crowd and the negative sample crowd, to obtain first Prediction model and learning training is carried out to the revised positive sample crowd and the negative sample crowd, to obtain second Prediction model；

Correcting module, for using the transfer learning model of structure to the positive sample crowd and the negative sample crowd into Row is corrected；

Prediction module respectively carries out the potential crowd using first prediction model and second prediction model It predicts to obtain target group.

To achieve these goals, the present invention provides a kind of information-pushing method, including：

Potential crowd is predicted based on first prediction model to obtain target group；

The target object is pushed to the target group.

Sample acquisition module, for obtaining positive sample crowd with the relevant essential information of target object according to input by user With negative sample crowd；

Training module, for carrying out learning training to the positive sample crowd and the negative sample crowd, to obtain first Prediction model；

Prediction module predicts potential crowd to obtain target group for being based on first prediction model；

Pushing module, for pushing the target object to the target group.

Obtain the input by user and relevant essential information of target object；

Positive sample crowd and negative sample crowd are obtained according to the essential information；

Learning training is carried out to the positive and negative sample crowd and the negative sample crowd, to obtain the first prediction model；

The potential crowd is predicted respectively using first prediction model and second prediction model to obtain To target group；

The target object is pushed to the target group.

First acquisition module, for obtaining the input by user and relevant essential information of target object；

Second acquisition module, for obtaining positive sample crowd and negative sample crowd according to the essential information；

Training module, for carrying out learning training to the positive and negative sample crowd and the negative sample crowd, to obtain first Prediction model and learning training is carried out to the revised positive sample crowd and the negative sample crowd, to obtain second Prediction model；

Prediction module, for utilizing first prediction model and second prediction model respectively to the potential crowd It is predicted to obtain target group；

Pushing module, for pushing the target object to the target group.

Information-pushing method provided by the invention, apparatus and system, by according to input by user related to target object Essential information obtain positive sample crowd and negative sample crowd, learning training is carried out to positive sample crowd and negative sample crowd, with The first prediction model is obtained, and positive sample crowd and negative sample crowd are modified using the transfer learning model of structure, it is right Revised positive sample crowd and negative sample crowd carry out learning training, to obtain the second prediction model and utilize first in advance It surveys model and the second prediction model is respectively predicted to obtain target group potential crowd, target pair is pushed to target group As.In the present embodiment, by having the machine learning module of supervision and transfer learning model, target person is got from potential crowd Group so that prediction result is more accurate, and pushes targeted advertisements to it.Due to prediction result, precisely then correspondingly target group is then It is the best crowd of advertisement delivery effect, and then advertiser can be helped to improve the effect that advertisement is launched.

Description of the drawings

Fig. 1 is the structure diagram of a kind of information transmission system that the embodiment of the present invention one provides；

Fig. 2 is a kind of structure diagram of information transmission system provided by Embodiment 2 of the present invention；

Fig. 3 is the structure diagram of a kind of interactive device that the embodiment of the present invention three provides；

Fig. 4 is the structure diagram of a kind of information push-delivery apparatus that the embodiment of the present invention four provides；

Fig. 5 is the flow diagram of a kind of model training method that the embodiment of the present invention five provides；

Fig. 6 is the structure diagram of a kind of model training apparatus that the embodiment of the present invention six provides；

Fig. 7 is the flow diagram of a kind of target group's acquisition methods that the embodiment of the present invention seven provides；

Fig. 8 is the structure diagram of a kind of target group's acquisition device that the embodiment of the present invention eight provides；

Fig. 9 is the flow diagram of a kind of information-pushing method that the embodiment of the present invention nine provides；

Figure 10 is the structure diagram of a kind of information push-delivery apparatus that the embodiment of the present invention ten provides；

Figure 11 is the flow diagram of a kind of information-pushing method that the embodiment of the present invention 11 provides；

Figure 12 is the flow diagram of a kind of information-pushing method that the embodiment of the present invention 12 provides；

Figure 13 is the application schematic diagram of a kind of information-pushing method that the embodiment of the present invention 12 provides；

Figure 14 is the structure diagram of a kind of information push-delivery apparatus that the embodiment of the present invention 13 provides；

Figure 15 is the structure diagram of information transmission system that the embodiment of the present invention 14 provides；

Figure 16 is one of application schematic diagram of information transmission system that the embodiment of the present invention 14 provides；

Figure 17 is the two of the application schematic diagram of information transmission system that the embodiment of the present invention 14 provides.

Specific embodiment

Below in conjunction with the accompanying drawings to information-pushing method provided in an embodiment of the present invention, apparatus and system, model prediction method And device, target group's acquisition methods and device, interactive device are described in detail.

Embodiment one

As shown in Figure 1, it is the structure diagram of a kind of information transmission system that the embodiment of the present invention one provides.The information Supplying system includes：Front end 1 and data processing platform (DPP) 2.

Wherein, front end 1 input by user be sent to data processing with the relevant essential information of target object and put down for receiving Platform 2.

In the present embodiment, target object can be advertisement to be pushed, or text, image, music, regards news Frequently, good friend even in instant messaging tools etc..In the present embodiment, by taking the advertisement that target object is to be pushed as an example, then input User with the relevant essential information of target object is advertiser, and receives the user i.e. target of the promotion message of target object The user of object is known as client in the present embodiment.Attempt to push a target pair to certain customers i.e. user in advertiser As when, it is necessary first to front end 1 input with the relevant essential information of the target object.For example, it can be set in front end 1 one wide Accuse main login interface and the input interface of essential information.After advertiser is logged in by login interface, essential information can be entered Input interface, on the input interface input target object essential information.

In the present embodiment, it can include in essential information：Uniform resource locator (the Uniform of target object Resource Locator, abbreviation URL), the user of target object can correspond to the classification found where commodity by the URL Identify (ID), brand ID, commodity ID and shop ID).Further, essential information can also include：The popularization of target object Platform, Extension Software Platform can include mobile terminal and computer end, and further, essential information can also include：Target object pushes away Wide keyword, for example, keyword is the contents such as discount, promotion.Further, essential information can also include：Target object Unit price, that is, commodity price.

Data processing platform (DPP) 2, for obtaining positive sample crowd and negative sample crowd according to essential information, to positive and negative sample crowd Learning training is carried out with negative sample crowd, to obtain the first prediction model, and using the transfer learning model built to positive sample Crowd and negative sample crowd are modified, and learning training are carried out to revised positive sample crowd and negative sample crowd, to obtain Second prediction model and potential crowd is predicted to obtain mesh respectively using the first prediction model and the second prediction model Mark crowd pushes target object to target group.

Data processing platform (DPP) 2, can basis after the user i.e. essential information of the target object of advertisement primary input is received These essential informations get positive sample crowd.Specifically, classification ID of the data processing platform (DPP) 2 in essential information, brand ID, commodity ID and corresponding Extension Software Platform, keyword such as commodity price of discount, target object etc., found in a period of time There is the client for buying, collecting or liking this class object, using these clients as potential positive sample crowd.

Further, data processing platform (DPP) 2 can also expand potential positive sample crowd according to certain rule, with Obtain positive sample crowd.For example, the loyal client in the shop can be found according to shop ID, by these loyal clients to latent Expanded in positive sample, and then form positive sample crowd.For another example the brand ID in essential information, obtains and the product There are associated commodity by board ID, then obtain to the interested client of these commodity, using these clients to potential positive sample people Group is expanded.For another example the classification ID in essential information, obtains and there are certain associated commodity with the classification, then To the interested client of these commodity, potential sample population is expanded using these clients.

Further, advertiser can also increase the special client for being biased to type manually by data processing platform (DPP) 2 Interpolation expands positive sample crowd.It is, for example, possible to use for a long time to certain class commodity there are preference client to it is potential just Sample population is expanded or is carried out the extreme user that any classification commodity are never bought in potential positive sample crowd It deletes or potential positive sample crowd is expanded using the client of frequent browse advertisements.For another example feature manually weights, Commodity will be bought in potential sample population, and at least client of 2 times or more is set as positive sample crowd.Or according to preset mark Label are modified the potential sample population got, such as after the potential sample population got, can choose part mark Such as women label, age label are signed, the crowd for meeting chosen label is chosen in potential sample population as positive sample people Group.Alternatively, can positive sample crowd be chosen with reference to the label chosen according to the industry where the classification.

In the present embodiment, data processing platform (DPP) 2 can according to number ratio preset between positive and negative sample population, at random from Certain customers are extracted in the whole network user as negative sample crowd, for example, the number ratio between general positive and negative sample population is 1:5 Or 1:10.Optionally, data processing platform (DPP) 2 according to above-mentioned essential information determine with the target object there are competitive relation its He pushes object, i.e., is under the jurisdiction of the commodity of same type with the target object, but the two belongs to different suppliers, two supplies It can vie each other between quotient and grab client.Data processing platform (DPP) 2 is pushed away with target object there are other of competitive relation getting After sending object, all clients there are network behavior to other push objects can be counted, using these clients as negative sample Crowd.Wherein, network behavior can include the operation behaviors such as purchase, click, browsing or collection.Optionally, data processing is put down Platform 2 extracts certain customers as negative sample crowd after positive sample crowd is got, at random from the whole network user, then, is based on Spy's algorithm obtains negative sample crowd, and the part negative sample people more similar to positive sample crowd is chosen from negative sample crowd Group, is adjusted positive sample crowd by the negative sample crowd of selection, to change sampling instances, so as to which sample is avoided to select The training difference brought when selecting.

In the present embodiment, a machine learning model is built in advance in data processing platform (DPP) 2, by by positive sample crowd It is input in the machine learning model and is trained with negative sample crowd, it is pre- using the model that trained performance is stablized as first Survey model.In the present embodiment, due to being trained using positive and negative sample population to machine learning model, there is supervision to learn as one kind It practises, is conducive to obtain more rational prediction model, so as to carry out more accurately predicting knot to the sample for being used to predict Fruit.

Wherein, the algorithm of machine learning can include two classification based training methods of standard, such as logistic regression (Logistic Regression, abbreviation LR) algorithm, support vector machines (Support Vector Machine, abbreviation SVM) algorithm and iteration Decision tree (Gradient Boosting Decision Tree, abbreviation GBDT) algorithm, single classification based training side can also be included Method, such as one-class support vector machine (One Class SVM).

In the present embodiment, data processing platform (DPP) 2 advances with the dispensing result after the advertisement dispensing of existing all industries The transfer learning model built in advance is trained, then using trained transfer learning model according to the throwing of all industries Put the relevance as a result, between excavating every profession and trade.For example, like the user's feature for buying shoes may be with liking buying cup User's feature there are certain relevances.

In the present embodiment, data processing platform (DPP) 2 repaiies the positive and negative sample population got by transfer learning model Just.Specifically, transfer learning model is using the relevance between the every profession and trade that analyzes, to the positive and negative sample population that gets into Row is corrected, and revised part negative sample may become positive sample, and part positive sample may become negative sample.

Wherein, transfer learning model can be constantly updated, and advertisement dispensing is carried out whenever there are one advertisers, and launching result will It flows back into transfer learning model, transfer learning model can be according to the dispensing result of advertisement again to its re -training with to model It is updated.

In the present embodiment, data processing platform (DPP) 2 will repair after revised positive sample crowd and negative sample crowd is got Positive sample crowd and negative sample crowd after just are input in preset machine learning model to be trained again, will be trained The model that performance is stablized is as the second prediction model.

Further, data processing platform (DPP) 2 obtains potential crowd according to default rule first, using potential crowd as use The potential customers of family advertisement.For example, default rule can be：Select have search for a period of time recently from the whole network user Or the client of this classification object is clicked as potential crowd.Alternatively, it is chosen from the whole network living on network in nearest a period of time The higher client of jerk is as potential crowd.Alternatively, the modes such as client are precisely oriented according to label selects potential crowd, such as select Client of the female age between 25~30 years old is taken as potential crowd.

Potential crowd is separately input in the first prediction model and the second prediction model by data processing platform (DPP) 2, utilizes One prediction model predicts each client in potential crowd, client each in potential crowd is carried out using the second prediction model pre- It surveys.Data processing platform (DPP) 2 is by the first prediction module to the prediction probability and the second prediction module pair of client each in potential crowd The prediction probability of each client is weighted in potential crowd, then according to weighted results therefrom selected part client as target Crowd.

Data processing platform (DPP) 2 is after target group is got, it is possible to push target object to target group.

Information transmission system provided in this embodiment, by according to the input by user and relevant essential information of target object Positive sample crowd and negative sample crowd are obtained, learning training is carried out to positive sample crowd and negative sample crowd, it is pre- to obtain first Survey model, and positive sample crowd and negative sample crowd be modified using the transfer learning model of structure, to it is revised just Sample population and negative sample crowd carry out learning training, to obtain the second prediction model and utilize the first prediction model and the Two prediction models are respectively predicted to obtain target group potential crowd, and target object is pushed to target group.This implementation In example, by having the machine learning module of supervision and transfer learning model, target group is got from potential crowd so that prediction As a result targeted advertisements more precisely, and to it are pushed.Due to prediction result, precisely then correspondingly target group is then that advertisement is launched The crowd of best results, and then advertiser can be helped to improve the effect that advertisement is launched.

Embodiment two

As shown in Fig. 2, it is a kind of schematic diagram of information transmission system provided by Embodiment 2 of the present invention.The information pushes System includes：The data processing platform (DPP) 2 in front end 1 and above-described embodiment in above-described embodiment.

Front end 1 includes：Line module 11 and the interactive module 12 for connecting line module and data processing platform (DPP) 2.

Data processing platform (DPP) 2 includes：Samples selection node 21, sample training node 22, prediction node 23 and push node 24。

Wherein, line module 11, for receiving essential information input by user.

Introduction about essential information can be found in the record of related content in above-described embodiment, and details are not described herein again.

Interactive module 12, for user to be issued data processing platform (DPP) samples selection node in essential information input by user 21。

For receiving the essential information of the transmission of interactive module 12, positive sample is selected using essential information for samples selection node 21 Positive sample crowd and negative sample crowd are sent to sample training node 22 by this crowd and negative sample crowd.

Sample chooses the process that node 21 chooses positive and negative sample population according to essential information, reference can be made to phase in above-described embodiment The record held inside the Pass, details are not described herein again.

Sample training node 22, it is pre- to obtain first for carrying out learning training to positive sample crowd and negative sample crowd It surveys model and positive sample crowd and negative sample crowd is modified based on transfer learning model, to revised positive sample Crowd and negative sample crowd carry out learning training, to obtain the second prediction model.

Specifically, during learning training is carried out to positive sample crowd and negative sample crowd, sample training node 22 Firstly the need of to trained positive and negative sample population, feature extraction is carried out based on the Feature Engineering in machine learning, to get just The fisrt feature information of negative sample crowd.After the fisrt feature information of positive and negative sample population is got, the spy in machine learning Sign engineering can also carry out feature including the processing procedures such as discretization, standardization, characteristic crossover and Feature Conversion, then obtain Fisrt feature information after processing, by the type parameter of positive and negative sample population and treated fisrt feature information, input It is trained into the machine learning model built in advance, the first prediction model is obtained after the completion of training.For example, using positive and negative The type parameter of sample population and treated feature form an input, and input includes UI, X₁, X₂,…..X_n, Y, wherein UI represents the ID of user, X₁-X_nThe feature of sample, Y represent the type parameter of positive negative sample after representative processing, generally represent negative with 0 Sample represents positive sample with 1, and the input of above-mentioned form is input to machine learning model is trained to form the first prediction mould Type.

Optionally, feature extraction is carried out to positive and negative sample population, obtains fisrt feature information, got according to essential information The target industry that target object is subordinate to carries out fisrt feature information according to the historic training data of target industry being subordinate to Adjustment, obtains first object characteristic information, for example, can be analyzed according to historic training data in the target industry has relatively The Partial Feature of conspicuousness, using this Partial Feature to the fisrt feature information that is extracted from positive sample crowd and negative sample crowd It is adjusted, obtains first object characteristic information.For example, these are filtered out from fisrt feature information with relative significance Partial Feature is as first object characteristic information.Alternatively, in order to simplify calculating process, can directly be sieved from fisrt feature information Partial Feature is chosen out as first object characteristic information.Further, using the first object characteristic information filtered out and just The type parameter of negative sample crowd is trained machine learning model, then can get the first prediction model.

Further, the positive and negative sample that sample training node 22 gets samples selection node 21 by transfer learning model This crowd is modified.Specifically, transfer learning model is using the relevance between the every profession and trade analyzed, to getting just Negative sample crowd is modified, and revised part negative sample may become positive sample, and part positive sample may become Negative sample.

It is similar with the process that the first prediction model obtains, after revised positive and negative sample population is got, after amendment Positive negative sample people carry out feature extraction, by the type parameter of revised positive and negative sample population and the second feature extracted Information in the specific format, is input in machine learning model and is trained, and obtains the second prediction model of advertiser.

Optionally, feature extraction is carried out to revised positive and negative sample population, obtains second feature information, according to basic letter Breath gets the target industry that target object is subordinate to, according to the historic training data of target industry being subordinate to second feature Information is adjusted, and obtains the second target signature information, for example, can be analyzed according to historic training data in the target industry Partial Feature with relative significance, using this Partial Feature to extracted from positive sample crowd and negative sample crowd second Characteristic information is adjusted, and obtains the second target signature information.For example, these are filtered out from second feature information has relatively The Partial Feature of conspicuousness is as the second target signature information.Alternatively, in order to simplify calculating process, it can be directly from second feature Go out Partial Feature as the second target signature information in being screened in information.Further, the second target signature filtered out is utilized The type parameter of information and positive and negative sample population is trained machine learning model, then can get the second prediction module.

Predict node 23, for being predicted respectively potential crowd using the first prediction model and the second prediction model, To obtain target group, target group is sent to push node 24.

Specifically, prediction node 23 obtains potential crowd according to default rule first, and potential crowd is pushed away as user Send the potential customers of advertisement.About default rule reference can be made in above-described embodiment related content record, details are not described herein again. Potential crowd is input in the first prediction model by prediction node 23, using first prediction model to each client in potential crowd It is predicted, obtains the first prediction probability of each client in the first prediction crowd and the first prediction people.It will be through in the present embodiment The potential crowd crossed after the prediction of the first prediction model is known as the first prediction crowd.

Further, potential crowd is input in the second prediction model by prediction node 23, utilizes second prediction model Each client in potential crowd is predicted, obtains the second prediction of each user in the second prediction crowd and the second prediction crowd Probability.The potential crowd after the prediction of the second prediction module is known as the second prediction crowd in the present embodiment.

In order to make advertisement pushing precision higher, in the present embodiment, need using the second prediction crowd to the first prediction people Group is modified, to get the client for being more suitable for advertisement.Predict that node 23 is pre- to first using the second prediction crowd Survey crowd is modified, to obtain target group.Specifically, prediction node 23 for each client by the first prediction probability and the Two prediction probability weighted averages obtain the final prediction probability of client, and all users are ranked up according to final prediction probability To obtain sequence crowd, target group is chosen from sequence crowd.

Node 24 is pushed, for pushing target object to target group.

Information transmission system provided in this embodiment gets the first prediction by the machine learning mode for having supervision first Crowd in order to make dispensing more accurate, based on the association between industry in transfer learning model, repaiies positive and negative sample population Just, it is then based on revised positive and negative sample and obtains the second prediction crowd, predict that crowd's is pre- based on the first prediction crowd and second After survey probability is weighted averagely, is chosen from potential crowd and obtain the target group for launching advertisement so that prediction result is more Precisely, and to it targeted advertisements are pushed.Due to prediction result, precisely then correspondingly target group is then that advertisement delivery effect is best Crowd, and then can help advertiser improve advertisement launch effect.

Embodiment three

As shown in figure 3, it is the structure diagram of a kind of interactive device that the embodiment of the present invention three provides.The interactive device Including：Front end 1 in above-described embodiment.Front end 1, will for receiving the input by user and relevant essential information of target object Essential information is sent to data processing platform (DPP), is additionally operable to and receives the target person that data processing platform (DPP) is obtained according to essential information Group, and push target object to target group.

Front end 1 includes line module 11 and the interactive module 12 for connecting line module and data processing platform (DPP).

Line module 11 obtains for receiving essential information input by user and receiving data platform according to essential information The target group taken, and push target object to target group.

Interactive module 12, for essential information to be issued data processing platform (DPP), so that data processing platform (DPP) is according to basic letter Breath obtains target group.

Interactive device provided in this embodiment receives the input by user and relevant essential information of target object, will be basic Information is sent to data processing platform (DPP) and receives the target group that data processing platform (DPP) is obtained according to the essential information, and Target object is pushed to target group.In the present embodiment, the user of target object is pushed, need to only be based on the interactive device to data Processing platform sends some essential informations, it is possible to obtain target group, then push target object to target group, improve User experience.

Example IV

As shown in figure 4, it is the structure diagram of a kind of information push-delivery apparatus that the embodiment of the present invention four provides.The information Pusher includes：Data processing platform (DPP) 2 in above-described embodiment.

In the present embodiment, data processing platform (DPP) 2 includes：Samples selection node 21, sample training node 22, prediction node 23 With push node 24.

Wherein, samples selection node 21 for receiving preceding essential information, using essential information selection positive sample crowd and is born Positive sample crowd and negative sample crowd are sent to sample training node 22 by sample population.

Node 24 is pushed, for pushing target object to target group.

About node each in data processing platform (DPP) 2 to the process of data processing, reference can be made in above-described embodiment mutually inside the Pass The record of appearance, details are not described herein again.

Information push-delivery apparatus provided in this embodiment gets the first prediction by the machine learning mode for having supervision first Crowd in order to make dispensing more accurate, based on the association between industry in transfer learning model, repaiies positive and negative sample population Just, it is then based on revised positive and negative sample and obtains the second prediction crowd, predict that crowd's is pre- based on the first prediction crowd and second After survey probability is weighted averagely, is chosen from potential crowd and obtain the target group for launching advertisement so that prediction result is more Precisely, and to it targeted advertisements are pushed.Due to prediction result, precisely then correspondingly target group is then that advertisement delivery effect is best Crowd, and then can help advertiser improve advertisement launch effect.

Embodiment five

As shown in figure 5, it is the flow diagram of a kind of model training method that the embodiment of the present invention five provides.The model Training method includes the following steps：

S101, feature extraction is carried out to positive sample crowd and negative sample crowd, to obtain the fisrt feature information of sample.

Specifically, it is relevant according to target object input by user, positive sample crowd and negative sample crowd are got, then Feature extraction is carried out to positive sample crowd and negative sample crowd.About the process that positive and negative sample population is obtained according to essential information, The record of related content in above-described embodiment is can be found in, details are not described herein again.

During learning training is carried out to positive sample crowd and negative sample crowd, it is necessary first to trained positive and negative sample This crowd carries out feature extraction based on the Feature Engineering in machine learning, to get the fisrt feature of positive and negative sample population letter Breath.After the fisrt feature information of positive and negative sample population is got, the Feature Engineering in machine learning can also wrap feature The processing procedures such as discretization, standardization, characteristic crossover and Feature Conversion are included, the fisrt feature letter after then being handled Breath by the type parameter of positive and negative sample population and treated fisrt feature information, is input to the machine learning built in advance It is trained in model, the first prediction model is obtained after the completion of training.For example, using the type parameter of positive and negative sample population with And treated feature forms an input, input includes UI, X₁, X₂,…..X_n, Y, wherein UI represent the ID of user, X₁-X_nGeneration The feature of sample after list processing, Y represent the type parameter of positive negative sample, generally represent negative sample with 0, positive sample is represented with 1, will The input of above-mentioned form is input to machine learning model and is trained to form the first prediction model.

In the present embodiment, using in shopping website as specific application scenarios, to model training method provided in this embodiment It is explained.Initially set up the data label that is possible under application scenarios, in this example, positive and negative sample population be by Client's composition, the feature of positive sample crowd and negative sample crowd include：The essential information of client, the behavior of client, client Condition of assets, the interest graph of client and abstract characteristics.Wherein, the essential information of client includes：Gender, age, marriage shape Whether condition has the information such as child, child's age.The behavior of client includes：It browses commodity, commodity is added in shopping cart, purchase quotient The behaviors such as product and collecting commodities or search commercial articles relevant information.The condition of assets of client includes：Whether there is vehicle, whether have The information such as room, mobile phone brand.The interest graph of client includes：The interested type of merchandise of client, the interested brand of client, visitor The interested classification in family.The interest graph of common customer can according to the historical data that can react Customer Shopping and in advance The computation model of structure is associated what analysis can obtain.Received by abstract characteristics refer to whether directly affect client clicks Advertisement factor, if client is to the favorable rating of advertisement, some users will not put advertisement completely in real life, and client's is upper Line frequent degree etc., such as the upper limit frequent degree by client, can predict the probability reached the standard grade in next week.

First foundation feature after extraction is subjected to discretization and standardization, discretization and standardization will be passed through First foundation feature afterwards carries out characteristic crossover derivation process, obtains fisrt feature information.

S102, it the type parameter of fisrt feature information and each sample is input in preset machine learning model carries out Training, to obtain the first prediction model.

According to the target industry that essential information acquisition target object is subordinate to, the historic training data of target industry is obtained, The fisrt feature information is adjusted according to historic training data, to obtain first object characteristic information, by first object It is trained in the type parameter of characteristic information and each sample input machine learning model, to obtain the first prediction model.

S103, revised positive sample crowd and negative sample crowd characteristic are extracted, to obtain the second feature of sample letter Breath.

It is modified using school's module is migrated to getting positive sample crowd and negative sample crowd, specific introduce can be found in The record of related content in above-described embodiment, details are not described herein again.

Further, feature extraction is carried out to revised positive sample crowd and negative sample crowd, obtains second feature letter Breath.Similar with the process of fisrt feature acquisition of information, details are not described herein again,

S104, it will be trained in the type parameter of second feature information and each sample input machine learning model, with Obtain the second prediction model.

According to the target industry that essential information acquisition target object is subordinate to, the historic training data of target industry is obtained, Second feature information is adjusted according to historic training data, to obtain the second target signature information, by the second target signature It is trained in the type parameter of information and each sample input machine learning model, to obtain the second prediction model.

In the present embodiment, bonding is converted from agreement buffer format to obtained fisrt feature information and second feature information Value is to form.

Model training method provided in this embodiment, by carrying out feature extraction to positive sample crowd and negative sample crowd, To obtain the fisrt feature information of sample, the type parameter of fisrt feature information and each sample is input to preset engineering It practises and being trained in model, to obtain the first prediction model, revised positive sample crowd and negative sample crowd characteristic are extracted, To obtain the second feature information of sample, the type parameter of second feature information and each sample is inputted in machine learning model It is trained, to obtain the second prediction model.In the present embodiment, first is obtained by machine mould in advance using positive and negative sample population Survey model, and using being modified to positive and negative sample population with reference to transfer learning model after, be again based on machine learning model Obtain the second prediction module so that the result predicted based on the first prediction model and the second prediction module is more accurate.

Embodiment six

As shown in fig. 6, it is the structure diagram of a kind of model training apparatus that the embodiment of the present invention six provides.The model Training device 3 includes：Characteristic extracting module 31 and training module 32.

Wherein, characteristic extracting module 31, for carrying out feature extraction to positive sample crowd and negative sample crowd, to obtain sample This fisrt feature information and the revised positive sample crowd and the negative sample crowd characteristic are extracted, to obtain The second feature information of sample.

Training module 32, for the type parameter of fisrt feature information and each sample to be input to preset machine learning It is trained in model, to obtain the first prediction model and input the type parameter of second feature information and each sample It is trained in machine learning model, to obtain the second prediction model.

Further, characteristic extracting module 31, specifically for obtaining target according to essential information relevant with target object The target industry that object is subordinate to obtains the historic training data of target industry, and fisrt feature is believed according to historic training data Breath is adjusted, and to obtain first object characteristic information, the type parameter of first object characteristic information and each sample is inputted It is trained in machine learning model, to obtain the first prediction model.

Further, training module 32, specifically for the target being subordinate to according to essential information acquisition target object Industry obtains the historic training data of target industry, second feature information is adjusted according to historic training data, to obtain The type parameter of second target signature information and each sample is inputted in machine learning model and carried out by the second target signature information Training, to obtain the second prediction model.

Further, training module 32 specifically for carrying out feature extraction to positive sample crowd and negative sample crowd, obtain First foundation feature after extraction is carried out discretization and standardization, will pass through discretization and standard by first foundation feature Change treated first foundation feature and carry out characteristic crossover derivation process, obtain fisrt feature information.

Further, characteristic extracting module 31, specifically for revised positive sample crowd and negative sample crowd progress Feature extraction obtains the second foundation characteristic, and the second foundation characteristic after extraction is carried out discretization and standardization, will be passed through The second foundation characteristic after discretization and standardization carries out characteristic crossover derivation process, obtains second feature information.

Further, characteristic extracting module 31 are additionally operable to fisrt feature information to obtaining and second feature information from association View buffer format is converted into key-value pair form.

Further, characteristic extracting module 31 are additionally operable to carrying out feature extraction to positive sample crowd and negative sample crowd, Before obtaining the fisrt feature information of sample, positive sample crowd and negative sample are obtained according to the essential information input by user Crowd.

Model training apparatus provided in this embodiment, by carrying out feature extraction to positive sample crowd and negative sample crowd, To obtain the fisrt feature information of sample, the type parameter of fisrt feature information and each sample is input to preset engineering It practises and being trained in model, to obtain the first prediction model, revised positive sample crowd and negative sample crowd characteristic are extracted, To obtain the second feature information of sample, the type parameter of second feature information and each sample is inputted in machine learning model It is trained, to obtain the second prediction model.In the present embodiment, first is obtained by machine mould in advance using positive and negative sample population Survey model, and using being modified to positive and negative sample population with reference to transfer learning model after, be again based on machine learning model Obtain the second prediction module so that the result predicted based on the first prediction model and the second prediction module is more accurate.

Embodiment seven

As shown in fig. 7, it is the flow diagram of a kind of target group's acquisition methods that the embodiment of the present invention seven provides.It should Target group's acquisition methods include the following steps：

S201, positive sample crowd and negative sample people are obtained with the relevant essential information of target object according to input by user Group.

S202, learning training is carried out to positive sample crowd and negative sample crowd, to obtain the first prediction model.

S203, positive sample crowd and negative sample crowd are modified using the transfer learning model of structure.

S204, learning training is carried out to revised positive sample crowd and negative sample crowd, to obtain the second prediction model.

About the introduction of S201~S204, reference can be made in above-described embodiment related content record, this is repeated no more.

S205, potential crowd is predicted to obtain target person respectively using the first prediction model and the second prediction model Group.

Specifically, the potential crowd is predicted to obtain the first prediction crowd based on the first prediction model, is based on Second prediction model is predicted potential crowd to obtain the second prediction crowd, using the second prediction crowd to the first prediction people Group is modified, to obtain target group.

In the present embodiment, the first prediction crowd is modified using the second prediction crowd, to obtain target group, specifically Including：It obtains the first prediction probability of each user in the first prediction crowd, obtains the of each user in the second prediction crowd Two prediction probabilities for each user, by the first prediction probability and the second prediction probability weighted average, obtain the final of the user All users according to final prediction probability are ranked up, to obtain sequence crowd, target are chosen in sequence crowd by prediction probability Crowd.

Target group's acquisition methods provided in this embodiment, by according to input by user relevant basic with target object Acquisition of information positive sample crowd and negative sample crowd carry out learning training to positive sample crowd and negative sample crowd, to obtain the One prediction model, and positive sample crowd and negative sample crowd are modified using the transfer learning model of structure, after amendment Positive sample crowd and negative sample crowd carry out learning training, with obtain the second prediction model and utilize the first prediction model Potential crowd is predicted to obtain target group respectively with the second prediction model, target object is pushed to target group.This In embodiment, by having the machine learning module of supervision and transfer learning model, target group is got from potential crowd so that Prediction result is more accurate, and pushes targeted advertisements to it.Due to prediction result, precisely then correspondingly target group is then advertisement The crowd of best results is launched, and then advertiser can be helped to improve the effect that advertisement is launched.

Embodiment eight

As shown in figure 8, it is the structure diagram of a kind of target group's acquisition device that the embodiment of the present invention eight provides.It should Target group's acquisition device 4 includes：Acquisition module 41, training module 42, correcting module 43 and prediction module 44.

Wherein, acquisition module 41, for obtaining positive sample with the relevant essential information of target object according to input by user Crowd and negative sample crowd.

Training module 42, for carrying out learning training to the positive sample crowd and the negative sample crowd, to obtain the One prediction model and learning training is carried out to the revised positive sample crowd and the negative sample crowd, to obtain the Two prediction models.

Correcting module 43, for utilizing the transfer learning model of structure to the positive sample crowd and the negative sample crowd It is modified.

Prediction module 44, using first prediction model and second prediction model respectively to the potential crowd into Row is predicted to obtain target group.

Further, prediction module 44 carry out in advance the potential crowd specifically for being based on first prediction model It surveys to obtain the first prediction crowd, the potential crowd is predicted based on second prediction model to obtain the second prediction Crowd is modified the first prediction crowd using the second prediction crowd, to obtain the target group.

Further, prediction module 44, specifically for obtaining the first prediction of each user in the first prediction crowd Probability obtains the second prediction probability of each user in the second prediction crowd, for each user, described first is predicted Probability and the second prediction probability weighted average, obtain the final prediction probability of the user, by all users according to it is described most Whole prediction probability is ranked up, and to obtain sequence crowd, the target group is chosen from the sequence crowd.

Target group's acquisition device provided in this embodiment, by according to input by user relevant basic with target object Acquisition of information positive sample crowd and negative sample crowd carry out learning training to positive sample crowd and negative sample crowd, to obtain the One prediction model, and positive sample crowd and negative sample crowd are modified using the transfer learning model of structure, after amendment Positive sample crowd and negative sample crowd carry out learning training, with obtain the second prediction model and utilize the first prediction model Potential crowd is predicted to obtain target group respectively with the second prediction model, target object is pushed to target group.This In embodiment, by having the machine learning module of supervision and transfer learning model, target group is got from potential crowd so that Prediction result is more accurate, and pushes targeted advertisements to it.Due to prediction result, precisely then correspondingly target group is then advertisement The crowd of best results is launched, and then advertiser can be helped to improve the effect that advertisement is launched.

Embodiment nine

As shown in figure 9, it is the flow diagram of a kind of information-pushing method that the embodiment of the present invention nine provides.The information Method for pushing includes the following steps：

S301, positive sample crowd and negative sample people are obtained with the relevant essential information of target object according to input by user Group.

In the present embodiment, target object can be advertisement to be pushed, or text, image, music, regards news Frequently, even if getting well even in means of communication has.In the present embodiment, by taking target object is the advertisement with push as an example, then input User with the relevant essential information of target object is advertiser, and receives the user i.e. target of the promotion message of target object The user of object is known as client in the present embodiment.Attempt to push a target pair to certain customers i.e. user in advertiser As when, it is necessary first to input with the relevant essential information of the target object.For example, one advertiser's login interface of setting, wide After accusing main login, advertiser can enter the input interface of essential information, and the basic of target object is inputted on input interface Information.

About the introduction of essential information, reference can be made in above-described embodiment related content record, details are not described herein again.

After the user i.e. essential information of the target object of advertisement primary input is received, it can be obtained according to these essential informations Get positive sample crowd.Specifically, the classification ID in essential information, brand ID, commodity ID and corresponding Extension Software Platform, Keyword such as commodity price of discount, target object etc., finds in a period of time and bought, collected or liked this kind of The client of object, using these clients as potential positive sample crowd.It further, can also be to potential positive sample crowd according to one Fixed rule is expanded, to obtain positive sample crowd.For example, the loyal client in the shop can be found according to shop ID, Potential positive sample is expanded by these loyalty clients, and then forms positive sample crowd.For another example according in essential information Brand ID, obtain there are associated commodity with brand ID, then obtain to the interested client of these commodity, use these Client expands potential positive sample crowd.For another example the classification ID in essential information, there are one with the classification for acquisition Fixed associated commodity, then to the interested client of these commodity, expand potential sample population using these clients.

Change it is possible to further to the special client for being biased to type manually increase, it is, for example, possible to use for a long time to certain Class commodity expand potential positive sample crowd there are the client of preference or will never be bought in potential positive sample crowd The extreme user of any classification commodity deleted or using frequent browse advertisements client to potential positive sample crowd into Row expands.For another example feature manually weights, by commodity were bought in potential sample population, at least client of 2 times or more was set as Positive sample crowd.Or the potential sample population got is modified according to preset label, such as latent what is got After sample population, part labels such as women label, age label etc. can be chosen, is chosen in potential sample population and meets institute The crowd of label is chosen as positive sample crowd.Alternatively, it can be chosen according to the industry where the classification with reference to the label chosen Positive sample crowd.

In the present embodiment, it can be taken out from the whole network user at random according to number ratio preset between positive and negative sample population Certain customers are taken as negative sample crowd, for example, the number ratio between general positive and negative sample population is 1:5 or 1:10.It can Selection of land, according to above-mentioned essential information determine with the target object there are competitive relation other push objects, i.e., with the target pair Commodity as being under the jurisdiction of same type, but the two belongs to different suppliers, the two, which can vie each other, grabs client.It is getting With target object there are competitive relation other push objects after, can count to other push object there are network behaviors All clients, using these clients as negative sample crowd.Wherein, network behavior can include purchase, click, browses or collect Etc. operation behaviors.Optionally, after positive sample crowd is got, certain customers are extracted from the whole network user at random as negative sample Then crowd, negative sample crowd is obtained based on spy's algorithm, is chosen from negative sample crowd more similar to positive sample crowd Part negative sample crowd is adjusted positive sample crowd by the negative sample crowd of selection, to change sampling instances, so as to The training difference brought during to avoid samples selection.

S302, learning training is carried out to positive sample crowd and negative sample crowd, to obtain the first prediction model.

In the present embodiment, a machine learning model is built in advance, by the way that positive sample crowd and negative sample crowd are inputted It is trained into the machine learning model, using the model that trained performance is stablized as the first prediction model.The present embodiment In, due to being trained machine learning model using positive and negative sample population, as a kind of supervised learning, be conducive to obtain more Add rational prediction model, so as to carry out more accurate prediction result to the sample for being used to predict.

Wherein, the algorithm of machine learning can be found in the record of related content in above-described embodiment, and details are not described herein again.

S303, potential crowd is predicted based on the first prediction model to obtain target group.

In the present embodiment, potential crowd is obtained according to default rule first, using potential crowd as user's advertisement Potential customers.For example, default rule can be：It selects to have search for a period of time recently from the whole network user or clicks this The client of classification object is as potential crowd.Alternatively, being chosen from the whole network in nearest a period of time, liveness is higher on network Client is as potential crowd.Alternatively, the modes such as client are precisely oriented according to label selects potential crowd, such as choose female age Client between 25~30 years old is as potential crowd.

Potential crowd is input in the first prediction model, using first prediction model to each client in potential crowd into Row prediction, to obtain target group.The potential crowd after the prediction of the first prediction model is known as first in advance in the present embodiment Survey crowd,, can be by repairing the first prediction crowd in order to make dispensing precision higher after the first prediction crowd is obtained Just, target group is obtained.Preferably, it is ranked up according to the prediction probability of client each in potential crowd, then from sequence Selected part client is as target group in potential crowd afterwards, for example, choosing after potential crowd sorts according to prediction probability Preceding 5,000,000 clients are as target group.It is alternatively possible to target object is subordinate to according to client each in the first prediction crowd The informativeness of the brand ID of category continues to be modified the first prediction crowd, chooses brand ID's from the first prediction crowd The classification ID's that loyal client is subordinate to as target group or according to user each in the first prediction crowd to target object Informativeness continues to be modified the first prediction crowd, and loyal client's conduct of classification ID is chosen from the first prediction crowd Target group.

S304, target object is pushed to target group.

Information-pushing method provided in this embodiment, by according to the input by user and relevant essential information of target object Positive sample crowd and negative sample crowd are obtained, learning training is carried out to positive sample crowd and negative sample crowd, it is pre- to obtain first Model is surveyed, potential crowd is predicted to obtain target group based on the first prediction model, target pair is pushed to target group As.In the present embodiment, interest level of the potential crowd to target object is predicted by way of supervised learning so that prediction As a result more precisely, it is then based on interested degree and the first prediction crowd is chosen from potential crowd as target group, and Targeted advertisements are pushed to it.Due to prediction result, precisely then correspondingly target group is then the best crowd of advertisement delivery effect, And then advertiser can be helped to improve the effect that advertisement is launched.

Embodiment ten

As shown in Figure 10, the structure diagram of a kind of information push-delivery apparatus provided for the embodiment of the present invention ten.The letter Breath pusher 5 includes：Sample acquisition module 51, training module 52, prediction module 53 and pushing module 54.

Wherein, sample acquisition module 51, for being obtained just with the relevant essential information of target object according to input by user Sample population and negative sample crowd.

In the present embodiment, target object can be advertisement to be pushed, input and the relevant essential information of target object User is advertiser, and receives the user i.e. user of target object of the promotion message of target object, is claimed in the present embodiment For client.When advertiser attempts to push a target object to certain customers i.e. user, it is necessary first to input and the target The relevant essential information of object.Introduction about essential information can be found in the record of related content in above-described embodiment, herein not It repeats again.

Sample acquisition module 51, can basis after the user i.e. essential information of the target object of advertisement primary input is received These essential informations get positive sample crowd.

Training module 52, for carrying out learning training to the positive sample crowd and the negative sample crowd, to obtain the One prediction model.

In the present embodiment, one machine learning model of structure in advance of training module 52, by by positive sample crowd and negative sample This crowd is input in the machine learning model and is trained, using the model that trained performance is stablized as the first prediction mould Type.In the present embodiment, training module 12 is had due to being trained using positive and negative sample population to machine learning model as one kind Supervised learning is conducive to obtain more rational prediction model, more accurate so as to be carried out to the sample for being used to predict Prediction result.Algorithm about machine learning can be found in the record of related content in above-described embodiment, and details are not described herein again.

Prediction module 53 predicts potential crowd to obtain target group for being based on first prediction model.

In the present embodiment, potential crowd is obtained according to default rule first, using potential crowd as user's advertisement Potential customers.About preset rules reference can be made in above-described embodiment related content record, details are not described herein again.

Further, potential crowd is input in the first prediction model by prediction module 53, utilizes first prediction model Each client in potential crowd is predicted, to obtain target group.It will be after the prediction of the first prediction model in the present embodiment Potential crowd be known as the first prediction crowd.After the first prediction crowd is obtained, in order to make dispensing precision higher, prediction module 53 It can be by being modified to the first prediction crowd, to obtain target group.Preferably, prediction module 13 is according in potential crowd The prediction probability of each client is ranked up, then from the potential crowd after sequence selected part client as target group, For example, 5,000,000 clients are as target group before choosing after potential crowd sorts according to prediction probability.Optionally, mould is predicted The informativeness of brand ID that block 13 can be subordinate to target object according to client each in the first prediction crowd, continues to first Prediction crowd is modified, chosen from the first prediction crowd the loyal client of brand ID as target group or according to The informativeness of classification ID that each user is subordinate to target object in first prediction crowd continues to carry out the first prediction crowd It corrects, the loyal client of classification ID is chosen from the first prediction crowd as target group.

Pushing module 54, for pushing the target object to the target group.

Information push-delivery apparatus provided in this embodiment, by according to the input by user and relevant essential information of target object Positive sample crowd and negative sample crowd are obtained, learning training is carried out to positive sample crowd and negative sample crowd, it is pre- to obtain first Model is surveyed, potential crowd is predicted to obtain target group based on the first prediction model, target pair is pushed to target group As.In the present embodiment, interest level of the potential crowd to target object is predicted by way of Supervised machine learning so that Prediction result is more accurate, is then based on interested degree and the first prediction crowd is chosen from potential crowd as target person Group, and push targeted advertisements to it.Due to prediction result, precisely then correspondingly target group is then that advertisement delivery effect is best Crowd, and then advertiser can be helped to improve the effect that advertisement is launched.

Embodiment 11

As shown in figure 11, the flow diagram of a kind of information-pushing method provided for the embodiment of the present invention 11.It should Information-pushing method includes：

S401, the input by user and relevant essential information of target object is obtained.

S402, positive sample crowd and negative sample crowd are obtained according to essential information.

After the user i.e. essential information of the target object of advertisement primary input is received, it can be obtained according to these essential informations Get positive sample crowd.Specifically, the classification ID in essential information, brand ID, commodity ID and corresponding Extension Software Platform, Keyword such as commodity price of discount, target object etc., finds in a period of time and bought, collected or liked this kind of The client of object, using these clients as potential positive sample crowd.It further, can also be to potential positive sample crowd according to one Fixed rule is expanded, to obtain positive sample crowd.Introduction about expansion can be found in related content in above-described embodiment It records, details are not described herein again.

In the present embodiment, it can be taken out from the whole network user at random according to number ratio preset between positive and negative sample population Certain customers are taken as negative sample crowd, for example, the number ratio between general positive and negative sample population is 1:5 or 1:10.It can Selection of land, according to above-mentioned essential information determine with the target object there are competitive relation other push objects, i.e., with the target pair Commodity as being under the jurisdiction of same type, but the two belongs to different suppliers, the two, which can vie each other, grabs client.Optionally, After positive sample crowd is got, certain customers are extracted from the whole network user at random as negative sample crowd, then, based on spy Algorithm obtains negative sample crowd, and the part negative sample crowd more similar to positive sample crowd is chosen from negative sample crowd, leads to It crosses the negative sample crowd chosen to be adjusted positive sample crowd, to change sampling instances, during so as to avoid samples selection The training difference brought.

S403, learning training is carried out to positive and negative sample crowd and negative sample crowd, to obtain the first prediction model.

Optionally, feature extraction is carried out to positive and negative sample population, obtains fisrt feature information, got according to essential information The target industry that target object is subordinate to carries out fisrt feature information according to the historic training data of target industry being subordinate to Adjustment, obtains first object characteristic information, for example, can be analyzed according to historic training data in the target industry has relatively The Partial Feature of conspicuousness, using this Partial Feature to the fisrt feature information that is extracted from positive sample crowd and negative sample crowd It is adjusted, obtains first object characteristic information.For example, these are filtered out from fisrt feature information with relative significance Partial Feature is as first object characteristic information.Alternatively, in order to simplify calculating process, can directly be sieved from fisrt feature information Partial Feature is chosen out as first object characteristic information.

Further, using the first object characteristic information and the type parameter of positive and negative sample population filtered out to engineering It practises model to be trained, then can get the first prediction model.

S404, positive sample crowd and negative sample crowd are modified using the transfer learning model of structure.

In the present embodiment, advance with the dispensing result after the advertisement dispensing of existing all industries and moved to what is built in advance Learning model is moved to be trained.Then using trained transfer learning model according to the dispensing of all industries as a result, excavating Relevance between every profession and trade.For example, like the user's feature for buying shoes that may be deposited with the user's feature for liking buying cup In certain relevance.

In the present embodiment, the positive and negative sample population got is modified by transfer learning model.Specifically, it migrates Learning model is modified the positive and negative sample population got, using the relevance between the every profession and trade analyzed after amendment Part negative sample may become positive sample, and part positive sample may become negative sample.Wherein, transfer learning model energy It is enough to constantly update, advertisement dispensing is carried out whenever there are one advertisers, result is launched and will back flow into transfer learning model, migration is learned Practising model again can be updated model its re -training according to the dispensing result of advertisement.

S405, learning training is carried out to revised positive sample crowd and negative sample crowd, to obtain the second prediction model.

Optionally, feature extraction is carried out to revised positive and negative sample population, obtains second feature information, according to basic letter Breath gets the target industry that target object is subordinate to, according to the historic training data of target industry being subordinate to second feature Information is adjusted, and obtains the second target signature information, for example, can be analyzed according to historic training data in the target industry Partial Feature with relative significance, using this Partial Feature to extracted from positive sample crowd and negative sample crowd second Characteristic information is adjusted, and obtains the second target signature information.For example, these are filtered out from second feature information has relatively The Partial Feature of conspicuousness is as the second target signature information.Alternatively, in order to simplify calculating process, it can be directly from second feature Go out Partial Feature as the second target signature information in being screened in information.

Further, using the second target signature information and the type parameter of positive and negative sample population filtered out to engineering It practises model to be trained, then can get the second prediction module.

S406, potential crowd is predicted to obtain target person respectively using the first prediction model and the second prediction model Group.

In the present embodiment, potential crowd is obtained according to default rule first, using potential crowd as user's advertisement Potential customers.About default rule reference can be made in above-described embodiment two related content record, details are not described herein again.It will be latent It is input in the first prediction model in crowd, each client in potential crowd is predicted using first prediction model, with To the first prediction crowd, the potential crowd after the prediction of the first prediction model is known as the first prediction crowd in the present embodiment.

Further, using potential crowd as the potential customers of user's advertisement.It is pre- that potential crowd is input to second It surveys in model, each client in potential crowd is predicted using second prediction model, it is pre- that second will be passed through in the present embodiment The potential crowd surveyed after module prediction is known as the second prediction crowd.

In order to make advertisement pushing precision higher, in the present embodiment, need using the second prediction crowd to the first prediction people Group is modified, to get the client for being more suitable for advertisement.

S407, target object is pushed to target group.

Information-pushing method provided in this embodiment gets the first prediction by the machine learning mode for having supervision first Crowd in order to make dispensing more accurate, based on the association between industry in transfer learning model, repaiies positive and negative sample population Just, it is then based on revised positive and negative sample and obtains the second prediction crowd, predict that crowd's is pre- based on the first prediction crowd and second After survey probability is weighted averagely, is chosen from potential crowd and obtain the target group for launching advertisement so that prediction result is more Precisely, and to it targeted advertisements are pushed.Due to prediction result, precisely then correspondingly target group is then that advertisement delivery effect is best Crowd, and then can help advertiser improve advertisement launch effect.

Embodiment 12

As shown in figure 12, the flow diagram for the information-pushing method of the embodiment of the present invention 12.The information pushes Method includes：

S501, positive sample crowd and negative sample people are obtained with the relevant essential information of target object according to input by user Group.

After the user i.e. essential information of the target object of advertisement primary input is received, it can be obtained according to these essential informations Get positive sample crowd.Detailed process can be found in the record of related content in above-described embodiment, and details are not described herein again.

S502, learning training is carried out to positive sample crowd and negative sample crowd, to obtain the first prediction model.

Detailed process can be found in the record of related content in above-described embodiment, and details are not described herein again.

S503, potential crowd is predicted based on the first prediction model to obtain the first prediction crowd.

In the present embodiment, potential crowd is obtained according to default rule first, using potential crowd as user's advertisement Potential customers.About default rule reference can be made in above-described embodiment related content record, details are not described herein again.It will be potential Crowd is input in the first prediction model, each client in potential crowd is predicted using first prediction model, to obtain Potential crowd after the prediction of the first prediction model is known as the first prediction crowd in the present embodiment by the first prediction crowd.

S504, positive sample crowd and negative sample crowd are modified using the transfer learning model of structure.

In the present embodiment, advance with the dispensing result after the advertisement dispensing of existing all industries and moved to what is built in advance Learning model is moved to be trained.Then using trained transfer learning model according to the dispensing of all industries as a result, excavating Relevance between every profession and trade.

The positive and negative sample population got is modified by transfer learning model.Specifically, transfer learning model profit With the relevance between the every profession and trade analyzed, the positive and negative sample population got is modified, sample is born in revised part Originally it may become positive sample, and part positive sample may become negative sample.Wherein, transfer learning model can constantly more Newly, advertisement dispensing is carried out whenever there are one advertisers, launches result and will back flow into transfer learning model, transfer learning model can With according to the dispensing result of advertisement again to its re -training to be updated to model.

S505, learning training is carried out to revised positive sample crowd and negative sample crowd, to obtain the second prediction model.

S506, potential crowd is predicted based on the second prediction model to obtain the second prediction crowd.

In the present embodiment, potential crowd is obtained according to default rule first, using potential crowd as user's advertisement Potential customers.Potential crowd is input in the second prediction model, using second prediction model to each visitor in potential crowd Family is predicted, the potential crowd after the prediction of the second prediction module is known as the second prediction crowd in the present embodiment.

S507, the first prediction crowd is modified using the second prediction crowd, to obtain target group.

In order to make advertisement pushing precision higher, in the present embodiment, need using the second prediction crowd to the first prediction people Group is modified, to get the client for being more suitable for advertisement.Specifically, when prediction model predicts potential crowd When, the prediction probability of each user can be got.In the present embodiment, potential crowd, can after the prediction of the first prediction model To obtain the first prediction probability of each user in the first prediction crowd, and potential crowd is after the prediction of the second prediction model, The second prediction probability of each user in the second prediction crowd can be obtained.For each user, the first prediction of the user is general The second prediction probability of rate and the user are weighted averagely, obtain the final prediction probability of the user.By all users according to Final prediction probability is ranked up, and to obtain sequence crowd, target group is got from sequence crowd.

S508, target object is pushed to target group.

S509, the dispensing result for counting target object.

Further, after targeted advertisements are pushed to target group, the dispensing of target object can also be counted as a result, Specifically, network behavior of the target group to target object, that is, targeted advertisements is obtained, by target group to the network of targeted advertisements Behavior is as dispensing as a result, wherein network behavior includes：Target group is to the browsing of commodity, collection, click, pass in targeted advertisements It closes and does shopping, add in the operation behaviors such as shopping cart.

S510, optimization is iterated to transfer learning model using launching result.

The transfer learning model built in advance using all industry datas is to constantly update, wide whenever there is advertiser to carry out It accuses after launching, transfer learning model is updated using result is launched, to optimize the transfer learning model, so as to preferably dig Dig the relevance between industry.Such as like buying the client of cap with liking buying the similar feature between the client of skirt, Parameter optimization is carried out in training preferably output to be suitble to the model of advertiser.It present embodiments provides a kind of full-automatic The dispensing environment of change does the Knowledge Integration between industry according to transfer learning model, and the estrangement between industry/user is weakened, Each mining effect after being adjusted using the dispensing result of other advertisers is more conducive to and provides the precision that advertisement is launched And effect.

Optionally, a knowledge base is built in the present embodiment, result will be launched and flowed back into knowledge base, utilize knowledge base pair Transfer learning model optimizes iteration, and the model of advertiser is suitble to better output.

Schematic diagram for the application scenarios of shopping website as shown in figure 13.Multiple classifications are frequently included on shopping website, Such as women's dress, men's clothing, Men's Shoes, women's shoes, fresh, household electrical appliances.The corresponding advertisement pushing flow of each classification, including：Samples selection, spy It levies engineering, model training, model prediction and launches the processes such as reflux.Under the application scenarios, advertiser can correspond to specifically The businessman of classification, such as women's dress businessman A, women's dress businessman B, Men's Shoes businessman A, fresh businessman A and household electrical appliances businessman A.The present embodiment In, entire shopping website corresponds to a transfer learning model, which can integrate the relevance between every profession and trade, for example, female Dress businessman A and the influence between women's dress businessman B and other businessmans is associated with.When launching falling, it can will launch result and return It falls in knowledge base, knowledge base can optimize iteration to transfer learning model.

Further, made according to client to receiving the dispensing of advertisement as a result, constantly optimizing transfer learning model The target group of the dispensing advertisement of advertiser can be more suitable for output by obtaining, so as to help to promote the effect of advertiser's dispensing Fruit.Information push whole process in the present embodiment is the closed loop of an automation, while closed loop constantly operates, pushes effect Also it can become better and better.

Embodiment 13

As shown in figure 14, the structure diagram of information push-delivery apparatus provided for the embodiment of the present invention 13.The information Pusher 6 includes：First acquisition module 61, the second acquisition module 62, training module 63, correcting module 64, prediction module 65 With pushing module 66.

First acquisition module 61, for obtaining the input by user and relevant essential information of target object.

Second acquisition module 62, for obtaining positive sample crowd and negative sample crowd according to essential information.

Training module 63, for carrying out learning training to positive and negative sample crowd and negative sample crowd, to obtain the first prediction mould Type and learning training is carried out to revised positive sample crowd and negative sample crowd, to obtain the second prediction model.

Correcting module 64, for being repaiied using the transfer learning model of structure to positive sample crowd and negative sample crowd Just.

Prediction module 65, for using the first prediction model and the second prediction model potential crowd is predicted respectively with Obtain target group.

Pushing module 66, for pushing target object to target group.

Wherein, prediction module 65, the potential crowd is predicted specifically for being based on first prediction module with The first prediction crowd is obtained, the potential crowd is predicted based on second prediction model to obtain the second prediction people Group, for being modified using the second prediction crowd to the first prediction crowd, to obtain target group.

Prediction module 65, specifically for obtaining the first prediction probability and second of each user in the first prediction crowd The second prediction probability of each user, for each user, the first prediction probability and the second prediction probability are added in prediction crowd Weight average obtains the final prediction probability of the user, is ranked up according to the final prediction probability of all users, to be sorted Crowd chooses the target group from sequence crowd.

Further, training module 63, specifically for positive sample crowd and the extraction of negative sample crowd characteristic, to obtain the One characteristic information, the target industry being subordinate to according to essential information acquisition target object and the acquisition target industry are gone through History training data is adjusted the fisrt feature information according to historic training data, to obtain first object characteristic information, The type parameter of first object feature and each sample is inputted in machine learning model and is trained, to obtain the first prediction mould Type.

Further, training module 63, specifically for being extracted to revised positive sample crowd and negative sample crowd characteristic, To obtain second feature information, according to the target industry that essential information acquisition target object is subordinate to, going through for target industry is obtained History training data is adjusted second feature information according to historic training data, to obtain the second target signature information, by It is trained in the type parameter of two target signature informations and each sample input machine learning model, to obtain the second prediction mould Type.

Further, the second acquisition module 11, specifically for extracting part from the user of the whole network according to default rule User is as potential crowd..

Further, described information pusher 6 further includes optimization module 67, for pushing module 66 to target group After pushing target object, the dispensing of target object is counted as a result, excellent using launching result transfer learning model being iterated Change.

Information push-delivery apparatus provided in this embodiment gets the first prediction by the machine learning mode for having supervision first Crowd in order to make dispensing more accurate, based on the association between industry in transfer learning model, repaiies positive and negative sample population Just, it is then based on revised positive and negative sample and obtains the second prediction crowd, based on every in the first prediction crowd and the second prediction crowd It after the prediction probability weighted average of a user, is chosen from potential crowd and obtains the target group for launching advertisement so that prediction knot Fruit is more accurate, and pushes targeted advertisements to it.Due to prediction result, precisely then correspondingly target group is then that effect is launched in advertisement The best crowd of fruit, and then advertiser can be helped to improve the effect that advertisement is launched.

Further, made according to client to receiving the dispensing of advertisement as a result, constantly optimizing transfer learning model The target group of the dispensing advertisement of advertiser can be more suitable for output by obtaining, so as to help to promote the effect of advertiser's dispensing Fruit.

Embodiment 14

As shown in figure 15, a kind of information transmission system schematic diagram provided for the embodiment of the present invention 14.In fig.15, The information transmission system includes：Web modules and open data processing service (Open Data Processing Service, letter Claim ODPS), wherein Web modules include front end and Web server.Wherein, front end is used to provide operation interface for advertiser, such as The application program of one advertisement on front end can be provided, the product page is provided in the application program, in product circle On face advertiser can be registered, be logged in, the input of targeted advertisements essential information and the calculating etc. for launching cost.When wide After some requests of announcement primary input, inspection parameter is carried out to request from the background, after the Verification of advertisement primary input is legal, can be returned Successful reminder message.For example, advertiser can input logging request, the identity of user can be carried in logging request correspondingly The parameters such as information.

Further, after essential information of the advertiser in front end input targeted advertisements, essential information can be sent to Web server, application programming interface (the Application Programming set on the Web server Interface, abbreviation API), which connect with ODPS, passes through API Calls OPDS, the workflow of OPDS performance objective crowds The workflow of journey, wherein ODPS includes：Samples selection, Feature Engineering, model training and model prediction.

Samples selection is used for the essential information according to advertisement primary input, obtains positive and negative sample population, then Feature Engineering pair Positive and negative sample population carries out feature extraction, to obtain the characteristic information of sample, then by characteristic information and the type parameter of sample It is input in model training, the pre-set machine learning algorithm in the model training is getting sample and characteristic information After proceed by training, machine learning model training after the completion of, it is possible to based on trained machine learning model to potential Crowd is predicted that the result that will prestore returns to API, and API is showing advertiser by the product page, and advertiser is according to pre- It surveys as a result, obtaining target group to the end.

Further, after target group is got, targeted advertisements can be pushed to target object by advertiser, optional Ground, advertiser can calculate dispensing cost in the product page, for example, it may be determined that the number of target group and cost it Between relationship, then determine need to how many target groups launch targeted advertisements.

ODPS obtains the processing procedure four levels as shown in figure 16 of target group, including：Foundation characteristic layer, sample process Layer, characteristic processing layer and model treatment layer.

Wherein, foundation characteristic layer：Initially set up the data label that is possible under application scenarios, in the present embodiment, with Shopping website is specific application scenarios, and method for pushing provided in this embodiment is explained.In this example, it is positive and negative Sample population is made of client, and the feature of positive sample crowd and negative sample crowd include：The essential information of client, client Behavior, the condition of assets of client, the interest graph of client and abstract characteristics.Wherein, the essential information of client includes：Gender, Whether age, marital status have the information such as child, child's age.The behavior of client includes：Commodity are browsed, commodity are added in and are purchased Object vehicle, purchase commodity and the behaviors such as collecting commodities or search commercial articles relevant information.The condition of assets of client includes：Whether There is vehicle, whether there are the information such as room, mobile phone brand.The interest graph of client includes：The interested type of merchandise of client, client's sense are emerging The brand of interest, the interested classification of client.The interest graph of common customer can be according to the history number that can react Customer Shopping According to and the computation model that builds in advance be associated what analysis can obtain.Abstract characteristics refer to whether directly affect client The factor of received advertisement is clicked, if client is to the favorable rating of advertisement, some users completely will not point in real life Advertisement, the frequent degree of reaching the standard grade of client etc., such as the upper limit frequent degree by client, can predict reach the standard grade in next week it is general Rate.

Sample process layer includes：Positive sample correcting module, negative sample choose module and potential crowd chooses module.Its In, positive sample correcting module be used for after based on acquisition of information to potential positive sample crowd, according to default rule to it is potential just Sample population is modified, for example, association brand client chooses, shop loyalty client screening and association classification client are chosen. Negative sample chooses module can be sampled selection negative sample crowd according to preset number ratio, can also pass through spy's algorithm Or there are competitive relation brand to carry out selection negative sample crowd.It is latent according to preset strategy acquisition that potential crowd chooses module As potential crowd or client's circle is enlivened to the interested client of the classification in crowd, such as selection a period of time recently The mode of choosing, which obtains potential customers or the modes such as client are precisely oriented according to label, selects potential crowd, such as choose women Age, the client between 25~30 years old was as potential crowd.

Characteristic processing layer includes：Discretization and standardized module, characteristic crossover derive module and Feature Conversion module.Its In, discretization and standardized module carry out feature extraction to the positive and negative sample population that sample process layer is got, and obtain basic spy Sign.In the present embodiment, after being extracted to positive and negative sample population, first foundation feature is obtained, to revised positive negative sample people After group carries out feature extraction, the second foundation characteristic is obtained.Discretization is carried out respectively to first foundation feature and the second foundation characteristic And standardization, such as frequency division, evidence component (Weight of Evidence, abbreviation WOE) or equivalence may be used It draws grading mode and discretization and standardization is carried out to feature.Further, continue to passing through discretization and standardization First foundation feature afterwards and the second foundation characteristic derive module discrete features combination and time window using characteristic crossover Grading mode is cut, carrying out characteristic crossover to the first foundation feature after discretization and standardization and the second foundation characteristic spreads out Raw processing, obtains fisrt feature information and second feature information.After characteristic crossover derivative, Feature Conversion model is by fisrt feature Information and second feature information are from agreement buffering area (Protocol Buffer, abbreviation PB) format conversion into key-value pair (Key- Value, abbreviation KV) form.

Model treatment layer includes：Model training module, model prediction module and result reflux module.Wherein, model is instructed Practice module to be trained using the feature and classification of positive and negative sample crowd, obtain trained prediction model.Wherein model training The machines such as LR, SVM or iteration decision tree (Gradient Boosting Decision Tree, abbreviation GBDT) are used in module The algorithm of device study is trained sample, and a prediction model is got after the completion of training.In the present embodiment, model instruction After transfer learning is modified positive and negative sample population in white silk module, re -training is carried out to revised positive and negative sample population, Obtain a prediction model.Model prediction module is predicted based on this two prediction model in potential crowd, obtains potential crowd Prediction probability, since two prediction models have differences, correspondingly, the prediction that two prediction models predict potential crowd Probability is also different.Further model prediction module carries out the prediction probability of potential crowd score distribution visualization processing, In, prediction probability is converted to score, and if prediction probability is 0.5, then score corresponds to 50.

Further, it after being weighted averagely to potential crowd according to the prediction probability that two prediction models obtain, obtains The sequence of one potential crowd therefrom selects part population as target group.As shown in figure 17, it is from potential crowd Screen the schematic diagram of target group.Wherein, dark parts are the target group that advertiser filters out from potential crowd as throwing Put the crowd of advertisement.

In the present embodiment, model prediction module can also be to offline clicking rate (Click-Through-Rate, abbreviation CTR) Index is estimated.For example, can sequence be shown targeted advertisements according to CTR.Or CTR and ad click can be passed through Price (Charge per Click, abbreviation CPC), which is estimated advertising income or filtered out, launches the bad advertisement keyword of result Word obtains the correlation of advertisement and the receptance of recommendation etc., optimizes advertisement delivery effect according to above- mentioned information.

After advertiser launches advertisement to target group, further, the module that as a result flows back can get target group To the network behavior of targeted advertisements, using network behavior as dispensing as a result, being then refluxed in the knowledge base of structure, then utilize Dispensing result in knowledge base optimizes iteration to transfer learning model, in order to get the target person for launching best results Group improves and launches effect.

Information-pushing method provided in this embodiment gets the first prediction people by the machine learning mode for having supervision Group, in order to make dispensing more accurate, based on the association between industry in transfer learning model, repaiies positive and negative sample population Just, it is then based on revised positive and negative sample and obtains the second prediction crowd, based on the first prediction crowd and second prediction crowd's weighting Afterwards, obtain launching the target group of advertisement so that prediction result is more accurate, and pushes targeted advertisements to it.Since prediction is tied Precisely then correspondingly target group is then the best crowd of advertisement delivery effect to fruit, and then advertiser can be helped to improve advertisement and thrown The effect put.

One of ordinary skill in the art will appreciate that：Realizing all or part of step of above-mentioned each method embodiment can lead to The relevant hardware of program instruction is crossed to complete.Aforementioned program can be stored in a computer read/write memory medium.The journey Sequence when being executed, performs the step of including above-mentioned each method embodiment；And aforementioned storage medium includes：ROM, RAM, magnetic disc or The various media that can store program code such as person's CD.

Finally it should be noted that：The above embodiments are only used to illustrate the technical solution of the present invention., rather than its limitations；To the greatest extent Pipe is described in detail the present invention with reference to foregoing embodiments, it will be understood by those of ordinary skill in the art that：Its according to Can so modify to the technical solution recorded in foregoing embodiments either to which part or all technical features into Row equivalent replacement；And these modifications or replacement, various embodiments of the present invention technology that it does not separate the essence of the corresponding technical solution The range of scheme.

Claims

1. a kind of information transmission system, which is characterized in that including：Front end and data processing platform (DPP)；

The front end input by user be sent to the data processing with the relevant essential information of target object and put down for receiving Platform；

The data processing platform (DPP), for obtaining positive sample crowd and negative sample crowd according to the essential information, to it is described just Negative sample crowd and the negative sample crowd carry out learning training, to obtain the first prediction model, and utilize the transfer learning of structure Model is modified the positive sample crowd and the negative sample crowd, to the revised positive sample crowd and described negative Sample population carries out learning training, to obtain the second prediction model and utilize first prediction model and described second in advance It surveys model respectively to predict to obtain target group the potential crowd, the target pair is pushed to the target group As.

2. system according to claim 1, which is characterized in that the data processing platform (DPP), including：Samples selection node, Sample training node, prediction node and push node；

The samples selection node for receiving the essential information that the front end is sent, is selected using the essential information The positive sample crowd and the negative sample crowd are sent to the sample by the positive sample crowd and the negative sample crowd Training node；

The sample training node, for carrying out learning training to the positive sample crowd and the negative sample crowd, to obtain First prediction model and positive sample crowd and negative sample crowd are modified based on transfer learning model, to revised The positive sample crowd and the negative sample crowd carry out learning training, to obtain the second prediction model；

The prediction node, for being carried out respectively to potential crowd using first prediction model and second prediction model Prediction, to obtain the target group, the push node is sent to by the target group；

The push node, for pushing the target object to the target group.

3. system according to claim 2, which is characterized in that the front end includes：

Line module and the interactive module for connecting the line module and the data processing platform (DPP)；

Wherein, the line module, for receiving the essential information input by user；The interactive module, for by user The samples selection node is issued in the essential information input by user.

4. system according to claim 3, which is characterized in that the sample training node is specifically used for：

Feature extraction is carried out to the positive sample crowd and the negative sample crowd, to obtain the fisrt feature information of sample；

The type parameter of the fisrt feature information and each sample is input in preset machine learning model and is trained, To obtain first prediction model；

The revised positive sample crowd and the negative sample crowd characteristic are extracted, to obtain the second feature of sample letter Breath；

The type parameter of the second feature information and each sample is inputted in the machine learning model and is trained, with To second prediction model.

5. system according to claim 3, which is characterized in that the prediction node is specifically used for：

The potential crowd is predicted using first prediction model, obtains the first prediction crowd and first prediction The first prediction probability of each user in crowd；

The potential crowd is predicted using second prediction model, obtains the second prediction crowd and second prediction The second prediction probability of each user in crowd；

For each user by first prediction probability and the second prediction probability weighted average, the final of the user is obtained Prediction probability；

All users are ranked up according to the final prediction probability to obtain sequence crowd；

The target group is chosen from the sequence crowd.

6. a kind of interactive device, which is characterized in that including：

The essential information for receiving the input by user and relevant essential information of target object, is sent to data by front end Processing platform and the target group that the data processing platform (DPP) is obtained according to the essential information is received, and to the target Crowd pushes the target object.

7. interactive device according to claim 6, which is characterized in that the front end includes：

The line module, for receiving the essential information input by user and receiving the data platform according to The target group that essential information obtains, and push the target object to the target group；

The interactive module, for the essential information to be issued the data processing platform (DPP), so that the data processing platform (DPP) The target group is obtained according to the essential information.

8. a kind of information push-delivery apparatus, which is characterized in that including：

Data processing platform (DPP), for obtaining positive sample crowd and negative sample crowd according to essential information relevant with target, to institute It states positive and negative sample crowd and the negative sample crowd carries out learning training, to obtain the first prediction model, and utilize the migration of structure Learning model is modified the positive sample crowd and the negative sample crowd, to the revised positive sample crowd and institute It states negative sample crowd and carries out learning training, to obtain the second prediction model and utilize first prediction model and described the Two prediction models are respectively predicted to obtain target group the potential crowd, and the target is pushed to the target group Object.

9. device according to claim 8, which is characterized in that the data processing platform (DPP), including：Samples selection node, Sample training node, prediction node and push node；

For receiving the essential information, the positive sample crowd is selected using the essential information for the samples selection node With the negative sample crowd, the positive sample crowd and the negative sample crowd are sent to the sample training node；

The push node, for pushing the target object to the target group.

10. a kind of model training method, which is characterized in that including：

The type parameter of the fisrt feature information and each sample is input in preset machine learning model and is trained, To obtain the first prediction model；

The type parameter of the second feature information and each sample is inputted in the machine learning model and is trained, with To the second prediction model.

It is 11. according to the method described in claim 10, it is characterized in that, described by the fisrt feature information and each sample Type parameter is input in preset machine learning model and is trained, to obtain the first prediction model, including：

The target industry being subordinate to according to the target object input by user relevant essential information acquisition target object；

Obtain the historic training data of the target industry；

The fisrt feature information is adjusted according to the historic training data, to obtain first object characteristic information；

The type parameter of the first object characteristic information and each sample is inputted in machine learning model and is trained, with To first prediction model；

Described input the type parameter of the second feature information and each sample in the machine learning model is trained, To obtain the second prediction model, including：

The target industry being subordinate to according to the essential information acquisition target object；

Obtain the historic training data of the target industry；

The second feature information is adjusted according to the historic training data, to obtain the second target signature information；

The type parameter of second target signature information and each sample is inputted in machine learning model and is trained, with To second prediction model.

12. according to the method for claim 11, which is characterized in that described that spy is carried out to positive sample crowd and negative sample crowd Sign extraction, to obtain the fisrt feature information of sample, including：

Feature extraction is carried out to the positive sample crowd and the negative sample crowd, obtains first foundation feature；

The first foundation feature after extraction is subjected to discretization and standardization；

The first foundation feature after discretization and standardization is subjected to characteristic crossover derivation process, is obtained described Fisrt feature information；

It is described that the revised positive sample crowd and the negative sample crowd characteristic are extracted, to obtain the second feature of sample Information, including：

To carrying out feature extraction to the revised positive sample crowd and the negative sample crowd, the second foundation characteristic is obtained；

Second foundation characteristic after extraction is subjected to discretization and standardization；

Second foundation characteristic after discretization and standardization is subjected to characteristic crossover derivation process, is obtained described Second feature information.

13. according to the method for claim 12, which is characterized in that further include：

Key assignments plaid matching is converted into from agreement buffer format to obtained the fisrt feature information and the second feature information Formula.

14. according to claim 10-13 any one of them methods, which is characterized in that described to positive sample crowd and negative sample Crowd carries out feature extraction, with before obtaining the fisrt feature information of sample, including：

The positive sample crowd and the negative sample crowd are obtained according to the essential information input by user.

15. a kind of model training apparatus, which is characterized in that including：

Characteristic extracting module, for carrying out feature extraction to positive sample crowd and negative sample crowd, to obtain the first of sample the spy Reference ceases and the revised positive sample crowd and the negative sample crowd characteristic is extracted, to obtain the second of sample Characteristic information；

Training module, for the type parameter of the fisrt feature information and each sample to be input to preset machine learning mould Be trained in type, with obtain the first prediction model and by the type parameter of the second feature information and each sample it is defeated Enter and be trained in the machine learning model, to obtain the second prediction model.

16. a kind of target group's acquisition methods, which is characterized in that including：

Learning training is carried out to the revised positive sample crowd and the negative sample crowd, to obtain the second prediction model；

Respectively the potential crowd is predicted to obtain mesh using first prediction model and second prediction model Mark crowd.

17. according to the method for claim 16, which is characterized in that described to utilize first prediction model and described second Prediction model respectively predicted to obtain target group the potential crowd, including：

The potential crowd is predicted based on first prediction module to obtain the first prediction crowd；

The potential crowd is predicted based on second prediction model to obtain the second prediction crowd；

The first prediction crowd is modified using the second prediction crowd, to obtain the target group.

18. according to the method for claim 17, which is characterized in that described to predict crowd to described first using described second Prediction crowd is modified, to obtain the target group, including：

Obtain the first prediction probability of each user in the first prediction crowd；

Obtain the second prediction probability of each user in the second prediction crowd；

For each user, by first prediction probability and the second prediction probability weighted average, the user is obtained most Whole prediction probability；

All users are ranked up according to the final prediction probability, to obtain sequence crowd；

The target group is chosen from the sequence crowd.

19. a kind of target group's acquisition device, which is characterized in that including：

Acquisition module, for obtaining positive sample crowd and negative sample with the relevant essential information of target object according to input by user Crowd；

Training module, for carrying out learning training to the positive sample crowd and the negative sample crowd, to obtain the first prediction Model and learning training is carried out to the revised positive sample crowd and the negative sample crowd, to obtain the second prediction Model；

Correcting module, for being repaiied using the transfer learning model of structure to the positive sample crowd and the negative sample crowd Just；

Prediction module respectively predicts the potential crowd using first prediction model and second prediction model To obtain target group.

20. a kind of information-pushing method, which is characterized in that including：

The target object is pushed to the target group.

21. according to the method for claim 20, which is characterized in that described to be based on first prediction model to potential crowd It is predicted to obtain target group, including：

The potential crowd is predicted using first prediction model, to obtain the first prediction crowd；

The first prediction crowd is modified to obtain the target group.

22. according to the method for claim 21, which is characterized in that described to be based on first prediction model to potential crowd It is predicted to obtain target group, including：

The prediction probability of each user in the potential crowd is obtained using first prediction model；

The prediction probability of each user is ranked up；

Selected part user is as the target group from the potential crowd of sequence.

23. a kind of information push-delivery apparatus, which is characterized in that including：

Sample acquisition module, for obtaining positive sample crowd according to the relevant essential information of input by user and target object and bearing Sample population；

Training module, for carrying out learning training to the positive sample crowd and the negative sample crowd, to obtain the first prediction Model；

Pushing module, for pushing the target object to the target group.

24. a kind of information-pushing method, which is characterized in that including：

Obtain the input by user and relevant essential information of target object；

Respectively the potential crowd is predicted to obtain mesh using first prediction model and second prediction model Mark crowd；

The target object is pushed to the target group.

25. according to the method for claim 24, which is characterized in that described to utilize first prediction model and described second Prediction model is respectively predicted to obtain target group the potential crowd, and the target pair is pushed to the target group As, including：

26. according to the method for claim 25, which is characterized in that described to predict crowd to described first using described second Prediction crowd is modified, to obtain the target group, including：

The target group is chosen from the sequence crowd.

27. according to the method for claim 24, which is characterized in that described to the positive sample crowd and the negative sample people Group carries out learning training, to obtain the first prediction model, including：

The positive sample crowd and the negative sample crowd characteristic are extracted, to obtain the fisrt feature information of sample；

Obtain the historic training data of the target industry；

It is described that learning training is carried out to the revised positive sample crowd and the negative sample crowd, to obtain the second prediction mould Type, including：

Obtain the historic training data of the target industry；

28. according to claim 24-27 any one of them methods, which is characterized in that further include：

Certain customers are extracted as the potential crowd from the user of the whole network according to default rule.

29. according to the method for claim 28, which is characterized in that the target group push the target object it Afterwards, it further includes：

Count the dispensing result of the target object；

Optimization is iterated to the transfer learning model using the dispensing result.

30. a kind of information push-delivery apparatus, which is characterized in that including：

Training module, for carrying out learning training to the positive and negative sample crowd and the negative sample crowd, to obtain the first prediction Model and learning training is carried out to the revised positive sample crowd and the negative sample crowd, to obtain the second prediction Model；

Prediction module, for being carried out respectively to the potential crowd using first prediction model and second prediction model It predicts to obtain target group；

Pushing module, for pushing the target object to the target group.