CN109801100A

CN109801100A - Advertisement placement method, device and computer readable storage medium

Info

Publication number: CN109801100A
Application number: CN201811604720.3A
Authority: CN
Inventors: 闫铭; 赵乙; 徐恪; 孔东营
Original assignee: Beijing Dajia Internet Information Technology Co Ltd
Current assignee: Beijing Dajia Internet Information Technology Co Ltd
Priority date: 2018-12-26
Filing date: 2018-12-26
Publication date: 2019-05-24

Abstract

The application is about a kind of advertisement placement method, device and computer storage medium.The advertisement placement method includes: the advertisement playing request for obtaining target user and initiating；The similarity being respectively compared between the attribute vector of the target user and the attribute vector of multiple historical users obtains multiple first historical users most like with the target user；History advertisement based on the multiple first historical user browses record, establishes about the ambient condition vector for launching targeted advertisements for the target user；Based on the ambient condition vector sum deep neural network model, obtain being worth corresponding to the dispensing of each targeted advertisements；And based on the corresponding dispensing value of each described targeted advertisements, corresponding dispensing is worth the highest targeted advertisements and is delivered to the target user.The advertisement placement method improves the accuracy and harmony of targeted advertisements dispensing.

Description

Advertisement placement method, device and computer readable storage medium

Technical field

The application belongs to computer software application field, especially advertisement placement method and device.

Background technique

Advertiser launches advertisement in advertising platform, and advertiser wishes that the advertisement of oneself is clicked or converts, advertising platform Wish to launch region, release time by adjusting advertisement and launch target group etc. with improve be launched advertisement clicking rate or Conversion ratio.During launching advertisement, advertiser can constantly extend new advertiser, and then constantly introduce new advertisement (namely new initiative advertisement).

Currently, launching field in new initiative advertisement, the most common drop mechanisms are random drop mechanisms.Specifically, right In the request for obtaining advertisement and playing that a user sends, advertising platform can randomly select centainly in numerous new initiative advertisements The advertisement of quantity is launched.Random drop mechanisms inevitably have ignored the attribute of advertisement itself, cause to launch accuracy It is difficult to ensure.The history impression information of new initiative advertisement is insufficient, and the discreet values such as the clicking rate estimated, the conversion ratio estimated are inaccurate Really, the accuracy that further reduced advertisement dispensing increases the lack of uniformity of advertisement dispensing.

Summary of the invention

To overcome the problems, such as that the dispensing of advertisement present in the relevant technologies is inaccurate and unbalanced, the application discloses a kind of advertisement Put-on method and device, the similitude judgement of the attribute vector based on user, obtain multiple first most like with target user Target user.It is insufficient to make up targeted advertisements historical information using the history advertisement browsing record of multiple first object user Problem is to improve the accuracy and harmony that targeted advertisements are launched.

According to the embodiment of the present application in a first aspect, providing a kind of advertisement placement method, comprising: include:

Obtain the advertisement playing request that target user initiates；

The similarity being respectively compared between the attribute vector of the target user and the attribute vector of multiple historical users, obtains To multiple first historical users most like with the target user；

History advertisement based on the multiple first historical user browses record, establishes and launches about for the target user The ambient condition vector of targeted advertisements；

Based on the ambient condition vector sum deep neural network model, the dispensing corresponding to each targeted advertisements is obtained Value；And

Based on the corresponding dispensing value of each described targeted advertisements, it is wide that corresponding dispensing is worth the highest target Announcement is delivered to the target user.

Optionally, the advertisement placement method, further includes:

It establishes about the deep neural network model for launching the targeted advertisements.

It is optionally, described to establish about before the deep neural network model for launching the targeted advertisements, comprising:

Word is carried out to the user property of each user and is embedded in squeeze operation, obtains the attribute of each user Vector.

Optionally, between the attribute vector for being respectively compared the target user and the attribute vector of multiple historical users Similarity, obtain multiple first historical users most like with the target user, comprising:

Calculate separately the Euclidean distance between the attribute vector of the target user and the attribute vector of multiple historical users；

Compare the corresponding Euclidean distance of each historical user and preset threshold, obtains having to the target user similar Multiple second historical users of property；

It sorts to the multiple second historical user according to the sequence of similitude from high to low, chooses the multiple of front Second historical user obtains the multiple first historical user；

Wherein, the multiple first historical user arranges according to the sequence of similitude from high to low.

Optionally, between the attribute vector for being respectively compared the target user and the attribute vector of multiple historical users Similarity, obtain multiple first historical users most like with the target user, further includes:

If the corresponding Euclidean distance of described each historical user is less than the preset threshold, each described history User and the target user have similitude；

If the corresponding Euclidean distance of described each historical user is more than or equal to the preset threshold, it is described each Historical user and the target user do not have similitude.

Optionally, the history advertisement based on the multiple first historical user browses record, establishes described in being The ambient condition vector of target user's dispensing targeted advertisements, comprising:

All targeted advertisements are filtered out from all advertisements of advertising platform；

Multiple targeted advertisements are selected at random in all targeted advertisements, as multiple first object advertisements；

The impressions for counting the multiple first object advertisement respectively obtain the throwing of the multiple first object advertisement Put overall condition vector；

By the history advertisement browsing record combination of the multiple first historical user and the target user, the mesh is obtained Mark the extension advertisement browsing record vector of user；And

The dispensing overall condition vector of the multiple first object advertisement, the extension of the target user is wide The attribute vector for accusing target user described in browsing record vector sum merges, and obtains the first object advertisement being delivered to institute State the ambient condition vector of target user.

Optionally, described to be based on the ambient condition vector sum deep neural network model, it obtains corresponding to each mesh Mark the dispensing value of advertisement, comprising:

The ambient condition vector is inputted into the deep neural network model, obtains each described first object advertisement Corresponding dispensing value.

Optionally, described based on the corresponding dispensing value of each described targeted advertisements, corresponding dispensing is worth highest The targeted advertisements be delivered to the target user, comprising:

Value of launching corresponding to the multiple first object advertisement sorts according to sequence from high to low, before coming most The first object advertisement in face is delivered to the target user as the second targeted advertisements, by second targeted advertisements.

It is optionally, described to establish about the deep neural network model for launching the targeted advertisements, comprising:

Establish deep neural network object module；

Training on line is carried out to the deep neural network object module；And

Based on backpropagation principle, the network parameter of the deep neural network object module is optimized.

It is optionally, described that training on line is carried out to the deep neural network object module, comprising:

Obtain the advertisement playing request that sample of users is initiated；

The similarity being respectively compared between the attribute vector of the sample of users and the attribute vector of multiple historical users, obtains To the multiple third historical users most like with the sample of users；

History advertisement based on the multiple third historical user browses record, establishes and launches about for the sample of users The ambient condition vector of sample advertisement；

Based on deep neural network object module described in the ambient condition vector sum, obtain wide corresponding to each sample The dispensing of announcement is worth；And

Based on the corresponding dispensing value of each described sample advertisement, it is wide that corresponding dispensing is worth the highest sample Announcement is delivered to the sample of users.

Optionally, between the attribute vector for being respectively compared the sample of users and the attribute vector of multiple historical users Similarity, obtain the multiple third historical users most like with the sample of users, comprising:

Calculate separately the Euclidean distance between the attribute vector of the sample of users and the attribute vector of multiple historical users；

Compare the corresponding Euclidean distance of each historical user and preset threshold, obtains having to the sample of users similar Multiple 4th historical users of property；

It sorts to the multiple 4th historical user according to the sequence of similitude from high to low, chooses the multiple of front 4th historical user obtains the multiple third historical user；

Wherein, the multiple third historical user arranges according to the sequence of similitude from high to low.

Optionally, between the attribute vector for being respectively compared the sample of users and the attribute vector of multiple historical users Similarity, obtain the multiple third historical users most like with the sample of users, further includes:

If the corresponding Euclidean distance of described each historical user is less than the preset threshold, each described history User and the sample of users have similitude；

If the corresponding Euclidean distance of described each historical user is more than or equal to the preset threshold, it is described each Historical user and the sample of users do not have similitude.

Optionally, the history advertisement based on the multiple third historical user browses record, establishes described in being The ambient condition vector of sample of users dispensing sample advertisement, comprising:

All sample advertisements are filtered out from all advertisements of advertising platform；

Multiple sample advertisements are selected at random in all sample advertisements, as multiple first sample advertisements；

The impressions for counting the multiple first sample advertisement respectively obtain the throwing of the multiple first sample advertisement Put overall condition vector；

By the history advertisement browsing record combination of the multiple third historical user and the sample of users, the sample is obtained The extension advertisement browsing record vector of this user；And

The dispensing overall condition vector of the multiple first sample advertisement, the extension of the sample of users is wide The attribute vector for accusing sample of users described in browsing record vector sum merges, and obtains the first sample advertisement being delivered to institute State the ambient condition vector of sample of users.

Optionally, described based on deep neural network object module described in the ambient condition vector sum, corresponded to The dispensing of each sample advertisement is worth, comprising:

The ambient condition vector is inputted into the deep neural network object module, obtains each described first sample The corresponding dispensing value of advertisement.

Optionally, described based on the corresponding dispensing value of each described sample advertisement, corresponding dispensing is worth highest The sample advertisement be delivered to the sample of users, comprising:

Value of launching corresponding to the multiple first sample advertisement sorts according to sequence from high to low, before coming most The first sample advertisement in face is delivered to the sample of users as the second sample advertisement, by the second sample advertisement.

It is optionally, described to establish about the deep neural network model for launching the targeted advertisements, further includes:

Based on the sample of users to the interbehavior for the sample advertisement being launched, the sample being launched is calculated Advertisement motivates the environmental feedback of the deep neural network object module；

It is motivated based on the environmental feedback, adjusts the network parameter of the deep neural network object module.

Optionally, it is described based on the sample of users to the interbehavior for the sample advertisement being launched, calculating is thrown The sample advertisement put motivates the environmental feedback of the deep neural network object module, comprising:

The second sample advertisement whether is clicked according to the sample of users, gives the deep neural network target mould Type clicks excitation；

The second sample advertisement is calculated separately to be delivered to before the sample of users and after being delivered to the sample of users The variance of the impressions of the multiple first sample advertisement；

Described in front of being delivered to the sample of users according to the second sample advertisement and being delivered to after the sample of users The deep neural network object module harmony excitation is given in the variation of variance；And

Based on click excitation and the harmonious excitation, the second sample advertisement being launched is to the depth Spend the environmental feedback excitation of neural network object module.

Optionally, if the second sample advertisement is delivered to the variance before the sample of users and is less than described second Sample advertisement is delivered to the variance after the sample of users, then gives the negative harmony of the deep neural network object module and swash It encourages；

If the second sample advertisement is delivered to the variance before the sample of users, to be greater than second sample wide Announcement is delivered to the variance after the sample of users, then gives the deep neural network object module positive harmonious excitation.

Optionally, the advertisement placement method, further includes:

Judge whether the deep neural network object module restrains；

If the deep neural network object module convergence, obtains the deep neural network model；

If the deep neural network object module is not restrained, new advertisement broadcasting is initiated when new sample of users and is asked When asking, the click excitation and the harmonious excitation are recalculated, the network of the deep neural network object module is optimized Parameter.

According to a second aspect of the embodiments of the present invention, a kind of advertisement delivery device is provided, comprising:

Data capture unit is configured as obtaining the advertisement playing request that target user initiates；

Comparing unit, be configured to target user described in comparison attribute vector and multiple historical users attribute to Similarity between amount obtains multiple first historical users most like with the target user；

Vector establishes unit, is configured as the history advertisement browsing record based on the multiple first historical user, establishes About the ambient condition vector for launching targeted advertisements for the target user；

Computing unit is configured as obtaining corresponding to every based on the ambient condition vector sum deep neural network model The dispensing value of one targeted advertisements；And

Unit is launched in advertisement, is configured as based on the corresponding dispensing value of each described targeted advertisements, by corresponding throwing It puts the highest targeted advertisements of value and is delivered to the target user.

Optionally, the advertisement delivery device, further includes:

Model foundation unit is configured as establishing the deep neural network model about the targeted advertisements are launched.

Embedded unit is configured as carrying out the user property of each user word insertion squeeze operation, obtain described The attribute vector of each user.

Establish deep neural network object module；

Training on line is carried out to the deep neural network object module；And

Obtain the advertisement playing request that sample of users is initiated；

Optionally, the advertisement delivery device, further includes:

Judge whether the deep neural network object module restrains；

If the deep neural network object module is not restrained, new advertisement broadcasting is initiated when new sample of users and is asked When asking, the environmental feedback excitation is recalculated, the network parameter of the deep neural network object module is optimized.

According to a third aspect of the embodiments of the present invention, a kind of server is provided, comprising:

Processor；

Memory for storage processor executable instruction；

Wherein, the processor is configured to executing advertisement placement method described in above-mentioned any one.

According to a fourth aspect of the embodiments of the present invention, a kind of computer readable storage medium is provided, which is characterized in that described Computer-readable recording medium storage has computer instruction, and the computer instruction, which is performed, realizes above-mentioned advertisement dispensing side Method.

According to a fifth aspect of the embodiments of the present invention, a kind of computer program product comprising instruction is provided, when it is being counted When being run on calculation machine, so that computer realizes method and step described in first aspect.

The technical solution that embodiments herein provides can include the following benefits:

From all historical users that there is similitude with the target user, select and most like specific of the target user The historical user of quantity, as certain amount of first historical user.By the dispensing of the certain amount of first object advertisement The attribute vector that the extension advertisement of overall condition vector, the target user browses the record vector sum target user merges, Obtain the ambient condition vector that the first object advertisement is delivered to the target user.It is used using certain amount of first history The history advertisement browsing record at family is to make up the problem of targeted advertisements historical information deficiency, to improve targeted advertisements dispensing Accuracy.

The second sample advertisement whether is clicked according to the sample of users, gives deep neural network object module click Excitation.According to the second sample advertisement be delivered to before the sample of users and be delivered to after the sample of users this certain amount of the Deep neural network object module harmony excitation is given in the variation of the variance of the impressions of one sample advertisement.According to Certain ratio motivates the click and the harmony incentive combination, obtains environmental feedback excitation.It is motivated using the environmental feedback To optimize the network parameter of the deep neural network object module.It is motivated based on the environmental feedback to the deep neural network target The dispensing environmental feedback of model adjusts the network parameter of the deep neural network object module, and dynamic adjusts advertisement serving policy, Improve the accuracy and harmony of advertisement dispensing.

It should be understood that above general description and following detailed description be only it is exemplary and explanatory, not The application can be limited.

Detailed description of the invention

Fig. 1 is the flow chart of advertisement placement method shown according to an exemplary embodiment.

Fig. 2 is the flow chart of advertisement placement method shown according to an exemplary embodiment.

Fig. 3 is the flow chart of advertisement placement method shown according to an exemplary embodiment.

Fig. 4 is the schematic diagram of advertisement delivery device shown according to an exemplary embodiment.

Fig. 5 is the schematic diagram of advertisement delivery device shown according to an exemplary embodiment.

Fig. 6 is a kind of block diagram of device for executing advertisement placement method shown according to an exemplary embodiment.

Fig. 7 is a kind of block diagram of device for executing advertisement placement method shown according to an exemplary embodiment.

Specific embodiment

Example embodiments are described in detail here, and the example is illustrated in the accompanying drawings.Following description is related to When attached drawing, unless otherwise indicated, the same numbers in different drawings indicate the same or similar elements.Following exemplary embodiment Described in embodiment do not represent all embodiments consistent with the application.On the contrary, they be only with it is such as appended The example of the consistent device and method of some aspects be described in detail in claims, the application.

Fig. 1 is the flow chart of advertisement placement method shown according to an exemplary embodiment, specifically includes the following steps:

In step s101, the advertisement playing request that target user initiates is obtained.

In this step, the platforms such as cell phone software or webpage receive the advertisement playing request that target user initiates.This is wide Accuse the trigger action of playing request, the e.g. access request to the cell phone software or webpage.The target user requests to play Advertisement include display advertising, video ads and interactive FLASH advertisement.

In step s 102, be respectively compared the target user attribute vector and multiple historical users attribute vector it Between similarity, obtain multiple first historical users most like with the target user.

The user property for registering or logging in the user of the platforms such as cell phone software, webpage includes: gender, age, area, mobile phone Model, network formats, concern list and bean vermicelli list；The advertisement behavior of user includes: browse advertisements, clicks advertisement, by advertisement Labeled as like and by it is AD tagged be disagreeable.User substantially similar for user property, the usual user are interested wide The classification of announcement is with uniformity, for example, young man user is generally interested in game series advertisements, if the hand that the user uses Machine is the content in the information such as a game machine or/and the concern list of the user about game, then judges that the user may It is interested in game series advertisements；The women of marriageable age is generally interested in wedding gauze kerchief series advertisements, if the concern list of the user Content in the information such as bean vermicelli list about wedding gauze kerchief, then judge that the user may be interested in wedding gauze kerchief series advertisements.So using The user property at family can be used as the foundation for determining to launch the optimal user group of advertisement.

In this step, it calculates separately between the attribute vector of the target user and the attribute vector of multiple historical users Distance, such as Euclidean distance.Compare the corresponding Euclidean distance of each historical user and preset threshold, obtains and the target user Multiple second historical users with similitude.If the corresponding Euclidean distance of each historical user is less than the preset threshold, Then each historical user and the target user have similitude.If the corresponding Euclidean distance of each historical user be greater than etc. In the preset threshold, then each historical user and the target user do not have similitude.

It sorts to multiple second historical user according to the sequence of similitude from high to low, chooses the certain number of front Second historical user of amount, obtains certain amount of first historical user.Wherein, certain amount of first historical user according to The sequence arrangement of similitude from high to low.

In step s 103, history advertisement based on the multiple first historical user browses record, establishes about for institute State the ambient condition vector that target user launches targeted advertisements.

New advertisement can be continually introduced during advertising platform operation.According to the request of target user, advertising platform is determined It is fixed that targeted advertisements are delivered to target user.For example, the targeted advertisements are the advertisements that spending on ads is less than certain amount.

In this step, all targeted advertisements are filtered out from all advertisements of advertising platform.It is wide in all targets Certain amount of targeted advertisements are selected in announcement at random, as certain amount of first object advertisement.The specific quantity is counted respectively First object advertisement impressions, obtain the dispensing overall condition vector of the certain amount of first object advertisement.

By the history advertisement browsing record combination of certain amount of first historical user and the target user, the mesh is obtained Mark the extension advertisement browsing record vector of user.

By the extension advertisement of the dispensing overall condition vector, the target user of the certain amount of first object advertisement The attribute vector of the browsing record vector sum target user merges, and obtains the first object advertisement being delivered to the target user Ambient condition vector.

In step S104, it is based on the ambient condition vector sum deep neural network model, obtains corresponding to each The dispensing of targeted advertisements is worth.

In the step, ambient condition vector input is used to the deep neural network mould of Q-Learning learning framework Type obtains the corresponding dispensing value of the corresponding Q value of each first object advertisement i.e. each first object advertisement.

In step s105, based on the corresponding dispensing value of each described targeted advertisements, most by corresponding dispensing value The high targeted advertisements are delivered to the target user.

In this step, to the certain amount of first object advertisement corresponding dispensing value according to sequence from high to low Second targeted advertisements using the first object advertisement for coming foremost as the second targeted advertisements, are delivered to the target by sequence User.

According to an embodiment of the present application, it from all historical users that there is similitude with the target user, selects and is somebody's turn to do The most like certain amount of historical user of target user, as certain amount of first historical user.This is certain amount of The dispensing overall condition vector of first object advertisement, extension advertisement browsing record vector sum target of the target user are used The attribute vector at family merges, and obtains the ambient condition vector that the first object advertisement is delivered to the target user.Utilize this The history advertisement browsing of certain amount of first historical user records to make up the problem of targeted advertisements historical information deficiency, thus Improve the accuracy of targeted advertisements dispensing.

Fig. 2 is the flow chart of advertisement placement method shown according to an exemplary embodiment, is more than previous embodiment Perfect embodiment.Specifically includes the following steps:

In step s 201, word is carried out to the user property of each user and is embedded in squeeze operation, obtained described each The attribute vector of user.

The user property of each user includes: gender, age, area, mobile phone model, network formats, concern list and powder Silk list.The user property of each user includes a large amount of discrete sparse attribute.

In this step, embedding operation is carried out to the discrete sparse attribute of the user property of each user, by each use Low-dimensional vector in the user property boil down to continuous space of the higher-dimension at family, obtains the attribute vector of each user.

In step S202, establish about the deep neural network model for launching the targeted advertisements.

In this step, it establishes about the deep neural network model for launching the targeted advertisements.

In step S203, the advertisement playing request that target user initiates is obtained.

In step S204, be respectively compared the target user attribute vector and multiple historical users attribute vector it Between similarity, obtain multiple first historical users most like with the target user.

In step S205, history advertisement based on the multiple first historical user browses record, establishes about for institute State the ambient condition vector that target user launches targeted advertisements.

In step S206, it is based on the ambient condition vector sum deep neural network model, obtains corresponding to each The dispensing of targeted advertisements is worth.

In step S207, based on the corresponding dispensing value of each described targeted advertisements, most by corresponding dispensing value The high targeted advertisements are delivered to the target user.

Step S203 to step S207 is consistent to step S105 with the step S101 in Fig. 1, just repeats no more here.

According to an embodiment of the present application, embedding operation is carried out to the discrete sparse attribute of the user property of each user, By the low-dimensional vector in the user property boil down to continuous space of the higher-dimension of each user, obtain the attribute of each user to Amount.History advertisement based on most similar certain amount of first historical user of the attribute vector with target user browses note The attribute vector of the dispensing overall condition vector sum target user of record, the certain amount of first object advertisement is established and is closed In the ambient condition vector for launching targeted advertisements for the target user.Excavate most similar spy of attribute vector with target user Relationship between the history advertisement browsing record and the attribute vector of target user of first historical user of fixed number amount, further mentions The accuracy that high targeted advertisements are launched.

Fig. 3 is the flow chart of advertisement placement method shown according to an exemplary embodiment, is established and closes in step S202 In the detailed process for the deep neural network model for launching the targeted advertisements.Specifically includes the following steps:

In step S301, deep neural network object module is established.

In this step, it is based on deep neural network algorithm, establishes deep neural network object module.

In step s 302, training on line is carried out to the deep neural network object module.

In this step, the advertisement playing request that sample of users is initiated is obtained.Be respectively compared the attribute of the sample of users to Similarity between amount and the attribute vector of multiple historical users obtains using with the most like multiple third history of the sample of users Family.For example, calculating separately the Euclidean distance between the attribute vector of the sample of users and the attribute vector of multiple historical users.Than Euclidean distance corresponding compared with each historical user and preset threshold obtain have similitude with the sample of users multiple four Historical user.If the corresponding Euclidean distance of each historical user be less than the preset threshold, each historical user and should Sample of users has similitude.If the corresponding Euclidean distance of each historical user is more than or equal to the preset threshold, each A historical user and the sample of users do not have similitude.Sequence to multiple 4th historical user according to similitude from high to low Sequence, chooses certain amount of 4th historical user of front, obtains certain amount of third historical user.Wherein, should Certain amount of third historical user arranges according to the sequence of similitude from high to low.

History advertisement based on multiple third historical user browses record, establishes and launches sample about for the sample of users The ambient condition vector of advertisement.For example, filtering out all sample advertisements from all advertisements of advertising platform.In all samples Certain amount of sample advertisement is selected in this advertisement at random, as certain amount of first sample advertisement.It is specific that this is counted respectively The impressions of the first sample advertisement of quantity, obtain the dispensing overall condition of the certain amount of first sample advertisement to Amount.By the history advertisement browsing record combination of the certain amount of third historical user and the sample of users, sample use is obtained The extension advertisement browsing record vector at family.And by the dispensing overall condition vector of the certain amount of first sample advertisement, The attribute vector of the extension advertisement browsing record vector sum sample of users of the sample of users merges, and obtains first sample This advertisement is delivered to the ambient condition vector of the sample of users.

Based on the ambient condition vector sum deep neural network object module, obtain corresponding to each sample advertisement Launch value.For example, the ambient condition vector is inputted the deep neural network object module, each first sample is obtained The corresponding dispensing value of advertisement.

Based on the corresponding dispensing value of each sample advertisement, corresponding dispensing is worth highest sample advertisement and is delivered to The sample of users.For example, value of launching corresponding to the certain amount of first sample advertisement is arranged according to sequence from high to low The second sample advertisement is delivered to the sample and used by sequence using the first sample advertisement for coming foremost as the second sample advertisement Family.

In step S303, be based on backpropagation principle, to the network parameter of the deep neural network object module into Row optimization.

In this step, it is based on backpropagation principle, the network parameter of the deep neural network object module is carried out excellent Change.

In step s 304, quilt is calculated to the interbehavior for the sample advertisement being launched based on the sample of users The sample advertisement launched motivates the environmental feedback of the deep neural network object module.

In this step, whether the second sample advertisement is clicked according to the sample of users, gives the deep neural network Object module clicks excitation R₁.For example, giving the deep neural network if the target user clicks the second sample advertisement Object module clicks excitation R₁.If the sample of users does not click the second sample advertisement, the deep neural network mesh is given It marks model and clicks excitation 0*R₁。

Before the second sample advertisement is delivered to the sample of users by calculating, the throwing of the certain amount of first sample advertisement Put the variance of number.After the second sample advertisement is delivered to the sample of users by calculating, the certain amount of first sample advertisement Impressions variance.It is delivered to before the sample of users and after being delivered to the sample of users according to the second sample advertisement Deep neural network object module harmony excitation R is given in the variation of the variance₂.For example, if the second sample advertisement is thrown It puts and is greater than the variance after the second sample advertisement is delivered to the sample of users to the variance before the sample of users, then give this The positive harmonious excitation R of deep neural network object module₂.If the second sample advertisement is delivered to this before the sample of users Variance is less than the second sample advertisement and is delivered to the variance after the sample of users, then gives the deep neural network object module Negative harmonious excitation-R₂.If the second sample advertisement, which is delivered to the variance before the sample of users, is equal to second sample Advertisement is delivered to the variance after the sample of users, then gives the deep neural network object module negative harmonious excitation 0* R₂。

According to certain ratio, which is motivated into R₁R is motivated with the harmony₂Combination obtains environmental feedback excitation R= r₁R₁+r₂R₂.Wherein, r₁, r₂For parameter.

In step S305, is motivated based on the environmental feedback, adjust the network of the deep neural network object module Parameter.

In this step, R=r is motivated using the environmental feedback₁R₁+r₂R₂To optimize the deep neural network object module Network parameter.

In step S306, judge whether the deep neural network object module restrains.

In this step, judge whether the deep neural network object module restrains.If the deep neural network target Model convergence, the then deep neural network model arrived will dispose on the deep neural network model full dose line.If the depth is refreshing It is not restrained through network objectives model, then when new sample of users initiates new advertisement playing request, recalculates click excitation It is motivated with the harmony, optimizes the network parameter of the deep neural network object module.

According to an embodiment of the present application, whether the second sample advertisement is clicked according to the sample of users, gives the depth Neural network object module clicks excitation.The sample is delivered to before the sample of users and is delivered to according to the second sample advertisement to use The variation of the variance of the impressions of the certain amount of first sample advertisement behind family, gives the deep neural network target The excitation of model harmony.According to certain ratio, which is motivated and the harmony incentive combination, environmental feedback is obtained and swashs It encourages.It is motivated using the environmental feedback to optimize the network parameter of the deep neural network object module.Swashed based on the environmental feedback The dispensing environmental feedback to the deep neural network object module is encouraged, the network ginseng of the deep neural network object module is adjusted Number, dynamic adjust advertisement serving policy, improve the accuracy and harmony of advertisement dispensing.

In an optional embodiment of the application, depth nerve is established using Actor-Critic learning strategy Network model.

Fig. 4 is the schematic diagram of advertisement delivery device shown according to an exemplary embodiment.As shown in figure 4, the device 40 It include: that data capture unit 401, comparing unit 402, vector establish unit 403, computing unit 404 and advertisement and launch unit 405。

Data capture unit 401 is configured as obtaining the advertisement playing request that target user initiates.

Comparing unit 402 is configured to the attribute vector of target user described in comparison and the category of multiple historical users Similarity between property vector, obtains multiple first historical users most like with the target user.

Vector establishes unit 403, is configured as the history advertisement browsing record based on the multiple first historical user, builds The vertical ambient condition vector about for target user dispensing targeted advertisements.

Computing unit 404 is configured as being corresponded to based on the ambient condition vector sum deep neural network model The dispensing of each targeted advertisements is worth.And

Unit 405 is launched in advertisement, is configured as based on the corresponding dispensing value of each described targeted advertisements, will be corresponding It launches the highest targeted advertisements of value and is delivered to the target user.

In embodiments herein, data capture unit 401 receives the advertisement playing request that target user initiates.It should The trigger action of advertisement playing request, the e.g. access request to the cell phone software or webpage.Target user request is broadcast The advertisement put includes display advertising, video ads and interactive FLASH advertisement.

Comparing unit 402 calculates separately between the attribute vector of the target user and the attribute vector of multiple historical users Distance, such as Euclidean distance.Compare the corresponding Euclidean distance of each historical user and preset threshold, obtains using with the target Family has multiple second historical users of similitude.If the corresponding Euclidean distance of each historical user is less than the default threshold Value, then each historical user and the target user have similitude.If the corresponding Euclidean distance of each historical user is big In being equal to the preset threshold, then each historical user and the target user do not have similitude.

Vector establishes unit 403, all targeted advertisements are filtered out from all advertisements of advertising platform.All Certain amount of targeted advertisements are selected in targeted advertisements at random, as certain amount of first object advertisement.The spy is counted respectively The impressions of the first object advertisement of fixed number amount, obtain the dispensing overall condition of the certain amount of first object advertisement to Amount.

Ambient condition vector input is used the depth nerve net of Q-Learning learning framework by computing unit 404 Network model obtains the corresponding dispensing value of each first object advertisement.

Unit 405 is launched in advertisement, is worth corresponding dispensings of the certain amount of first object advertisement according to from high to low Sequence sequence, using the first object advertisement for coming foremost as the second targeted advertisements, which is delivered to The target user.

Fig. 5 is the schematic diagram of advertisement delivery device shown according to an exemplary embodiment.It is more than previous embodiment Perfect embodiment.As shown in figure 5, the device 50 includes: embedded unit 501, model foundation unit 502, data capture unit 503, comparing unit 504, vector establish unit 505, computing unit 506 and advertisement and launch unit 507.

Embedded unit 501 is configured as carrying out the user property of each user word insertion squeeze operation, obtains institute State the attribute vector of each user.

Model foundation unit 502 is configured as establishing the deep neural network mould about the targeted advertisements are launched Type.

Data capture unit 503 is configured as obtaining the advertisement playing request that target user initiates.

Comparing unit 504 is configured to the attribute vector of target user described in comparison and the category of multiple historical users Similarity between property vector, obtains multiple first historical users most like with the target user.

Vector establishes unit 505, is configured as the history advertisement browsing record based on the multiple first historical user, builds The vertical ambient condition vector about for target user dispensing targeted advertisements.

Computing unit 506 is configured as being corresponded to based on the ambient condition vector sum deep neural network model The dispensing of each targeted advertisements is worth.And

Unit 507 is launched in advertisement, is configured as based on the corresponding dispensing value of each described targeted advertisements, will be corresponding It launches the highest targeted advertisements of value and is delivered to the target user.

In embodiments herein, embedded unit 501, to the discrete sparse attribute of the user property of each user into Low-dimensional vector in the user property boil down to continuous space of the higher-dimension of each user is obtained each use by row embedding operation The attribute vector at family.Model foundation unit 502 is established about the deep neural network model for launching the targeted advertisements.

In an optional embodiment of the application, model foundation unit 502 uses Actor-Critic learning strategy To establish the deep neural network model.

In an optional embodiment of the application, establish about the depth nerve net for launching the targeted advertisements The detailed process of network model, comprising: be based on deep neural network algorithm, establish deep neural network object module.

Training on line is carried out to the deep neural network object module.Specifically, the advertisement that sample of users is initiated is obtained Playing request.The similarity being respectively compared between the attribute vector of the sample of users and the attribute vector of multiple historical users, obtains To the multiple third historical users most like with the sample of users.For example, calculating separately the attribute vector of the sample of users and more Euclidean distance between the attribute vector of a historical user.Compare the corresponding Euclidean distance of each historical user and default threshold Value obtains multiple 4th historical users for having similitude with the sample of users.If the corresponding Euclidean of each historical user Distance is less than the preset threshold, then each historical user and the sample of users have similitude.If each historical user Corresponding Euclidean distance is more than or equal to the preset threshold, then each historical user and the sample of users do not have similitude.To this Multiple 4th historical users sort according to the sequence of similitude from high to low, choose certain amount of 4th history of front User obtains certain amount of third historical user.Wherein, the certain amount of third historical user according to similitude by height to Low sequence arrangement.

The second sample advertisement whether is clicked according to the sample of users, gives deep neural network object module click Motivate R₁.For example, giving deep neural network object module click if the sample of users clicks the second sample advertisement Motivate R₁.If the sample of users does not click the second sample advertisement, gives the deep neural network object module and click and swash Encourage 0*R₁。

R=r is motivated using the environmental feedback₁R₁+r₂R₂To optimize the network parameter of the deep neural network object module.

Judge whether the deep neural network object module restrains.If the deep neural network object module is restrained, The deep neural network model arrived will dispose on the deep neural network model full dose line.If the deep neural network target Model is not restrained, then when new sample of users initiates new advertisement playing request, recalculates click excitation and the harmony Excitation, optimizes the network parameter of the deep neural network object module.

Fig. 6 is a kind of block diagram of device 1200 for executing advertisement placement method shown according to an exemplary embodiment.Example Such as, interactive device 1200 can be mobile phone, computer, digital broadcasting terminal, messaging device, game console, put down Panel device, Medical Devices, body-building equipment, personal digital assistant etc..

Referring to Fig. 6, device 1200 may include following one or more components: processing component 1202, memory 1204, electricity Source component 1206, multimedia component 1208, audio component 1210, the interface 1212 of input/output (I/O), sensor module 1214 and communication component 1216.

The integrated operation of the usual control device 1200 of processing component 1202, such as with display, telephone call, data communication, Camera operation and record operate associated operation.Processing component 1202 may include one or more processors 1220 to execute Instruction, to perform all or part of the steps of the methods described above.In addition, processing component 1202 may include one or more moulds Block, convenient for the interaction between processing component 1202 and other assemblies.For example, processing component 1202 may include multi-media module, To facilitate the interaction between multimedia component 1208 and processing component 1202.

Memory 1204 is configured as storing various types of data to support the operation in equipment 1200.These data Example includes the instruction of any application or method for operating on device 1200, contact data, telephone book data, Message, picture, video etc..Memory 1204 can by any kind of volatibility or non-volatile memory device or they Combination is realized, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), it is erasable can Program read-only memory (EPROM), programmable read only memory (PROM), read-only memory (ROM), magnetic memory, flash memory Reservoir, disk or CD.

Power supply module 1206 provides electric power for the various assemblies of device 1200.Power supply module 1206 may include power management System, one or more power supplys and other with for device 1200 generate, manage, and distribute the associated component of electric power.

Multimedia component 1208 includes the screen of one output interface of offer between described device 1200 and user.? In some embodiments, screen may include liquid crystal display (LCD) and touch panel (TP).If screen includes touch panel, Screen may be implemented as touch screen, to receive input signal from the user.Touch panel includes that one or more touch passes Sensor is to sense the gesture on touch, slide, and touch panel.The touch sensor can not only sense touch or sliding is dynamic The boundary of work, but also detect duration and pressure associated with the touch or slide operation.In some embodiments, more Media component 1208 includes a front camera and/or rear camera.When equipment 1200 is in operation mode, as shot mould When formula or video mode, front camera and/or rear camera can receive external multi-medium data.Each preposition camera shooting Head and rear camera can be a fixed optical lens system or have focusing and optical zoom capabilities.

Audio component 1210 is configured as output and/or input audio signal.For example, audio component 1210 includes a wheat Gram wind (MIC), when device 1200 is in operation mode, when such as call mode, recording mode, and voice recognition mode, microphone quilt It is configured to receive external audio signal.The received audio signal can be further stored in memory 1204 or via communication Component 1216 is sent.In some embodiments, audio component 1210 further includes a loudspeaker, is used for output audio signal.

I/O interface 1212 provides interface, above-mentioned peripheral interface module between processing component 1202 and peripheral interface module It can be keyboard, click wheel, button etc..These buttons may include, but are not limited to: home button, volume button, start button and Locking press button.

Sensor module 1214 includes one or more sensors, and the state for providing various aspects for device 1200 is commented Estimate.For example, sensor module 1214 can detecte the state that opens/closes of equipment 1200, the relative positioning of component, such as institute The display and keypad that component is device 1200 are stated, sensor module 1214 can be with detection device 1200 or device 1,200 1 The position change of a component, the existence or non-existence that user contacts with device 1200,1200 orientation of device or acceleration/deceleration and dress Set 1200 temperature change.Sensor module 1214 may include proximity sensor, be configured in not any physics It is detected the presence of nearby objects when contact.Sensor module 1214 can also include optical sensor, as CMOS or ccd image are sensed Device, for being used in imaging applications.In some embodiments, which can also include acceleration sensing Device, gyro sensor, Magnetic Sensor, pressure sensor or temperature sensor.

Communication component 1216 is configured to facilitate the communication of wired or wireless way between device 1200 and other equipment.Dress The wireless network based on communication standard, such as WiFi can be accessed by setting 1200, carrier network (such as 2G, 3G, 4G or 5G) or they Combination.In one exemplary embodiment, communication component 1216 receives via broadcast channel and comes from external broadcasting management system Broadcast singal or broadcast related information.In one exemplary embodiment, the communication component 1216 further includes near-field communication (NFC) module, to promote short range communication.For example, radio frequency identification (RFID) technology, Infrared Data Association can be based in NFC module (IrDA) technology, ultra wide band (UWB) technology, bluetooth (BT) technology and other technologies are realized.

In the exemplary embodiment, device 1200 can be by one or more application specific integrated circuit (ASIC), number Signal processor (DSP), digital signal processing appts (DSPD), programmable logic device (PLD), field programmable gate array (FPGA), controller, microcontroller, microprocessor or other electronic components are realized, for executing the above method.

In the exemplary embodiment, a kind of non-transitorycomputer readable storage medium including instruction, example are additionally provided It such as include the memory 1204 of instruction, above-metioned instruction can be executed by the processor 1220 of device 1200 to complete the above method.Example Such as, the non-transitorycomputer readable storage medium can be ROM, random access memory (RAM), CD-ROM, tape, soft Disk and optical data storage devices etc..

Fig. 7 is a kind of block diagram of device 1300 for executing advertisement placement method shown according to an exemplary embodiment.Example Such as, device 1300 may be provided as a server.Referring to Fig. 7, device 1300 includes processing component 1322, further comprises One or more processors, and the memory resource as representated by memory 1332, can be by processing component 1322 for storing Execution instruction, such as application program.The application program stored in memory 1332 may include one or more Each corresponds to the module of one group of instruction.In addition, processing component 1322 is configured as executing instruction, to execute above- mentioned information column Table display methods.

Device 1300 can also include that a power supply module 1326 be configured as the power management of executive device 1300, and one Wired or wireless network interface 1350 is configured as device 1300 being connected to network and input and output (I/O) interface 1358.Device 1300 can be operated based on the operating system for being stored in memory 1332, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM or similar.

The application be also implemented as some or all equipment for executing method as described herein or Program of device (for example, computer program and computer program product).Such program for realizing the application, which can store, to be counted On calculation machine readable medium, or it may be in the form of one or more signals.Such signal can be from Internet platform Upper downloading obtains, and is perhaps provided on the carrier signal or is provided in any other form.

Those skilled in the art after considering the specification and implementing the invention disclosed here, will readily occur to its of the application Its embodiment.This application is intended to cover any variations, uses, or adaptations of the application, these modifications, purposes or Person's adaptive change follows the general principle of the application and including the undocumented common knowledge in the art of the application Or conventional techniques.The description and examples are only to be considered as illustrative, and the true scope and spirit of the application are by following Claim is pointed out.

It should be understood that the application is not limited to the precise structure that has been described above and shown in the drawings, and And various modifications and changes may be made without departing from the scope thereof.Scope of the present application is only limited by the accompanying claims.

Claims

1. a kind of advertisement placement method characterized by comprising

Obtain the advertisement playing request that target user initiates；

The similarity being respectively compared between the attribute vector of the target user and the attribute vector of multiple historical users, obtain with Most like multiple first historical users of the target user；

History advertisement based on the multiple first historical user browses record, establishes and launches target about for the target user The ambient condition vector of advertisement；

Based on the ambient condition vector sum deep neural network model, the dispensing valence corresponding to each targeted advertisements is obtained Value；And

Based on the corresponding dispensing value of each described targeted advertisements, the highest targeted advertisements of corresponding dispensing value are thrown It puts to the target user.

2. advertisement placement method according to claim 1, which is characterized in that further include:

It establishes about the deep neural network model for launching the targeted advertisements；

It is preferably, described to establish about before the deep neural network model for launching the targeted advertisements, comprising:

Word is carried out to the user property of each user and is embedded in squeeze operation, obtain the attribute of each user to Amount.

3. advertisement placement method according to claim 2, which is characterized in that the category for being respectively compared the target user Similarity between the attribute vector of the property multiple historical users of vector sum, obtains multiple first most like with the target user Historical user, comprising:

Compare the corresponding Euclidean distance of each historical user and preset threshold, obtains that there is similitude with the target user Multiple second historical users；

It sorts to the multiple second historical user according to the sequence of similitude from high to low, chooses multiple the second of front Historical user obtains the multiple first historical user；

Wherein, the multiple first historical user arranges according to the sequence of similitude from high to low；

Preferably, the phase between the attribute vector for being respectively compared the target user and the attribute vector of multiple historical users Like degree, multiple first historical users most like with the target user are obtained, further includes:

If the corresponding Euclidean distance of described each historical user is less than the preset threshold, each described historical user There is similitude with the target user；

If the corresponding Euclidean distance of described each historical user is more than or equal to the preset threshold, each described history User and the target user do not have similitude；

Preferably, the history advertisement based on the multiple first historical user browses record, establishes about for the target The ambient condition vector of user's dispensing targeted advertisements, comprising:

The impressions for counting the multiple first object advertisement respectively, the dispensing for obtaining the multiple first object advertisement are whole Body situation vector；

By the history advertisement browsing record combination of the multiple first historical user and the target user, obtains the target and use The extension advertisement browsing record vector at family；And

The dispensing overall condition vector of the multiple first object advertisement, the extension advertisement of the target user is clear Look at record vector sum described in target user the attribute vector merge, obtain the first object advertisement being delivered to the mesh Mark the ambient condition vector of user；

Preferably, described to be based on the ambient condition vector sum deep neural network model, it obtains wide corresponding to each target The dispensing of announcement is worth, comprising:

The ambient condition vector is inputted into the deep neural network model, it is corresponding to obtain each described first object advertisement Dispensing value；

Preferably, described based on the corresponding dispensing value of each described targeted advertisements, corresponding dispensing is worth highest institute It states targeted advertisements and is delivered to the target user, comprising:

Value of launching corresponding to the multiple first object advertisement sorts according to sequence from high to low, will come foremost The first object advertisement is delivered to the target user as the second targeted advertisements, by second targeted advertisements.

4. advertisement placement method according to claim 2, which is characterized in that described to establish about the dispensing targeted advertisements The deep neural network model, comprising:

Establish deep neural network object module；

Training on line is carried out to the deep neural network object module；And

Based on backpropagation principle, the network parameter of the deep neural network object module is optimized；

It is preferably, described that training on line is carried out to the deep neural network object module, comprising:

Obtain the advertisement playing request that sample of users is initiated；

The similarity being respectively compared between the attribute vector of the sample of users and the attribute vector of multiple historical users, obtain with The most like multiple third historical users of the sample of users；

History advertisement based on the multiple third historical user browses record, establishes and launches sample about for the sample of users The ambient condition vector of advertisement；

Based on deep neural network object module described in the ambient condition vector sum, obtain corresponding to each sample advertisement Launch value；And

Based on the corresponding dispensing value of each described sample advertisement, the highest sample advertisement of corresponding dispensing value is thrown It puts to the sample of users；

Preferably, between the attribute vector for being respectively compared the sample of users and the attribute vector of multiple historical users Similarity obtains the multiple third historical users most like with the sample of users, comprising:

Compare the corresponding Euclidean distance of each historical user and preset threshold, obtains that there is similitude with the sample of users Multiple 4th historical users；

It sorts to the multiple 4th historical user according to the sequence of similitude from high to low, chooses multiple the four of front Historical user obtains the multiple third historical user；

Wherein, the multiple third historical user arranges according to the sequence of similitude from high to low；

Preferably, the phase between the attribute vector for being respectively compared the sample of users and the attribute vector of multiple historical users Like degree, the multiple third historical users most like with the sample of users are obtained, further includes:

If the corresponding Euclidean distance of described each historical user is less than the preset threshold, each described historical user There is similitude with the sample of users；

If the corresponding Euclidean distance of described each historical user is more than or equal to the preset threshold, each described history User and the sample of users do not have similitude；

Preferably, the history advertisement based on the multiple third historical user browses record, establishes about for the sample The ambient condition vector of user's dispensing sample advertisement, comprising:

The impressions for counting the multiple first sample advertisement respectively, the dispensing for obtaining the multiple first sample advertisement are whole Body situation vector；

By the history advertisement browsing record combination of the multiple third historical user and the sample of users, obtains the sample and use The extension advertisement browsing record vector at family；And

The dispensing overall condition vector of the multiple first sample advertisement, the extension advertisement of the sample of users is clear Look at record vector sum described in sample of users the attribute vector merge, obtain the first sample advertisement being delivered to the sample The ambient condition vector of this user；

Preferably, described based on deep neural network object module described in the ambient condition vector sum, it obtains corresponding to each The dispensing of a sample advertisement is worth, comprising:

The ambient condition vector is inputted into the deep neural network object module, obtains each described first sample advertisement Corresponding dispensing value；

Preferably, described based on the corresponding dispensing value of each described sample advertisement, corresponding dispensing is worth highest institute It states sample advertisement and is delivered to the sample of users, comprising:

Value of launching corresponding to the multiple first sample advertisement sorts according to sequence from high to low, will come foremost The first sample advertisement is delivered to the sample of users as the second sample advertisement, by the second sample advertisement.

5. advertisement placement method according to claim 4, which is characterized in that described to establish about the dispensing targeted advertisements The deep neural network model, further includes:

Based on the sample of users to the interbehavior for the sample advertisement being launched, the sample advertisement being launched is calculated Environmental feedback excitation to the deep neural network object module；

Preferably, it is described based on the sample of users to the interbehavior for the sample advertisement being launched, calculate and be launched The sample advertisement motivates the environmental feedback of the deep neural network object module, comprising:

The second sample advertisement whether is clicked according to the sample of users, gives the deep neural network object module point Hit excitation；

Calculate separately the second sample advertisement be delivered to before the sample of users and be delivered to it is described after the sample of users The variance of the impressions of multiple first sample advertisements；

The variance after being delivered to before the sample of users and being delivered to the sample of users according to the second sample advertisement Variation, give deep neural network object module harmony excitation；And

Based on click excitation and the harmonious excitation, the second sample advertisement being launched is to the depth mind Environmental feedback excitation through network objectives model；

Preferably, if the second sample advertisement is delivered to the variance before the sample of users and is less than second sample Advertisement is delivered to the variance after the sample of users, then gives the deep neural network object module negative harmonious excitation；

If the second sample advertisement is delivered to the variance before the sample of users and throws greater than the second sample advertisement It puts to the variance after the sample of users, then gives the deep neural network object module positive harmonious excitation；

Preferably, the advertisement placement method, which is characterized in that further include:

Judge whether the deep neural network object module restrains；

If the deep neural network object module is not restrained, when new sample of users initiates new advertisement playing request When, the click excitation and the harmonious excitation are recalculated, the network ginseng of the deep neural network object module is optimized Number.

6. a kind of advertisement delivery device characterized by comprising

Comparing unit, be configured to target user described in comparison attribute vector and multiple historical users attribute vector it Between similarity, obtain multiple first historical users most like with the target user；

Vector establishes unit, be configured as based on the multiple first historical user history advertisement browsing record, establish about The ambient condition vector of targeted advertisements is launched for the target user；

Computing unit is configured as obtaining corresponding to each based on the ambient condition vector sum deep neural network model The dispensing of targeted advertisements is worth；And

Unit is launched in advertisement, is configured as based on the corresponding dispensing value of each described targeted advertisements, by corresponding dispensing valence It is worth the highest targeted advertisements and is delivered to the target user.

7. advertisement delivery device according to claim 6, which is characterized in that further include:

Embedded unit is configured as carrying out the user property of each user word insertion squeeze operation, obtain described each The attribute vector of user.

8. advertisement delivery device according to claim 7, which is characterized in that the category for being respectively compared the target user Similarity between the attribute vector of the property multiple historical users of vector sum, obtains multiple first most like with the target user Historical user, comprising:

Value of launching corresponding to the multiple first object advertisement sorts according to sequence from high to low, will come foremost The first object advertisement is delivered to the target user as the second targeted advertisements, by second targeted advertisements；

It is preferably, described to establish about the deep neural network model for launching the targeted advertisements, comprising:

Establish deep neural network object module；

Training on line is carried out to the deep neural network object module；And

Obtain the advertisement playing request that sample of users is initiated；

It is preferably, described to establish about the deep neural network model for launching the targeted advertisements, further includes:

It is motivated based on the environmental feedback, adjusts the network parameter of the deep neural network object module；

Preferably, the advertisement delivery device, which is characterized in that further include:

Judge whether the deep neural network object module restrains；

If the deep neural network object module is not restrained, when new sample of users initiates new advertisement playing request When, the environmental feedback excitation is recalculated, the network parameter of the deep neural network object module is optimized.

9. a kind of server characterized by comprising

Processor；

Memory for storage processor executable instruction；

Wherein, the processor is configured to executing advertisement placement method described in 1 to 5 any one of the claims.

10. a kind of computer readable storage medium, which is characterized in that the computer-readable recording medium storage has computer to refer to It enables, the computer instruction is performed realization such as advertisement placement method described in any one of claim 1 to 5.