CN110458220A

CN110458220A - Crowd's orientation method, device, server and storage medium

Info

Publication number: CN110458220A
Application number: CN201910714826.7A
Authority: CN
Inventors: 杨春风
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2019-07-31
Filing date: 2019-07-31
Publication date: 2019-11-15
Anticipated expiration: 2039-07-31
Also published as: CN110458220B

Abstract

The embodiment of the invention discloses a kind of crowd's orientation method, device, server and media, and wherein method includes: the reference user collection and candidate user collection for obtaining targeted advertisements；Optimization is trained to prediction model with reference to user's collection and the candidate user collection using described, the prediction model optimized；It calls the prediction model of the optimization to concentrate the attribute data of each candidate user to carry out advertisement orientation to each candidate user according to the candidate user to estimate, the orientation probability of each candidate user is obtained, the orientation probability refers to that candidate user generates the probability of positive feedback to the targeted advertisements；The attribute data for filtering out directional user is concentrated from the candidate user according to the orientation probability of each candidate user, and the attribute data of the directional user is added in the orientation demographic data of the targeted advertisements.The embodiment of the present invention can preferably carry out crowd's orientation, improve the accuracy of orientation demographic data.

Description

Crowd's orientation method, device, server and storage medium

Technical field

The present invention relates to Internet technical fields, and in particular to technical field is launched in advertisement more particularly to a kind of crowd is fixed To method, a kind of crowd's orienting device, a kind of server and a kind of computer storage medium.

Background technique

Advertisement informs certain part things to social the public as the term suggests being exactly to publicize widely；The advertisement of narrow sense refers to Advertiser propagates the means of commodity or information on services with way of paying by advertising media's platform to consumer or user.Currently, During launching targeted advertisements, it will usually first carry out crowd's directional process to targeted advertisements to determine the targeted advertisements Orientation crowd includes potential audience relevant to targeted advertisements in orientation crowd；Then target is launched in orientation crowd Advertisement.It can be seen that crowd's orientation is a very important link in advertisement release process, orient the accuracy of crowd with Advertisement delivery effect is closely bound up；Therefore, how preferably to carry out crowd's orientation with improve the accuracy of orientation crowd gradually at For research hotspot.

Summary of the invention

The embodiment of the invention provides a kind of crowd's orientation method, device, server and computer storage medium, Ke Yigeng Crowd's orientation is carried out well, improves the accuracy of orientation demographic data.

On the one hand, the embodiment of the invention provides a kind of crowd's orientation method, which includes:

Obtain the reference user collection and candidate user collection of targeted advertisements；It is described to collect with reference to user including multiple with reference to user's Attribute data, it is described with reference to user be refer to the targeted advertisements generate positive feedback user；The candidate user collection Attribute data including multiple candidate users, the candidate user are users to be oriented；

Optimization is trained to prediction model with reference to user's collection and the candidate user collection using described, what is optimized is pre- Estimate model；

The prediction model of the optimization is called to concentrate the attribute data of each candidate user to described according to the candidate user Each candidate user carries out advertisement orientation and estimates, and obtains the orientation probability of each candidate user, the orientation probability refers to candidate User generates the probability of positive feedback to the targeted advertisements；

The attribute number for filtering out directional user is concentrated from the candidate user according to the orientation probability of each candidate user According to, and the attribute data of the directional user is added in the orientation demographic data of the targeted advertisements.

On the other hand, the embodiment of the invention provides a kind of crowd's orienting device, which includes:

Acquiring unit, the reference user for obtaining targeted advertisements collects and candidate user collection；It is described to include with reference to user's collection Multiple attribute datas with reference to user, it is described with reference to user be refer to the targeted advertisements generate positive feedback user； The candidate user collection includes the attribute data of multiple candidate users, and the candidate user is user to be oriented；

Optimize unit, it is excellent for being trained using the reference user collection and the candidate user collection to prediction model Change, the prediction model optimized；

Processing unit, for calling the prediction model of the optimization to concentrate according to the candidate user category of each candidate user Property data advertisement orientation carried out to each candidate user estimate, obtain the orientation probability of each candidate user, the orientation Probability refers to that candidate user generates the probability of positive feedback to the targeted advertisements；

The processing unit, for being filtered out according to the orientation probability of each candidate user from candidate user concentration The attribute data of directional user, and the attribute data of the directional user is added to the orientation demographic data of the targeted advertisements In.

In another aspect, the server includes communication interface the embodiment of the invention provides a kind of server, the service Device further include:

Processor is adapted for carrying out one or more instruction；And

Computer storage medium, the computer storage medium are stored with one or more instruction, and described one or more Instruction is suitable for being loaded by the processor and executing following steps:

In another aspect, the embodiment of the invention provides a kind of computer storage medium, the computer storage medium storage There is one or more instruction, one or more instruction is suitable for being loaded by processor and executing following steps:

The embodiment of the present invention can first obtain the reference user collection and candidate user of targeted advertisements during crowd orients Collection；It wherein, include multiple attribute datas with reference to user with reference to user's collection, candidate user collection includes the attribute of multiple candidate users Data.Since reference user is the user for referring to generate target user positive feedback, and candidate user is use to be oriented Family；Therefore optimization first can be trained to prediction model using reference user collection and candidate user collection, then calls the pre- of optimization Estimate model to estimate each candidate user progress advertisement orientation according to the attribute data of each candidate user, obtains determining for each candidate user To probability；The training space and prediction space that may make prediction model in this way are consistent, to improve determining for each candidate user To the accuracy of probability.It, can basis since orientation probability refers to that candidate user generates positive feedback probability to targeted advertisements The orientation probability of each candidate user concentrates the attribute data for screening and adding directional user to determine to targeted advertisements from candidate user Into demographic data；It, can be to improve the accuracy for orienting demographic data by improving the accuracy of orientation probability.

Detailed description of the invention

Technical solution in order to illustrate the embodiments of the present invention more clearly, below will be to needed in embodiment description Attached drawing is briefly described, it should be apparent that, drawings in the following description are some embodiments of the invention, general for this field For logical technical staff, without creative efforts, it is also possible to obtain other drawings based on these drawings.

Fig. 1 a is the system architecture diagram that a kind of advertisement provided in an embodiment of the present invention is launched；

Fig. 1 b is a kind of schematic diagram of crowd's orientation scheme provided in an embodiment of the present invention；

Fig. 2 is a kind of flow diagram of crowd's orientation method provided in an embodiment of the present invention；

Fig. 3 is a kind of schematic diagram of spatial offset phenomenon provided in an embodiment of the present invention；

Fig. 4 be another embodiment of the present invention provides a kind of crowd's orientation method flow diagram；

Fig. 5 a is the application scenario diagram that a kind of advertisement provided in an embodiment of the present invention is launched；

Fig. 5 b is the application scenario diagram that another advertisement provided in an embodiment of the present invention is launched；

Fig. 5 c is the flow diagram of another crowd's orientation method provided in an embodiment of the present invention；

Fig. 5 d is the application scenario diagram that another advertisement provided in an embodiment of the present invention is launched；

Fig. 5 e is the application scenario diagram that another advertisement provided in an embodiment of the present invention is launched；

Fig. 6 is a kind of structural schematic diagram of crowd's orienting device provided in an embodiment of the present invention；

Fig. 7 is a kind of structural schematic diagram of server provided in an embodiment of the present invention.

Specific embodiment

Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete Site preparation description.

Crowd's orientation, which refers to the process of, to be filtered out and the maximally related potential audience of targeted advertisements.Wherein, potential audience again may be used Referred to as directional user, in particular to the potential recipient for thering is greater probability to receive targeted advertisements.Studies have shown that launching target During advertisement, the accuracy of crowd's orientation is usually directly proportional up to rate to the touching of targeted advertisements；That is the accuracy of crowd's orientation Higher, the touching of targeted advertisements is also higher up to rate.So-called touching is alternatively referred to as exposure rate up to rate, refers to actual exposure in orientation crowd The ratio of number of users and total number of users of orientation crowd, actual exposure number of users herein refer within a preset period of time to target There are the quantity of the user of advertisement exposure behavior (seeing the behavior of targeted advertisements) for advertisement；Touching can be used for measuring orientation crowd up to rate The user how many ratio is middle reality understand obtains actual advertisement and launches, i.e. touching can be used for measuring and actually see in orientation crowd up to rate See user's ratio of targeted advertisements.For example, total number of users of orientation crowd is 500 people, if wide to target within a preset period of time Accusing there are the quantity of the user of advertisement exposure behavior is 450 people, then touching is equal to 450/500=90% up to rate (exposure rate)； If being within a preset period of time 200 people there are the quantity of the user of advertisement exposure behavior to targeted advertisements, then touching (is exposed up to rate Rate) it is equal to 200/500=40%.Preset time period herein can be arranged according to practical business demand.

It can be seen that crowd's orientation is a very important link in advertisement release process；Either from advertiser The angle of (having advertisement to launch the user of demand) all be unable to do without crowd's orientation still from the angle of flow side's (such as ad system) This process.For example, the budget that low exposure rate means that advertisement is launched cannot be consumed effectively, nothing for advertiser Method touching reaches the user group of enough scales；And for flow side, low exposure rate will affect its income (especially according to exposure The advertisement deducted fees).Therefore, it orients to preferably carry out crowd to improve subsequent advertisement delivery effect, the embodiment of the present invention Propose a kind of crowd's orientation scheme；Crowd's orientation scheme can be applied in ad system, and ad system herein refers to energy Advertising media's platform is enough provided, is the system that advertiser launches advertisement on advertising media's platform with charge method, such as Tencent's public affairs The SPA MI system of department.

In one embodiment, which can be the ad system that human-computer interaction is carried out based on webpage, It substantially may include front end and backstage two parts；Wherein, front end refers to the foreground partition of ad system, runs on the browsing of terminal In device and show advertiser browse webpage；Backstage refers to for carrying out a series of data management operations to front end to realize The background server of advertising service is provided for advertiser.In another embodiment, which can also be that one is based on Client carries out the ad system of human-computer interaction, substantially may include client and server-side two parts；Wherein, client refers to It installs and runs in terminal, and provide the application program (APP) of advertising service for advertiser；Server-side refers to and client phase It is corresponding, it is used to support client and the background server of advertising service is provided.For convenient for elaboration, unless stated otherwise, subsequent implementation Ad system mentioned in example for carrying out the ad system of human-computer interaction based on webpage to be illustrated.It is above-mentioned mentioned Terminal may include but be not limited to: the mobile devices such as smart phone, laptop computer, tablet computer and desktop computer, etc. Deng；Targeted advertisements can include but is not limited to: the advertisement of the types such as television advertising, cinema sign, the web advertisement, video ads； Advertising media's platform may include but be not limited to: TV, film, webpage, video playing client (such as Tencent's videoconference client) with And instant communication client (such as Tencent QQ client, wechat client) multimedia platform.

It, can be by the front end of ad system to background server when advertiser wants to determine the orientation crowd of targeted advertisements (hereinafter referred to as server) sends crowd and orients request, as shown in Figure 1a.Server receive the crowd orient request after, It can orient and request in response to the crowd, execute above-mentioned crowd's orientation scheme to determine orientation crowd's number of targeted advertisements According to as shown in Figure 1 b；It include the attribute data of multiple directional users in the orientation demographic data, multiple directional users constitute target The orientation crowd of advertisement.In the specific implementation process, server can first obtain the seed demographic data collection of targeted advertisements and big Disk user data set；Wherein, seed demographic data concentration may include the attribute data of multiple seed users, and so-called seed is used Family is the user for referring to generate targeted advertisements positive feedback；Deep bid user data concentration may include multiple deep bid users Attribute data, so-called deep bid user refer to user to be oriented.Secondly, can be from seed demographic data collection and deep bid number of users According to the training set determined for being trained to prediction model is concentrated, prediction model herein refers to can be to user's (such as seed User, deep bid user etc.) carry out advertisement orientation estimate, with obtain user to targeted advertisements generate positive feedback probability mould Type.It is then possible to be trained optimization to prediction model using training set, and using the prediction model of optimization to deep bid user into Row advertisement orientation is estimated, and the probability that each deep bid user generates positive feedback to targeted advertisements is obtained.It finally can be general based on this Rate is ranked up each deep bid user to determine directional user, and the attribute data of directional user is added to orientation crowd In data.

It can be seen that the prediction model of optimization is instructed to prediction model real-time perfoming in above-mentioned crowd's orientation scheme Practice what optimization obtained.It is concentrated by elder generation from seed demographic data collection and deep bid user data and determines training set, and with training set pair Prediction model is trained optimization；Then the prediction model of optimization is called to carry out advertisement orientation to each deep bid user again again pre- The mode estimated, may make prediction model training space and prediction space be consistent, thus may make estimate it is each Deep bid user is higher to the accuracy for the probability that targeted advertisements generate positive feedback, and then the accurate of orientation demographic data can be improved Property.During the subsequent orientation demographic data based on the high accuracy launches targeted advertisements, targeted advertisements can be improved Touching reach rate, to improve the dispensing effect of targeted advertisements.

Based on the description above, the embodiment of the present invention proposes a kind of crowd's orientation method, which can be by Above-mentioned mentioned server executes.Fig. 2 is referred to, which may include following steps S201-S204:

S201 obtains the reference user collection and candidate user collection of targeted advertisements.

When advertiser wants to carry out targeted advertisements orientation crowd of crowd's orientation to determine targeted advertisements, advertisement can be passed through The front end of system orients request to the crowd that server sends targeted advertisements.Correspondingly, server is receiving crowd orientation After request, it can orient and request in response to the crowd, obtain the reference user collection and candidate user collection of targeted advertisements.Wherein, Candidate user collection may include the attribute data of multiple candidate users, and so-called candidate user is that user to be oriented is (i.e. aforementioned Mentioned deep bid user).It may include multiple attribute datas with reference to user with reference to user's collection, it is so-called to refer to reference to user Targeted advertisements can be generated with the user of positive feedback, positive feedback herein refers to during being launched targeted advertisements, It sees targeted advertisements or clicks the feedback of targeted advertisements；Correspondingly, not seeing that target is wide during being launched targeted advertisements The feedback of announcement can be described as negative-feedback.For example, targeted advertisements have been launched to user A in No. 7.20 this days, and in other words, user A At No. 7.20, this day has been launched targeted advertisements；If user A has seen the targeted advertisements in No. 7.20 this days or clicks The targeted advertisements then show that user A produces positive feedback to targeted advertisements, and the user A in the case of this can be for reference to user；If User A does not see the targeted advertisements in this day at No. 7.20, then shows that user A produces negative sense feedback to targeted advertisements, this In the case of user A be not with reference to user.

Wherein, above-mentioned mentioned attribute data can include at least the user identifier and user's portrait of user, i.e., with reference to use The attribute data at family includes with reference to the user identifier of user and user's portrait, and the attribute data of candidate user includes candidate user User identifier and user's portrait.User identifier herein refers to the number that can be used for unique identification user identity, may include with It is at least one of lower: social account (such as QQ number code, wechat account), EIC equipment identification code (such as IMEI of Android system equipment The IFA of (International Mobile Equipment Identity, international mobile equipment identification number), IOS system equipment (a kind of EIC equipment identification code)), telephone number, ID card No., etc..User's portrait includes for embodying user image Label may include at least one of following: gender, age, personality, interest (hobby), etc..

S202 is trained optimization to prediction model using reference user collection and candidate user collection, and what is optimized estimates Model.

In order to preferably be trained optimization to prediction model, so that the better performances of the prediction model of optimization；The present invention The history that embodiment first uses targeted advertisements launches data and has carried out trained optimization to prediction model, and history is launched wraps in data Include the attribute data of multiple historical users；And performance test also is carried out to the prediction model of optimization.But practice discovery uses History launch data prediction model is trained optimize the prediction model of obtained optimization performance it is poor, using the optimization Prediction model carry out advertisement and launch to will appear the touchings of targeted advertisements problem lower up to rate (i.e. exposure rate).The embodiment of the present invention This is analyzed, it is found that the main reason for low exposure rate problem occur is as follows:

When being trained optimization to prediction model using history dispensing data, selects there are advertisement exposure behavior and deposited The ad click behavior behavior of advertisement (click) historical user attribute data as positive sample, selected that there are advertisement exposures Light behavior and there is no ad click behavior historical user attribute data be used as negative sample, with this come to prediction model progress Training optimization.It can be seen that training set used by this training optimal way include to be entirely that history had been launched target wide The attribute data of the historical user of announcement, i.e., positive and negative sample standard deviation are exposed；And forecast set is each candidate user (i.e. deep bid User), but candidate user is not necessarily present advertisement exposure behavior.Optimization is trained to prediction model with such training set, Be the equal of assuming that all candidate users in forecast set are all bound to expose, that is, assume all times in forecast set Family is selected all to be bound to existing advertisement exposure behavior for targeted advertisements；And due to actually and not can guarantee each time Selecting family to be bound to, there are advertisement exposure behaviors, therefore this hypothesis is invalid in crowd's orientation.In other words, this to adopt The training method of data training prediction model is launched with history, what training space included is that there are the users of advertisement exposure behavior (the inside contains there are advertisement exposure behavior and there are the historical user of ad click behavior and there are advertisement exposures for group Behavior and the historical user that ad click behavior is not present), and predicting space is that whole corresponding to entire candidate user collection is waited Family is selected, is equivalent between trained space and prediction space and spatial offset (Sample Selection Bias) has occurred, such as scheme Shown in 3.

Studies have shown that the spatial offset phenomenon between above-mentioned mentioned training space and prediction space is to lead to low exposure The main reason for the problem of rate；Also, low exposure rate problem also results in low clicking rate problem, and clicking rate (CTR) herein is The ratio of actual click number of users and actual exposure number of users into crowd is specified, actual click number of users herein refers to pre- If to targeted advertisements, there are the quantity of the user of ad click behavior after seeing targeted advertisements in the period；For example, orientation crowd Total number of users be 500 people, if within a preset period of time to targeted advertisements there are the quantity of the user of advertisement exposure behavior be 400 People, then the actual exposure number of users of orientation crowd is just 400 people；If there is 200 people click in this 400 people within a preset period of time Targeted advertisements, i.e., with the presence of 200 people to targeted advertisements ad click behavior, then the actual click number of users for orienting crowd is 200 people, then clicking rate is equal to 200/400=50%.Clicking rate can be used for measuring targeted advertisements and carry out in orientation crowd Click effect after exposure；Clicking rate is bigger, then shows that the click effect after targeted advertisements are exposed in orientation crowd is got over It is good；Clicking rate is smaller, then shows that the click effect after targeted advertisements are exposed in orientation crowd is poorer.Therefore, of the invention Embodiment is in order to avoid training space and predicts the phenomenon that space shifts, and selection is using with reference to user's collection and candidate user collection Optimization is trained to prediction model, so that the training space of prediction model and prediction space are consistent, to guarantee to train The prediction model for optimizing obtained optimization is with good performance, so that the subsequent prediction model using the optimization carries out advertisement The exposure rate or clicking rate of targeted advertisements can be improved when dispensing.

S203 calls the prediction model of optimization to concentrate the attribute data of each candidate user to each candidate use according to candidate user Family carries out advertisement orientation and estimates, and obtains the orientation probability of each candidate user.

After the prediction model optimized, the prediction model of the optimization can first be called to concentrate each time according to candidate user It selects the attribute data at family to carry out advertisement orientation to each candidate user to estimate, obtains the orientation probability of each candidate user, herein Orientation probability refers to that candidate user generates the probability of positive feedback to targeted advertisements.From the foregoing it will be appreciated that positive feedback refers in quilt During launching targeted advertisements, sees targeted advertisements or click the feedback of targeted advertisements；As it can be seen that if the orientation of candidate user is general Rate is bigger, then the probability for showing that targeted advertisements are seen or clicked by the candidate user is bigger, which becomes targeted advertisements Directional user probability it is also bigger.It therefore, can be according to each candidate use after obtaining the orientation probability of each candidate user The orientation probability at family concentrates the attribute data for filtering out directional user from candidate user, and the attribute data of directional user is added Into the orientation demographic data of targeted advertisements, i.e. execution step S204.It should be noted that being launched when positive feedback refers to During targeted advertisements, when seeing the feedback of targeted advertisements, orientation probability is then exposure probability；When positive feedback refers in quilt During launching targeted advertisements, when clicking the feedback of targeted advertisements, orientation probability is then to click probability.

S204 concentrates the attribute data for filtering out directional user according to the orientation probability of each candidate user from candidate user, And the attribute data of directional user is added in the orientation demographic data of targeted advertisements.

In order to further increase the exposure rate or clicking rate of targeted advertisements, server may be selected when executing step S204 will The biggish candidate user of probability is oriented as directional user, and concentrate the attribute data for obtaining the directional user to add from candidate user It adds in orientation demographic data.In the specific implementation process, crowd, which orients request, can carry orientation crowd's quantity；So, it services Device can concentrate the attribute data of each candidate user to be ranked up candidate user first according to the sequence of probability from big to small is oriented, Obtain ordered set；Then, the attribute of corresponding candidate user can be successively chosen from ordered set according to orientation crowd's quantity Attribute data of the data as directional user.

In practical applications, advertiser also may choose whether that each reference user that will be concentrated with reference to user uses as orientation Family.If advertiser selects each reference user for concentrating reference user as directional user, crowd orients can also in request It carries and targeted advertisements is directed to each instruction information with reference to user；So in the case, server is according to orientation crowd When quantity successively chooses attribute data of the attribute data of corresponding candidate user as directional user from ordered set, it can ask The difference between orientation crowd quantity and the quantity of reference user is taken, screening quantity is obtained；And according to screening quantity successively from row Attribute data of the attribute data as directional user of corresponding candidate user is successively chosen in ordered sets.Also, server is also Each attribute data with reference to user that reference user can be concentrated is added in the orientation demographic data of targeted advertisements.If advertiser It selects each reference user not concentrated reference user as directional user, then just will not carry this in the request of crowd's orientation and refer to Show information；So in the case, server can directly successively be chosen from ordered set according to orientation crowd's quantity corresponding Candidate user attribute data of the attribute data as directional user.

For example, if it is 50 that crowd, which orients the entrained orientation crowd's quantity of request, concentrating with reference to user includes 10 A attribute data with reference to user, candidate user concentrate include 60 candidate users attribute data, and according to orientation probability from Small sequence is arrived greatly, the attribute data of each candidate user is ranked up, and obtaining ordered set successively includes: candidate user 1 Attribute data, the attribute data of candidate user 2, candidate user 3 attribute data ... candidate user 60 attribute data.If crowd It is carried in orientation request and targeted advertisements is directed to each instruction information with reference to user, then it is 50- that screening quantity, which can be calculated, 10=40, then the attribute data that can successively choose 40 candidate users from ordered set is added to orientation demographic data In, and 10 attribute datas with reference to user are also added to and are oriented in demographic data；Orientation crowd's number i.e. in the case of this According to the attribute data and the attribute data of candidate user 1, the attribute data of candidate user 2, candidate for including 10 reference users The attribute data of attribute data ... the candidate user 40 of user 3.If crowd orients not carry in request and is directed to targeted advertisements The instruction information of user is respectively referred to, then the attribute data that 50 candidate users can be successively chosen directly from ordered set is added to It orients in demographic data；Orientation demographic data i.e. in the case of this includes the category of the attribute data of candidate user 1, candidate user 2 The attribute data of property data, attribute data ... the candidate user 50 of candidate user 3.

Fig. 4 is referred to, is the flow diagram of another crowd's orientation method provided in an embodiment of the present invention.The crowd is fixed It can be executed to method by above-mentioned mentioned server.Fig. 4 is referred to, which may include following steps S401-S406:

S401 obtains the reference user collection and candidate user collection of targeted advertisements.

Server after the crowd that receives orients request, can obtain the reference user collection and candidate user collection of targeted advertisements； The crowd, which orients request, can at least carry the crowd's directional aim and platform identification of targeted advertisements, platform identification packet herein It includes: launching the platform identification of advertising media's platform of targeted advertisements or the platform identification of ad system.Specifically, if advertiser The not specified advertising media's platform for launching targeted advertisements, then it is the flat of ad system that crowd, which orients the entrained platform identification of request, Station identification；If advertiser specifies the advertising media's platform for launching targeted advertisements, crowd orients the entrained platform mark of request Know for the platform identification of advertising media's platform specified by advertiser.

Correspondingly, server obtain targeted advertisements candidate user collection when, if platform identification include launch targeted advertisements Advertising media's platform platform identification, then can first determine advertising media's platform corresponding to platform identification；Then by advertisement matchmaker All registration users in body platform obtain the attribute data of all registration users as candidate user as candidate user Attribute data is added to candidate user concentration.Alternatively, being used the active registration user of advertising media's platform near-mid term as candidate Family, and the attribute data for obtaining registration user active in the recent period is added to candidate user collection as the attribute data of candidate user In；Wherein, registration user active in the recent period is referred in the preset time period calculated forward based on present system time, logs in advertisement The number of media platform is greater than the registration user of preset times.If platform identification includes the platform identification of ad system, can incite somebody to action In the advertising user list stored in ad system all advertising users (be such as launched any advertisement historical user, Launched the advertiser of any advertisement) as candidate user, and the attribute data of all advertising users is obtained as candidate user Attribute data be added to candidate user concentration.Alternatively, by the recent work in the advertising user list stored in ad system The advertising user of jump obtains attribute of the attribute data of advertising user active in the recent period as candidate user as candidate user Data are added to candidate user concentration.

The reference user collection of targeted advertisements can be advertiser oneself upload, that is, referring to user is the use that advertiser specifies Family；Or targeted advertisements reference user collection can be third client active reporting, i.e., with reference to user be on line is offline Convert user.Wherein, third party's client is the client for referring to carry out targeted advertisements corresponding audience business processing End, audience refer to the object propagated by targeted advertisements, can be commodity, application program (APP), website etc., line User has been converted on offline refers to the historical user for carrying out business operation to target object by targeted advertisements；For example, target Advertisement is the advertisement about A commodity, then audience is then A commodity, then third party's client refers to purchase A commodity Client (such as shopping client), has converted user on line is offline and has referred to the history for buying A class commodity by targeted advertisements User；For another example, targeted advertisements are the advertisements about social APP, then audience is then social activity APP, then third party client End is the client (such as application shop client) for referring to download social activity APP, has converted user on line is offline and refers to and passes through Targeted advertisements and the historical user, etc. for downloading social activity APP.Correspondingly, obtaining the specific of the reference user collection of targeted advertisements Embodiment may is that the crowd for receiving targeted advertisements orients request, which orients request and carry seed user list；Herein Seed user list in include multiple seed users attribute data, user list is uploaded by advertiser either by the Tripartite's client reports.Add the attribute data of seed user each in seed user list as the attribute data with reference to user The reference user for adding to targeted advertisements concentrates.

In another embodiment, the reference user collection of targeted advertisements can also be server according to the history of targeted advertisements What input situation was collected automatically.Correspondingly, obtain targeted advertisements reference user collection specific embodiment may is that firstly, The crowd that can receive targeted advertisements orients request, which orients crowd's directional aim that request carries targeted advertisements.Secondly, can The history for obtaining targeted advertisements launches flowing water table, and history launches attribute data and the behavior that historical user is included at least in flowing water table Data, the behavioral data of historical user are used to indicate historical user for targeted advertisements with the presence or absence of advertisement exposure behavior and advertisement Click behavior；Then, it can be launched according to crowd's directional aim from history and obtain the attribute data for referring to user in flowing water table, and will The reference user concentration of targeted advertisements is added to reference to the attribute data of user.

Wherein, when launching the attribute data obtained in flowing water table with reference to user from history according to crowd's directional aim, if Crowd's directional aim is that crowd's orientation is carried out on the basis of exposure rate, then by there are advertisement exposure behaviors in history dispensing flowing water table Historical user attribute data as refer to user attribute data；If crowd's directional aim is to be carried out on the basis of clicking rate Crowd's orientation, or it is equal to or less than clicking rate for the weight of joint exposure rate and clicking rate progress crowd's orientation and exposure rate Weight, then using there are the attribute datas of the historical user of ad click behavior as the category for referring to user in history dispensing flowing water table Property data；If crowd's directional aim is joint exposure rate and clicking rate carries out crowd's orientation and the weight of exposure rate is greater than clicking rate Weight, then will launch in flowing water table that there are the attribute datas of the historical user of advertisement exposure behavior to be sampled to obtain to history Attribute data as refer to user attribute data, and by history launch flowing water table in there are the history of ad click behavior use The attribute data at family is as the attribute data for referring to user.

It should be noted that history launches the time tag that may also include the behavioral data of historical user in flowing water table, it should Time tag can be used for identifying the time for generating behavioral data, i.e., there are advertisement exposure behavior or ad clicks by mark historical user The time of behavior；Correspondingly, server is launching the category obtained in flowing water table with reference to user according to crowd's directional aim from history Property data when, may also be combined with the time tag and crowd's directional aim, launch in flowing water table and obtained with reference to user from history Attribute data.For example, history can be launched flowing water if crowd's directional aim is to carry out crowd's orientation on the basis of exposure rate There are the categories of the historical user of advertisement exposure behavior for table near-mid term (being calculated in preset time period forward based on present system time) Property data as refer to user attribute data.

In another embodiment, the reference user collection of targeted advertisements can also include that following at least two mode is got Reference user attribute data: advertiser oneself upload, third client active reporting and server according to target What the history input situation of advertisement was collected automatically.50 attribute datas with reference to user are needed for example, collecting with reference to user, and advertisement The attribute data of 30 seed users has been only included in the seed user list that Your Majesty passes；So server is in addition to will be on advertiser The attribute data of 30 seed users in the seed user list of biography is added to as the attribute data with reference to user with reference to use Family is concentrated, and the attribute data that 20 seed users can be also obtained from the seed user list of third party's client active reporting is made It is added to for the attribute data with reference to user and is concentrated with reference to user；Or server can also be according to crowd's directional aim from going through The attribute data that history launches 20 historical users of acquisition in flowing water table is added to as the attribute data with reference to user with reference to user It concentrates.For another example, collect with reference to user and need 100 attribute datas with reference to user, and in the seed user list that advertiser uploads It only included the attribute data of 30 seed users, only included 20 kinds in the seed user list that third party's client reports The attribute data of child user；So category of the server in addition to 30 seed users in the seed user list that uploads advertiser Property the seed user list that reports of data and third party's client in 20 seed users as the attribute number for referring to user It is concentrated according to being added to reference to user, can also be launched according to crowd's directional aim from history and obtain 50 in flowing water table with reference to user's Attribute data is added to reference to user's concentration, etc..

It should be noted that can be that above-mentioned three kinds of modes are (logical previously according to practical business demand in the specific implementation process Cross that the seed user list of advertiser's upload obtains, the seed user list reported by third client obtains and according to The history input situation of targeted advertisements is collected automatically) setting priority, in order to it is subsequent can be successively independent according to priority orders The attribute data with reference to user is obtained using or in conjunction with using above-mentioned three kinds of modes.

S402 is trained optimization to prediction model using reference user collection and candidate user collection, and what is optimized estimates Model.

In embodiments of the present invention, prediction model, which is one, to estimate user's progress advertisement orientation, to obtain user couple Targeted advertisements generate the model of the probability of positive feedback.In order to improve the performance of prediction model, the embodiment of the present invention uses people Machine learning (Machine Learning, ML) technology pair in work intelligence field (Artificial Intelligence, AI) Prediction model is trained optimization, estimates, mentions so that the prediction model of optimization preferably can carry out advertisement orientation to user High accuracy.Artificial intelligence herein is machine simulation, extension and the expansion controlled using digital computer or digital computer The intelligence of people is opened up, perception environment obtains knowledge and theory, method, technology and application system using Knowledge Acquirement optimum； Machine learning in artificial intelligence field is the core of artificial intelligence, specifically can be understood as a multi-field intersection and learns Section is related to the multiple subjects such as probability theory, statistics, Approximation Theory, convextiry analysis, algorithm complexity theory.Using machine learning skill During art is trained optimization to prediction model, prediction model can be by studying the mankind's simulated or realized to computer how Learning behavior obtains new knowledge or skills, then reorganizes the existing structure of knowledge and is allowed to constantly improve the property of itself Energy.In the specific implementation process, step S402 may include following steps s11-s13:

Reference user is concentrated each attribute data with reference to user as positive sample by s11.

It, can be by reference user's since reference user is the user for referring to generate targeted advertisements positive feedback Attribute data is as positive sample；Specifically, merging features can be carried out to the attribute data of reference user, to obtain positive sample This.The quantity of positive sample is identical with the quantity with reference to user；In other embodiments, if the quantity with reference to user is more, Progress random sampling can be collected to reference user according to actual needs and obtain positive sample, the quantity of the positive sample in the case of this is less than With reference to the quantity of user.

S12 is sampled to obtain more according to crowd's directional aim of targeted advertisements to reference user collection and candidate user collection A negative sample.

Candidate user collection may also include the behavioral data of each candidate user, and the behavioral data of the candidate user is used to indicate time Family is selected to whether there is advertisement exposure behavior and ad click behavior for targeted advertisements.In practical application scene, negative sample The attribute data that the reference user for concentrating sampling to obtain from reference user can be only included, can also only include from candidate user collection The attribute data of middle obtained candidate user of sampling.Specifically, if crowd's directional aim only values exposure rate, in order to avoid instruction Practice space and prediction space generates spatial offset phenomenon, the candidate user that advertisement exposure behavior is not present can be concentrated from candidate user Attribute data in sampling obtain negative sample；That is, if crowd's directional aim is to carry out crowd on the basis of exposure rate to determine To the attribute data for the candidate user that then candidate user concentration can be not present advertisement exposure behavior carries out random sampling, obtain Negative sample.If crowd's directional aim only values clicking rate, can not concentrated there are advertisement exposure behavior and not from reference user directly There are sampling in the attribute data of the reference user of ad click behavior to obtain negative sample；That is, if crowd's directional aim To carry out crowd's orientation on the basis of clicking rate, then reference user can be concentrated there are advertisement exposure behavior and advertisement point is not present The attribute data for hitting the reference user of behavior carries out random sampling, obtains negative sample.

Due to being born if directly concentrating to sample from reference user in the case where crowd's directional aim only values clicking rate Sample then may result in and generate spatial offset phenomenon between trained space and prediction space, leads to the clicking rate of targeted advertisements It is lower.And from the foregoing it will be appreciated that the main reason for causing this problem be since include is candidate user in prediction space, and it is candidate User cannot be guaranteed expose；Therefore, in order to solve this problem, the embodiment of the present invention is being examined during crowd orients Candidate user is considered to targeted advertisements with the presence or absence of before ad click behavior, also further contemplates candidate user to targeted advertisements With the presence or absence of advertisement exposure behavior.Based on this, the embodiment of the present invention carries out people on the basis of crowd's directional aim is by clicking rate In the case where group's orientation, crowd's orientation mesh of progress crowd's orientation on the basis of joint clicking rate and exposure rate is further provided Mark；Under this crowd's directional aim, to candidate user carry out click estimate when, need first to estimate candidate user to targeted advertisements There are the exposure probability of advertisement exposure behavior, then predict that there are the click of ad click behavior is general to targeted advertisements after its exposure Rate carries out advertisement orientation to candidate user and estimates and (estimate herein to click) obtained orientation probability=exposure probability * click Probability.

In order to further prove that, when crowd's directional aim is joint exposure rate and clicking rate carries out crowd's orientation, candidate uses The orientation probability at family is equal to exposure probability and clicks this relationship of the product of probability, and the embodiment of the present invention has carried out mathematics to this and pushed away Lead: assuming that x represents candidate user, y is there are advertisement exposure behavior (y=1 indicates exposure), and z is that there are click behavior (z=1 tables Show click)；So, the exposure probability of candidate user x and click probability are as follows:

The exposure probability of candidate user x are as follows:

The click probability of candidate user x are as follows:

It calculates exposure probability and clicks the product of probability: It can be seen that the final orientation probability that ad click behavior occurs is just for candidate user x It is the product for exposing probability and clicking probability both.

Therefore, when crowd's directional aim is to carry out crowd's orientation on the basis of clicking rate, the embodiment of the present invention can also be into Crowd's directional aim is transformed into joint exposure rate to one step and clicking rate carries out crowd's orientation；Based on crowd's directional aim, It also proposed and optimization is trained to prediction model up to the method that rate (exposure rate) and clicking rate joint training optimize using touching.Tool Body, use there are advertisement exposure behavior and there are the attribute datas of the reference user of ad click behavior as positive sample, it is right Candidate user, which is concentrated, to be carried out random sampling and obtains negative sample, then using under this mode positive sample and negative sample to prediction model It is trained optimization.It is trained the prediction model for optimizing obtained optimization using this training set, candidate user is carried out wide The desired value for accusing the orientation probability that orientation is estimated is as follows:

It, can be to a certain degree since the candidate user of sampling is to carry out random sampling to candidate user collection to obtain It is upper to indicate all candidate users.Therefore, it is desirable to which being worth can further indicate that are as follows:

Wherein, exposure click number of users refers to there are advertisement exposure behavior and there are the reference users of ad click behavior Quantity, the candidate user quantity of sampling refer to candidate user collection carry out the obtained candidate user of random sampling quantity. It is used it can be seen that any one candidate can be estimated using this mode for obtaining negative sample to the progress random sampling of candidate user collection Family eventually generates the probability of ad click behavior to targeted advertisements.(rate (exposure rate) is reached using touching in this training method With the mode of clicking rate joint training) under, the training space and prediction space of prediction model are consistent, and can effectively solve low exposure The problem of light rate and low clicking rate.The orientation crowd that the prediction model of obtained optimization is determined in the case where calling this training method Data are more accurate, so that subsequent during launching targeted advertisements, the overall exposing rate and clicking rate of targeted advertisements are obtained Effective promotion is arrived, especially exposure rate integrally improves 60%.

Based on the description above, when crowd's directional aim is joint clicking rate and exposure rate carries out crowd's orientation, negative sample Originally it can only include from candidate user and concentrate the attribute data of candidate user being sampled, be also possible to that there are advertisement exposures Light behavior and there is no the attribute data of the reference user of ad click behavior and the times that are sampled to candidate user collection The mixing for selecting the attribute data at family, the accounting of which part may depend on more greatly the relatively heavy of clicking rate and exposure rate in mixing Want degree.In the specific implementation, if clicking rate and exposure rate is of equal importance or crowd's directional aim is more biased to exposure rate, Directly sampling can be concentrated to obtain negative sample from candidate user；That is, if crowd's directional aim is joint exposure rate and point Hit rate and carry out the weight of crowd's orientation and exposure rate and be equal to or more than the weight of clicking rate, then candidate user collection can be carried out with Machine sampling, obtains negative sample.If crowd's directional aim is more biased to clicking rate, can be carried out at random to candidate user collection On the basis of sampling, be added it is some there are advertisement exposure behavior and there is no ad click behavior reference user attribute number According to；That is, if crowd's directional aim is joint exposure rate and clicking rate carries out crowd's orientation and the weight of exposure rate is less than The weight of clicking rate then carries out the attribute data of candidate user that random sampling obtains as negative sample for candidate user collection, And by reference user concentration is there are advertisement exposure behavior and there is no the progress of the attribute data of the reference user of ad click behavior The attribute data that random sampling obtains is as negative sample.

It should be noted that in one embodiment, when the quantity of candidate user be much larger than with reference to user quantity (just The quantity of sample) when；For the quantitative proportion of balanced positive negative sample, so as to preferably carry out subsequent training optimization；This hair Bright embodiment is sampled to obtain multiple in crowd's directional aim according to targeted advertisements to reference user collection and candidate user collection When negative sample, the quantity of positive sample and crowd's directional aim of targeted advertisements may also be combined with, to reference user collection and candidate user Collection is sampled to obtain multiple negative samples, so that the quantitative proportion between negative sample and positive sample meets default ratio；This is pre- If ratio can be arranged according to actual business demand or empirical value, such as it is 1,1.5 etc. that default ratio, which is arranged,.

In another embodiment, due to the presence of candidate user to the behavior of targeted advertisements advertisement exposure and may deposit In ad click behavior, therefore in order to avoid when being sampled to candidate user collection, the attribute of such candidate user is arrived in sampling Data cause prediction model mistake in training optimization to learn the data characteristics to negative sample, to influence pre- as negative sample Estimate the training effect of optimization of model；The embodiment of the present invention in crowd's directional aim according to targeted advertisements, to reference user collection and When candidate user collection is sampled to obtain multiple negative samples, can also first removing candidate user concentration, there are advertisement exposures to targeted advertisements Light behavior and there are the attribute datas of the candidate user of ad click behavior, obtains remaining users collection；Then according to targeted advertisements Crowd's directional aim be sampled to remaining users collection and with reference to user's collection, obtain multiple negative samples, or combine positive sample Quantity and targeted advertisements crowd's directional aim, to remaining users collection and with reference to user collection be sampled to obtain multiple negative samples This.

The related content as documented by step S401-S402 is as it can be seen that the embodiment of the present invention can be according to crowd's directional aim Difference selects different positive negative samples；Under different crowd's directional aims, the selection mode of positive negative sample can be as shown in table 1:

Table 1

Wherein, exposure data refers to that there are the attribute numbers of the historical user of advertisement exposure behavior in advertisement dispensing flowing water table According to click data refers to that advertisement is launched there are the attribute data of the historical user of ad click behavior in flowing water table, and sampling exposes Obtained by data refer to being sampled in history dispensing flowing water table there are the attribute data of the historical user of advertisement exposure behavior Attribute data, unexposed data refer to candidate user concentrate there is no advertisement exposure behavior candidate user attribute data, Exposure do not click on data refer to advertisement launch flowing water table in there are advertisement exposure behavior and be not present ad click behavior history The attribute data of user, data from the sample survey, which refers to, is sampled obtained attribute data to candidate user collection.

S13 is trained optimization to prediction model using multiple positive samples and multiple negative samples, and what is optimized estimates mould Type.

In the specific implementation process, after obtaining multiple positive samples and multiple negative samples, model instruction can be directly based upon Practice algorithm, optimization is trained to prediction model using multiple positive samples and multiple negative samples, the prediction model optimized.Its In, model training algorithm may include but be not limited to: xgboost algorithm (extreme gradient ascent algorithm), GBDT algorithm (Gradient Boosting Decision Tree, gradient promote decision Tree algorithms), etc..Wherein, xgboost algorithm is a kind of using ladder Degree enhancing frame and the ensemble machine learning algorithm based on decision tree, can specifically be made of, decision herein multiple decision trees Tree is Taxonomy and distribution (Classification and regression tree, CART), and CART decision is one two Fork tree, the value of internal node feature are "Yes" and "No", can be using the branch that the value of each node is "Yes" as the node Left branch, using value be "No" branch as the right branch of the node；The basic thought of xgboost algorithm are as follows: according to sample This feature gradually constructs multiple decision trees, when one decision tree of every building, is intended to so that the overall effect of model is promoted, The decision tree for declining the functional value of loss function, and currently constructing is fitted caused by the decision tree of previous building Residual error.GBDT algorithm is a kind of decision Tree algorithms of iteration, which is made of more decision trees, and algorithm is exported final As a result it is obtained by the conclusion of all decision trees is cumulative.

In one embodiment, due in negative sample there may be negative sample similar with positive sample, and it is such with just The similar negative sample of sample actually and is not belonging to real negative sample；Therefore in order to avoid using positive sample and negative sample pair During prediction model is trained optimization, prediction model is accidentally using such negative sample similar with positive sample as really negative Sample is learnt, and causes prediction model that can not accurately distinguish the difference between positive negative sample, and the training for influencing prediction model is excellent Change effect；The embodiment of the present invention is being trained optimization to prediction model using multiple positive samples and multiple negative samples, obtains excellent During the specific implementation of the prediction model of change, also first multiple negative samples can be carried out at sample cleaning according to multiple positive samples Reason, sample cleaning treatment herein refer to the processing for removing negative sample similar with positive sample in multiple negative samples；Then it uses Multiple positive samples and the cleaned negative sample of sample are trained optimization to prediction model, the prediction model optimized.

Wherein, it may is that according to a kind of embodiment that multiple positive samples carry out sample cleaning treatment to multiple negative samples It firstly, multiple positive samples can be arbitrarily divided into the first positive sample set and the second positive sample set, and is the first positive sample The first label of set addition (value of the first label is 1)；Using in the second positive sample set positive sample and multiple negative samples Construct negative sample set, and the sample set that is negative adds the second label (value of the second label is -1).Secondly, just using first Sample set and the first label and negative sample set and the second label train classification models；Using the disaggregated model after training Sample class prediction is carried out to each negative sample, obtains the prediction probability that each negative sample is positive sample.Since prediction probability is got over Greatly, indicate that the similarity of negative sample and positive sample is bigger；Prediction probability is smaller, indicates that the similarity of negative sample and positive sample is got over It is small；Therefore, after obtaining the prediction probability that each negative sample is positive sample, prediction probability can be less than to the negative sample of probability threshold value As the negative sample after sample cleaning.

Alternatively, can be with according to the another embodiment that multiple positive samples carry out sample cleaning treatment to multiple negative samples Be: firstly, construct positive sample set using multiple positive samples, and the sample set that is positive adds the first label (first label takes 1) value is；Negative sample set is constructed using multiple negative samples, and sample set addition the second label (value of the second label that is negative It is -1).Secondly, using positive sample set and the first label and negative sample set and the second label train classification models；Using Disaggregated model after training carries out sample class prediction to each negative sample, and it is general to obtain the prediction that each negative sample is positive sample Rate.Then, prediction probability can be less than to the negative sample of probability threshold value as the negative sample after sample cleaning.

Alternatively, may is that according to a kind of embodiment that multiple positive samples carry out sample cleaning treatment to multiple negative samples Using 1-DNF algorithm (a kind of sample cleaning algorithm), sample cleaning treatment is carried out to multiple negative samples according to multiple positive samples.

S403 calls the prediction model of optimization to concentrate the attribute data of each candidate user to each candidate use according to candidate user Family carries out advertisement orientation and estimates, and obtains the orientation probability of each candidate user.

After the prediction model optimized, the prediction model that can call directly optimization concentrates each time according to candidate user It selects the attribute data at family to carry out advertisement orientation to each candidate user to estimate, obtains the orientation probability of each candidate user.

It should be noted that being directed to above-mentioned steps S402-S403: in other embodiments, when crowd's directional aim is connection When closing exposure rate and clicking rate progress crowd's orientation, prediction model can respectively include exposure prediction model and click prediction model etc. Two models.In this case, can be used with reference to user's collection and candidate user collection in step S402 respectively to exposure prediction model It is trained optimization with prediction model is clicked, the click prediction model of the exposure prediction model and optimization that are optimized.Specific In implementation process, for exposure prediction model: can from history launch flowing water table in choose there are the history of advertisement exposure behavior The attribute data of user concentrates the candidate user there is no advertisement exposure behavior as exposure positive sample, and to candidate user Attribute data carries out random sampling and obtains exposure negative sample；Using exposure positive sample and exposure negative sample to exposure prediction model into Row training optimization, the exposure prediction model optimized.For click prediction model: can launch in flowing water table and choose from history There are the attribute datas of the historical user of ad click behavior as click positive sample, and there are advertisements to reference user concentration Exposure behavior and there is no ad click behavior reference user attribute data carry out random sampling obtain click negative sample；It adopts It is trained optimization to prediction model is clicked with clicking positive sample and clicking negative sample, the click prediction model optimized.

Correspondingly, for any candidate user, the exposure prediction model of optimization can be called according to the time in step S403 It selects the attribute data at family to be exposed candidate user to estimate, obtains the exposure rate of the candidate user；And call optimization Click prediction model click to candidate user and be estimated according to the attribute data of the candidate user, obtains the point of the candidate user Hit rate；Then the product for calculating the exposure probability of candidate user and the click probability of the candidate user, obtains the candidate user Orient probability.Repeat above-mentioned steps, the orientation probability of available each candidate user.

S404 concentrates the attribute data for filtering out directional user according to the orientation probability of each candidate user from candidate user, And the attribute data of directional user is added in the orientation demographic data of targeted advertisements.

S405 stores the orientation demographic data of targeted advertisements, and the attribute information of output directional demographic data.

After obtaining the orientation demographic data of targeted advertisements, server can store the orientation demographic data, with Advertisement dispensing is carried out based on the orientation demographic data convenient for subsequent.In one embodiment, server can be directly by target The orientation demographic data of advertisement is stored into database；In another embodiment, server can also determine targeted advertisements Compression processing is carried out to demographic data, the orientation demographic data after compression processing is stored into database, to save database Memory space.

In addition, server after obtaining the orientation demographic data of targeted advertisements, can also obtain the orientation demographic data Attribute information, and export the attribute information.Wherein, attribute information includes at least: data name and data bulk；Data herein Title can be what server was randomly provided, it is pre- to be also possible to advertiser for identifying the orientation demographic data this time obtained First it is arranged, such as data 1, data a.Data bulk refers to the quantity of directional user corresponding to orientation demographic data；For example, Orientation demographic data includes the attribute data of 5 directional users, then the quantity of directional user corresponding to orientation demographic data It is then 5, i.e. data bulk is just 5；For another example, orientation demographic data includes the attribute data of 100 directional users, then orientation people The quantity of directional user corresponding to group's data is then 100, i.e., data bulk is just 100.

S406, if detecting the trigger event for carrying out advertisement dispensing based on orientation demographic data, in orientation demographic data Targeted advertisements are launched in corresponding orientation crowd.

In one embodiment, advertiser can choose when to carry out advertisement dispensing according to the demand of itself；This situation Under, the trigger event that advertisement is launched may include: to detect that advertiser's selection carries out advertisement dispensing based on the orientation demographic data Advertisement launch operation event.In another embodiment, server can also carry out advertisement dispensing automatically；In this case, The trigger event that advertisement is launched may include: to detect the event for the advertisement fixed condition for meeting targeted advertisements；Advertisement herein Fixed condition may include but be not limited to: getting the condition of orientation demographic data or gets orientation demographic data and process Condition after preset duration.

Server upon detecting a triggering event, can be in response to the trigger event, corresponding to orientation demographic data Orientation crowd in launch targeted advertisements.Specifically, user's mark of each directional user in available orientation demographic data Know, targeted advertisements are issued in user account associated by the user identifier of each directional user, to realize in orientation crowd Middle dispensing targeted advertisements.

Based on the description above, the embodiment of the present invention also proposed a kind of based on the progress advertisement throwing of above-mentioned crowd's orientation method The application scenarios put；In this application scenarios, prediction model is carried out in a manner of using exposure rate and clicking rate joint training It is illustrated for training optimization.When advertiser wants to launch targeted advertisements A, ad system can be logged in, as shown in Figure 5 a. After successfully logging in ad system, advertiser can click the creation button in user interface to enter the setting of stereotactic conditions Interface, as shown in Figure 5 b.In the set interface, it may include multiple stereotactic conditions: seed crowd, orientation crowd's quantity, whether Include seed crowd, extension tendency (crowd's directional aim), advertising media's platform, the type of number, etc..Wherein, seed crowd Three options can at least be corresponded to: advertiser uploads, third party's client reports, server automatically determines.Extension tendency is for true Determine crowd's directional aim, such as when extension tendency is advertisement exposure, then crowd's directional aim is that crowd is carried out on the basis of exposure rate Orientation；When extension tendency is ad click, then crowd's directional aim is that crowd's orientation, etc. is carried out on the basis of clicking rate.

Advertiser is configured this multiple stereotactic conditions in combination with itself want advertisement.For example, advertiser's setting Stereotactic conditions are as follows: server automatically determine seed crowd, orientation crowd's quantity be 500 people, not comprising seed crowd, by making by oneself Right way of conduct formula setting extension tendency is that advertisement exposure is clicked, release platform is unlimited, the type of number is QQ number.Advertiser is setting this After a little stereotactic conditions, submitting button can be clicked, the front end of ad system can generate people according to the setting information of advertiser at this time Group's orientation request, and the crowd is oriented into request and is sent to server.Crowd herein orients request and at least carries following information: (joint exposure rate and clicking rate carry out crowd's orientation and exposure for orientation crowd quantity (500 people), crowd's directional aim of targeted advertisements The weight of light rate is equal to the weight of clicking rate), the platform identification of ad system etc..Correspondingly, server can receive the crowd Orientation request, and orient and request in response to the crowd, subsequent crowd's directional process is carried out to obtain the orientation people of targeted advertisements Group's data；The specific implementation process of its crowd's directional process can be together referring to shown in Fig. 5 c:

Server can obtain crowd's directional aim of targeted advertisements；Specifically, the side of request can be oriented by parsing crowd Formula obtains crowd's directional aim.After getting crowd's directional aim, flowing water can be launched from advertisement according to crowd's directional aim Positive sample is obtained in table；Since crowd's directional aim is the weight that joint exposure rate and clicking rate carry out crowd's orientation and exposure rate Equal to the weight of clicking rate, therefore can be by there are the attribute datas of the historical user of ad click behavior in history dispensing flowing water table As the attribute data of reference user, merging features are carried out to the attribute data of reference user and obtain positive sample.Then in order to equal Quantity between the positive sample that weighs and negative sample, can determine the quantity of negative sample according to the quantity of positive sample.Since crowd orients Request carries the platform identification of ad system, therefore the attribute data structure of the active advertising user of ad system near-mid term can be used Build candidate user collection.After obtaining the quantity of negative sample, request can be oriented according to crowd and concentrate sampling to obtain from candidate user The negative sample of respective numbers.Specifically, being that joint exposure rate and clicking rate carry out crowd's orientation and exposure since crowd orients request The weight of light rate is equal to the weight of clicking rate, therefore can carry out random sampling to candidate user collection, obtains negative sample.

Due to being likely to be present in the similar negative sample of positive sample in negative sample, can also sample cleaning be carried out to negative sample To remove negative sample similar with positive sample in negative sample.It, can be using negative after sample cleaning after carrying out sample cleaning Sample and positive sample carry out the optimization training of model, advertisement orientation is estimated, orient a series of processing of demographic data generation.Specifically , negative sample and positive sample after sample cleaning can be used are trained optimization to prediction model, the prediction model optimized； It is estimated secondly, the prediction model of optimization is called to carry out advertisement orientation to each candidate user according to the attribute data of each candidate user, Obtain the orientation probability of each candidate user.Then, each time can be concentrated to candidate user according to the sequence of orientation probability from big to small It selects the attribute data at family to be ranked up, obtains ordered set；According to orientation crowd's quantity (i.e. 500 people) successively from ordered set Attribute data of the middle attribute data for choosing corresponding candidate user as directional user, and used obtained each orientation is chosen The attribute data at family is added in orientation demographic data.Server can store crowd orientation after obtaining orientation demographic data Data, and the attribute information of output directional demographic data, as fig 5d.

After advertiser can get this orientation demographic data, if thinking directly to carry out advertisement dispensing, it can pass through a little It hits " dispensing " button trigger the server and is directly based upon the advertisement dispensing that orientation crowd's number carries out targeted advertisements.If advertiser is at this time It is not desired to carry out advertisement dispensing, button " can be exited " by click and exit ad system, it can also be by clicking the Back button The set interface of stereotactic conditions is again returned to, and is once again set up different stereotactic conditions.Resetting different orientation bars After part, front end can be still triggered by click " submission " button and sends crowd's orientation request to server again；Correspondingly, service Device can be oriented again according to the crowd transmitted by front end and be requested, and carried out crowd's directional process again and obtained another orientation crowd's number According to, and store the orientation demographic data obtained again and export the attribute information of the orientation demographic data obtained again (the entitled orientation crowd 2 of such as data, data bulk are 100 people) ... and so on.

When advertiser wants to launch targeted advertisements, it can choose at least one orientation demographic data and carry out advertisement dispensing, As depicted in fig. 5e；If front end detects that advertiser clicks the operation of " dispensing " button, it can send what advertisement was launched to server Trigger event.Correspondingly, server is after detecting the trigger event, it can be according to the selected orientation demographic data of advertiser Carry out advertisement dispensing；Specifically, server can obtain each directional user's in the selected orientation demographic data of advertiser Targeted advertisements are issued in user account associated by the user identifier of each directional user by user identifier, to realize fixed The targeted advertisements are launched into crowd.Optionally, after server successfully launches targeted advertisements, exportable feedback information is such as schemed Shown in 5e.

It, can be first using positive sample and negative sample to estimating it can be seen that the embodiment of the present invention is during advertisement is launched Model is trained optimization, then call the prediction model of optimization according to the attribute data of each candidate user to each candidate user into Row advertisement orientation is estimated, and the orientation probability of each candidate user is obtained.Since negative sample is to concentrate sampling to obtain from candidate user, Therefore it can avoid training space and predict the phenomenon that space generates spatial offset, it is ensured that the training space of prediction model and prediction Space is consistent, to improve the accuracy of the orientation probability of each candidate user.Since orientation probability refers to candidate user pair Targeted advertisements generate positive feedback probability, therefore can be concentrated according to the orientation probability of each candidate user from candidate user and screen and add Add the attribute data of directional user into the orientation demographic data of targeted advertisements；It, can be from by improving the accuracy of orientation probability And the accuracy of orientation demographic data is improved, and then the exposure rate and clicking rate of targeted advertisements can be improved.

Based on the description of above-mentioned crowd's orientation method embodiment, orients and fill the embodiment of the invention also discloses a kind of crowd It sets, crowd's orienting device can be operate in a computer program (including program code) in server.The crowd Orienting device can execute Fig. 2 or method shown in Fig. 4.Fig. 6 is referred to, crowd's orienting device, which can be run, such as to place an order Member:

Acquiring unit 101, the reference user for obtaining targeted advertisements collects and candidate user collection；It is described to refer to user Ji Bao Include multiple attribute datas with reference to user, described with reference to user is the use for referring to generate the targeted advertisements positive feedback Family；The candidate user collection includes the attribute data of multiple candidate users, and the candidate user is user to be oriented；

Optimize unit 102, for being trained with reference to user's collection and the candidate user collection to prediction model using described Optimization, the prediction model optimized；

Processing unit 103, for calling the prediction model of the optimization to concentrate each candidate user according to the candidate user Attribute data advertisement orientation carried out to each candidate user estimate, obtain the orientation probability of each candidate user, it is described Orientation probability refers to that candidate user generates the probability of positive feedback to the targeted advertisements；

The processing unit 103 is concentrated from the candidate user for the orientation probability according to each candidate user and is sieved The attribute data of directional user is selected, and the attribute data of the directional user is added to the orientation crowd of the targeted advertisements In data.

In one embodiment, acquiring unit 101 is specific to use in the reference user collection for obtaining targeted advertisements In:

The crowd for receiving targeted advertisements orients request, and the crowd orients request and carries seed user list, the seed It include the attribute data of multiple seed users in user list, the user list is to be uploaded either by advertiser by third What square client reported, third party's client is to refer to carry out at business the corresponding audience of the targeted advertisements The client of reason；

It is added to the attribute data of each seed user in the seed user list as the attribute data with reference to user The reference user of targeted advertisements concentrates.

In another embodiment, acquiring unit 101 is specific to use in the reference user collection for obtaining targeted advertisements In:

The crowd for receiving targeted advertisements orients request, and the crowd orients crowd's orientation that request carries the targeted advertisements Target；

The history for obtaining the targeted advertisements launches flowing water table, and the history, which is launched, includes at least historical user in flowing water table Attribute data and behavioral data, the behavioral data of the historical user be used to indicate the historical user for the target it is wide It accuses and whether there is advertisement exposure behavior and ad click behavior；

It is launched according to crowd's directional aim from the history and obtains the attribute data for referring to user in flowing water table, and will The reference user that the attribute data with reference to user is added to the targeted advertisements concentrates.

In another embodiment, acquiring unit 101 according to crowd's directional aim from the history for launching When obtaining the attribute data for referring to user in flowing water table, it is specifically used for:

If crowd's directional aim is to carry out crowd's orientation on the basis of exposure rate, the history is launched into flowing water table It is middle that there are the attribute datas of the historical user of advertisement exposure behavior as the attribute data for referring to user；

If crowd's directional aim is progress crowd's orientation on the basis of clicking rate, or is joint exposure rate and click The weight of rate progress crowd's orientation and exposure rate is equal to or less than the weight of clicking rate, then launches the history and deposit in flowing water table Ad click behavior historical user attribute data as refer to user attribute data；

If crowd's directional aim is joint exposure rate and clicking rate carries out crowd's orientation and the weight of exposure rate is greater than The weight of clicking rate, then by the history launch flowing water table in there are the attribute data of the historical user of advertisement exposure behavior into The attribute data that line sampling obtains as refer to user attribute data, and by the history launch flowing water table in there are advertisement points The attribute data of the historical user of behavior is hit as the attribute data for referring to user.

In another embodiment, optimization unit 102 is for using the reference user collection and the candidate user collection Optimization is trained to prediction model, when the prediction model optimized, is specifically used for:

Concentrate each attribute data with reference to user as positive sample, the quantity and ginseng of the positive sample with reference to user for described The quantity for examining user is identical；

According to crowd's directional aim of the targeted advertisements, taken out to described with reference to user's collection and the candidate user collection Sample obtains multiple negative samples；

Optimization is trained to the prediction model using multiple positive samples and the multiple negative sample, what is optimized is pre- Estimate model.

In another embodiment, the candidate user collection further includes the behavioral data of each candidate user, the candidate use The behavioral data at family is used to indicate the candidate user for the targeted advertisements with the presence or absence of advertisement exposure behavior and advertisement point Hit behavior；Correspondingly, optimization unit 102 refers to user to described for crowd's directional aim according to the targeted advertisements When collection and the candidate user collection are sampled to obtain multiple negative samples, it is specifically used for:

If crowd's directional aim is to carry out crowd's orientation on the basis of exposure rate, the candidate user is concentrated not There are the attribute datas of the candidate user of advertisement exposure behavior to carry out random sampling, obtains negative sample；

If crowd's directional aim is to carry out crowd's orientation on the basis of clicking rate, described concentrate with reference to user is deposited In advertisement exposure behavior and there is no the attribute datas of the reference user of ad click behavior to carry out random sampling, obtains negative sample This；

If crowd's directional aim is joint exposure rate and clicking rate carries out crowd's orientation and the weight of exposure rate is equal to Or the weight greater than clicking rate, then random sampling is carried out to the candidate user collection, obtains negative sample；

If crowd's directional aim is joint exposure rate and clicking rate carries out crowd's orientation and the weight of exposure rate is less than The weight of clicking rate then will carry out the attribute data for the candidate user that random sampling obtains to the candidate user collection as negative sample This, and by described with reference to user's concentration is there are advertisement exposure behavior and there is no the attribute numbers of the reference user of ad click behavior The attribute data obtained according to progress random sampling is as negative sample.

In another embodiment, optimization unit 102 is for using multiple positive samples and the multiple negative sample to institute It states prediction model and is trained optimization, when the prediction model optimized, be specifically used for:

Sample cleaning treatment is carried out to the multiple negative sample according to the multiple positive sample, the sample cleaning treatment is Refer to the processing for removing negative sample similar with positive sample in the multiple negative sample；

Optimization is trained to the prediction model using the multiple positive sample and the cleaned negative sample of sample, The prediction model optimized.

In another embodiment, the crowd orients request and carries orientation crowd quantity；Correspondingly, processing unit 103 The attribute data for filtering out directional user is being concentrated from the candidate user for the orientation probability according to each candidate user When, it is specifically used for:

According to the sequence of orientation probability from big to small, the attribute data of each candidate user is concentrated to carry out the candidate user Sequence, obtains ordered set；

The attribute data of corresponding candidate user is successively chosen from the ordered set according to the orientation crowd quantity Attribute data as directional user.

In another embodiment, the targeted advertisements are directed to each with reference to user by crowd's orientation request also carrying Instruction information；Correspondingly, processing unit 103 according to the orientation crowd quantity from the ordered set for successively selecting When taking attribute data of the attribute data of corresponding candidate user as directional user, it is specifically used for:

The difference between the orientation crowd quantity and the quantity with reference to user is sought, screening quantity is obtained；And root The attribute data of corresponding candidate user is successively chosen from the ordered set as directional user's according to the screening quantity Attribute data；

Processing unit 103 can also be used in: each attribute data with reference to user concentrated with reference to user is added to institute It states in the orientation demographic data of targeted advertisements.

In another embodiment, processing unit 103 can also be used in:

The orientation demographic data of the targeted advertisements is stored, and exports the attribute information of the orientation demographic data, it is described Attribute information includes at least: data name and data bulk；

If the trigger event for carrying out advertisement dispensing based on the orientation demographic data is detected, in the orientation crowd number According to launching the targeted advertisements in corresponding orientation crowd.

According to one embodiment of present invention, each step involved in Fig. 2 or method shown in Fig. 4 may each be by scheming Each unit in crowd's orienting device shown in 6 is performed.For example, step S201-S202 shown in Fig. 2 can distinguish The acquiring unit 101 shown in Fig. 6 is executed with optimization unit 102, and step S203 and S204 can be handled as shown in Fig. 6 Unit 103 executes；For another example, step S401-S402 shown in Fig. 4 can 101 He of acquiring unit as shown in Fig. 6 respectively Optimization unit 102 executes, and step S403-S406 processing unit 103 shown in Fig. 6 executes.

According to another embodiment of the invention, each unit in crowd's orienting device shown in fig. 6 can respectively or All one or several other units are merged into constitute or some (a little) unit therein can also be split as function again Smaller multiple units are constituted on energy, this may be implemented similarly to operate, and the technology without influencing the embodiment of the present invention is imitated The realization of fruit.Said units are logic-based function divisions, and in practical applications, the function of a unit can also be by multiple Unit is realized or the function of multiple units is realized by a unit.In other embodiments of the invention, fixed based on crowd It also may include other units to device, in practical applications, these functions can also be assisted to realize by other units, and can It is realized with being cooperated by multiple units.

It according to another embodiment of the invention, can be by including central processing unit (CPU), random access memory It is transported on the universal computing device of such as computer of the processing elements such as medium (RAM), read-only storage medium (ROM) and memory element Row is able to carry out the computer program (including program code) of each step involved in the correlation method as shown in Fig. 2 or Fig. 4, Construct crowd's orienting device equipment as shown in Figure 6, and come crowd's orientation method for realizing the embodiment of the present invention.It is described Computer program can be recorded in such as computer readable recording medium, and be loaded by computer readable recording medium It states and calculates in equipment, and run wherein.

Description based on above method embodiment and Installation practice, the embodiment of the present invention also provide a kind of server； The server can be the background server of above-mentioned mentioned ad system.Fig. 7 is referred to, which includes at least processing Device 201, communication interface 202 and computer storage medium 203.Wherein, communication interface 202 may include RF transceiver, service Device can be carried out data transmission by the RF transceiver with other equipment.Processor 201, communication interface 202 in server with And computer storage medium 203 can be connected by bus or other modes.

Computer storage medium 203 can store in the memory of server, and the computer storage medium 203 is used for Computer program is stored, the computer program includes program instruction, and the processor 201 is for executing the computer storage The program instruction that medium 203 stores.Processor 201 (or CPU (Central Processing Unit, central processing unit)) It is the calculating core and control core of server, is adapted for carrying out one or more instruction, is particularly adapted to load and executes one Item or a plurality of instruction are to realize correlation method process or corresponding function；In one embodiment, described in the embodiment of the present invention Processor 201 can be used for carrying out targeted advertisements a series of crowd's directional process, specific can include:

The embodiment of the invention also provides a kind of computer storage medium (Memory), the computer storage medium is clothes The memory device being engaged in device, for storing program and data.It is understood that computer storage medium herein both can wrap Include the built-in storage medium in server, naturally it is also possible to the expansion storage medium supported including server.Computer storage Medium provides memory space, which stores the operating system of server.Also, it is also housed in the memory space Suitable for by one or more instruction that processor 201 loads and executes, these instructions can be one or more meter Calculation machine program (including program code).It should be noted that computer storage medium herein can be high speed RAM memory, It is also possible to non-labile memory (non-volatile memory), for example, at least a magnetic disk storage；It is optional to go back It can be at least one computer storage medium for being located remotely from aforementioned processor.

In one embodiment, it can be loaded by processor 201 and execute one stored in computer storage medium or more Item instruction, to realize the above-mentioned corresponding steps in relation to the method in crowd's orientation method embodiment；In the specific implementation, computer is deposited One or more instruction in storage media is loaded by processor 201 and executes following steps:

In one embodiment, in the reference user collection for obtaining targeted advertisements, one or more instruction may be used also It is loaded by processor 201 and is specifically executed:

In another embodiment, in the reference user collection for obtaining targeted advertisements, one or more instruction may be used also It is loaded by processor 201 and is specifically executed:

In another embodiment, referred to launching to obtain in flowing water table from the history according to crowd's directional aim When the attribute data of user, one or more instruction can also be loaded by processor 201 and specifically be executed:

In another embodiment, prediction model is instructed with reference to user's collection and the candidate user collection using described Practice optimization, when the prediction model optimized, one or more instruction can also be loaded by processor 201 and specifically execution:

In another embodiment, the candidate user collection further includes the behavioral data of each candidate user, the candidate use The behavioral data at family is used to indicate the candidate user for the targeted advertisements with the presence or absence of advertisement exposure behavior and advertisement point Hit behavior；Correspondingly, in crowd's directional aim according to the targeted advertisements, to described with reference to user's collection and the candidate user When collection is sampled to obtain multiple negative samples, one or more instruction can also be loaded by processor 201 and specifically be executed:

In another embodiment, the prediction model is instructed using multiple positive samples and the multiple negative sample Practice optimization, when the prediction model optimized, one or more instruction can also be loaded by processor 201 and specifically execution:

In another embodiment, the crowd orients request and carries orientation crowd quantity；Correspondingly, according to described each When the orientation probability of candidate user filters out the attribute data of directional user from candidate user concentration, described one or more Instruction can also be loaded by processor 201 and specifically be executed:

In another embodiment, the targeted advertisements are directed to each with reference to user by crowd's orientation request also carrying Instruction information；Correspondingly, corresponding candidate use is successively chosen from the ordered set according to the orientation crowd quantity When attribute data of the attribute data at family as directional user, one or more instruction can also be loaded simultaneously by processor 201 It is specific to execute:

It is described one or more instruction can also by processor 201 load and specifically execute: by it is described with reference to user concentration Respectively it is added in the orientation demographic data of the targeted advertisements with reference to the attribute data of user.

In another embodiment, one or more instruction can also be loaded by processor 201 and specifically be executed:

The above disclosure is only the preferred embodiments of the present invention, cannot limit the right model of the present invention with this certainly It encloses, therefore equivalent changes made in accordance with the claims of the present invention, is still within the scope of the present invention.

Claims

1. a kind of crowd's orientation method characterized by comprising

Obtain the reference user collection and candidate user collection of targeted advertisements；It is described to include multiple attributes with reference to user with reference to user's collection Data, it is described with reference to user be refer to the targeted advertisements generate positive feedback user；The candidate user collection includes The attribute data of multiple candidate users, the candidate user are users to be oriented；

Optimization is trained to prediction model with reference to user's collection and the candidate user collection using described, what is optimized estimates mould Type；

The prediction model of the optimization is called to concentrate the attribute data of each candidate user to each time according to the candidate user It selects family to carry out advertisement orientation to estimate, obtains the orientation probability of each candidate user, the orientation probability refers to candidate user The targeted advertisements are generated with the probability of positive feedback；

The attribute data for filtering out directional user is concentrated from the candidate user according to the orientation probability of each candidate user, and The attribute data of the directional user is added in the orientation demographic data of the targeted advertisements.

2. the method as described in claim 1, which is characterized in that the reference user collection for obtaining targeted advertisements, comprising:

The crowd for receiving targeted advertisements orients request, and the crowd orients request and carries seed user list, the seed user It include the attribute data of multiple seed users in list, the user list is uploaded by advertiser either by third party visitor Family end reports, and third party's client refers to carry out business processing to the corresponding audience of the targeted advertisements Client；

Target is added to using the attribute data of each seed user in the seed user list as the attribute data with reference to user The reference user of advertisement concentrates.

3. the method as described in claim 1, which is characterized in that the reference user collection for obtaining targeted advertisements, comprising:

The crowd for receiving targeted advertisements orients request, and the crowd orients the crowd for requesting to carry the targeted advertisements and orients mesh Mark；

The history for obtaining the targeted advertisements launches flowing water table, and the history launches the category that historical user is included at least in flowing water table Property data and behavioral data, the behavioral data of the historical user, which is used to indicate the historical user for the targeted advertisements, is It is no that there are advertisement exposure behaviors and ad click behavior；

It is launched according to crowd's directional aim from the history and obtains the attribute data for referring to user in flowing water table, and will be described The reference user concentration of the targeted advertisements is added to reference to the attribute data of user.

4. method as claimed in claim 3, which is characterized in that described to be launched according to crowd's directional aim from the history The attribute data for referring to user is obtained in flowing water table, comprising:

If crowd's directional aim is to carry out crowd's orientation on the basis of exposure rate, the history is launched in flowing water table and is deposited Advertisement exposure behavior historical user attribute data as refer to user attribute data；

If crowd's directional aim be crowd's orientation is carried out on the basis of clicking rate, or for joint exposure rate and clicking rate into The weight of pedestrian group orientation and exposure rate is equal to or less than the weight of clicking rate, then launches the history in flowing water table in the presence of wide The attribute data of the historical user of click behavior is accused as the attribute data for referring to user；

If crowd's directional aim is joint exposure rate and clicking rate carries out crowd's orientation and the weight of exposure rate is greater than click The weight of rate, then will be to there are the attribute datas of the historical user of advertisement exposure behavior to take out in history dispensing flowing water table The attribute data that sample obtains as refer to user attribute data, and by the history launch flowing water table in there are ad click rows For historical user attribute data as refer to user attribute data.

5. the method as described in claim 1, which is characterized in that described to be collected and the candidate user collection using described with reference to user Optimization is trained to prediction model, the prediction model optimized, comprising:

Concentrate each attribute data with reference to user as positive sample, the quantity of the positive sample and with reference to use with reference to user for described The quantity at family is identical；

According to crowd's directional aim of the targeted advertisements, it is sampled to described with reference to user's collection and the candidate user collection To multiple negative samples；

Optimization is trained to the prediction model using multiple positive samples and the multiple negative sample, what is optimized estimates mould Type.

6. method as claimed in claim 5, which is characterized in that the candidate user collection further includes the behavior number of each candidate user According to the behavioral data of the candidate user is used to indicate the candidate user for the targeted advertisements with the presence or absence of advertisement exposure Behavior and ad click behavior；

Crowd's directional aim according to the targeted advertisements is taken out to described with reference to user's collection and the candidate user collection Sample obtains multiple negative samples, comprising:

If crowd's directional aim is to carry out crowd's orientation on the basis of exposure rate, candidate user concentration is not present The attribute data of the candidate user of advertisement exposure behavior carries out random sampling, obtains negative sample；

If crowd's directional aim is to carry out crowd's orientation on the basis of clicking rate, concentrated with reference to user in the presence of wide to described It accuses exposure behavior and there is no the attribute datas of the reference user of ad click behavior to carry out random sampling, obtain negative sample；

If the weight that crowd's directional aim is joint exposure rate and clicking rate progress crowd's orientation and exposure rate is equal to or greatly In the weight of clicking rate, then random sampling is carried out to the candidate user collection, obtain negative sample；

If crowd's directional aim is joint exposure rate and clicking rate carries out crowd's orientation and the weight of exposure rate is less than click The weight of rate then carries out the attribute data of candidate user that random sampling obtains as negative sample for the candidate user collection, And by described with reference to user's concentration is there are advertisement exposure behavior and there is no the attribute datas of the reference user of ad click behavior The attribute data that progress random sampling obtains is as negative sample.

7. method as claimed in claim 5, which is characterized in that described to use multiple positive samples and the multiple negative sample to institute It states prediction model and is trained optimization, the prediction model optimized, comprising:

Sample cleaning treatment is carried out to the multiple negative sample according to the multiple positive sample, the sample cleaning treatment refers to Except the processing of negative sample similar with positive sample in the multiple negative sample；

Optimization is trained to the prediction model using the multiple positive sample and the cleaned negative sample of sample, is obtained The prediction model of optimization.

8. such as the described in any item methods of claim 2-7, which is characterized in that the crowd orients request and carries orientation crowd number Amount；The orientation probability according to each candidate user concentrates the attribute number for filtering out directional user from the candidate user According to, comprising:

According to the sequence of orientation probability from big to small, the attribute data of each candidate user is concentrated to arrange the candidate user Sequence obtains ordered set；

The attribute data conduct of corresponding candidate user is successively chosen from the ordered set according to the orientation crowd quantity The attribute data of directional user.

9. method according to claim 8, which is characterized in that the crowd, which orients request and also carries, determines the targeted advertisements To each instruction information with reference to user；

The attribute data for successively choosing corresponding candidate user from the ordered set according to the orientation crowd quantity Attribute data as directional user, comprising: seek the difference between the orientation crowd quantity and the quantity with reference to user Value obtains screening quantity；And the category of corresponding candidate user is successively chosen from the ordered set according to the screening quantity Attribute data of the property data as directional user；

The method also includes: each attribute data with reference to user concentrated with reference to user is added to the targeted advertisements Orientation demographic data in.

10. the method according to claim 1 to 7, which is characterized in that the method also includes:

The orientation demographic data of the targeted advertisements is stored, and exports the attribute information of the orientation demographic data, the attribute Information includes at least: data name and data bulk；

If the trigger event for carrying out advertisement dispensing based on the orientation demographic data is detected, in orientation demographic data institute The targeted advertisements are launched in corresponding orientation crowd.

11. a kind of crowd's orienting device characterized by comprising

Acquiring unit, the reference user for obtaining targeted advertisements collects and candidate user collection；It is described to collect with reference to user including multiple With reference to the attribute data of user, it is described with reference to user be refer to the targeted advertisements generate positive feedback user；It is described Candidate user collection includes the attribute data of multiple candidate users, and the candidate user is user to be oriented；

Optimize unit, for being trained optimization to prediction model with reference to user's collection and the candidate user collection using described, obtains To the prediction model of optimization；

Processing unit, for calling the prediction model of the optimization to concentrate according to the candidate user attribute number of each candidate user It is estimated according to advertisement orientation is carried out to each candidate user, obtains the orientation probability of each candidate user, the orientation probability Refer to that candidate user generates the probability of positive feedback to the targeted advertisements；

The processing unit, for filtering out orientation from candidate user concentration according to the orientation probability of each candidate user The attribute data of user, and the attribute data of the directional user is added in the orientation demographic data of the targeted advertisements.

12. a kind of server, including communication interface, which is characterized in that further include:

Processor is adapted for carrying out one or more instruction；And

Computer storage medium, the computer storage medium are stored with one or more instruction, one or more instruction Suitable for being loaded by the processor and being executed such as the described in any item crowd's orientation methods of claim 1-10.

13. a kind of computer storage medium, which is characterized in that the computer storage medium is stored with one or more instruction, One or more instruction is suitable for being loaded by processor and being executed such as the described in any item crowd orientation sides claim 1-10 Method.