CN109783539A - Usage mining and its model building method, device and computer equipment - Google Patents
Usage mining and its model building method, device and computer equipment Download PDFInfo
- Publication number
- CN109783539A CN109783539A CN201910010777.9A CN201910010777A CN109783539A CN 109783539 A CN109783539 A CN 109783539A CN 201910010777 A CN201910010777 A CN 201910010777A CN 109783539 A CN109783539 A CN 109783539A
- Authority
- CN
- China
- Prior art keywords
- user
- users
- user group
- behavior
- target
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000005065 mining Methods 0.000 title claims abstract description 98
- 238000000034 method Methods 0.000 title claims abstract description 50
- 230000006399 behavior Effects 0.000 claims abstract description 149
- 238000010801 machine learning Methods 0.000 claims abstract description 49
- 238000012216 screening Methods 0.000 claims abstract description 20
- 230000003542 behavioural effect Effects 0.000 claims description 62
- 238000012549 training Methods 0.000 claims description 43
- 230000006870 function Effects 0.000 claims description 28
- 238000000605 extraction Methods 0.000 claims description 17
- 238000010276 construction Methods 0.000 claims description 16
- 238000012545 processing Methods 0.000 claims description 15
- 238000004458 analytical method Methods 0.000 claims description 7
- 230000004927 fusion Effects 0.000 claims description 6
- 238000009412 basement excavation Methods 0.000 claims description 2
- 230000009471 action Effects 0.000 description 16
- 238000005267 amalgamation Methods 0.000 description 13
- 238000010586 diagram Methods 0.000 description 9
- 230000008569 process Effects 0.000 description 8
- 238000004364 calculation method Methods 0.000 description 7
- 230000000694 effects Effects 0.000 description 5
- 210000001320 hippocampus Anatomy 0.000 description 5
- 239000000047 product Substances 0.000 description 5
- 239000002537 cosmetic Substances 0.000 description 4
- 238000003066 decision tree Methods 0.000 description 4
- 230000003993 interaction Effects 0.000 description 4
- 239000008267 milk Substances 0.000 description 4
- 210000004080 milk Anatomy 0.000 description 4
- 235000013336 milk Nutrition 0.000 description 4
- 239000000843 powder Substances 0.000 description 4
- 239000013589 supplement Substances 0.000 description 4
- 238000004422 calculation algorithm Methods 0.000 description 3
- 238000004891 communication Methods 0.000 description 3
- 238000004590 computer program Methods 0.000 description 3
- 238000009826 distribution Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000003860 storage Methods 0.000 description 3
- 241001269238 Data Species 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 238000002372 labelling Methods 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 230000003749 cleanliness Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 238000013442 quality metrics Methods 0.000 description 1
- 238000007637 random forest analysis Methods 0.000 description 1
- 238000003786 synthesis reaction Methods 0.000 description 1
- 230000001755 vocal effect Effects 0.000 description 1
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
本发明公开了一种用户挖掘及其模型构建方法、装置及计算机设备，所述方法包括：根据目标用户的语义特征表达对待挖掘用户群中的用户进行筛选，得到所包含用户符合语义特征表达的初选用户群；获取所述初选用户群中用户在不同场景的行为数据；基于所述初选用户群中用户在不同场景的行为数据，对所述初选用户群中用户进行多场景行为融合的综合评分；根据所述初选用户群中用户的综合分值，从所述初选用户群中用户在不同场景的行为数据中选取正样本和负样本；通过所述正样本和所述负样本引导机器学习模型进行训练，由完成训练的机器学习模型构建得到用于实现目标用户预测的用户挖掘模型。采用本发明解决了现有技术中用户挖掘的准确率不高的问题。
The invention discloses a user mining and its model building method, device and computer equipment. The method includes: screening users in a user group to be mined according to the semantic feature expression of the target user, and obtaining the users that conform to the semantic feature expression. Preliminary selection of user groups; obtain behavior data of users in the primary selection user group in different scenarios; based on the behavior data of users in the primary selection user group in different scenarios, perform multi-scenario behaviors on the users in the primary selection user group The integrated comprehensive score; according to the comprehensive scores of users in the primary selection user group, select positive samples and negative samples from the behavior data of users in different scenarios in the primary selection user group; The negative samples guide the machine learning model to train, and the user mining model used to predict the target user is constructed from the trained machine learning model. The invention solves the problem of low accuracy of user mining in the prior art.
Description
Technical field
The present invention relates to big data technical field more particularly to a kind of usage mining and its model building method, device and
Computer equipment.
Background technique
With the development of big data, for advertisement putting field, how efficiently to obtain and make a big purchase the latent of product in large quantities
In user, the urgent need of advertisement putting business is become.
Currently, artificial intelligence technology is widely used in the usage mining of advertisement dispensing, that is, be based on machine learning model
Prediction obtains target user from users, and then promotes product to target user.
However, in target user, it would still be possible to exist and not feel completely due to the self reasons such as hobby to popularization product
The user of interest, then, this kind of users are bound to that the potential user of purchase product can not be become, and increase advertisement dispensing instead
The dispensing cost of quotient.
From the foregoing, it will be observed that the problem that existing usage mining is not high there are still accuracy.
Summary of the invention
Accuracy rate in order to solve the problems, such as usage mining present in the relevant technologies is not high, and various embodiments of the present invention provide
A kind of usage mining and its model building method, device and computer equipment.
Wherein, the technical scheme adopted by the invention is as follows:
Disclosed one side according to the present invention, a kind of model building method applied to usage mining, comprising: according to target
The user in user group to be excavated is screened in the semantic feature expression of user, obtains included user and meets semantic feature table
The primary election user group reached；User is obtained in the primary election user group in the behavioral data of different scenes；Based on the primary election user
User in the behavioral data of different scenes, comment by the synthesis for carrying out more scene action amalgamations to user in the primary election user group in group
Point；According to the comprehensive scores of user in the primary election user group, behavior of the user in different scenes from the primary election user group
Positive sample and negative sample are chosen in data；Machine learning model is guided to be trained by the positive sample and the negative sample,
Machine learning model by completing training constructs to obtain the usage mining model for realizing target user's prediction.
Disclosed another aspect according to the present invention, a kind of usage mining method, comprising: obtain user in user group to be excavated
In the behavioral data of different scenes；Call usage mining model, to user in the user group to be excavated different scenes row
Target user's prediction is carried out for data, obtains potential user group；Wherein, the usage mining model is by positive sample and negative sample
The training of this guidance machine learning model obtains, and user exists in the positive sample and the negative sample and the user group to be excavated
Comprehensive scores in more scene action amalgamations are related.
Disclosed another aspect according to the present invention, a kind of model construction device applied to usage mining, comprising: first uses
Family screening module is obtained for being screened according to the expression of the semantic feature of target user to the user in user group to be excavated
Included user meets the primary election user group of semantic feature expression；Behavioral data obtains module, for obtaining the primary election user
Behavioral data of the user in different scenes in group；User's grading module is used for based on user in the primary election user group in difference
The behavioral data of scene carries out more scene comprehensive scores to user in the primary election user group；Second user screening module, is used for
According to the comprehensive scores of user in the primary election user group, behavioral data of the user in different scenes from the primary election user group
Middle selection positive sample and negative sample；Model construction module, for guiding machine learning by the positive sample and the negative sample
Model is trained, and the machine learning model by completing training constructs to obtain the usage mining mould for realizing target user's prediction
Type.
Disclosed another aspect according to the present invention, a kind of usage mining device, comprising: behavioral data obtains module, is used for
User is obtained in user group to be excavated in the behavioral data of different scenes；Target user's prediction module, for calling usage mining
Model carries out target user's prediction in the behavioral data of different scenes to user in the user group to be excavated, obtains target use
Family group；Wherein, the usage mining model is to guide machine learning model training to obtain by positive sample and negative sample, described
Positive sample and the negative sample are related with comprehensive scores of the user in more scene action amalgamations in the user group to be excavated.
Disclosed another aspect, a kind of computer equipment, including processor and memory according to the present invention, the memory
On be stored with computer-readable instruction, side as described above is realized when the computer-readable instruction is executed by the processor
Method.
Disclosed another aspect, a kind of storage medium are stored thereon with computer program, the computer according to the present invention
Method as described above is realized when program is executed by processor.
In the above-mentioned technical solutions, when based on machine learning model building usage mining model, for user to be excavated
User in group is repeatedly screened, and to obtain pure positive sample and negative sample, thus solves usage mining in the prior art
The not high problem of accuracy.
Specifically, being based on user group to be excavated, is expressed according to the semantic feature of target user and carry out first time user sieve
Choosing, and second of user's screening is carried out according to comprehensive scores of the user in more scene action amalgamations, obtain positive sample and negative sample
This, and then guide machine learning model to be trained by positive sample and negative sample, thus building obtains for realizing target use
The usage mining model of family prediction, the positive sample being trained as a result, not only conform with the semantic feature expression of target user, and
The comprehensive scores with higher in more scene action amalgamations, fully ensure that has real meaning for trained sample
Forward direction, and then fully ensure that the accuracy of usage mining.
It should be understood that above general description and following detailed description be only it is exemplary and explanatory, not
It can the limitation present invention.
Detailed description of the invention
The drawings herein are incorporated into the specification and forms part of this specification, and shows and meets implementation of the invention
Example, and in specification together principle for explaining the present invention.
Fig. 1 is the schematic diagram of related implementation environment according to the present invention.
Fig. 2 is a kind of hardware block diagram of server shown according to an exemplary embodiment.
Fig. 3 is a kind of process of model building method applied to usage mining shown according to an exemplary embodiment
Figure.
Fig. 4 be in Fig. 3 corresponding embodiment step 310 in the flow chart of one embodiment.
Fig. 5 be in Fig. 3 corresponding embodiment step 310 in the flow chart of another embodiment.
Fig. 6 be in Fig. 3 corresponding embodiment step 350 in the flow chart of one embodiment.
Fig. 7 be in Fig. 6 corresponding embodiment step 351 in the flow chart of one embodiment.
Fig. 8 be in Fig. 3 corresponding embodiment step 390 in the flow chart of one embodiment.
Fig. 9 be in Fig. 8 corresponding embodiment step 393 in the flow chart of one embodiment.
Figure 10 be in Fig. 8 corresponding embodiment step 393 in the flow chart of another embodiment.
Figure 11 is a kind of flow chart of usage mining method shown according to an exemplary embodiment.
Figure 12 be in Figure 11 corresponding embodiment step 430 in the flow chart of one embodiment.
Figure 13 is a kind of specific implementation schematic diagram of usage mining method in an application scenarios.
Figure 14 is a kind of block diagram of model construction device applied to usage mining shown according to an exemplary embodiment.
Figure 15 is a kind of block diagram of usage mining device shown according to an exemplary embodiment.
Figure 16 is a kind of block diagram of computer equipment shown according to an exemplary embodiment.
Through the above attached drawings, it has been shown that the specific embodiment of the present invention will be hereinafter described in more detail, these attached drawings
It is not intended to limit the scope of the inventive concept in any manner with verbal description, but is by referring to specific embodiments
Those skilled in the art illustrate idea of the invention.
Specific embodiment
Here will the description is performed on the exemplary embodiment in detail, the example is illustrated in the accompanying drawings.Following description is related to
When attached drawing, unless otherwise indicated, the same numbers in different drawings indicate the same or similar elements.Following exemplary embodiment
Described in embodiment do not represent all embodiments consistented with the present invention.On the contrary, they be only with it is such as appended
The example of device and method being described in detail in claims, some aspects of the invention are consistent.
Fig. 1 is a kind of schematic diagram of implementation environment involved in usage mining method.The implementation environment includes user terminal 110
With server end 130.
Specifically, user terminal 110 is configured with display screen, to show recommendation to user by display screen, for example,
Recommendation includes advertisement etc..User terminal 110 can be desktop computer, laptop, tablet computer, smart phone, palm
Computer etc., herein without limiting.
Server end 130 can be a server, is also possible to the server cluster being made of multiple servers, may be used also
To be the cloud computing center being made of multiple servers.Wherein, server is set for providing a user the electronics of background service
It is standby, for example, background service includes usage mining and its model construction service, ad placement services etc..
Certainly, according to the actual demand of application scenarios, different background services can be deployed on different server, can also
To be deployed on same server, herein and without specifically limiting.
User terminal 110 and server end 130 by it is wireless or it is wired pre-establish network connection, to be connected by this network
It connects and realizes that the data between user terminal 110 and server end 130 are transmitted.For example, the data transmitted include recommendation, behavior
Data etc..
For server end 130, to advertisement putting business provide usage mining service, i.e., by usage mining model by
User group to be excavated obtains potential user group, and then user launches recommendation into potential user group.
By the interaction between user terminal 110 and server end 130, for user in potential user group, by with
The recommendation ready to receive sent to server end 130 of family end 110, and then the display screen based on configuration carries out recommendation
Displaying.
Fig. 2 is a kind of hardware block diagram of server shown according to an exemplary embodiment.This kind of server is applicable in
Server end 130 in the implementation environment shown by Fig. 1.
It should be noted that this kind of server, which is one, adapts to example of the invention, it must not believe that there is provided right
Any restrictions of use scope of the invention.This kind of server can not be construed to need to rely on or must have in Fig. 2
One or more component in illustrative server 200 shown.
The hardware configuration of server 200 can generate biggish difference due to the difference of configuration or performance, as shown in figure 3,
Server 200 include: power supply 210, interface 230, at least a memory 250 and an at least central processing unit (CPU,
Central Processing Units)270。
Specifically, power supply 210 is used to provide operating voltage for each hardware device on server 200.
Interface 230 includes an at least wired or wireless network interface, for interacting with external equipment.For example, carrying out Fig. 1 institute
Interaction in implementation environment between user terminal 110 and server end 130 is shown.
Certainly, in the example that remaining present invention is adapted to, interface 230 can further include an at least serioparallel exchange and connect
233, at least one input/output interface 235 of mouth and at least USB interface 237 etc., as shown in Fig. 2, herein not to this composition
It is specific to limit.
The carrier that memory 250 is stored as resource, can be read-only memory, random access memory, disk or CD
Deng the resource stored thereon includes operating system 251, application program 253 and data 255 etc., and storage mode can be of short duration
It stores or permanently stores.
Wherein, operating system 251 be used for manage and control server 200 on each hardware device and application program 253,
To realize operation and processing of the central processing unit 270 to mass data 255 in memory 250, Windows can be
ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM etc..
Application program 253 is the computer program based at least one of completion particular job on operating system 251, can
To include an at least module (being not shown in Fig. 2), each module can separately include the series of computation to server 200
Machine readable instruction.For example, usage mining and its model construction device can be considered the application program 253 for being deployed in server 200.
Data 255 can be stored in photo, picture in disk etc., can also be behavioral data, recommendation etc., deposit
It is stored in memory 250.
Central processing unit 270 may include the processor of one or more or more, and be set as total by least one communication
Line is communicated with memory 250, to read the computer-readable instruction stored in memory 250, and then is realized in memory 250
The operation and processing of mass data 255.For example, reading the series of computation stored in memory 250 by central processing unit 270
The form of machine readable instruction completes usage mining and its model building method.
In addition, also can equally realize the present invention by hardware circuit or hardware circuit combination software, therefore, this hair is realized
The bright combination for being not limited to any specific hardware circuit, software and the two.
Referring to Fig. 3, in one exemplary embodiment, a kind of model building method applied to usage mining is suitable for figure
The structure of the server end of implementation environment shown in 1, the server end can be as shown in Figure 2.
The model building method that this kind is applied to usage mining can be executed by server end, may comprise steps of:
Step 310, the user in user group to be excavated is screened according to the expression of the semantic feature of target user, is obtained
Included user meets the primary election user group of semantic feature expression.
Target user refers to the user for carrying out recommendation dispensing, it is understood that be, to the interested use of recommendation
Family.
The semantic feature of target user is expressed, and is substantially the text for characterizing target user's distinguishing feature, for example, crucial
Word etc..
The user for meeting semantic feature expression, then refer to have the user of target user's distinguishing feature, that is to say, to recommendation
The interested user of content.
As an example it is assumed that recommendation is mother and baby's articles, then, target user, which then refers to, is likely to purchase mother and baby's articles
Potential customers, for example pregnant woman or family have the parent etc. of baby.Correspondingly, the semantic feature of the target user, which is expressed, includes
But it is not limited to: baby, child-bearing, milk powder, feeding bottle, diaper, trolley, diatery supplement, early education etc..
As a result, for the user in user group to be excavated, meet the user of semantic feature expression, for example, once browsing
The user of diatery supplement related article is likely to become target user, by phase alternatively, once watching the user etc. of early education video
It is added to primary election user group with answering.
And those do not meet the user of semantic feature expression, for example, the user of child-bearing related article was never browsed, or
Person never bought the user etc. of milk powder, will be regarded as to the completely uninterested user of recommendation, and was not added into primary election
User group.
That is, the first time that user in user group to be excavated carries out is used in the semantic feature expression based on target user
Family screening, eliminates to the completely uninterested user of recommendation due to the self reasons such as hobby, so that choosing training
The user group of sample is contracted to primary election user group by user group to be excavated, and then it is advantageously ensured that model training accuracy.
Step 330, user is obtained in the primary election user group in the behavioral data of different scenes.
For a user, scene is accordingly constructed with the operation of client in user terminal.Correspondingly, behavior
Data are then based on the client run in user terminal, the various manipulation behaviors that user is carried out and generate.
For example, client is browser client, and user can search for video, browsing webpage by browser client
In article, news etc., at this point, the search behavior as performed by user, browsing behavior just generate user in browser scene
Behavioral data.
Alternatively, client is social client, user can search for public platform by social client, browsing subscribes to a number push
Article etc., at this point, the search behavior as performed by user, browsing behavior just generate user in the behavior number of social scene
According to.
Or client is electric business platform client, user can search for by electric business platform client admires commodity
Picture, and then buy the commodity etc., at this point, the search behavior as performed by user, buying behavior just generate user in electric business
The behavioral data of scene.
From the foregoing, it will be observed that behavioral data, substantially has recorded the manipulation behavior that user is carried out in the scene, referred to as user
Behavior.Behavior data include but is not limited to: type, time of origin of user behavior of user behavior etc..Wherein, user behavior
Type further comprise: search behavior, browsing behavior, buying behavior etc..
It should be noted that the input module that user terminal is configured is different, then the various manipulation behaviors that user is carried out
By different from.For example, input module is mouse, then the various manipulation behaviors that user is carried out, which can be, to be clicked, double-clicks, pulling
Equal mechanically actuateds, alternatively, input module is Touch Screen, then the various manipulation behaviors that user is carried out can be sliding, click
Equal gesture operations, the present embodiment are not limited this.
So, user is generated after the behavioral data of various scenes in user terminal, with user terminal and server end
Between interaction can receive the user in the behavioral data of various scenes, and store for server end, so as to
It is used when providing usage mining service.
Step 350, based on user in the primary election user group different scenes behavioral data, to the primary election user group
Middle user carries out the comprehensive score of more scene action amalgamations.
It is opposite to limit to if being appreciated that and only considering sole user behavior of the user in some single scene, do not have not only
There is the other users behavior for considering user in other scenes, is easy to cause mode input to have partially, and simply just depend on
The foundation that the behavior frequency of user behavior is screened as user, not can guarantee the degree of purity of mode input.
For this purpose, carrying out more scene action amalgamations first in the present embodiment for user, then determine user in more scene behaviors
Comprehensive score in fusion, in this, as the foundation of subsequent second of user screening.
Wherein, more scene action amalgamations, refer to collect primary election user group in user different scenes behavioral data, so as to
In comprehensively considering user in the various actions of various scenes.
Comprehensive score, then reflect whether user in primary election user group can become the user for carrying out recommendation dispensing,
It that is to say, comprehensive score is higher, and the probability as target user is bigger.
Optionally, it about comprehensive score, can be carried out from the different dimensions of user behavior, for example, in conjunction with user's row
For the behavior frequency and time of origin, the single dimension in user behavior is avoided relying on this, and then guarantee the pure of mode input
Cleanliness.
Step 370, according to the comprehensive scores of user in the primary election user group, user is not from the primary election user group
With selection positive sample and negative sample in the behavioral data of scene.
Specifically, positive sample refers to that comprehensive scores are not less than the user of score threshold in different scenes in primary election user group
Behavioral data, that is to say, by the behavioral data of first kind label for labelling, it is corresponding which is used to indicate positive sample
User belong to target user's classification.
Correspondingly, negative sample then refers to that comprehensive scores are less than the user of score threshold in different scenes in primary election user group
Behavioral data, that is to say, by the behavioral data of the second class label for labelling, it is corresponding which is used to indicate negative sample
User belong to non-targeted class of subscriber.
It wherein, about the selection of negative sample, is rejected in the behavioral data of different scenes from user in primary election user group
After positive sample, choose immediately.
It is noted that being consistent for the positive sample of model training and the quantity of negative sample.
Step 390, it guides machine learning model to be trained by the positive sample and the negative sample, is trained by completing
Machine learning model construct to obtain the usage mining model for realizing target user's prediction.
Wherein, training, adds the parameter of machine learning model essentially by training sample (positive sample and negative sample)
With iteration optimization, so that the assignment algorithm function convergence based on the building of this parameter.
Optionally, machine learning model can be Random Forest model, Logic Regression Models, decision-tree model etc..Wherein,
Decision-tree model further comprises adaptive boosting decision-tree model, gradient promotion decision-tree model.
Optionally, assignment algorithm function is including but not limited to greatest hope function, loss function etc..
That is, based on the training that training sample carries out machine learning model, usage mining model is substantially expert at
Mapping relations are constructed between data and target user.
So, based on mapping relations constructed by usage mining model, by user in user group to be excavated in different scenes
Behavioral data be input to usage mining model, can predict to obtain the user in user group to be excavated whether be target use
Family.
By the above process, comprehensive scores of the selection of positive sample and negative sample based on user in primary election user group are different
In the behavior frequency of the prior art based on the sole user's behavior in some single scene of user in user group to be excavated, avoid out
Existing sample bias problem, has fully ensured that the degree of purity of training sample (positive sample and negative sample), be more advantageous to training obtain it is pre-
The high usage mining model of accuracy rate is surveyed, and then has sufficiently ensured the accuracy of usage mining model construction.
Referring to Fig. 4, in one exemplary embodiment, the semantic feature expression of target user includes positive keyword.
It is baby, child-bearing, milk powder, feeding bottle, diaper, hand push for mother and baby's articles as previously mentioned, for recommendation
Vehicle, diatery supplement, early education etc. can be used as positive keyword.
Correspondingly, step 310 may comprise steps of:
Step 311, the media data that user browsed in the user group to be excavated is obtained.
Wherein, media data can be article, picture, video etc..
As previously mentioned, user can carry out the browsing of media data by the different clients run in user terminal, for example, with
Family is by article, the news etc. in browser client viewing video, browsing webpage, alternatively, user searches by social client
Rope public platform, browsing subscribe to number article of push etc., or, search admires the picture of commodity, and then buys the commodity etc..
It should be appreciated that no matter user based on any client carry out media data browsing, client all will be user
Save corresponding historical viewings record.
So, with the interaction between user terminal and server end, for server end, historical viewings can be passed through
The reception of record gets the media data that user browsed.
Step 313, in the media data, the matched and searched of the positive keyword is carried out.
It, can be by positive keyword in the media data received after receiving the media data that user browsed
Middle carry out matched and searched, whether to determine in the media data received comprising positive keyword.
As previously mentioned, media data can be article, picture, video etc., it will be understood that article is text representation, that is, is wrapped
Containing keyword, and for picture, video, then firstly the need of text conversion processing is carried out, for example, to the master in video
Topic, word segment content carry out keyword extraction, alternatively, to picture carry out text identification so that picture, Video Quality Metric be comprising
There is the text representation of keyword.
Based on this, matched and searched is substantially the keyword for being included for media data, searches whether to exist and close with forward direction
The matched keyword of keyword.
If finding the keyword with positive Keywords matching, it is determined that comprising positive keyword in media data, i.e.,
It jumps and executes step 315.
, whereas if not finding the keyword with positive Keywords matching, it is determined that do not include forward direction in media data
Keyword determines that the user in user group to be excavated does not meet the semantic feature expression of target user.
Step 315, if determined in the user group to be excavated comprising the positive keyword in the media data
The user meet semantic feature expression, will meet semantic feature expression the user be added to the primary election user group.
As an example it is assumed that target user is defined as the potential customers of purchase cosmetics A, then, the target user
Semantic feature expression include at least positive keyword " cosmetics A ".
As a result, for the user in user group to be excavated for, as long as browsed the article about " cosmetics A ", picture,
The media datas such as video can determine that the use by matched and searched of the positive keyword " cosmetics A " in aforementioned media data
Family meets semantic feature expression.
As a result, under the action of positive keyword, it ensure that in the primary election user group screened by positive keyword
User is real meaning " forward direction ", that is, meets the semantic feature expression of target user, and then is conducive to based on usage mining mould
Target user's prediction that type carries out, has fully ensured that the accuracy of usage mining.
Referring to Fig. 5, in one exemplary embodiment, the semantic feature expression of target user includes negative sense keyword.
It is still illustrated so that recommendation is mother and baby's articles as an example, it is assumed that some user once browsed about hippocampus baby
Article, although this article relates to " baby " two word, can not indicate the user be buy mother and baby's articles potential customers,
Therefore, " hippocampus baby " can be used as negative sense keyword.
Correspondingly, step 310 can with the following steps are included:
Step 312, the media data browsed based on user in the primary election user group, carries out the negative sense keyword
Matched and searched.
Similarly in positive keyword, matched and searched refers to the keyword for being included for media data, searches whether exist
With the keyword of negative sense Keywords matching.
If finding the keyword with negative sense Keywords matching, it is determined that include negative sense keyword in media data, i.e.,
It jumps and executes step 314.
, whereas if not finding the keyword with negative sense Keywords matching, it is determined that do not include negative sense in media data
The semantic feature expression that keyword, the i.e. user in judgement primary election user group meet target user, remains in just selection
In the group of family.
Step 314, if including the negative sense keyword in the media data, determine in the primary election user group
The user is noise user, and the noise user is rejected from the primary election user group.
Noise user refers to and is present in primary election user group, and not to the real interested user of recommendation.
As an example it is assumed that target user is defined as the potential customers of purchase mother and baby's articles, then, the target user
The included negative sense keyword of semantic feature expression can have " hippocampus baby " etc..
As a result, for the user in primary election user group for, if browsed the article about " hippocampus baby ", picture,
The media datas such as video determine the use by matched and searched of the negative sense keyword " hippocampus baby " in aforementioned media data
Family does not meet semantic feature expression, belongs to noise user.
That is, the effect by negative sense keyword is rejected so that the noise user in primary election user group is removed
There may be ambiguity or " forward direction " of the non-real meaning of ambiguity, the pure of user in primary election user group is further ensured
Degree, and then be conducive to the target user's prediction carried out based on usage mining model, sufficiently ensure the accuracy of usage mining.
Referring to Fig. 6, in one exemplary embodiment, step 350 may comprise steps of:
Step 351, for user in the primary election user group Same Scene behavioral data, at least from user behavior
One dimension determines scoring coefficient.
As previously mentioned, user behavior is the manipulation behavior that the user recorded by behavioral data carries out in the scene, that
, at least one dimension of user behavior can then refer to the number of same subscriber behavior, the time of origin of user behavior, user
The importance etc. of behavior.
The scoring coefficient includes but is not limited to as a result: the behavior frequency, time of the act pad value and behavior weighted value.
Step 353, according to the scoring coefficient, the row that user in the primary election user group corresponds to Same Scene is calculated
For score value.
Specifically, multiplying is carried out to the behavior frequency, time of the act pad value and behavior weighted value, obtained described
User corresponds to the behavior score value of Same Scene in primary election user group.
For user i in primary election user group, the behavior score value calculating process such as calculation formula (1) of corresponding Same Scene
It is shown:
scorei=weightsource×exp-ln2×time×sigmoid(actioncnt) (1)。
Wherein, score indicates that user i corresponds to the behavior score value of scene source.weightsourceIndicate corresponding scene
The behavior weighted value of source, behavior weighted value is higher, and comprehensive scores are higher, conversely, behavior weighted value is lower, comprehensive scores are got over
It is low.
actioncntThe expression behavior frequency, sigmoid () are then indicated to actioncntIt is normalized, that is to say,
The behavior frequency is higher, and comprehensive scores are higher, conversely, the behavior frequency is lower, comprehensive scores are lower.
Time indicates time of the act pad value, further, time=| t1-t2|。
Wherein, t1Indicate the time of origin of user behavior in behavioral data, t2Indicate the current time of progress usage mining,
It is also understood that the current time in system.
For this purpose, what time substantially characterized is time of the act absolute value of the difference, then, the time of origin system of distance of behavior is worked as
The preceding time is closer, and comprehensive scores are higher, conversely, the time of origin system of distance current time of behavior is remoter, comprehensive scores are lower.
Step 355, the behavior score value that user in the primary election user group corresponds to different scenes is added, obtains the primary election
The comprehensive scores of user in user group.
For some user in primary election user group, shown in the calculating process of comprehensive scores such as calculation formula (2):
Wherein, Score indicates comprehensive scores, and source0~N indicates 0~N of scene, scoreiIndicate that user i corresponds to not
With the behavior score value of scene.
Under the action of above-described embodiment, the comprehensive score for carrying out more scene action amalgamations for user is realized, is second
Secondary user's screening provides foundation, and the accuracy of usage mining is ensured with this.
Referring to Fig. 7, in one exemplary embodiment, step 351 may comprise steps of:
Step 3511, for the historical user group for having carried out recommendation dispensing, exist to user in the historical user group
The behavioral data of different scenes carries out significance analysis, obtains user in the historical user group and divides in the conspicuousness of different scenes
Value.
Significance analysis, it is therefore intended which user is more suitable for carrying out the dispensing of recommendation in verifying historical user group,
Essence is to calculate corresponding conspicuousness score value in the behavioral data of different scenes based on user in historical user group.
Optionally, conspicuousness score value that is to say user's conversion ratio, can be probability of transaction, clicking rate, searching rate, view rate
In any one or several combinations.
Step 3513, user in the historical user group is normalized in the conspicuousness score value of different scenes,
Multiple behavior weighted values are obtained, each behavior weighted value corresponds to a kind of scene.
It should be appreciated that the user is because own interests are liked, behavior is practised for each of historical user group user
The reasons such as used are different in the behavioral data of different scenes, then, which will also have in the conspicuousness score value of different scenes
It is distinguished.
For example, the user prefers to do shopping by browser client, rather than it is based on electric business platform client, then,
For the user, for the row generated in the behavioral data and electric business platform client generated in browser client
For data, significance analysis is carried out, which will be apparently higher than the significant of electric business scene in the conspicuousness score value of browser scene
Property score value.
For this purpose, the user will be normalized in the conspicuousness score value of different scenes, in the present embodiment to obtain
Corresponding to the behavior weighted value of scene, and then more accurately reflect user for the preference of different scenes, so that it is guaranteed that user
The accuracy of excavation.
For example, for the user A in historical user group, it is assumed that user A is m1 in the conspicuousness score value of scene C1,
The conspicuousness score value of scene C2 is m2, is m3 in the conspicuousness score value of scene C3.
So, for user A,
Certainly, behavior weighted value can be respectively configured for the conspicuousness score value of different user, can also be useful based on institute
The conspicuousness mean scores at family are configured, and the present embodiment not constitutes specific limit to this.
As a result, when carrying out more scene action amalgamations, different scenes are based on the historical user for having carried out recommendation dispensing
Group and be configured with corresponding behavior weighted value, realized independent of artificial, the variation of recommendation release time will be followed
And correspondingly change, further ensure that the accuracy of usage mining.
Referring to Fig. 8, in one exemplary embodiment, step 390 may comprise steps of:
Step 391, the feature extraction of the positive sample and negative sample is carried out respectively.
As previously mentioned, behavioral data, has recorded the user behavior of user in the scene, behavioral data includes but is not limited to:
The time of origin of the type of user behavior, user behavior.
For this purpose, feature, is the string number set converted based on behavioral data, and then in digital form
The user behavior of user that behavioral data is recorded in the scene is uniquely identified, that is, realizes user recorded to behavioral data
The accurate description of user behavior in the scene.
About feature extraction, the time of origin by the type of the user behavior in behavioral data according to user behavior can be
Sequential concatenation is also possible to carry out accumulating operation to the type of user behavior in behavioral data, is also based on user behavior
Frequency makees further statistics.
For example, the type of user behavior is ranked up according to the time of origin of user behavior first, at this point, behavior
Data include: the Class1 of user behavior, the type 2 of user behavior, the type of user behavior 3, the Class1 of user behavior, user
The Class1 of behavior.
Under this explanation, for the type of user behavior by digital representation, the type of user behavior is identical, indicates that user exists
Different moments, there are identical user behaviors in scene.
Based on this, the type of the user behavior after sequence is spliced, that is, extracts and obtains feature=[1,2,3,1,1].
It is, of course, also possible to based on the identical user behavior recurred, feature is further converted to [(1,1), (2,
1),(3,1),(1,2)]。
Step 393, feature selecting processing is carried out to the feature extracted, obtains target signature.
Optionally, feature selecting processing can be carried out based on information gain, Mutual information entropy, similarity, consistency etc., this reality
It applies example and does not constitute specific limit to this.
For example, the feature that behavioral data A is extracted is A1, the feature that behavioral data B is extracted is B1, it is assumed that by behavior number
It predicts to obtain user according to A and belongs to target user, predict to obtain user by behavioral data B and belong to non-targeted user, and feature A1=
Feature B1.
So, the feature selecting processing based on consistency, feature A1 and feature B1 can not be used as target signature.
By feature selecting, the target signature for being input to usage mining model is purer, is conducive to based on usage mining mould
Target user's prediction that type carries out, substantially ensures the accuracy of target user's prediction.
Step 395, loss function is constructed according to the parameter of the target signature and the machine learning model.
Step 397, it when the loss function is restrained by the backpropagation of the machine learning model, is trained by completing
Machine learning model construct to obtain the usage mining model.
Now illustrate the training process of machine learning model using assignment algorithm function as loss function.
Specifically, the parameter of random initializtion machine learning model is based on according to the feature for working as previous training sample
The parameter of random initializtion by propagated forward carry out probability calculation, the probabilistic forecasting user by calculating be target user or
Non-targeted user, the Dice distance between the class of subscriber obtained based on prediction and correct mark (the true classification of user) are constructed
Loss function, and further calculate the penalty values of the loss function.
If the penalty values of loss function are not up to minimum, the parameter of machine learning model is updated by backpropagation,
And according to the feature of the latter training sample, the parameter based on update carries out probability calculation, is predicted again by the probability of calculating
Class of subscriber rebuilds loss function based on the Dice distance between the class of subscriber predicted again and correct mark,
And the penalty values of the loss function rebuild are calculated again.
Such iterative cycles are considered as loss function convergence until the penalty values of constructed loss function reach minimum, this
When, machine learning model also restrains, and meets default required precision, then stops iteration.
Otherwise, iteration updates the parameter of machine learning model, and according to the parameter of the feature of remaining training sample and update,
The penalty values of the loss function thus constructed are calculated, until loss function is restrained.
It is noted that will also stop if the number of iterations has reached iteration threshold before loss function convergence
Iteration guarantees the efficiency of machine learning model training with this.
When machine learning model restrains and meets default required precision, indicate that machine learning model completes training,
It constructs to obtain the usage mining model for being applied to target user's prediction based on the machine learning model for completing training.
Referring to Fig. 9, in one exemplary embodiment, step 393 may comprise steps of:
Step 3931, each feature obtained for extraction, calculates corresponding information gain.
It step 3933, is more than described by the information gain if the information gain being calculated is more than gain threshold
The feature of gain threshold is as the target signature.
The information gain of feature, for the validity of characteristic feature, calculation formula such as (3-1), (3-2), (3-3) is shown.
Wherein, Y indicates the characteristic set constituted by extracting feature, and the feature in Y includes { y1、y2、y3、……ym,
Each feature yiThe probability of appearance is Pi, then H (Y) indicates the comentropy of Y.
Wherein, xiIt indicates to extract another obtained feature and be not present in Y, and H (Y | X=xi) it indicates by feature xiStructure
At characteristic set Y comentropy, then H (Y | X) is indicated with xiThe conditional information entropy of Y when as additional conditions, it is understood that
For, be Y be added feature xiComentropy.
IG (Y | X)=H (Y)-H (Y | X) (3-3).
Wherein, the comentropy of H (Y) expression Y, and H (Y | X) indicate that feature x is added in YiComentropy, then IG (Y | X) indicate plus
Enter feature xiThe information gain of front and back Y, it is understood that being is feature xiInformation gain.
That is, if feature xiInformation gain it is bigger, this feature xiY is set to tend to deterministic degree higher, also
I.e. so that in Y the distribution of each feature it is purer, then, this feature xiIt is more effective, be more conducive to when being input to usage mining model
Improve the accuracy rate of target user's prediction.
For this purpose, the feature that information gain is more than gain threshold can be used as target signature and be input to usage mining model.
Referring to Fig. 10, in one exemplary embodiment, step 393 can with the following steps are included:
Step 3932, the Mutual information entropy between different target feature is calculated.
It step 3934, is more than the entropy by the Mutual information entropy if the Mutual information entropy being calculated is more than entropy threshold
The target signature of threshold value is as redundancy feature.
Step 3936, the redundancy feature is rejected from the target signature.
Mutual information entropy, for the degree of redundancy between characteristic feature, it is understood that be the correlation characterized between feature
Property, calculation formula (4) is as follows:
Wherein, X indicates a target signature, and Y indicates another target signature, and P (x) indicates characteristic component x in target spy
The probability occurred in sign X, P (y) indicate that the probability that characteristic component y occurs in target signature Y, P (x, y) indicate the joint of x, y
Distribution probability, then I (X；Y the Mutual information entropy between target signature X, Y) is indicated.
That is, if Mutual information entropy is bigger, the degree of redundancy between feature is bigger, at this point, being input to usage mining
It is unfavorable for improving the accuracy rate of target user's prediction instead when model.
For this purpose, the target signature that Mutual information entropy is more than entropy threshold will be removed as redundancy feature, target is guaranteed with this
The validity of feature.
Cooperation through the foregoing embodiment realizes the feature selecting based on information gain and Mutual information entropy, relative to existing
There is the extensive style feature input in technology, i.e., all features extracted all are used for the prediction of target user, are avoided in vain, no
True feature is input to usage mining model, substantially ensures that inputted target signature plays positive acting to model, thus
It is more advantageous to the accuracy rate for improving target user's prediction.
Figure 11 is please referred to, in one exemplary embodiment, a kind of usage mining model method is suitable for implementing ring shown in Fig. 1
The structure of the server end in border, the server end can be as shown in Figure 2.
This kind of usage mining model method can be executed by server end, may comprise steps of:
Step 410, user is obtained in user group to be excavated in the behavioral data of different scenes.
About user in user group to be excavated in the acquisition of the behavioral data of different scenes, and used in aforementioned primary election user group
User is consistent in the acquisition process of the behavioral data of different scenes in the group of family, not repeated description herein.
Step 430, call usage mining model, to user in the user group to be excavated different scenes behavioral data
Target user's prediction is carried out, potential user group is obtained.
Wherein, the usage mining model is to guide machine learning model training to obtain by positive sample and negative sample,
The positive sample and the negative sample have with comprehensive scores of the user in more scene action amalgamations in the user group to be excavated
It closes.
In one embodiment, as shown in figure 12, step 430 may comprise steps of:
Step 431, feature extraction is carried out in the behavioral data of different scenes to user in the user group to be excavated.
Herein, characteristic extraction procedure is consistent with the characteristic extraction procedure in model training, not repeated description.
Step 433, the feature extracted is input to the usage mining model, calculates and is used in the user group to be excavated
Family belongs to the probability of different user classification.
In the present embodiment, target user's prediction, is realized based on the classifier being arranged in machine learning, i.e., using classification
Device calculates the probability that user in candidate user group belongs to different user classification.
Wherein, class of subscriber includes target user's classification and non-targeted class of subscriber.
Step 435, according to the probability being calculated, determine whether the user in the user group to be excavated belongs to target use
The user for belonging to target user's classification in the user group to be excavated is added to the potential user group by family classification.
For example, it for the user A in user group to be excavated, calculates separately user A and belongs to different user classification
Probability, it is assumed that user A belong to target user's classification probability be P1, user A belong to non-targeted class of subscriber probability be P2.
So, if P1 > P2, then it represents that user A belongs to target user's classification, that is, determines user A in user group to be excavated
For target user, whereas if P1 < P2, then it represents that user A belongs to non-targeted class of subscriber, that is, determines in user group to be excavated
User A is not target user.
As a result, if user A is target user in user group to be excavated, user A is added to potential user group.
Certainly, in other embodiments, in order to further increase the accuracy of usage mining, probability threshold can also be set
Value, then, only when user A belong to target user's classification probability P 1 be more than probability threshold value, target user can be considered as, in turn
Potential user group is added to as target user.
Wherein, probability threshold value can neatly be adjusted according to the actual needs of application scenarios, for example, digging to user
In the higher application scenarios of the accuracy requirement of pick, probability threshold value is set as 0.75, does not constitute specific limit herein.
By process as described above, efficient usage mining is realized.
Figure 13 is a kind of specific implementation schematic diagram of usage mining method in an application scenarios.The application scenarios are for advertisement
It launches quotient and carries out usage mining, to obtain the potential customers with purchase the recommended mother and baby's product of mother and baby's series advertisements.
In the application scenarios, including two-way branch: training branch and predicted branches.
Training branch:
Based on training sample, i.e. positive sample and negative sample, the building of usage mining model is realized.Wherein, positive sample and negative
Sample is to be screened based on multiple user, so as to improve the accuracy rate of usage mining.
Specifically, as shown in figure 13, it by executing step 700, is expressed according to the semantic feature of target user to be excavated
User in user group carries out first time user screening, obtains primary election user group.
Wherein, the semantic feature expression of target user can not only indicate the foundation characteristic of target user, including age, property
Not, region, educational background, occupation, online duration, online scene etc., also may indicate that the interest characteristics of target user, including interested
Electric business platform, the search of interested topic, interested information, interested general entertainment selection etc., or even be also represented by and be used for
KL divergence (Kullback-Leibler divergence) feature of the keyword of first time user screening, including baby, educate
Youngster, milk powder, feeding bottle, diaper, trolley, diatery supplement, early education etc., and then fully ensured that the reliability of training sample.
It is true in the behavioral data institute of different scenes based on user in primary election user group by execution step 701 to step 702
Fixed comprehensive scores carry out second of user's screening, obtain positive sample and negative sample 703, ensure the pure of training sample with this
Degree, and then be conducive to improve model training effect.
By executing step 704, feature extraction and selection are carried out based on positive sample and negative sample, obtain target signature, with
This ensures to be input to the validity of the feature of machine learning model, and then is conducive to improve model training effect.
By executing step 705, machine learning model is trained using target signature, and store and obtain usage mining
Model.
Predicted branches:
Potential user group is predicted in real time for advertisement putting business, it is fixed using the user in potential user group as potential customers
To dispensing mother and baby's series advertisements.
Specifically, as shown in figure 13, by execution step 706 to step 707, the row based on user in user group to be excavated
For data, feature extraction is carried out.
By executing step 708 to step 709, obtained feature will be extracted and be input to usage mining model to carry out target pre-
It surveys, obtains potential user group.
The overall process for completing usage mining as a result, enables advertisement putting business based on the user in potential user group
The orientation for carrying out mother and baby's series advertisements is launched.
In this application scene, realize model self-training, i.e., with the increase of training samples, the predictive ability of target user
It will constantly enhance therewith, and be not necessarily to manual maintenance, not only contribute to reduce cost of labor, improve the intelligence of target user's prediction
Energyization, and effectively improve the accuracy rate rate of target prediction.
In addition, not only contributing to improve model in conjunction with the acquisition of the positive negative sample of high-purity and the input of validity feature
Training effect, and be conducive to improve the prediction effect of model, the accuracy rate of usage mining is effectively promoted with this.
Following is apparatus of the present invention embodiment, can be used for executing usage mining and its model construction according to the present invention
Method.For undisclosed details in apparatus of the present invention embodiment, usage mining and its model according to the present invention are please referred to
The embodiment of the method for construction method.
Figure 14 is please referred to, in one exemplary embodiment, a kind of model construction device 900 applied to usage mining includes
But be not limited to: first user's screening module 910, behavioral data obtain module 930, user's grading module 950, second user screening
Module 970 and model construction module 990.
Wherein, first user's screening module 910, for being expressed according to the semantic feature of target user to user group to be excavated
In user screen, obtain included user meet semantic feature expression primary election user group.
Behavioral data obtains module 930, for obtaining in the primary election user group user in the behavioral data of different scenes.
User's grading module 950, for based on user in the primary election user group different scenes behavioral data, to institute
State the comprehensive score that user in primary election user group carries out more scene action amalgamations.
Second user screening module 970, for the comprehensive scores according to user in the primary election user group, from the primary election
User chooses positive sample and negative sample in the behavioral data of different scenes in user group.
Model construction module 990, for guiding machine learning model to be instructed by the positive sample and the negative sample
Practice, the machine learning model by completing training constructs to obtain the usage mining model for realizing target user's prediction.
In one exemplary embodiment, the semantic feature expression of the target user includes positive keyword.
Correspondingly, the first user screening module 910 includes but is not limited to: media data acquiring unit, the first matching
Searching unit and first user's adding unit.
Wherein, media data acquiring unit, for obtaining the media data that user browsed in the user group to be excavated.
First matched and searched unit, in the media data, carrying out the matched and searched of the positive keyword.
First user's adding unit, if for described in comprising the positive keyword, determining in the media data
The user in user group to be excavated meets semantic feature expression, the user for meeting semantic feature expression is added to described first
Select user group.
In one exemplary embodiment, the semantic feature expression of the target user further includes negative sense keyword.
Correspondingly, the first user screening module 910 further includes but is not limited to: the second matched and searched unit and user pick
Except unit.
Wherein, the second matched and searched unit, the media data for being browsed based on user in the primary election user group, into
The matched and searched of the row negative sense keyword.
User's culling unit, if determining the primary election for including the negative sense keyword in the media data
The user in user group is noise user, and the noise user is rejected from the primary election user group.
In one exemplary embodiment, user's grading module 950 includes but is not limited to: determining scoring coefficient elements,
Behavior score value computing unit and behavior score value addition unit.
Wherein it is determined that scoring coefficient elements, for for user in the primary election user group Same Scene behavior number
According to from the determining scoring coefficient of at least one dimension of user behavior.
Behavior score value computing unit, for user couple in the primary election user group to be calculated according to the scoring coefficient
Answer the behavior score value of Same Scene.
Behavior score value addition unit, for user in the primary election user group to be corresponded to the behavior score value phase of different scenes
Add, obtains the comprehensive scores of user in the primary election user group.
In one exemplary embodiment, the scoring coefficient includes behavior weighted value.
Correspondingly, the determining scoring coefficient elements include but is not limited to: significance analysis subelement and normalized
Subelement.
Wherein, significance analysis subelement, for being gone through to described for the historical user group for having carried out recommendation dispensing
User carries out significance analysis in the behavioral data of different scenes in history user group, obtains in the historical user group user not
With the conspicuousness score value of scene.
Normalized subelement, for being carried out to user in the historical user group in the conspicuousness score value of different scenes
Normalized obtains multiple behavior weighted values, each behavior weighted value corresponds to a kind of scene.
In one exemplary embodiment, the scoring coefficient includes the behavior frequency, time of the act pad value and behavior weight
Value.
Correspondingly, the behavior score value computing unit includes but is not limited to: multiplying subelement.
Wherein, multiplying subelement, for being carried out to the behavior frequency, time of the act pad value and behavior weighted value
Multiplying obtains the behavior score value that user in the primary election user group corresponds to Same Scene.
In one exemplary embodiment, the second user screening module 970 includes but is not limited to: positive sample selection unit
With negative sample selection unit.
Wherein, positive sample selection unit, if the comprehensive scores for user in the primary election user group are more than score threshold
Value, then using comprehensive scores not less than score threshold user different scenes behavioral data as the positive sample.
Negative sample selection unit, for being less than the user of score threshold from comprehensive scores in the behavioral data of different scenes
It randomly selects to obtain the negative sample.
In one exemplary embodiment, the model construction module 990 includes but is not limited to: feature extraction unit, feature
Selecting unit, function construction unit and function convergence unit.
Wherein, feature extraction unit, for being carried out to user in the candidate user group in the behavioral data of different scenes
Feature extraction.
Feature selection unit obtains target signature for carrying out feature selecting processing to the feature extracted.
Function construction unit, for constructing loss letter according to the parameter of the target signature and the machine learning model
Number.
Function convergence unit, for when the loss function is restrained by the backpropagation of the machine learning model,
Machine learning model by completing training constructs to obtain the usage mining model.
In one exemplary embodiment, the feature selection unit includes but is not limited to: information gain computation subunit and
Feature adding unit.
Wherein, information gain computation subunit, for calculating corresponding information and increasing for each obtained feature is extracted
Benefit.
Feature adding unit, if the information gain for being calculated is more than gain threshold, by the information gain
More than the gain threshold feature as the target signature.
In one exemplary embodiment, the feature selection unit further includes but is not limited to: Mutual information entropy computation subunit,
Characterizing definition subelement and feature reject subelement.
Wherein, Mutual information entropy computation subunit, for calculating the Mutual information entropy between different target feature.
Characterizing definition subelement, if the Mutual information entropy for being calculated is more than entropy threshold, by the Mutual information entropy
More than the entropy threshold target signature as redundancy feature.
Feature rejects subelement, for rejecting the redundancy feature from the target signature.
Figure 15 is please referred to, in one exemplary embodiment, a kind of usage mining device 1100 includes but is not limited to: behavior number
According to acquisition module 1110 and target user's prediction module 1130.
Wherein, behavioral data obtains module 1110, for obtaining in user group to be excavated user in the behavior of different scenes
Data.
Target user's prediction module 1130 exists to user in the user group to be excavated for calling usage mining model
The behavioral data of different scenes carries out target user's prediction, obtains potential user group.
Wherein, the usage mining model is to guide machine learning model training to obtain by positive sample and negative sample,
The positive sample and the negative sample have with comprehensive scores of the user in more scene action amalgamations in the user group to be excavated
It closes.
In one exemplary embodiment, target user's prediction module 1130 includes but is not limited to: feature extraction unit, probability
Computing unit and target user's judging unit.
Wherein, feature extraction unit, for user in the user group to be excavated different scenes behavioral data into
Row feature extraction.
Probability calculation unit calculates described to be excavated for the feature extracted to be input to the usage mining model
User belongs to the probability of different user classification in user group.
Target user's judging unit, for determining the user in the user group to be excavated according to the probability being calculated
Whether belong to target user's classification, the user that target user's classification is belonged in the user group to be excavated is added to the target
User group.
In one exemplary embodiment, described device 1100 further includes but is not limited to: content putting module.
Wherein, the recommendation is thrown to the target user for obtaining recommendation by content putting module
Group.
It should be noted that device provided by above-described embodiment is when carrying out relevant treatment, only with above-mentioned each function mould
The division progress of block can according to need and for example, in practical application by above-mentioned function distribution by different functional modules
It completes, i.e., the internal structure of device will be divided into different functional modules, to complete all or part of the functions described above.
In addition, apparatus and method embodiment provided by above-described embodiment belongs to same design, wherein modules are executed
The concrete mode of operation is described in detail in embodiment of the method, and details are not described herein again.
Figure 16 is please referred to, in one exemplary embodiment, a kind of computer equipment 1000, including an at least processor
1001, an at least memory 1002 and at least a communication bus 1003.
Wherein, computer-readable instruction is stored on memory 1002, processor 1001 is read by communication bus 1003
The computer-readable instruction stored in memory 1002.
It is realized when the computer-readable instruction is executed by processor 1001 in the various embodiments described above and is applied to usage mining
Model building method.
In one exemplary embodiment, a kind of storage medium, is stored thereon with computer program, which is located
Manage the model building method applied to usage mining realized in the various embodiments described above when device executes.
Above content, preferable examples embodiment only of the invention, is not intended to limit embodiment of the present invention, this
Field those of ordinary skill central scope according to the present invention and spirit can be carried out very easily corresponding flexible or repaired
Change, therefore protection scope of the present invention should be subject to protection scope required by claims.
Claims (15)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910010777.9A CN109783539A (en) | 2019-01-07 | 2019-01-07 | Usage mining and its model building method, device and computer equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910010777.9A CN109783539A (en) | 2019-01-07 | 2019-01-07 | Usage mining and its model building method, device and computer equipment |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109783539A true CN109783539A (en) | 2019-05-21 |
Family
ID=66499135
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910010777.9A Pending CN109783539A (en) | 2019-01-07 | 2019-01-07 | Usage mining and its model building method, device and computer equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109783539A (en) |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110569428A (en) * | 2019-08-08 | 2019-12-13 | 阿里巴巴集团控股有限公司 | recommendation model construction method, device and equipment |
| CN111046952A (en) * | 2019-12-12 | 2020-04-21 | 深圳市随手金服信息科技有限公司 | Method and device for establishing label mining model, storage medium and terminal |
| CN111047011A (en) * | 2019-12-07 | 2020-04-21 | 深圳前海黑顿科技有限公司 | Scene variable automatic deep mining engine system based on machine learning model |
| CN111242239A (en) * | 2020-01-21 | 2020-06-05 | 腾讯科技（深圳）有限公司 | Training sample selection method and device and computer storage medium |
| CN112150199A (en) * | 2020-09-21 | 2020-12-29 | 北京小米松果电子有限公司 | Method and device for determining vertical characteristics and storage medium |
| WO2021081914A1 (en) * | 2019-10-31 | 2021-05-06 | 深圳市欢太科技有限公司 | Pushing object determination method and apparatus, terminal device and storage medium |
| CN112784074A (en) * | 2021-02-05 | 2021-05-11 | 腾讯科技（深圳）有限公司 | Multimedia data processing method, computer and readable storage medium |
| CN113908555A (en) * | 2020-07-09 | 2022-01-11 | 北京奇虎科技有限公司 | Target user mining method and system |
| CN114036398A (en) * | 2021-11-30 | 2022-02-11 | 北京百度网讯科技有限公司 | Content recommendation and ranking model training method, device, equipment and storage medium |
| CN114422835A (en) * | 2021-12-29 | 2022-04-29 | 上海数即数据科技有限公司 | Advertisement directional promotion platform based on big data analysis |
| CN116186585A (en) * | 2023-02-28 | 2023-05-30 | 广州朝辉智能科技有限公司 | User behavior intention mining method and device based on big data analysis |
| CN116468096A (en) * | 2023-03-30 | 2023-07-21 | 之江实验室 | A model training method, device, equipment and readable storage medium |
| CN118474180A (en) * | 2024-07-11 | 2024-08-09 | 每日互动股份有限公司 | User screening method, device, equipment and medium |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108205766A (en) * | 2016-12-19 | 2018-06-26 | 阿里巴巴集团控股有限公司 | Information-pushing method, apparatus and system |
| CN108734327A (en) * | 2017-04-20 | 2018-11-02 | 腾讯科技（深圳）有限公司 | A kind of data processing method, device and server |
-
2019
- 2019-01-07 CN CN201910010777.9A patent/CN109783539A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108205766A (en) * | 2016-12-19 | 2018-06-26 | 阿里巴巴集团控股有限公司 | Information-pushing method, apparatus and system |
| CN108734327A (en) * | 2017-04-20 | 2018-11-02 | 腾讯科技（深圳）有限公司 | A kind of data processing method, device and server |
Cited By (22)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110569428A (en) * | 2019-08-08 | 2019-12-13 | 阿里巴巴集团控股有限公司 | recommendation model construction method, device and equipment |
| CN110569428B (en) * | 2019-08-08 | 2023-10-20 | 创新先进技术有限公司 | Recommendation model construction method, device and equipment |
| WO2021081914A1 (en) * | 2019-10-31 | 2021-05-06 | 深圳市欢太科技有限公司 | Pushing object determination method and apparatus, terminal device and storage medium |
| CN114223012A (en) * | 2019-10-31 | 2022-03-22 | 深圳市欢太科技有限公司 | Push object determination method and device, terminal equipment and storage medium |
| CN111047011A (en) * | 2019-12-07 | 2020-04-21 | 深圳前海黑顿科技有限公司 | Scene variable automatic deep mining engine system based on machine learning model |
| CN111047011B (en) * | 2019-12-07 | 2023-05-23 | 深圳前海黑顿科技有限公司 | Scene variable automatic deep mining engine system based on machine learning model |
| CN111046952B (en) * | 2019-12-12 | 2023-11-14 | 深圳市铭数信息有限公司 | Method and device for establishing label mining model, storage medium and terminal |
| CN111046952A (en) * | 2019-12-12 | 2020-04-21 | 深圳市随手金服信息科技有限公司 | Method and device for establishing label mining model, storage medium and terminal |
| CN111242239A (en) * | 2020-01-21 | 2020-06-05 | 腾讯科技（深圳）有限公司 | Training sample selection method and device and computer storage medium |
| CN113908555A (en) * | 2020-07-09 | 2022-01-11 | 北京奇虎科技有限公司 | Target user mining method and system |
| CN112150199A (en) * | 2020-09-21 | 2020-12-29 | 北京小米松果电子有限公司 | Method and device for determining vertical characteristics and storage medium |
| CN112784074B (en) * | 2021-02-05 | 2022-04-22 | 腾讯科技（深圳）有限公司 | Multimedia data processing method, computer and readable storage medium |
| CN112784074A (en) * | 2021-02-05 | 2021-05-11 | 腾讯科技（深圳）有限公司 | Multimedia data processing method, computer and readable storage medium |
| CN114036398A (en) * | 2021-11-30 | 2022-02-11 | 北京百度网讯科技有限公司 | Content recommendation and ranking model training method, device, equipment and storage medium |
| CN114036398B (en) * | 2021-11-30 | 2024-02-13 | 北京百度网讯科技有限公司 | Content recommendation and ranking model training method, device, equipment and storage medium |
| CN114422835A (en) * | 2021-12-29 | 2022-04-29 | 上海数即数据科技有限公司 | Advertisement directional promotion platform based on big data analysis |
| CN116186585A (en) * | 2023-02-28 | 2023-05-30 | 广州朝辉智能科技有限公司 | User behavior intention mining method and device based on big data analysis |
| CN116186585B (en) * | 2023-02-28 | 2023-10-31 | 省广营销集团有限公司 | User behavior intention mining method and device based on big data analysis |
| CN116468096A (en) * | 2023-03-30 | 2023-07-21 | 之江实验室 | A model training method, device, equipment and readable storage medium |
| CN116468096B (en) * | 2023-03-30 | 2024-01-02 | 之江实验室 | Model training method, device, equipment and readable storage medium |
| CN118474180A (en) * | 2024-07-11 | 2024-08-09 | 每日互动股份有限公司 | User screening method, device, equipment and medium |
| CN118474180B (en) * | 2024-07-11 | 2024-09-20 | 每日互动股份有限公司 | User screening method, device, equipment and medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109783539A (en) | Usage mining and its model building method, device and computer equipment | |
| CN112163676B (en) | Method, device, equipment and storage medium for training multitasking service prediction model | |
| EP4181026A1 (en) | Recommendation model training method and apparatus, recommendation method and apparatus, and computer-readable medium | |
| CN109299994B (en) | Recommendation method, device, equipment and readable storage medium | |
| CN111325571B (en) | Automatic generation method, device and system for commodity comment labels for multitask learning | |
| US11061990B2 (en) | Generating feedback for a target content item based on published content items | |
| CN111241394B (en) | Data processing methods, devices, computer-readable storage media and electronic equipment | |
| US20120066073A1 (en) | User interest analysis systems and methods | |
| CN111859160A (en) | A method and system for session sequence recommendation based on graph neural network | |
| CN112380453B (en) | Article recommendation method and device, storage medium and equipment | |
| CN110110205A (en) | Method and device for generating recommendation information | |
| CN112989174A (en) | Information recommendation method and device, medium and equipment | |
| WO2020215977A1 (en) | System, method and device for displaying information | |
| CN113781149B (en) | Information recommendation method and device, computer readable storage medium and electronic equipment | |
| CN111967924A (en) | Commodity recommendation method, commodity recommendation device, computer device, and medium | |
| CN116894711A (en) | Commodity recommendation reason generation method and device and electronic equipment | |
| CN112749330A (en) | Information pushing method and device, computer equipment and storage medium | |
| CN115238191A (en) | Object recommendation method and device | |
| CN115760315A (en) | Commodity recommendation method, commodity recommendation device, commodity recommendation equipment and commodity recommendation medium | |
| CN116521937A (en) | Video form generation method, device, equipment, storage medium and program product | |
| CN113344648B (en) | Advertisement recommendation method and system based on machine learning | |
| CN111767459A (en) | Method and device for article recommendation | |
| CN115114519A (en) | Recommendation method, device, electronic device and storage medium based on artificial intelligence | |
| CN115221954A (en) | User portrait method, device, electronic equipment and storage medium | |
| CN113641900A (en) | Information recommendation method and device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190521 |