CN107562875A - A kind of update method of model, apparatus and system - Google Patents
A kind of update method of model, apparatus and system Download PDFInfo
- Publication number
- CN107562875A CN107562875A CN201710775357.0A CN201710775357A CN107562875A CN 107562875 A CN107562875 A CN 107562875A CN 201710775357 A CN201710775357 A CN 201710775357A CN 107562875 A CN107562875 A CN 107562875A
- Authority
- CN
- China
- Prior art keywords
- targeted customer
- information
- interest
- model
- candidate information
- 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
Abstract
The embodiment of the present application discloses a kind of update method of model, apparatus and system, and this method includes:Target candidate information is recommended to targeted customer according to the interest model of targeted customer;Obtain first behavior feedback data of the targeted customer to target candidate information;Obtain based on the interesting data for influenceing association user, update the influence relational matrix between association user and targeted customer, association user is to establish the user for having predetermined association relation with targeted customer;Respectively according to the first behavior feedback data and influence relational matrix, the interest model of targeted customer is updated.Utilize the embodiment of the present application, it can realize and interest model is updated in a manner of online updating, so that subsequently when carrying out information recommendation, recommendation results can more precisely match user's current interest focus and read scene, so as to, the subject area of the information of recommendation is more suitable for user's request, enriches the approach of information source, improves the accuracy rate of recommendation.
Description
Technical field
The application is related to field of computer technology, more particularly to a kind of update method of model, apparatus and system.
Background technology
With the continuous development of the popularization of terminal device, especially mobile terminal device, it is emerging that its sense is obtained by internet
The information such as the information of interest turn into the Main Means that people obtain information.In order to improve Consumer's Experience, generally, the net such as each information class
Standing-meeting is provided with the interest model of user, for recommending its information interested to the user.
Currently, the interest model of user is typically to be generated under offline or line using the historical data of user, specifically,
The collection historical information relevant with certain user (video that such as user watched, the news checked), then, is gone through using this
History information is trained to the interest model pre-established, the interest model after being trained.Finally, by the interest mould after training
Type is put into the service system of information recommendation so that the service system of information recommendation can based on the interest model after training to
User recommends its possible information interested.
However, generally user reading behavior lacks continuity, so, the focus for allowing for user lacks chronicity and bright
Really point to, easily shifted with the change of current events focus, therefore, the interest of user is built using offline or line under type
The model of (or training) can only catch the most important interest direction of user, the subject area of the information of recommendation can be caused narrow, limit
The approach of information source has been made, has reduced the accuracy rate of recommendation.
The content of the invention
The purpose of the embodiment of the present application is to provide a kind of update method of model, apparatus and system, to realize with online more
New mode updates interest model so that subsequently when carrying out information recommendation, it is current that recommendation results can more precisely match user
Interest focus and reading scene, so as to which the subject area of the information of recommendation is more suitable for user's request, enriches the way of information source
Footpath, improve the accuracy rate of recommendation.
In order to solve the above technical problems, what the embodiment of the present application was realized in:
A kind of update method for model that the embodiment of the present application provides, methods described include:
Target candidate information is recommended to the targeted customer according to the interest model of targeted customer;
Obtain first behavior feedback data of the targeted customer to the target candidate information;
Obtain based on the interesting data for influenceing association user, update the shadow between the association user and the targeted customer
Ring relational matrix;The association user is to establish the user for having predetermined association relation with the targeted customer;
Respectively according to the first behavior feedback data and the influence relational matrix, to the interest mould of the targeted customer
Type is updated.
Alternatively, methods described also includes:
The content of predetermined information based on the targeted customer, feature extraction is carried out to the predetermined information, obtained described
The content characteristic of predetermined information;
The interest model of the targeted customer is built according to the content characteristic.
Alternatively, the interest model according to targeted customer recommends target candidate information to the targeted customer, including:
According to the interest model, target candidate information is chosen from candidate information storehouse;
Give the target candidate information recommendation to the targeted customer.
Alternatively, it is described according to the interest model, the selection target candidate information from candidate information storehouse, including:
According to the interest model, the sequencing that each single item candidate information in the candidate information storehouse is recommended is calculated
Ranking score;
The candidate that predetermined number is chosen according to the descending order of the ranking score of the recommended sequencing believes
Breath is used as the target candidate information.
Alternatively, the ranking score of the recommended sequencing is that each single item in the candidate information storehouse is waited
Select the feedback data of estimating of information, and targeted customer imply the factor estimate confidential interval and each single item candidate information
The implicit factor of content estimates confidential interval, and is determined by way of weight;
The targeted customer, which implies the factor and the implicit factor of the content, to be determined based on the interest model.
Alternatively, it is described according to the first behavior feedback data, the interest model of the targeted customer is updated,
Including:
By predetermined alternating least-squares to interacting square between the targeted customer and the target candidate information
Battle array carries out matrix decomposition, obtains the interest information of the targeted customer for the target candidate information, wherein, the interaction
Matrix is determined by the targeted customer to the interest relation of target candidate information;
For the target candidate information, the prediction interest information of the targeted customer is obtained;
According to the interest information of the targeted customer and prediction interest information, the interest model of the targeted customer is carried out
Renewal.
Alternatively, the interest information according to the targeted customer and prediction interest information, to the targeted customer's
Interest model is updated, including:
Obtain the error between the interest information of the targeted customer and prediction interest information;
Based on the error, the interest model of the targeted customer is updated.
Alternatively, methods described also includes:
According to the first behavior feedback data, the influence relational matrix is updated;
According to the influence relational matrix after renewal, the interest model for the association user that the targeted customer is influenceed is updated.
Alternatively, methods described also includes:
According to the related information between the targeted customer prestored and association user, build the targeted customer with
Influence relational model between association user;
According to the influence relational model, the influence relational matrix between the targeted customer and association user is determined.
Alternatively, methods described also includes:
According to the first behavior feedback data, target candidate corresponding to the first behavior feedback data prestored is updated
The content hot statistics data of information.
A kind of updating device for model that the embodiment of the present application provides, described device include:
Recommending module, recommend target candidate information to the targeted customer for the interest model according to targeted customer;
Acquisition module is fed back, for obtaining first behavior feedback coefficient of the targeted customer to the target candidate information
According to;
Matrix update module, for obtaining based on the interesting data for influenceing association user, update the association user and institute
State the influence relational matrix between targeted customer;The association user is to establish to have predetermined association relation with the targeted customer
User;
Model modification module, for respectively according to the first behavior feedback data and the influence relational matrix, to institute
The interest model for stating targeted customer is updated.
Alternatively, described device also includes:
Characteristic extracting module, for the content of the predetermined information based on the targeted customer, the predetermined information is carried out
Feature extraction, obtain the content characteristic of the predetermined information;
Model construction module, for building the interest model of the targeted customer according to the content characteristic.
Alternatively, the model modification module, including:
Matrix decomposition unit, for by predetermined alternating least-squares to the targeted customer and the target candidate
Interactive matrix between information carries out matrix decomposition, obtains the interest letter for the targeted customer of the target candidate information
Breath, wherein, the Interactive matrix is determined by the targeted customer to the interest relation of target candidate information;
Interest unit is predicted, for for the target candidate information, obtaining the prediction interest information of the targeted customer;
Model modification unit, for the interest information according to the targeted customer and prediction interest information, to the target
The interest model of user is updated.
Alternatively, described device also includes:
Temperature update module, for according to the first behavior feedback data, updating the first behavior feedback prestored
The content hot statistics data of target candidate information corresponding to data.
A kind of more new system for model that the embodiment of the present application separately provides, the system include the interest mould of multiple users
Type, the interest model of the multiple user include the interest model of targeted customer and the interest model of association user, wherein:
The interest model of the targeted customer recommends first object candidate information to the targeted customer, and obtains the mesh
Mark first behavior feedback data of the user to the first object candidate information;
The interest model of the association user obtains the data for the interest for influenceing association user;Institute is updated based on the data
State the influence relational matrix between association user and the targeted customer;
The interest model of the targeted customer is based respectively on the first behavior feedback data and the influence relational matrix
It is updated.
The technical scheme provided from above the embodiment of the present application, the embodiment of the present application is according to the interest mould of targeted customer
Type recommends target candidate information to targeted customer, then, obtains first behavior feedback coefficient of the targeted customer to target candidate information
According to;In addition, obtain the influence between the targeted customer of the second behavior feedback data renewal based on association user and association user
Relational matrix, finally, respectively according to the first behavior feedback data and influence relational matrix, the interest model of targeted customer is carried out
Renewal, so, the feedback of the target candidate information by obtaining recommendation in real time, and the shadow between targeted customer and association user
Relational matrix is rung, interest model is updated in a manner of online updating so that subsequently when carrying out information recommendation, recommendation results can be with
More precisely match user's current interest focus and read scene, so as to, the subject area of the information of recommendation is more suitable for user's request,
The approach of information source is enriched, improves the accuracy rate of recommendation.
Brief description of the drawings
, below will be to embodiment or existing in order to illustrate more clearly of the embodiment of the present application or technical scheme of the prior art
There is the required accompanying drawing used in technology description to be briefly described, it should be apparent that, drawings in the following description are only this
Some embodiments described in application, for those of ordinary skill in the art, do not paying the premise of creative labor
Under, other accompanying drawings can also be obtained according to these accompanying drawings.
Fig. 1 is a kind of update method embodiment of model of the application;
Fig. 2 is a kind of structural representation of the more new system of model of the application;
Fig. 3 is a kind of update method embodiment of model of the application;
Fig. 4 is a kind of updating device embodiment of model of the application;
Fig. 5 is a kind of renewal apparatus embodiments of model of the application.
Embodiment
The embodiment of the present application provides a kind of update method of model, apparatus and system.
In order that those skilled in the art more fully understand the technical scheme in the application, it is real below in conjunction with the application
The accompanying drawing in example is applied, the technical scheme in the embodiment of the present application is clearly and completely described, it is clear that described implementation
Example only some embodiments of the present application, rather than whole embodiments.It is common based on the embodiment in the application, this area
The every other embodiment that technical staff is obtained under the premise of creative work is not made, it should all belong to the application protection
Scope.
Embodiment one
As shown in figure 1, the embodiment of the present application provides a kind of update method of model, this method can apply to to user
In the processing that the model of recommendation information is updated.The executive agent of this method can be terminal device or server, wherein, should
Terminal device can such as personal computer terminal device, server can be single server or by multiple clothes
The server cluster of business device composition.It is described in detail in the embodiment of the present application by taking server as an example, for executive agent for eventually
The situation of end equipment, the situation for being referred to server are performed, will not be repeated here.This method can specifically include following step
Suddenly:
In step s 102, target candidate information is recommended to the targeted customer according to the interest model of targeted customer.
Wherein, targeted customer can treat any user to its recommendation information, such as the user of certain video website, or
Person, the user of certain news website etc..Interest model can be the relevant informations such as the hobby based on user structure model, base
Its information interested, such as video or news can be recommended to user in the interest model, the interest model of targeted customer can
With the interest model for being the initial interest model established or having been subjected to renewal.Target candidate information can be it is any to
The information that targeted customer recommends, target candidate information can include a kind of classification, can be with for example, video class or news category etc.
Including plurality of classes, for example, video class, news category, microblogging class and question and answer class etc..
In force, with the continuous development of the popularization of terminal device, especially mobile terminal device, obtained by internet
Take the information such as its information interested turn into people obtain information Main Means, for example, user usually spend one's leisure,
Stand-by period or other trifling times before handling a certain business etc. browse and read the information such as news in internet, video
Content.In order to improve Consumer's Experience, generally, the website such as each information class can be provided with the interest model of user, for the user
Recommend its information interested.Wherein, the interest model of setting can be set or root respectively according to the difference of user
Set according to the groups of users that multiple users with same or analogous interest are formed, if for example, by the emerging of user
Interest statistics, it is found that 10 users be present has same or analogous interest, then can be that 10 users establish an interest mould
Type, it is each with the interest model can be used per family in 10 users.
The foundation of the interest model of user can include a variety of implementations, for example, (i.e. target is used for a certain user
Family), the historical data of targeted customer can be obtained (for example, type (such as social news, the amusement of the news that targeted customer browses
News or financial and economic news etc.), and the quantity of the news of each type, and/or, the video that targeted customer watches can be obtained
Historical data (for example, the type (such as TV play, film or variety show) of the video of targeted customer's viewing, and each
The quantity of the video of type), the type for the news that can be browsed based on targeted customer in the historical data of targeted customer, and
The quantity of the news of each type, and/or, targeted customer watches in the historical data for the video for obtaining targeted customer's viewing
The type of video, and the data such as quantity of the video of each type are calculated, and obtain the interest for meeting the targeted customer
The interest model of hobby, specific processing can be as:Interest model expression formula is pre-established, wherein multiple ginsengs undetermined can be included
Number, can be according to each historical data (including the feedback data of targeted customer) of targeted customer and the interest pre-established
Model expression, by way of iteration (such as 5000 feedback data, iteration 10 times, totally 5 ten thousand renewals) constantly to interest
Model is updated, so that it approaches the true interest of targeted customer, finally gives the interest model of targeted customer.
When needing to targeted customer's recommendation information, can be chosen using above-mentioned interest model from information bank to be recommended
One or more information (i.e. target candidate information) to be recommended, it is then possible to one or more letters to be recommended by selection
Breath (i.e. target candidate information) is sent to terminal device (such as mobile phone, tablet personal computer or personal computer that targeted customer uses
Deng), terminal device can show that the target candidate information of recommendation or prompting user check the server push of certain application program in time
The target candidate information sent.It should be noted that the processing to targeted customer's recommendation information can be logged in accordingly in user
Performed during application program or offline in user or perform when being not logged in corresponding application programs, the embodiment of the present application is to this
Do not limit.
In step S104, first behavior feedback data of the above-mentioned targeted customer to above-mentioned target candidate information is obtained.
Wherein, the first behavior feedback data can be related data of the targeted customer to the feedback of the information of recommendation, for example,
Targeted customer to the information of recommendation click on and checked, then the first behavior feedback data can be click on checking corresponding to operation
Data, or, targeted customer is collected to the information of recommendation, then the first behavior feedback data can be that collection operation is corresponding
Data, or, targeted customer gives the Information Sharing of recommendation to its good friend, then the first behavior feedback data can share behaviour
Data corresponding to work etc., in actual applications, the first behavior feedback data can also include a variety of in addition to including above-mentioned data
Data, it can specifically be determined according to actual conditions, for example, thumbing up data etc. corresponding to data corresponding to operation or forwarding operation.
In force, target candidate information, targeted customer are recommended to targeted customer by above-mentioned steps S102 processing
It can be believed by its terminal device in the title of the correspondence position display target candidate information of corresponding application programs and/or summary etc.
Breath, after user checks above- mentioned information, if it find that including the title and/or summary of wanting to check its particular content, Ke Yidian
The title or summary are hit, now, terminal device can detect clicking operation of the user to the title or summary, can be by this
Mark of clicking operation and corresponding target candidate information etc. generates the first behavior feedback data, can feed back first behavior
Data are sent to server, so as to which it is anti-that server can receive first behavior of the targeted customer to above-mentioned target candidate information
Present data.Meanwhile after targeted customer clicks on the title or summary, terminal device can obtain target corresponding to the title or summary
Candidate information, and the target candidate information got is shown, targeted customer may browse through the particular content of the target candidate information.
In addition, if targeted customer has carried out such as thumbing up, forward or sharing operation when browsing the target candidate information, terminal is set
It is standby to continue to generate corresponding first behavior feedback data, and the first behavior feedback data is sent to server.
In step s 106, obtain based on the interesting data for influenceing association user, update above-mentioned association user and used with target
Influence relational matrix between family, the association user are to establish the user for having predetermined association relation with the targeted customer.
Wherein, the interesting data for influenceing association user can be the data of any interest that can influence association user,
The interesting data that association user is influenceed in practical application can be including a variety of, for example, can include association user and targeted customer
Between " concern " relation, friend relation or the data such as relation in same communication group, association can also be included and used
The second behavior feedback data at family, the second feedback data can be related data of the association user to the feedback of the information of recommendation,
Checked for example, association user to the information of recommendation click on, then the second behavior feedback data can be click on checking operation
Corresponding data, or, association user is collected to the information of recommendation, then the second behavior feedback data can be collection behaviour
Data corresponding to work, or, association user gives the Information Sharing of recommendation to its good friend, then the second behavior feedback data can be
Share data etc. corresponding to operation, in actual applications, the second behavior feedback data can also wrap in addition to including above-mentioned data
A variety of data are included, can specifically be determined according to actual conditions, for example, thumbing up number corresponding to data corresponding to operation or forwarding operation
According to etc..It can be relational matrix of a certain user to the influence degree of another user to influence relational matrix, for example, such as Fig. 2 institutes
Show, a certain user has shared certain information to its good friend, after the good friend receives the information shared, also gives the Information Sharing to other use
Family, so, the behavioral implications of the user behavior of its good friend, the matrix determined with this is to influence relational matrix.Influence
Relational matrix can be determined by the behavior feedback data of one or two user in interactional two users.Association is used
Family specifically can such as above-mentioned user good friend.
In force, the interesting data for influenceing association user can be from the data related to association user prestored
Obtain (such as concern relation or friend relation), or, based on the acquisition in above-mentioned steps S102 such as the first behavior feedback data
Mode obtains (such as the second behavior feedback data).It is true that interactional relation between targeted customer and association user can be based on
The influence relational model to set the goal between user and association user, closed to simplify the influence between targeted customer and association user
System, it can be determined to influence relational matrix by the influence relational model., can be with after obtaining influenceing the data of interest of association user
Relational matrix is influenceed based on the interesting data renewal for influenceing association user, so as to the targeted customer after being updated and association user
Between influence relational matrix.
In step S108, respectively according to above-mentioned first behavior feedback data and above-mentioned influence relational matrix, to above-mentioned mesh
The interest model of mark user is updated.
In force, in order to be updated to the interest model of targeted customer, can start with from many aspects, the application is real
Apply from the first behavior feedback data in example and start with terms of influenceing relational matrix two and the interest model of targeted customer is carried out more
Newly, specifically, as shown in Fig. 2 the side of iteration can be passed through according to the first behavior feedback data and interest model of targeted customer
Formula is constantly iterated renewal to interest model, so that it further approaches the true interest of targeted customer, so as to reach pair
The purpose that the interest model of targeted customer is updated, further, it is also possible to using influence relational matrix obtained above to target
The interest model of user is updated, and can obtain being updated by the first behavior feedback data and influence relational matrix respectively
The interest model of targeted customer afterwards.
It should be noted that the interest model of targeted customer can carry out circular treatment by said process, obtained with continuous
To the interest model of the targeted customer of renewal.
The embodiment of the present application provides a kind of update method of model, is pushed away according to the interest model of targeted customer to targeted customer
Target candidate information is recommended, then, obtains first behavior feedback data of the targeted customer to target candidate information;In addition, obtain base
Influence relational matrix between the targeted customer and association user that the second behavior feedback data of association user updates, finally,
Respectively according to the first behavior feedback data and influence relational matrix, the interest model of targeted customer is updated, so, passed through
The feedback for the target candidate information recommended, and the influence relational matrix between targeted customer and association user are obtained in real time, with
The mode of online updating updates interest model so that subsequently when carrying out information recommendation, recommendation results can more precisely match use
Family current interest focus and reading scene, so as to which the subject area of the information of recommendation is more suitable for user's request, enriches information
The approach in source, improve the accuracy rate of recommendation.
Embodiment two
As shown in figure 3, the embodiment of the present application provides a kind of update method of model, this method can apply to to user
In the processing that the model of recommendation information is updated.The executive agent of this method can be terminal device or server, wherein, should
Terminal device can such as personal computer terminal device, server can be single server or by multiple clothes
The server cluster of business device composition.It is described in detail in the embodiment of the present application by taking server as an example, for executive agent for eventually
The situation of end equipment, the situation for being referred to server are performed, will not be repeated here.This method can specifically include following step
Suddenly:
In the embodiment of the present application, the interest model of targeted customer can be updated by way of on-line study, its
The interest model of middle targeted customer can be built in several ways, a kind of optional processing mode presented below, specifically can be with
Referring to following steps S302~step S306.
In step s 302, the content of the predetermined information of targeted customer is pre-processed, obtains pretreated target
Information.
Wherein, predetermined information can be any information, for example, the information for the news that targeted customer checks, targeted customer are seen
Information for the video seen etc..Target information can be the information that predetermined information obtains after pretreatment.
In force, the implementation pre-processed to the content of information can be including a variety of, for example, it is contemplated that arriving target
Redundancy may be included in the predetermined information of user, then the content of the predetermined information of targeted customer can be detected,
Determine whether to include redundancy, if comprising can pre-process the redundancy, from predetermined information
Above-mentioned redundancy is deleted, so as to obtain pretreated target information.For another example in view of in the predetermined information of targeted customer
It may include for building the skimble-skamble information of interest model, then the content of the predetermined information of targeted customer can be carried out
Detection, determines whether to include insignificant information, if comprising can be located in advance to the insignificant information
Reason, deletes above-mentioned redundancy, so as to obtain pretreated target information from predetermined information., can be with if do not included
It is without any processing to predetermined information.So, by being pre-processed to the predetermined information of targeted customer, predetermined letter can be removed
Impurity information and error message in breath etc., so as to be utilized the higher target information of rate.
In step s 304, the content based on above-mentioned target information, feature extraction is carried out to the target information, obtains the mesh
Mark the content characteristic of information.
In force, the content of each target information of targeted customer can be analyzed, and passes through predetermined spy
Levy extracting rule and feature extraction is carried out to the content of each target information, obtain the content characteristic of each target information.
In practical application, feature extraction can be carried out using different modes for different types of content, be believed with the target of news category
It exemplified by breath, can first be segmented for the textual portions in the target information, obtain multiple words, it is then possible to using hidden
The topic model algorithms such as the distribution LDA of Cray containing Di Li (Latent Dirichlet Allocation, document subject matter generation model)
Or vector corresponding to the theme of the word embedding grammar such as word2vec target information that obtains news category.For in the target information
Picture section, the picture in the target information can be identified and be classified using convolutional neural networks CNN, obtained corresponding
Content tab, in addition, in addition to information obtained above can be used as content characteristic, can also be by the length of text, picture
Quantity, the source web etc. of content can serve as content characteristic.
It should be noted that in actual applications, the processing of feature extraction is carried out to the content of target information not only can be with
LDA topic models algorithm, word2vec words embedding grammar and convolutional neural networks CNN by above-mentioned offer etc., can also lead to
Various ways realization is crossed, can specifically be set according to actual conditions, the embodiment of the present application is not limited this.
In step S306, according to the interest model of the above feature construction targeted customer.
In force, in order to simplify subsequent processes, the interest model of targeted customer can be built, so as to based on
The interest model of targeted customer is to targeted customer's recommendation information.Different target informations, the interest model of the targeted customer of structure
Difference is also tended to, for example, the target information of certain user is mainly news category information, then the interest model of the user can be based on being somebody's turn to do
Interactive relation structure between user and news category information., should if the target information of certain user is mainly video category information
The interest model of user can be based on interactive relation structure between the user and video category information etc..
By taking the matrix decomposition based on model that offline commending system is commonly used as an example, if the target information of targeted customer is mainly
News category information, targeted customer can pass throughRepresent, news category information can pass throughRepresent, targeted customer and news category information can be respectively mapped to an implicit factor space RL, then
The model parameter of the interest model of targeted customer is the implicit factor matrix P ∈ R of targeted customerM×L, above-mentioned target can be based on and used
The implicit factor matrix at family determines the interest model of targeted customer.Wherein, u1,u2…uMRepresent that each targeted customer's is hidden respectively
Containing factor vector, a1,a2…aNThe implicit factor vector of each news category information is represented respectively, and M represents targeted customer's quantity, N tables
Show the quantity of news category information.
In step S308, according to above-mentioned interest model, target candidate information is chosen from candidate information storehouse.
Wherein, candidate information storehouse can be information that the targeted customer that collects in advance may be interested or in advance
The set for the possible information interested of multiple users collected, including the possible information interested of targeted customer.
In force, a part of candidate can be screened from the global candidate information pond pre-set according to business demand
The candidate information storehouse of information structure targeted customer, a variety of different types of candidate informations can be included in candidate information storehouse, with new
Exemplified by news, hot news, the news of each theme, and the news that other offline recommended models are recommended can be included in candidate information storehouse
Deng.Each single item candidate information in candidate information storehouse can be input in the interest model of targeted customer and be calculated, obtained
Targeted customer estimates numerical value to each single item candidate information.Wherein, estimating the magnitude range of numerical value can set according to actual conditions
It is fixed, for example, the maximum for estimating numerical value is 1 and minimum value is 0, or, the maximum for estimating numerical value is 10 and minimum value is 1
Deng.The magnitude relationship of numerical value is estimated according to this, the candidate information in candidate information storehouse can be ranked up, can be chosen and estimate
Numerical value is more than the candidate information of predetermined threshold as target candidate information.
Above-mentioned steps S308 processing can be accomplished in several ways, a kind of optional processing mode presented below, tool
Body may comprise steps of one and step 2.
Step 1, according to above-mentioned interest model, calculate what each single item candidate information in above-mentioned candidate information storehouse was recommended
The ranking score of sequencing.
Wherein, the ranking score of the recommended sequencing can be each single item candidate in the candidate information storehouse
Information estimates feedback data, and targeted customer imply the factor estimate confidential interval and what content implied the factor estimates confidence
Section, and determined by way of weight.Feedback data therein of estimating can be to the anti-of certain candidate information to targeted customer
Data (such as clicking on probability etc.) are presented, targeted customer implies the factor and content implies the factor (i.e. in the interest model of targeted customer
Model parameter) be based on above-mentioned interest model determine.Targeted customer therein implies the factor and the implicit factor of content can be with table
Targeted customer's potential interest that may be present is levied, also characterizes whether interested in the content of a certain information not true of targeted customer
It is qualitative.So, on the one hand targeted customer accurate can be determined whether in corresponding information by estimating feedback data
Hold (showing as the recommendation to interest known to targeted customer) interested, on the other hand implying estimating for the factor by targeted customer puts
Letter section and content imply the factor estimate confidential interval come goal seeking user may the content of information interested (show as
Exploration to the unknown interest of targeted customer), based in terms of above-mentioned two to targeted customer's recommendation information.
Specifically, for example, being all news category candidate information in candidate information storehouse, (originally show including three candidate informations
Candidate's news is referred to as in example) a, b and c, it has the behavior feedback data of 300,2 and 50 users, can calculated respectively
The respective implicit factor.Assuming that targeted customer has produced the behavior feedback data to 100 news, then can calculate its it is implicit because
Son, and click probability of the targeted customer to candidate's news a, b and c is calculated according to this, so as to be estimated feedback data accordingly.
Row matrix point can be entered to the target Interactive matrix between the targeted customer prestored and the news category information content
Solution, obtain targeted customer and imply the factor and the implicit factor of content.For different types of content, it is usual that targeted customer implies the factor
A species normal distribution can be met, accordingly, for different user, content implies the factor and generally will suffice for a species normal state point
Cloth, it is assumed that targeted customer or content imply factor Normal Distribution N (μ, σ2), wherein, μ is the phase that targeted customer implies the factor
Hope, and as the targeted customer being collected into increases the feedback of different types of content, interest model gradually updates so as to mesh
Mark user, which implies estimating for the factor, can move closer to real μ, and corresponding σ can be also gradually reduced, therefore can be weighed with σ therein
The uncertainty degree of discreet value is measured, and can determine that targeted customer is implied estimating for the factor and put by the uncertainty degree of discreet value
Letter section or content, which imply the factor, estimates confidential interval, such as (- k σ, k σ), wherein, k>0, and k is constant, specifically such as, σ is
0.1, then the uncertainty degree estimated to the corresponding implicit factor can be represented with 0.1k, to participate in the calculating of ranking score.It is logical
Above-mentioned example is crossed, because above-mentioned candidate's news a, b and c have the behavior feedback data of 300,2 and 50 users respectively, then to waiting
Select news a estimating for the implicit factor of content can be relatively more accurate, candidate's news c takes second place, and candidate's news b accuracy is worst.
When calculating ranking score, known portions and unknown portions are balanced, candidate's news a ranking score is essentially from targeted customer to waiting
News a feedback data of estimating is selected, and candidate's news b ranking score is implied the σ of the distribution of the factor by candidate's news b content
Influence bigger.
It should be noted that above-mentioned floating numerical value is determined based on the half for estimating confidential interval size, in practical application
In can also determine otherwise, can specifically be determined according to actual conditions, specifically such as estimate confidential interval the upper bound or
Lower bound, or, estimate the upper bound of confidential interval or the prearranged multiple (such as 0.1 times, 0.5 times or 2 times) of lower bound etc..
In force, MAB (Multi-armed bandit, multi-arm Slot Machine) problem model can be based on, will be by emerging
Interesting model carries out recommending to be combined with carrying out exploration to the interest of user.In the embodiment of the present application, in selection to targeted customer's
During recommendation information, can calculate each single item candidate information in candidate information storehouse UCB (Upper confidence bound,
The confidential interval upper bound) numerical value, i.e. model parameter in the interest model of targeted customer the numerical value for estimating the confidential interval upper bound.Base
In above-mentioned steps S306 example, targeted customer can pass throughRepresent, news category information can pass throughRepresent, then the UCB numerical value of candidate information can be obtained by below equation (1)
Wherein, t represents interaction times, and Section 1 is to recommend target candidate information a's to estimate feedback coefficient to targeted customer u
According to second and third is embodied respectivelyWithUncertainty, and with interest explore increasing for number be gradually reduced, make UCB
Numerical value converges on true feedback data,WithIt is illustrated respectively in the t times and recommends target candidate information (i.e. to targeted customer u
Above-mentioned news category information) when, the implicit factor vector p of the targeted customer uuDiscreet value and target candidate information a it is implicit because
Subvector qaDiscreet value.WithRepresent respectivelyWithUncertainty degree.α1And α2, can be with for two constants
Control respectivelyWithTo UCB influence degree.
Based on the above, the news of prioritizing selection while interest known to targeted customer is met, to targeted customer its
He is explored unknown interest, ensure that the freshness and diversity of recommendation information.
Step 2, predetermined number is chosen according to the descending order of the ranking score of above-mentioned recommended sequencing
Candidate information is as target candidate information.
Wherein, predetermined number can be set according to actual conditions, specific such as 10 or 20.
In step S310, by above-mentioned target candidate information recommendation to targeted customer.
In step S312, first behavior feedback data of the targeted customer to above-mentioned target candidate information is obtained.
Wherein, the first behavior feedback data packet includes but is not limited to browse, click on, score, scroll through pages, thumbs up, steps on, receiving
Hide, forward, sharing, skip this song, check the lyrics, being circulated into details page, into special edition page and single etc..
Above-mentioned steps S312 step content is identical with the step content of the step S104 in above-described embodiment one, specifically may be used
Referring to the processing procedure of the step S104 in above-described embodiment one, will not be repeated here.
It should be noted that after getting targeted customer to the first behavior feedback data of above-mentioned target candidate information, can
So that the first behavior feedback data is sent in the service system of information recommendation, the service system of information recommendation is collected into target use
After the first behavior feedback data at family, different types of feedback can be united, obtain rua∈ [0 ,+∞), wherein, 0 table
Show that targeted customer loses interest in completely to the target candidate information of recommendation, uuaIt is bigger to represent target candidate of the targeted customer to recommendation
Information is interested.For example, targeted customer clicked to enter recommend target candidate information where the page after, read about 30%
The list of the target candidate information of recommendation is returned afterwards, then can use ruaTarget candidate of=0.3 approximate representation targeted customer to recommendation
The fancy grade of information.
In actual applications, it can start with from the first behavior feedback data and in terms of influenceing relational matrix two and target is used
The interest model at family is updated, wherein, the relevant treatment for influenceing relational matrix may refer to following step S314 and step
S316。
In step S314, according to the related information between the targeted customer and association user prestored, target is built
Influence relational model between user and association user.
Wherein, related information can be for characterizing certain existing incidence relation between targeted customer and association user
Information, such as the relevant information of friend relation, relevant information of concern relation etc., it can specifically include such as in above-mentioned relevant information
Information for the degree that influences each other etc..
In force, the service system of information recommendation needs to carry out the interest of targeted customer the exploration of a period of time,
Effective recommendation can be produced.And during above-mentioned exploration, if the overlong time explored can cause the loss of user.This
Application is to degree (the i.e. targeted customer that interacts in terms of the interest hobby between all users in the service system of information recommendation
Related information between association user) establish model S ∈ RM×M, i.e., the influence relation mould between targeted customer and association user
Type.
In step S316, according to above-mentioned influence relational model, determine that the influence between targeted customer and association user is closed
It is matrix.
In force, the example based on above-mentioned steps S314, the influence relational model between targeted customer and association user
For S ∈ RM×M, then influence relational matrix of the user i to user j can be expressed as
Wherein, SijFor the degree of correlation between user i and user j, i, j, k ∈ [0, M).
It should be noted that different types of information can be determined using different methods targeted customer and association user it
Between influence relational matrix.For example, for microblogging and socialization music service, the concern relation between user can be utilized to establish
Influence relational matrix between targeted customer and association user., can with reference to user's reading histories and article feature in the present embodiment
To go out targeted customer's Long-term Interest I using Logic Regression Models off-line calculationu∈RH, then between targeted customer and association user
Influence relational model
Wherein, H is the dimension of targeted customer's Long-term Interest vector.
In actual applications, the first behavior feedback data can include explicit feedback data and implicit feedback data, explicitly
Feedback data can be the feedback data that can be directly viewable by some methods or means, and explicit feedback data can include
Score, thumb up, implicit feedback data can be feedback data that can not be by plain mode to be directly viewable, implicit anti-
Feedback data can include browsing, clicking on, scroll through pages, collecting and share.Based on explicit feedback data and implicit feedback data,
The processing for carrying out model modification may comprise steps of S318~step S320.
In step S318, obtain based on the interesting data for influenceing association user, renewal association user and targeted customer it
Between influence relational matrix, association user is to establish to have the user of predetermined association relation with targeted customer.
Above-mentioned steps S318 step content is identical with the step content of the step S106 in above-described embodiment one, specifically may be used
Referring to the processing procedure of the step S106 in above-described embodiment one, will not be repeated here.
In step s 320, by predetermined alternating least-squares to the friendship between targeted customer and target candidate information
Mutual matrix carries out matrix decomposition, obtains the interest information of the targeted customer for target candidate information, wherein, Interactive matrix is by mesh
User is marked to determine the interest relation of target candidate information.
In step S322, for above-mentioned target candidate information, the prediction interest information of targeted customer is obtained.
In force, inaccurate model can be established according to the interest information of targeted customer, inaccurate model can be used for
The interest of targeted customer is predicted, hobby of the targeted customer to target candidate information can be predicted by inaccurate model,
So as to obtain the prediction interest information of targeted customer.
In step S324, according to the interest information of targeted customer and prediction interest information, to the interest mould of targeted customer
Type is updated.
The processing for the above-mentioned steps S324 for needing to illustrate can include a variety of implementations, a kind of feasible place presented below
Reason mode, specifically may comprise steps of one and step 2.
Step 1, obtain the error between the interest information of targeted customer and prediction interest information.
Wherein, the error can be characterized by target loss function.
Step 2, based on above-mentioned error, the interest model of targeted customer is updated.
Above-mentioned steps S320 and step S324 specific processing may refer to the description below:
In force, by taking news category information as an example, the service system of information recommendation is generally difficult to collect the aobvious of targeted customer
Formula feedback data (such as scores), in order to make full use of implicit feedback number present in the first behavior feedback data of targeted customer
According to (such as browse, read, share), A weighting LS (Alternating least squares, alternately least square can be used
Method) carry out matrix decomposition.Target loss function if necessary to optimization is
Above-mentioned expression formula is solved, following result can be obtained:
Pu=(QTWuQ+λ2I)-1QTWuru, qa=(PTWaP+λ1I)-1PTWara... ... (5)
Wherein, λ1And λ2Respectively puAnd qaL2 regular terms coefficient.P and Q is respectively targeted customer and target candidate letter
The implicit factor matrix of breath, WuWeight matrix for targeted customer u to the feedback data of all target candidate information, wherein,Weight for targeted customer u to target candidate information a feedback data, similarly, WaIt is all targeted customers to target
The weight matrix of candidate information a feedback data.I is unit matrix, ruAnd raRespectively targeted customer is believed all target candidates
The feedback matrix of the feedback matrix of breath and all targeted customers to target candidate information.
In the case of on-line study, above-mentioned expression formula can be converted to the form of iteration renewal, meanwhile, it can utilize
Sherman-Morrison formula optimization matrix inversion operations, can respectively obtain following result:
Wherein, AuAnd buIt is that p is calculated by ALS algorithmsu,tIntermediate variable.
It can also similarly obtain:
Now,
Wherein, wuaFor the weight of a certain item feedback data, for positive feedback data (such as user exist click on), then wua
Take 1.CaAnd daIt is that q is calculated by ALS algorithmsa,tIntermediate variable.Wherein, Au、bu、CaAnd daInitialization:A1=λ1I1, b1=
0L, C1=λ2I2, d1=0L.For negative factor according to (such as user does not click on), the service system of information recommendation is typically by unified
Weight processing, the embodiment of the present application makes full use of the implicit feedback data of targeted customer, to different negative factors according to setting not
Same weight,
Wherein, above-mentioned wu,aValue is between 0-1.W therein0For the criteria weights of negative factor evidence,For according to implicit
The content temperature for the target candidate information a that feedback data is recommended, that is, the news CTR obtained (Click through rate, are clicked on
Rate) value, β is the significance level of control popular information, and as β > 1, the importance of popular information is reinforced, popular as β < 1
The importance of information is suppressed.Equivalent to w during β=0u,aTake unified weight w0/N.By controlling the weight of negative factor evidence, this
Application can be to user interest modeling it is more accurate.
In addition, collaborative filtering needs to be learnt according to user feedback data, for ageing stronger information, each is new
Content will face the problem of cold start-up-i.e. not no or seldom pushed away in the case of its behavior feedback data of user feedback
Recommend.The embodiment of the present application introduces the characteristic vector based on the information content, and the characteristic vector of the information content is updated to, wherein,For known features, va∈RlCan be random value during initialization to imply the factor.Accordingly
, user characteristics vector is updated to pu=(xu, vu).By the extraction and processing to explicit features therein, can reduce to pushing away
The uncertainty for estimating feedback data of information is recommended, accelerates pace of learning.
In step S326, according to the influence relational matrix of renewal, to the interest of the targeted customer updated in step S324
Model is updated.
Above-mentioned steps S326 specific processing may refer to the related content in above-described embodiment one in step S108, herein
Repeat no more.
In addition, the targeted customer that can be not only updated by the data based on the interest for influenceing association user and association user
Between influence relational matrix, can also according to the first behavior feedback data update influence relational matrix, and and then update association
The interest model of user, it specifically may refer to following step S402.
In step S402, according to above-mentioned first behavior feedback data, above-mentioned influence relational matrix is updated.
In force, as shown in Fig. 2 the first obtained behavior feedback data can be input to above-mentioned steps S316 public affairs
Formula is calculated in (2), to update influence relational matrix.
In step s 404, the association user influenceed according to the influence relational matrix after renewal, renewal targeted customer
Interest model.
Above-mentioned steps S404 concrete processing procedure, it may refer to phase in the step S106 in above-described embodiment one inside the Pass
Hold, will not be repeated here.
In addition, can not only be updated according to the first behavior feedback data influences relational matrix, and and then renewal association user
Interest model, corresponding content hot statistics data can also be updated according to the first behavior feedback data, specifically may refer to
Following step S502.
In step S502, according to above-mentioned first behavior feedback data, the first behavior feedback data prestored is updated
The content hot statistics data of corresponding target candidate information.
Wherein, content hot statistics data can be used to indicate that hobby journey of the current goal user to certain target candidate information
Degree, specific to represent that user is interested in corresponding target candidate information all the time such as content hot statistics data for 1, content temperature system
Count and corresponding target candidate information is lost interest in all the time for 0 expression user.Content hot statistics data can basis
The click of user, the behavior fed back statistics such as thumb up, share, commenting on and obtain., can be with use information content in the embodiment of the present application
CTR is as corresponding content hot statistics data ha=CTRa=Clicksa/Impressionsa, i.e. clicking rate=click time
Number/exposure frequency.
The embodiment of the present application provides a kind of update method of model, is pushed away according to the interest model of targeted customer to targeted customer
Target candidate information is recommended, then, obtains first behavior feedback data of the targeted customer to target candidate information;In addition, obtain base
Influence relational matrix between the targeted customer and association user that the second behavior feedback data of association user updates, finally,
Respectively according to the first behavior feedback data and influence relational matrix, the interest model of targeted customer is updated, so, passed through
The feedback for the target candidate information recommended, and the influence relational matrix between targeted customer and association user are obtained in real time, with
The mode of online updating updates interest model so that subsequently when carrying out information recommendation, recommendation results can more precisely match use
Family current interest focus and reading scene, so as to which the subject area of the information of recommendation is more suitable for user's request, enriches information
The approach in source, improve the accuracy rate of recommendation.
Embodiment three
The update method of the model provided above for the embodiment of the present application, based on same thinking, the embodiment of the present application is also
A kind of updating device of model is provided, as shown in Figure 4.
The updating device of the model includes:Recommending module 601, feedback acquisition module 602, the and of matrix acquisition module 603
Model modification module 604, wherein:
Recommending module 601, recommend target candidate information to the targeted customer for the interest model according to targeted customer;
Acquisition module 602 is fed back, for obtaining first behavior feedback of the targeted customer to the target candidate information
Data;
Matrix update module 603, for obtain based on influence association user interesting data, update the association user with
Influence relational matrix between the targeted customer;The association user is to establish to have predetermined association relation with the targeted customer
User;
Model modification module 604 is right for respectively according to the first behavior feedback data and the influence relational matrix
The interest model of the targeted customer is updated.
In the embodiment of the present application, described device also includes:
Characteristic extracting module, for the content of the predetermined information based on the targeted customer, the predetermined information is carried out
Feature extraction, obtain the content characteristic of the predetermined information;
Model construction module, for building the interest model of the targeted customer according to the content characteristic.
In the embodiment of the present application, described device also includes:
Pretreatment module, the content for the predetermined information to targeted customer pre-process, and obtain pretreated mesh
Mark information;
Accordingly, the characteristic extracting module, for the content based on the target information, the target information is carried out
Feature extraction, obtain the content characteristic of the target information.
In the embodiment of the present application, the recommending module 601, including:
Information extracting unit, for according to the interest model, target candidate information to be chosen from candidate information storehouse;
Recommendation unit, for giving the target candidate information recommendation to the targeted customer.
In the embodiment of the present application, described information chooses unit, for according to the interest model, calculating candidate's letter
Cease the ranking score for the sequencing that each single item candidate information in storehouse is recommended;According to the row of the recommended sequencing
The descending order of sequence fraction chooses the candidate information of predetermined number as the target candidate information.
In the embodiment of the present application, the ranking score of the recommended sequencing is according in the candidate information storehouse
Each single item candidate information estimate feedback data, and targeted customer imply the factor estimate confidential interval and content it is implicit because
Son estimates confidential interval, and is determined by way of weight;
The targeted customer, which implies the factor and the implicit factor of the content, to be determined based on the interest model.
In the embodiment of the present application, the model modification module 604, including:
Matrix decomposition unit, for by predetermined alternating least-squares to the targeted customer and the target candidate
Interactive matrix between information carries out matrix decomposition, obtains the interest letter for the targeted customer of the target candidate information
Breath, wherein, the Interactive matrix is determined by the targeted customer to the interest relation of target candidate information;
Interest unit is predicted, for for the target candidate information, obtaining the prediction interest information of the targeted customer;
Model modification unit, for the interest information according to the targeted customer and prediction interest information, to the target
The interest model of user is updated.
In the embodiment of the present application, the model modification unit, for obtaining the interest information of the targeted customer and pre-
The error surveyed between interest information;Based on the error, the interest model of the targeted customer is updated.
In the embodiment of the present application, the error is characterized by target loss function.
In the embodiment of the present application, described device also includes:
Matrix update module, for according to the first behavior feedback data, updating the influence relational matrix;
The model modification module, for according to the influence relational matrix after renewal, updating the targeted customer to be influenceed
Association user interest model.
In the embodiment of the present application, described device also includes:
Relational model builds module, for associating letter according between the targeted customer and association user prestored
Breath, builds the influence relational model between the targeted customer and association user;
Matrix deciding module, for according to the influence relational model, determining between the targeted customer and association user
Influence relational matrix.
In the embodiment of the present application, described device also includes:
Temperature update module, for according to the first behavior feedback data, updating the first behavior feedback prestored
The content hot statistics data of target candidate information corresponding to data.
In the embodiment of the present application, the first behavior feedback data packet include browse, click on, scroll through pages, thumb up, step on,
Collect and share.
The embodiment of the present application provides a kind of updating device of model, is pushed away according to the interest model of targeted customer to targeted customer
Target candidate information is recommended, then, obtains first behavior feedback data of the targeted customer to target candidate information;In addition, obtain base
Influence relational matrix between the targeted customer and association user that the second behavior feedback data of association user updates, finally,
Respectively according to the first behavior feedback data and influence relational matrix, the interest model of targeted customer is updated, so, passed through
The feedback for the target candidate information recommended, and the influence relational matrix between targeted customer and association user are obtained in real time, with
The mode of online updating updates interest model so that subsequently when carrying out information recommendation, recommendation results can more precisely match use
Family current interest focus and reading scene, so as to which the subject area of the information of recommendation is more suitable for user's request, enriches information
The approach in source, improve the accuracy rate of recommendation.
Example IV
The updating device of the model provided above for the embodiment of the present application, based on same thinking, the embodiment of the present application is also
A kind of more new system of model is provided, as shown in Figure 2.
The more new system of the model includes the interest model of multiple users, and the interest model of the multiple user includes
The interest model of targeted customer and the interest model of association user, as shown in Fig. 2 the interest model of association user can include closing
The interest model of the interest model at combination family 1, interest model ... the association user n of association user 2, wherein:
The interest model of the targeted customer recommends first object candidate information to the targeted customer, and obtains the mesh
Mark first behavior feedback data of the user to the first object candidate information;
The interest model of the association user obtains the data for the interest for influenceing association user;Institute is updated based on the data
State the influence relational matrix between association user and the targeted customer;
The interest model of the targeted customer is based respectively on the first behavior feedback data and the influence relational matrix
It is updated.
Alternatively, the content of predetermined information of the interest model of the targeted customer based on the targeted customer, to described
Predetermined information carries out feature extraction, obtains the content characteristic of the predetermined information;The target is built according to the content characteristic
The interest model of user.
Alternatively, the interest model of the targeted customer pre-processes to the content of the predetermined information of targeted customer, obtains
To pretreated target information;Accordingly, the content of the predetermined information based on the targeted customer, to the predetermined letter
Breath carries out feature extraction, obtains the content characteristic of the predetermined information, including:Based on the content of the target information, to described
Target information carries out feature extraction, obtains the content characteristic of the target information.
Alternatively, the interest model of the targeted customer chooses target according to the interest model from candidate information storehouse
Candidate information;Give the target candidate information recommendation to the targeted customer.
Alternatively, the interest model of the targeted customer is calculated in the candidate information storehouse according to the interest model
The ranking score of the recommended sequencing of each single item candidate information;According to the ranking score of the recommended sequencing by
Small order is arrived greatly chooses the candidate information of predetermined number as the target candidate information.
Alternatively, the ranking score of the recommended sequencing is that each single item in the candidate information storehouse is waited
Select the feedback data of estimating of information, and targeted customer imply the factor estimate confidential interval and content implies estimating for the factor and put
Believe section, and determined by way of weight;
The targeted customer, which implies the factor and the implicit factor of the content, to be determined based on the interest model.
Alternatively, the interest model of the targeted customer by predetermined alternating least-squares to the targeted customer with
Interactive matrix between the target candidate information carries out matrix decomposition, obtains the target for the target candidate information
The interest information of user, wherein, the Interactive matrix is determined by the targeted customer to the interest relation of target candidate information;Pin
To the target candidate information, the prediction interest information of the targeted customer is obtained;According to the interest information of the targeted customer
With prediction interest information, the interest model of the targeted customer is updated.
Alternatively, the interest information according to the targeted customer and prediction interest information, to the targeted customer's
Interest model is updated, including:
Obtain the error between the interest information of the targeted customer and prediction interest information;
Based on the error, the interest model of the targeted customer is updated.
Alternatively, the error is characterized by target loss function.
Alternatively, the interest model of the targeted customer updates the influence and closed according to the first behavior feedback data
It is matrix;According to the influence relational matrix after renewal, the interest model for the association user that the targeted customer is influenceed is updated.
Alternatively, the interest model of the targeted customer is according between the targeted customer prestored and association user
Related information, build the influence relational model between the targeted customer and association user;According to the influence relational model,
Determine the influence relational matrix between the targeted customer and association user.
Alternatively, the interest model of the targeted customer updates what is prestored according to the first behavior feedback data
The content hot statistics data of target candidate information corresponding to first behavior feedback data.
Alternatively, the first behavior feedback data packet, which includes, browses, clicks on, scroll through pages, thumbing up, step on, collect and share.
The embodiment of the present application provides a kind of more new system of model, is pushed away according to the interest model of targeted customer to targeted customer
Target candidate information is recommended, then, obtains first behavior feedback data of the targeted customer to target candidate information;In addition, obtain base
Influence relational matrix between the targeted customer and association user that the second behavior feedback data of association user updates, finally,
Respectively according to the first behavior feedback data and influence relational matrix, the interest model of targeted customer is updated, so, passed through
The feedback for the target candidate information recommended, and the influence relational matrix between targeted customer and association user are obtained in real time, with
The mode of online updating updates interest model so that subsequently when carrying out information recommendation, recommendation results can more precisely match use
Family current interest focus and reading scene, so as to which the subject area of the information of recommendation is more suitable for user's request, enriches information
The approach in source, improve the accuracy rate of recommendation.
Embodiment five
The more new system of the model provided above for the embodiment of the present application, based on same thinking, the embodiment of the present application is also
A kind of more new equipment of model is provided, as shown in Figure 5.
The more new equipment of the model can be server that above-described embodiment provides etc..
The more new equipment of model can produce bigger difference because configuration or performance are different, can include one or one
More than processor 701 and memory 702, can be stored with memory 702 one or more storage application programs or
Data.Wherein, memory 702 can be of short duration storage or persistently storage.Being stored in the application program of memory 702 can include
One or more modules (diagram is not shown), each module can include to the series of computation in the more new equipment of model
Machine executable instruction.Further, processor 701 could be arranged to communicate with memory 702, in model more on new equipment
Perform the series of computation machine executable instruction in memory 702.The more new equipment of model can also include one or one with
Upper power supply 703, one or more wired or wireless network interfaces 704, one or more input/output interfaces 705,
One or more keyboards 706.
Specifically in the present embodiment, the more new equipment of model includes memory, and one or more program,
One of them or more than one program storage is in memory, and one or more than one program can include one or one
Individual with upper module, and each module can include to the series of computation machine executable instruction in the more new equipment of model, and pass through
Configuration is so that by one, either more than one computing device this or more than one program bag contain for carrying out following calculate
Machine executable instruction:
Target candidate information is recommended to the targeted customer according to the interest model of targeted customer;
Obtain first behavior feedback data of the targeted customer to the target candidate information;
Obtain based on the interesting data for influenceing association user, update the shadow between the association user and the targeted customer
Ring relational matrix;The association user is to establish the user for having predetermined association relation with the targeted customer;
Respectively according to the first behavior feedback data and the influence relational matrix, to the interest mould of the targeted customer
Type is updated.
Alternatively, the executable instruction when executed, can also make the processor:
The content of predetermined information based on the targeted customer, feature extraction is carried out to the predetermined information, obtained described
The content characteristic of predetermined information;
The interest model of the targeted customer is built according to the content characteristic.
Alternatively, the executable instruction when executed, can also make the processor:
The content of the predetermined information of targeted customer is pre-processed, obtains pretreated target information;
Accordingly, the content of the predetermined information based on the targeted customer, feature is carried out to the predetermined information and carried
Take, obtain the content characteristic of the predetermined information, including:
Based on the content of the target information, feature extraction is carried out to the target information, obtains the target information
Content characteristic.
Alternatively, the executable instruction when executed, can also make the processor:
According to the interest model, target candidate information is chosen from candidate information storehouse;
Give the target candidate information recommendation to the targeted customer.
Alternatively, the executable instruction when executed, can also make the processor:
According to the interest model, the sequencing that each single item candidate information in the candidate information storehouse is recommended is calculated
Ranking score;
The candidate that predetermined number is chosen according to the descending order of the ranking score of the recommended sequencing believes
Breath is used as the target candidate information.
Alternatively, the ranking score of the recommended sequencing is that each single item in the candidate information storehouse is waited
Select the feedback data of estimating of information, and targeted customer imply the factor estimate confidential interval and content implies estimating for the factor and put
Believe section, and determined by way of weight;
The targeted customer, which implies the factor and the implicit factor of the content, to be determined based on the interest model.
Alternatively, the executable instruction when executed, can also make the processor:
By predetermined alternating least-squares to interacting square between the targeted customer and the target candidate information
Battle array carries out matrix decomposition, obtains the interest information of the targeted customer for the target candidate information, wherein, the interaction
Matrix is determined by the targeted customer to the interest relation of target candidate information;
For the target candidate information, the prediction interest information of the targeted customer is obtained;
According to the interest information of the targeted customer and prediction interest information, the interest model of the targeted customer is carried out
Renewal.
Alternatively, the executable instruction when executed, can also make the processor:
Obtain the error between the interest information of the targeted customer and prediction interest information;
Based on the error, the interest model of the targeted customer is updated.
Alternatively, the error is characterized by target loss function.
Alternatively, the executable instruction when executed, can also make the processor:
According to the first behavior feedback data, the influence relational matrix is updated;
According to the influence relational matrix after renewal, the interest model for the association user that the targeted customer is influenceed is updated.
Alternatively, the executable instruction when executed, can also make the processor:
According to the related information between the targeted customer prestored and association user, build the targeted customer with
Influence relational model between association user;
According to the influence relational model, the influence relational matrix between the targeted customer and association user is determined.
Alternatively, the executable instruction when executed, can also make the processor:
According to the first behavior feedback data, target candidate corresponding to the first behavior feedback data prestored is updated
The content hot statistics data of information.
Alternatively, the first behavior feedback data packet, which includes, browses, clicks on, scroll through pages, thumbing up, step on, collect and share.
The embodiment of the present application provides a kind of more new equipment of model, is pushed away according to the interest model of targeted customer to targeted customer
Target candidate information is recommended, then, obtains first behavior feedback data of the targeted customer to target candidate information;In addition, obtain base
Influence relational matrix between the targeted customer and association user that the second behavior feedback data of association user updates, finally,
Respectively according to the first behavior feedback data and influence relational matrix, the interest model of targeted customer is updated, so, passed through
The feedback for the target candidate information recommended, and the influence relational matrix between targeted customer and association user are obtained in real time, with
The mode of online updating updates interest model so that subsequently when carrying out information recommendation, recommendation results can more precisely match use
Family current interest focus and reading scene, so as to which the subject area of the information of recommendation is more suitable for user's request, enriches information
The approach in source, improve the accuracy rate of recommendation.
Embodiments herein is the foregoing is only, is not limited to the application.For those skilled in the art
For, the application can have various modifications and variations.All any modifications made within spirit herein and principle, it is equal
Replace, improve etc., it should be included within the scope of claims hereof.
Claims (15)
1. a kind of update method of model, it is characterised in that methods described includes:
Target candidate information is recommended to the targeted customer according to the interest model of targeted customer;
Obtain first behavior feedback data of the targeted customer to the target candidate information;
The influence based on the interesting data for influenceing association user, updated between the association user and the targeted customer is obtained to close
It is matrix;The association user is to establish the user for having predetermined association relation with the targeted customer;
Respectively according to the first behavior feedback data and the influence relational matrix, the interest model of the targeted customer is entered
Row renewal.
2. according to the method for claim 1, it is characterised in that methods described also includes:
The content of predetermined information based on the targeted customer, feature extraction is carried out to the predetermined information, obtained described predetermined
The content characteristic of information;
The interest model of the targeted customer is built according to the content characteristic.
3. method according to claim 1 or 2, it is characterised in that the interest model according to targeted customer is to described
Targeted customer recommends target candidate information, including:
According to the interest model, target candidate information is chosen from candidate information storehouse;
Give the target candidate information recommendation to the targeted customer.
4. according to the method for claim 3, it is characterised in that it is described according to the interest model, from candidate information storehouse
Target candidate information is chosen, including:
According to the interest model, the row of the recommended sequencing of each single item candidate information in the candidate information storehouse is calculated
Sequence fraction;
The candidate information that predetermined number is chosen according to the descending order of the ranking score of the recommended sequencing is made
For the target candidate information.
5. according to the method for claim 4, it is characterised in that the ranking score of the recommended sequencing is basis
Each single item candidate information in the candidate information storehouse estimates feedback data, and targeted customer imply the factor estimate confidence
What section and content implied the factor estimates confidential interval, and is determined by way of weight;
The targeted customer, which implies the factor and the implicit factor of the content, to be determined based on the interest model.
6. according to the method for claim 1, it is characterised in that it is described according to the first behavior feedback data, to described
The interest model of targeted customer is updated, including:
The Interactive matrix between the targeted customer and the target candidate information is entered by predetermined alternating least-squares
Row matrix is decomposed, and obtains the interest information of the targeted customer for the target candidate information, wherein, the Interactive matrix
The interest relation of target candidate information is determined by the targeted customer;
For the target candidate information, the prediction interest information of the targeted customer is obtained;
According to the interest information of the targeted customer and prediction interest information, the interest model of the targeted customer is carried out more
Newly.
7. according to the method for claim 6, it is characterised in that the interest information and prediction according to the targeted customer
Interest information, the interest model of the targeted customer is updated, including:
Obtain the error between the interest information of the targeted customer and prediction interest information;
Based on the error, the interest model of the targeted customer is updated.
8. according to the method for claim 1, it is characterised in that methods described also includes:
According to the first behavior feedback data, the influence relational matrix is updated;
According to the influence relational matrix after renewal, the interest model for the association user that the targeted customer is influenceed is updated.
9. according to the method for claim 8, it is characterised in that methods described also includes:
According to the related information between the targeted customer prestored and association user, the targeted customer is built with associating
Influence relational model between user;
According to the influence relational model, the influence relational matrix between the targeted customer and association user is determined.
10. according to the method for claim 1, it is characterised in that methods described also includes:
According to the first behavior feedback data, target candidate information corresponding to the first behavior feedback data prestored is updated
Content hot statistics data.
11. a kind of updating device of model, it is characterised in that described device includes:
Recommending module, recommend target candidate information to the targeted customer for the interest model according to targeted customer;
Acquisition module is fed back, for obtaining first behavior feedback data of the targeted customer to the target candidate information;
Matrix update module, for obtaining based on the interesting data for influenceing association user, update the association user and the mesh
Mark the influence relational matrix between user;The association user is to establish the use for having predetermined association relation with the targeted customer
Family;
Model modification module, for respectively according to the first behavior feedback data and the influence relational matrix, to the mesh
The interest model of mark user is updated.
12. device according to claim 11, it is characterised in that described device also includes:
Characteristic extracting module, for the content of the predetermined information based on the targeted customer, feature is carried out to the predetermined information
Extraction, obtains the content characteristic of the predetermined information;
Model construction module, for building the interest model of the targeted customer according to the content characteristic.
13. device according to claim 11, it is characterised in that the model modification module, including:
Matrix decomposition unit, for by predetermined alternating least-squares to the targeted customer and the target candidate information
Between Interactive matrix carry out matrix decomposition, obtain for the target candidate information the targeted customer interest information,
Wherein, the Interactive matrix is determined by the targeted customer to the interest relation of target candidate information;
Interest unit is predicted, for for the target candidate information, obtaining the prediction interest information of the targeted customer;
Model modification unit, for the interest information according to the targeted customer and prediction interest information, to the targeted customer
Interest model be updated.
14. device according to claim 11, it is characterised in that described device also includes:
Temperature update module, for according to the first behavior feedback data, updating the first behavior feedback data prestored
The content hot statistics data of corresponding target candidate information.
15. the more new system of a kind of model, it is characterised in that the system includes the interest model of multiple users, the multiple
The interest model of user includes the interest model of targeted customer and the interest model of association user, wherein:
The interest model of the targeted customer recommends first object candidate information to the targeted customer, and obtains the target and use
First behavior feedback data of the family to the first object candidate information;
The interest model of the association user obtains the data for the interest for influenceing association user;The pass is updated based on the data
The influence relational matrix being combined between family and the targeted customer;
The interest model of the targeted customer is based respectively on the first behavior feedback data and the influence relational matrix is carried out
Renewal.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710775357.0A CN107562875A (en) | 2017-08-31 | 2017-08-31 | A kind of update method of model, apparatus and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710775357.0A CN107562875A (en) | 2017-08-31 | 2017-08-31 | A kind of update method of model, apparatus and system |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107562875A true CN107562875A (en) | 2018-01-09 |
Family
ID=60977724
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710775357.0A Pending CN107562875A (en) | 2017-08-31 | 2017-08-31 | A kind of update method of model, apparatus and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107562875A (en) |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108595493A (en) * | 2018-03-15 | 2018-09-28 | 腾讯科技(深圳)有限公司 | Method for pushing and device, storage medium, the electronic device of media content |
CN108804619A (en) * | 2018-05-31 | 2018-11-13 | 腾讯科技(深圳)有限公司 | Interest preference prediction technique, device, computer equipment and storage medium |
CN108874998A (en) * | 2018-06-14 | 2018-11-23 | 华东师范大学 | A kind of dialog mode music recommended method indicated based on composite character vector |
CN109002490A (en) * | 2018-06-26 | 2018-12-14 | 腾讯科技(深圳)有限公司 | User's portrait generation method, device, server and storage medium |
CN109102127A (en) * | 2018-08-31 | 2018-12-28 | 杭州贝购科技有限公司 | Method of Commodity Recommendation and device |
CN109591825A (en) * | 2018-11-29 | 2019-04-09 | 北京新能源汽车股份有限公司 | A kind of driving fatigue detection method, device and vehicle |
CN110457576A (en) * | 2019-07-08 | 2019-11-15 | 深圳壹账通智能科技有限公司 | Account-classification method, device, computer equipment and storage medium |
CN110851734A (en) * | 2019-11-01 | 2020-02-28 | 广州视源电子科技股份有限公司 | Content recommendation method and device |
CN111159541A (en) * | 2019-12-11 | 2020-05-15 | 微民保险代理有限公司 | Method and device for determining account behavior preference |
CN111357014A (en) * | 2018-09-19 | 2020-06-30 | 华为技术有限公司 | AI model development method and device |
CN111385659A (en) * | 2018-12-29 | 2020-07-07 | 广州市百果园信息技术有限公司 | Video recommendation method, device, equipment and storage medium |
CN111753184A (en) * | 2019-03-29 | 2020-10-09 | 北京达佳互联信息技术有限公司 | Information recommendation method and device |
CN111859188A (en) * | 2020-07-03 | 2020-10-30 | 美味不用等(上海)信息科技股份有限公司 | Restaurant recommendation method and system based on hierarchical analysis of large-scale matrix |
CN111914159A (en) * | 2019-05-10 | 2020-11-10 | 招商证券股份有限公司 | Information recommendation method and terminal |
CN112132608A (en) * | 2020-09-15 | 2020-12-25 | 北京慧辰资道资讯股份有限公司 | Charging service intelligent recommendation method and device based on big data of power internet of things |
CN112424763A (en) * | 2019-04-30 | 2021-02-26 | 北京字节跳动网络技术有限公司 | Object recommendation method and device, storage medium and terminal equipment |
CN113573106A (en) * | 2020-04-28 | 2021-10-29 | 北京达佳互联信息技术有限公司 | Model updating method and device for multimedia information and server |
CN113688295A (en) * | 2021-10-26 | 2021-11-23 | 北京达佳互联信息技术有限公司 | Data determination method and device, electronic equipment and storage medium |
CN114140802A (en) * | 2022-01-29 | 2022-03-04 | 北京易真学思教育科技有限公司 | Text recognition method and device, electronic equipment and storage medium |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102135989A (en) * | 2011-03-09 | 2011-07-27 | 北京航空航天大学 | Normalized matrix-factorization-based incremental collaborative filtering recommending method |
CN103678647A (en) * | 2013-12-20 | 2014-03-26 | Tcl集团股份有限公司 | Method and system for recommending information |
CN104572797A (en) * | 2014-05-12 | 2015-04-29 | 深圳市智搜信息技术有限公司 | Individual service recommendation system and method based on topic model |
-
2017
- 2017-08-31 CN CN201710775357.0A patent/CN107562875A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102135989A (en) * | 2011-03-09 | 2011-07-27 | 北京航空航天大学 | Normalized matrix-factorization-based incremental collaborative filtering recommending method |
CN103678647A (en) * | 2013-12-20 | 2014-03-26 | Tcl集团股份有限公司 | Method and system for recommending information |
CN104572797A (en) * | 2014-05-12 | 2015-04-29 | 深圳市智搜信息技术有限公司 | Individual service recommendation system and method based on topic model |
Cited By (29)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108595493A (en) * | 2018-03-15 | 2018-09-28 | 腾讯科技(深圳)有限公司 | Method for pushing and device, storage medium, the electronic device of media content |
CN108804619A (en) * | 2018-05-31 | 2018-11-13 | 腾讯科技(深圳)有限公司 | Interest preference prediction technique, device, computer equipment and storage medium |
CN108804619B (en) * | 2018-05-31 | 2020-11-03 | 腾讯科技(深圳)有限公司 | Interest preference prediction method, device, computer equipment and storage medium |
CN108874998A (en) * | 2018-06-14 | 2018-11-23 | 华东师范大学 | A kind of dialog mode music recommended method indicated based on composite character vector |
CN108874998B (en) * | 2018-06-14 | 2021-10-19 | 华东师范大学 | Conversational music recommendation method based on mixed feature vector representation |
CN109002490A (en) * | 2018-06-26 | 2018-12-14 | 腾讯科技(深圳)有限公司 | User's portrait generation method, device, server and storage medium |
CN109102127A (en) * | 2018-08-31 | 2018-12-28 | 杭州贝购科技有限公司 | Method of Commodity Recommendation and device |
CN111357014A (en) * | 2018-09-19 | 2020-06-30 | 华为技术有限公司 | AI model development method and device |
CN109591825A (en) * | 2018-11-29 | 2019-04-09 | 北京新能源汽车股份有限公司 | A kind of driving fatigue detection method, device and vehicle |
CN111385659A (en) * | 2018-12-29 | 2020-07-07 | 广州市百果园信息技术有限公司 | Video recommendation method, device, equipment and storage medium |
CN111753184B (en) * | 2019-03-29 | 2024-01-02 | 北京达佳互联信息技术有限公司 | Information recommendation method and device |
CN111753184A (en) * | 2019-03-29 | 2020-10-09 | 北京达佳互联信息技术有限公司 | Information recommendation method and device |
CN112424763B (en) * | 2019-04-30 | 2023-09-12 | 抖音视界有限公司 | Object recommendation method and device, storage medium and terminal equipment |
CN112424763A (en) * | 2019-04-30 | 2021-02-26 | 北京字节跳动网络技术有限公司 | Object recommendation method and device, storage medium and terminal equipment |
CN111914159B (en) * | 2019-05-10 | 2024-03-12 | 招商证券股份有限公司 | Information recommendation method and terminal |
CN111914159A (en) * | 2019-05-10 | 2020-11-10 | 招商证券股份有限公司 | Information recommendation method and terminal |
CN110457576A (en) * | 2019-07-08 | 2019-11-15 | 深圳壹账通智能科技有限公司 | Account-classification method, device, computer equipment and storage medium |
CN110851734A (en) * | 2019-11-01 | 2020-02-28 | 广州视源电子科技股份有限公司 | Content recommendation method and device |
CN110851734B (en) * | 2019-11-01 | 2023-09-05 | 广州视源电子科技股份有限公司 | Content recommendation method and device |
CN111159541B (en) * | 2019-12-11 | 2023-08-25 | 微民保险代理有限公司 | Method and device for determining account behavior preference |
CN111159541A (en) * | 2019-12-11 | 2020-05-15 | 微民保险代理有限公司 | Method and device for determining account behavior preference |
CN113573106A (en) * | 2020-04-28 | 2021-10-29 | 北京达佳互联信息技术有限公司 | Model updating method and device for multimedia information and server |
CN113573106B (en) * | 2020-04-28 | 2023-03-21 | 北京达佳互联信息技术有限公司 | Model updating method and device for multimedia information and server |
CN111859188A (en) * | 2020-07-03 | 2020-10-30 | 美味不用等(上海)信息科技股份有限公司 | Restaurant recommendation method and system based on hierarchical analysis of large-scale matrix |
CN111859188B (en) * | 2020-07-03 | 2024-03-12 | 美之味(上海)信息技术有限公司 | Restaurant recommendation method and system based on hierarchical analysis of large-scale matrix |
CN112132608A (en) * | 2020-09-15 | 2020-12-25 | 北京慧辰资道资讯股份有限公司 | Charging service intelligent recommendation method and device based on big data of power internet of things |
CN112132608B (en) * | 2020-09-15 | 2024-03-19 | 北京慧辰资道资讯股份有限公司 | Charging service intelligent recommendation method and device based on big data of electric power Internet of things |
CN113688295A (en) * | 2021-10-26 | 2021-11-23 | 北京达佳互联信息技术有限公司 | Data determination method and device, electronic equipment and storage medium |
CN114140802A (en) * | 2022-01-29 | 2022-03-04 | 北京易真学思教育科技有限公司 | Text recognition method and device, electronic equipment and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107562875A (en) | A kind of update method of model, apparatus and system | |
WO2021203819A1 (en) | Content recommendation method and apparatus, electronic device, and storage medium | |
US9934515B1 (en) | Content recommendation system using a neural network language model | |
CN108431833A (en) | End-to-end depth collaborative filtering | |
CN110781391A (en) | Information recommendation method, device, equipment and storage medium | |
CN111444428A (en) | Information recommendation method and device based on artificial intelligence, electronic equipment and storage medium | |
CN110162698B (en) | User portrait data processing method, device and storage medium | |
CN110266745B (en) | Information flow recommendation method, device, equipment and storage medium based on deep network | |
CN111259263B (en) | Article recommendation method and device, computer equipment and storage medium | |
CN106845644A (en) | A kind of heterogeneous network of the contact for learning user and Mobile solution by correlation | |
CN109087178A (en) | Method of Commodity Recommendation and device | |
CN111400603A (en) | Information pushing method, device and equipment and computer readable storage medium | |
CN111143684B (en) | Artificial intelligence-based generalized model training method and device | |
CN112559896B (en) | Information recommendation method, device, equipment and computer readable storage medium | |
CN111723292B (en) | Recommendation method, system, electronic equipment and storage medium based on graph neural network | |
Chen et al. | Next POI recommendation based on location interest mining with recurrent neural networks | |
CN109656541A (en) | Exploitative recommended method, device, storage medium and electronic equipment | |
Wang et al. | Deep user modeling for content-based event recommendation in event-based social networks | |
Ge et al. | Estimating local information trustworthiness via multi-source joint matrix factorization | |
CN113344648B (en) | Advertisement recommendation method and system based on machine learning | |
CN111767459A (en) | Item recommendation method and device | |
CN109190040A (en) | Personalized recommendation method and device based on coevolution | |
CN111652673B (en) | Intelligent recommendation method, device, server and storage medium | |
CN114358807A (en) | User portrayal method and system based on predictable user characteristic attributes | |
CN114817692A (en) | Method, device and equipment for determining recommended object and computer storage medium |
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 | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20180109 |