CN108965938A - Potential paying customer prediction technique and system in smart television - Google Patents
Potential paying customer prediction technique and system in smart television Download PDFInfo
- Publication number
- CN108965938A CN108965938A CN201810879462.3A CN201810879462A CN108965938A CN 108965938 A CN108965938 A CN 108965938A CN 201810879462 A CN201810879462 A CN 201810879462A CN 108965938 A CN108965938 A CN 108965938A
- Authority
- CN
- China
- Prior art keywords
- program
- user
- feature
- watched
- paying customer
- 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.)
- Granted
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
- H04N21/20—Servers specifically adapted for the distribution of content, e.g. VOD servers; Operations thereof
- H04N21/25—Management operations performed by the server for facilitating the content distribution or administrating data related to end-users or client devices, e.g. end-user or client device authentication, learning user preferences for recommending movies
- H04N21/251—Learning process for intelligent management, e.g. learning user preferences for recommending movies
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
- H04N21/20—Servers specifically adapted for the distribution of content, e.g. VOD servers; Operations thereof
- H04N21/25—Management operations performed by the server for facilitating the content distribution or administrating data related to end-users or client devices, e.g. end-user or client device authentication, learning user preferences for recommending movies
- H04N21/258—Client or end-user data management, e.g. managing client capabilities, user preferences or demographics, processing of multiple end-users preferences to derive collaborative data
- H04N21/25866—Management of end-user data
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
- H04N21/40—Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
- H04N21/45—Management operations performed by the client for facilitating the reception of or the interaction with the content or administrating data related to the end-user or to the client device itself, e.g. learning user preferences for recommending movies, resolving scheduling conflicts
- H04N21/4508—Management of client data or end-user data
- H04N21/4532—Management of client data or end-user data involving end-user characteristics, e.g. viewer profile, preferences
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
- H04N21/40—Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
- H04N21/45—Management operations performed by the client for facilitating the reception of or the interaction with the content or administrating data related to the end-user or to the client device itself, e.g. learning user preferences for recommending movies, resolving scheduling conflicts
- H04N21/466—Learning process for intelligent management, e.g. learning user preferences for recommending movies
- H04N21/4667—Processing of monitored end-user data, e.g. trend analysis based on the log file of viewer selections
Landscapes
- Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Multimedia (AREA)
- Signal Processing (AREA)
- Computing Systems (AREA)
- Computer Graphics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses potential paying customer prediction technique and systems in smart television, comprising: building is positively correlated data set and negatively correlated data set；Feature is positively correlated from being positively correlated to extract in data set；Negatively correlated feature is extracted from negatively correlated data set；Data characteristics is derivative；The Embeded Wide&Deep model built in advance is trained using fused positive correlation new feature and fused negatively correlated new feature, obtains trained Embeded Wide&Deep model；The new feature of user to be predicted is input in trained Embeded Wide&Deep model, exporting user to be predicted is potential paying customer or paying customer non-potential.It is effective to excavate potential paying customer, the solution currently in intelligent television field is proposed, promotes firms profitability in order to which enterprise formulates corresponding business strategy.
Description
Technical field
The present invention relates to ntelligent television technolog field, more particularly to potential paying customer prediction technique in smart television and
System.
Background technique
With daily household electrical appliances such as internet+and the artificial intelligence technology applications in household electrical appliances, including TV, refrigerator and air-conditioning
Design, manufacture kimonos does honest work and develops to intelligent and networked services directions.There is predictive display, arrives the year two thousand twenty, intelligence
TV market occupation rate is up to 90% or more, and global artificial intelligence system will be brought for household appliances enterprise more than 47,000,000,000 dollars
Income, shows powerful market potential.Here smart television may be defined as to connect radio and television broadband programme content and
It can charge and video program given by free video display set meal and network video provider provided by video on demand producer.
It due to the improvement of people's living standards, is statisticallyd analyze according to the user of Hisense of Largest In China TV producer TV, is ready that viewing is paid
The user for taking program is more and more.A challenging project is brought therewith, i.e., how to predict that a user can be from mesh
Preceding non-payment user becomes paying customer.Obviously, the solution of the above problem will help enterprise to carry out user behavior analysis, business wide
It accuses and the accurate dispensing of favor information, the help enterprises such as TV producer and video website brings more interests.Also it helps simultaneously
TV user purchase value-added service and obtain better user experience etc..
Fewer and fewer for the research of potential paying customer forecasting problem both at home and abroad at present, current recommender system is for inhaling
It quotes family to be known as paying the scheme of member, if red-letter day is invigorated dynamic, the mode extensively casted net provides discount coupon etc., and this mode is for root
Originally it is unwilling for the user of payment to be a kind of " bothering ", while inefficiency.
Due to the complexity of user behavior in smart television and the deficient unicity of data information, with traditional electronic commerce
Middle user behavior modeling has essential distinction, for the associated solutions in traditional electronic commerce field, is not suitable for smart television
Field.
Summary of the invention
In order to solve the deficiencies in the prior art, the present invention provides potential paying customer prediction technique in smart television and it is
System；
As the first aspect of the present invention, potential paying customer prediction technique in smart television is provided；
Potential paying customer prediction technique in smart television, comprising:
Data set building: building is positively correlated data set and negatively correlated data set；
Data characteristics is extracted: being positively correlated feature from being positively correlated to extract in data set；Negative is extracted from negatively correlated data set
Close feature；
Data characteristics is derivative: deriving to feature is positively correlated, generates and be positively correlated new feature；Feature and positive will be positively correlated
It closes new feature and carries out Fusion Features, obtain fused positive correlation feature；Negatively correlated feature is derived, is generated negatively correlated new
Feature；Negatively correlated feature and negatively correlated new feature are subjected to Fusion Features, obtain fused negatively correlated feature；
Model training step: using fused positive correlation new feature and fused negatively correlated new feature to building in advance
Embeded Wide&Deep model be trained, obtain trained Embeded Wide&Deep model；
Potential paying customer prediction: user itself behavioural characteristic of user to be predicted, the spy of the watched program of user are extracted
It seeks peace by the feature of viewing program；To the quilt of user itself behavioural characteristic of the user to be predicted of extraction, the watched program of user
It moves feature and carries out feature by the active features of viewing program and derive, form the new feature of user to be predicted；By user to be predicted
New feature be input in trained Embeded Wide&Deep model, export user to be predicted be potential paying customer also
It is non-potential paying customer.
Further, the positive correlation data set, comprising: the sponsored program of paying customer's history viewing, paying customer go through
History viewing non-payment program, paying customer's history pay records, paying customer's history viewing sponsored program recommended information and
The recommended information of paying customer's history viewing non-payment program；Paying customer's history pay records, comprising: paying customer goes through
The sponsored program and corresponding paid-for time of history viewing.
Further, the negatively correlated data set, comprising: the program and non-paid user of non-paid user history viewing are gone through
The recommended information of history viewing program.
Further, the positive correlation feature, comprising: program viewing period, the history pay records frequency, charges paid section
Purpose spending amount, the type of program, the performer of program, the director of program, program country, section object language, program it is upper
Reflect time, the scoring of program and the corresponding set meal ID of program.
Further, the negatively correlated feature, comprising: program viewing period, the type of program, the performer of program, section
Purpose director, the country of program, section object language, the show time of program, the scoring of program and the corresponding set meal ID of program.
Further, the positive correlation new feature, comprising: continuous to be positively correlated feature and continuous-discrete positive correlation feature；
Derivative continuous positive correlation feature: charges paid program is watched from user in set period of time is gone out derived from positive correlation feature
Number of days, the average number of programs watched daily of user and the average charges paid program of viewing daily of user duration characteristics；
Derivative continuous-discrete positive correlation feature: paid from going out to watch certain seed type daily in derived from positive correlation feature
Take the quantity of program and watches within one day the duration characteristics of type charges paid program in certain per hour.
Further, the negatively correlated new feature, comprising: continuous negative correlation feature and continuous-discrete negatively correlated feature；
Derivative continuous negatively correlated feature: non-paid program is watched from user in set period of time is gone out derived from negatively correlated feature
Number of days, the average non-paid number of programs watched daily of user and the average duration characteristics for watching non-paid program daily of user；
Derivative continuous-discrete negatively correlated feature: unpaid from going out to watch certain seed type daily in derived from negatively correlated feature
Take the quantity of program and watches within one day the duration characteristics of the non-paid program of type in certain per hour.
Further, itself behavioural characteristic of the user of user to be predicted, comprising: the time of user's viewing TV；
Further, the feature of the watched program of user to be predicted, comprising: program is watched within the scope of user's setting time
Number of days, the average number of programs watched daily of user, the average duration watched daily of user；
Further, by the feature of viewing program, comprising: the program viewing period, the type of program, program performer,
The director of program, the country of program, section object language, the show time of program, the scoring of program and the corresponding set meal ID of program；
Further, to the passive spy of user itself behavioural characteristic of the user to be predicted of extraction, the watched program of user
It seeks peace derivative by the active features progress feature of viewing program, forms the new feature of user to be predicted are as follows:
User watches program number of days, the number of programs watched daily that is averaged, the average time watched daily, is averaged one week seven days
Viewing number of days accounting, the one week seven days average number of programs watched daily, the one week seven days average time watched daily, every kind daily
Time that type programs are averagely watched, the average number watched of each type program, each type program is averaged one week seven days
The average number watched of the time of viewing, one week seven days each type program, each type program are averaged by viewing time, often
A performer uses language by watched time, each program belonging country watched time, every kind of program by watched time, each director
Watched time, each program show date watched time, different scoring program by watched time or watch program and belong to set meal
Number.
Further, the Embeded Wide&Deep model, comprising: linear model and depth model two parts；
The linear model is that Logic Regression Models learn user's low order feature representation, expression are as follows:
xembed=wembed*x
Y=wTxembed+b
Wherein, xembedTo be embedded in vector, x is input layer user vector, wembedFor embeding layer map vector, w is model power
Weight vector, b are bigoted item, and y is predicted value.
The depth model part is multiple perceptron model, learns user's Higher Order Abstract feature, expression are as follows:
Y=f[l](w[l]*(f[l-1]*(...f[1](w[1]*x+b[1])...))+b[l])
Wherein, l is network number of plies number, and f is the activation primitive of current layer, and x is user characteristics vector, and w is Model Weight
Vector, b are bigoted.
Joint training linear model and depth model splice the input of depth model the last layer and linear model one
It rises, exports predicted value after training.
Joint training will optimize all parameters of linear model and depth model simultaneously, by by linear model and depth mould
The mode that type is weighted summation in training carries out；
Embeded Wide&Deep model tormulation are as follows:
Wherein, y is the label of two classification, and σ () is sigmoid activation primitive, and x is user characteristics vector, and b is bigoted
?.wwideIt is the weight of linear model, wdeepIt is the weight of depth model.
Note in the training process of joint training linear model and depth model: loss loss function is intersection entropy function；
When predicted value is 0 or negative, predicted value is substituted for a minimum, to prevent from Nan value occur in the calculating of Loss loss function
Trap；
Depth model is designed as the design of tower layer, and halve principle: wherein bottom is widest, and upper layer articulamentum successively subtracts
Few, the width of tower layer successively halves from bottom to top；Use Xavier weights initialisation solution.
As a second aspect of the invention, potential paying customer forecasting system in smart television is provided；
Potential paying customer forecasting system in smart television, comprising: memory, processor and storage are on a memory simultaneously
The computer instruction run on a processor when the computer instruction is run by processor, is completed described in any of the above-described method
The step of.
Compared with prior art, the beneficial effects of the present invention are:
(1) feature of paying customer and non-paid user are sufficiently excavated, and is potential by feature derivative and Fusion Features
The accurate prediction of paying customer provides foundation；
(2) potential paying customer is effectively excavated, the solution currently in intelligent television field is proposed, in order to look forward to
Industry formulates corresponding business strategy and promotes firms profitability.
Detailed description of the invention
The accompanying drawings constituting a part of this application is used to provide further understanding of the present application, and the application's shows
Meaning property embodiment and its explanation are not constituted an undue limitation on the present application for explaining the application.
Fig. 1 is flow chart of the invention；
Fig. 2 is the division that user of the present invention watches log information paying customer (positive example) log；
Fig. 3 is the division that user of the present invention watches log information non-payment user (negative example) log；
Fig. 4 is Embeded Wide&Deep model support composition of the present invention.
Specific embodiment
It is noted that following detailed description is all illustrative, it is intended to provide further instruction to the application.Unless another
It indicates, all technical and scientific terms used herein has usual with the application person of an ordinary skill in the technical field
The identical meanings of understanding.
It should be noted that term used herein above is merely to describe specific embodiment, and be not intended to restricted root
According to the illustrative embodiments of the application.As used herein, unless the context clearly indicates otherwise, otherwise singular
Also it is intended to include plural form, additionally, it should be understood that, when in the present specification using term "comprising" and/or " packet
Include " when, indicate existing characteristics, step, operation, device, component and/or their combination.
As one embodiment of the invention, potential paying customer prediction technique in smart television is provided；
As shown in Figure 1, potential paying customer prediction technique in smart television, comprising:
Data set construction step: building is positively correlated data set and negatively correlated data set；
The positive correlation data set, comprising: the sponsored program of paying customer's history viewing, paying customer's history are watched non-
Sponsored program, paying customer's history pay records, the recommended information of paying customer's history viewing sponsored program and paying customer go through
The recommended information of history viewing non-payment program；Paying customer's history pay records, comprising: the viewing of paying customer's history is paid
Take program and corresponding paid-for time；
The negative correlation data set, comprising: the program and non-paid user history viewing section of non-paid user history viewing
Purpose recommended information；
Data characteristics extraction step:
Feature is positively correlated from being positively correlated to extract in data set；The positive correlation feature, comprising: the program viewing period is gone through
The history pay records frequency, the spending amount of charges paid program, the type of program, the performer of program, the director of program, section country of destination
Family, section object language, the show time of program, the scoring of program and the corresponding set meal ID of program；
Negatively correlated feature is extracted from negatively correlated data set；The negative correlation feature, comprising: program viewing period, section
Purpose type, the performer of program, the director of program, the country of program, section object language, the show time of program, program are commented
Divide set meal ID corresponding with program；
Data characteristics derives step:
Derive to feature is positively correlated, generates and be positively correlated new feature；Feature will be positively correlated and be positively correlated new feature and carried out
Fusion Features obtain fused positive correlation feature；
The positive correlation new feature, comprising: continuous to be positively correlated feature and continuous-discrete positive correlation feature；
Derivative continuous positive correlation feature: charges paid program is watched from user in set period of time is gone out derived from positive correlation feature
Number of days, the average number of programs watched daily of user and the average charges paid program of viewing daily of user duration characteristics；
Derivative continuous-discrete positive correlation feature: paid from going out to watch certain seed type daily in derived from positive correlation feature
Take the quantity of program and watches within one day the duration characteristics of type charges paid program in certain per hour；
Negatively correlated feature is derived, negatively correlated new feature is generated；Negatively correlated feature and negatively correlated new feature are carried out
Fusion Features obtain fused negatively correlated feature；
The negative correlation new feature, comprising: continuous negative correlation feature and continuous-discrete negatively correlated feature；
Derivative continuous negatively correlated feature: non-paid program is watched from user in set period of time is gone out derived from negatively correlated feature
Number of days, the average non-paid number of programs watched daily of user and the average duration characteristics for watching non-paid program daily of user；
Derivative continuous-discrete negatively correlated feature: unpaid from going out to watch certain seed type daily in derived from negatively correlated feature
Take the quantity of program and watches within one day the duration characteristics of the non-paid program of type in certain per hour；
Model training step: using fused positive correlation new feature and fused negatively correlated new feature to building in advance
Embeded Wide&Deep model be trained, obtain trained Embeded Wide&Deep model；
User charges prediction steps: user itself behavioural characteristic of user to be predicted, the quilt of the watched program of user are extracted
Move feature and by the active features of viewing program；User itself behavioural characteristic of the user to be predicted of extraction, user are watched
The passive feature of program and to carry out feature by the active features of viewing program derivative, forms the new feature of user to be predicted；It will be to
The new feature of prediction user is input in trained Embeded Wide&Deep model, exports setting for user to be predicted
Payment probability in period.
The Embeded Wide&Deep model, comprising: linear model and depth model two parts；
The linear model is that Logic Regression Models learn user's low order feature representation, expression are as follows:
xembed=wembed*x
Y=wTxembed+b
Wherein, xembedTo be embedded in vector, x is input layer user vector, wembedFor embeding layer map vector, w is model power
Weight vector, b are bigoted item, and y is predicted value.
The depth model part is multiple perceptron model, learns user's Higher Order Abstract feature, expression are as follows:
Y=f[l](w[l]*(f[l-1]*(...f[1](w[1]*x+b[1])...))+b[l])
Wherein, l is network number of plies number, and f is the activation primitive of current layer, and x is user characteristics vector, and w is Model Weight
Vector, b are bigoted.
Joint training linear model and depth model splice the input of depth model the last layer and linear model one
It rises, exports predicted value after training.
Joint training will optimize all parameters of linear model and depth model simultaneously, by by linear model and depth mould
The mode that type is weighted summation in training carries out；
Model is as shown in figure 4, Embeded Wide&Deep model tormulation are as follows:
Wherein, y is the label of two classification, and σ () is sigmoid activation primitive, and x is user characteristics vector, and b is bigoted
?.wwiDe is the weight of linear model, wdeepIt is the weight of depth model.
Note in the training process of joint training linear model and depth model: loss loss function is intersection entropy function；
When predicted value is 0 or negative, predicted value is substituted for a minimum, to prevent from Nan value occur in the calculating of Loss loss function
Trap；
Depth model is designed as the design of tower layer, and halve principle: wherein bottom is widest, and upper layer articulamentum successively subtracts
Few, the width of tower layer successively halves from bottom to top；Use Xavier weights initialisation solution.
As second embodiment of the invention, potential paying customer forecasting system in smart television is provided；
Potential paying customer forecasting system in smart television, comprising: memory, processor and storage are on a memory simultaneously
The computer instruction run on a processor when the computer instruction is run by processor, is completed described in any of the above-described method
The step of.
The application proposes characteristic processing frame derived from Information expansion and feature, and the Embeded Wide& being applicable in
Deep model.
Step 1: Information expansion includes expanding user according to by original user viewing log and user's purchaser record
The information of itself behavior watches extended log files by original user and goes out the passive information of the watched program of user, and by original
Log and auxiliary information are expanded by viewing programme information initiative information；
Step 2: feature derivative, which refers to, to be combined by existing data to find new meaning, and individual features are increased；
Step 3:Embeded Wide&Deep model refers to linear according to being designed by corresponding parameter with frame
The model of model and depth model.
The Embeded Wide&Deep model refers to according to the multilayer sense being designed by corresponding parameter and frame
Know machine model；
User is watched log information to divide according to payment and non-payment user, paying customer is specifically watched into log letter
Log information before breath payment day extracts, and non-payment user's whole user journal information extraction is come out, for doing
The potential paying customer task stated；
Further, raw information described in the step 1 includes: user watches log information and auxiliary information,
Middle auxiliary information includes film information and user's purchaser record.
It is wherein described that user is watched into log information according to payment and non-payment user division such as Fig. 2, shown in Fig. 3.Wherein
Fig. 2 indicates that user has purchased set meal member at some time point, and the user journal before this user purchase is extracted as number
According to the object of excavation, and label is 1, is expressed as positive example.Wherein Fig. 3 indicates that user did not bought set meal member eventually, by this its institute
There is user's viewing log to extract the object as data mining, and label is 0, indicates the example that is negative.
Itself behavioural information of user described in the step 1 refers to the field watched in log information according to original user
Or user's purchaser record Information expansion information relevant to user itself behavior in auxiliary information, addition is in phase after being quantified
In the journal entries answered.Itself behavioural information of the user refers to: being extended according to time field information in original log information
Which period occurred in one day for the current behavior of user out, occurred in the temporal informations such as week is several, because of user's watching behavior
With temporal regularity, this time cycle be can be one day, is also possible to one week, and addition is in journal entries after being quantified
In；
It is described extended log files watched by original user go out the passive information of the watched program of user refer to from the log
The statistics passive information of the watched program of extending user is carried out in information, addition is in journal entries after being quantified.
The watched programme information of user is expanded with auxiliary information by original log to refer to: day is watched according to original user
The watched program of user is associated with auxiliary information (programme information) in will, expands the watched program initiative information of user,
Addition is in journal entries after being quantified.
User, which is expanded, in user's purchaser record information in auxiliary information has bought information, such as: user has bought
Set meal member's frequency, frequency etc. is bought in spending amount, single video payment.
It is described the watched programme information of user is expanded by original log to refer to: section is first counted from all logs
Mesh calculates program popularity by the viewing frequency, and addition is in journal entries after being quantified.
It expands the watched programme information of user with programme information in auxiliary information by original log to refer to: according to original
The watched program ID of user is associated with auxiliary information (programme information) in log, expands program category, label, classification,
Performer, director, country, language are shown the date, video scoring, set meal ID.Because above-mentioned every kind of information has multiple attributes, Mei Gejie
Mesh may also correspond to multiple attributes in every kind of information, first quantify every kind of information attribute, to each attribute in every kind of information
The continuous unique encodings since 0 are carried out, are stored in dictionary, so as to later period quick obtaining attribute coding.Similar to Pareto
Rule (Pareto ' s principle) is also a bar Lai Tedinglv, and in mass data learning process, effective influence factor is certain
That relative quantity is biggish, if amount seldom if think its for can not Consideration, therefore, for certain auxiliary informations (director, drill
Member) the too many situation of number of attributes, the highest top n attribute of frequency of occurrence is selected according to testing sieve, remaining ranges one
A attribute, is then quantified.
Further, feature derivative is by extracting to data in the step 2, and combination finds new data meaning
Justice increases individual features, and specific includes derivative continuous feature, the derivative continuous feature of discrete features &.
The wherein continuous feature of the derivative: extracting the continuity Characteristics of user, passes through traversal and counts each user journal
Data entries obtain.It is specific: to extract user and watch number of days, the average number of programs watched daily of user, user are average daily
The continuous feature such as the duration of viewing.Wherein the user watches number of days: within the time span of model data processing, user
Open the number of days of television set viewing program.The average number of programs watched daily of user described in wherein: cumulative user each out sees
It sees number of programs sum, watches total number of days divided by user and obtain.The average duration watched daily of user described in wherein: cumulative every out
A user watches program total duration, watches total number of days divided by user and obtains.
The continuous feature of derivative discrete features described in wherein: the temporal information similar expanded is combined and is expanded
Other information combine to obtain the continuous feature of discrete &.First by the discrete dimension to the equivalent time of feature, for feature per one-dimensional
Degree, traversal counts each user journal data entries in obtained data, extracts the continuity Characteristics of respective dimensions.
The continuous feature of derivative discrete features described in wherein: because user's watching behavior has temporal regularity, this
A time cycle can be one day, be also possible to one week, and the similar temporal information expanded is combined other expanded
Information consolidation obtains the continuous feature of discrete &.Such as: 7 days one week with the group of viewing program category (n attribute, i.e. n dimension) quantity
It closes, obtaining 7*n dimensional vector indicates that user user watches the feature of the quantity of certain type programs for one week daily.24 hours one day
With the combination of viewing program category (n attribute, i.e. n are tieed up) duration, obtaining 24*n dimensional vector is indicated user user 24 hours one day
The feature of the duration of certain type programs is watched per hour.
User's information of drawing a portrait conventionally (discretization or serialization) quantified, above-mentioned is derived above
Continuous feature, after the discrete continuous feature of & merges with the quantization of user's Figure Characteristics, final we obtain the user characteristics vector of n dimension,
And the corresponding label of assignment, all user characteristics are merged into matrix and form complete data collection.After data set is upset sequence, press
Training set and test set are generated according to ratio；
User's portrait information is conventionally quantified, the above above-mentioned continuous feature derived, discrete & connects
After continuous feature merges with user's three kinds of information of portrait, final we obtain the user characteristics vector of 2161 dimensions, and assignment is corresponding
Label forms complete data collection.After data set is upset sequence, training set and test set are generated according to the ratio of 8:2；
The foregoing is merely preferred embodiment of the present application, are not intended to limit this application, for the skill of this field
For art personnel, various changes and changes are possible in this application.Within the spirit and principles of this application, made any to repair
Change, equivalent replacement, improvement etc., should be included within the scope of protection of this application.
Claims (10)
1. potential paying customer prediction technique in smart television, characterized in that include:
Data set building: building is positively correlated data set and negatively correlated data set；
Data characteristics is extracted: being positively correlated feature from being positively correlated to extract in data set；It is extracted from negatively correlated data set negatively correlated special
Sign；
Data characteristics is derivative: deriving to feature is positively correlated, generates and be positively correlated new feature；Feature will be positively correlated and be positively correlated new
Feature carries out Fusion Features, obtains fused positive correlation feature；Negatively correlated feature is derived, is generated negatively correlated new special
Sign；Negatively correlated feature and negatively correlated new feature are subjected to Fusion Features, obtain fused negatively correlated feature；
Model training step: using fused positive correlation new feature and fused negatively correlated new feature to building in advance
Embeded Wide&Deep model is trained, and obtains trained Embeded Wide&Deep model；
Potential paying customer prediction: extract user itself behavioural characteristic of user to be predicted, the feature of the watched program of user and
By the feature of viewing program；To the passive spy of user itself behavioural characteristic of the user to be predicted of extraction, the watched program of user
It seeks peace derivative by the active features progress feature of viewing program, forms the new feature of user to be predicted；By the new of user to be predicted
Feature is input in trained Embeded Wide&Deep model, and exporting user to be predicted is potential paying customer also right and wrong
Potential paying customer.
2. potential paying customer prediction technique in smart television as described in claim 1, characterized in that the positive correlation data
Collection, comprising: the sponsored program of paying customer's history viewing, the non-payment program of paying customer's history viewing, paying customer's history
The introduction of pay records, the recommended information of paying customer's history viewing sponsored program and paying customer's history viewing non-payment program
Information；Paying customer's history pay records, comprising: when the sponsored program of paying customer's history viewing and corresponding payment
Between；
The negative correlation data set, comprising: the program and non-paid user history of non-paid user history viewing watch program
Recommended information.
3. potential paying customer prediction technique in smart television as described in claim 1, characterized in that
The positive correlation feature, comprising: the program viewing period, the history pay records frequency, charges paid program spending amount,
The type of program, the performer of program, the director of program, the country of program, section object language, the show time of program, program
Score set meal ID corresponding with program；
The negative correlation feature, comprising: program viewing period, the type of program, the performer of program, the director of program, program
Country, section object language, the show time of program, the scoring of program and the corresponding set meal ID of program.
4. potential paying customer prediction technique in smart television as described in claim 1, characterized in that described to be positively correlated new spy
Sign, comprising: continuous to be positively correlated feature and continuous-discrete positive correlation feature；
Derive continuous positive correlation feature: from the day for going out user's viewing charges paid program in set period of time derived from positive correlation feature
The duration characteristics of number, the average number of programs watched daily of user and the average charges paid program of viewing daily of user；
Derivative continuous-discrete positive correlation feature: certain seed type charges paid section is watched daily from being positively correlated to go out one week derived from feature
Purpose quantity and the duration characteristics for watching type charges paid program in certain for one day per hour.
5. potential paying customer prediction technique in smart television as described in claim 1, characterized in that described negatively correlated new special
Sign, comprising: continuous negative correlation feature and continuous-discrete negatively correlated feature；
Derivative continuous negatively correlated feature: from the day for going out user in set period of time derived from negatively correlated feature and watching non-paid program
Number, the average non-paid number of programs watched daily of user and the average duration characteristics for watching non-paid program daily of user；
Derivative continuous-discrete negatively correlated feature: the non-paid section of certain seed type is watched daily from going out one week derived from negatively correlated feature
Purpose quantity and the duration characteristics for watching the non-paid program of type in certain for one day per hour.
6. potential paying customer prediction technique in smart television as described in claim 1, characterized in that the use of user to be predicted
Family itself behavioural characteristic, comprising: the time of user's viewing TV；
The feature of the watched program of user to be predicted, comprising: the number of days of program is watched within the scope of user's setting time, user is average
Number of programs, the average duration watched daily of user watched daily；
By the feature of viewing program, comprising: program viewing period, the type of program, the performer of program, the director of program, section
Purpose country, section object language, the show time of program, the scoring of program and the corresponding set meal ID of program.
7. potential paying customer prediction technique in smart television as described in claim 1, characterized in that
To user itself behavioural characteristic of the user to be predicted of extraction, the passive feature of the watched program of user and by viewing program
Active features to carry out feature derivative, form the new feature of user to be predicted are as follows:
User watches program number of days, the average number of programs watched daily, the average time watched daily, average daily one week seven days
Viewing number of days accounting, the one week seven days average number of programs watched daily, the one week seven days average time watched daily, each type
Time that program is averagely watched, the average number watched of each type program, each type program is averagely watched one week seven days
Time, the average number watched of each type program, each type program are average by viewing time, Mei Geyan one week seven days
Member is watched by watched time, each program belonging country watched time, every kind of program using language by watched time, each director
Number, each program show date watched time, different scoring program by watched time or watch time that program belongs to set meal
Number.
8. potential paying customer prediction technique in smart television as described in claim 1, characterized in that
The Embeded Wide&Deep model, comprising: linear model and depth model two parts；
The linear model is that Logic Regression Models learn user's low order feature representation, expression are as follows:
xembed=wembed*x
Y=wTxembed+b
Wherein, xembedTo be embedded in vector, x is input layer user vector, wembedFor embeding layer map vector, w be Model Weight to
Amount, b are bigoted item, and y is predicted value；
The depth model part is multiple perceptron model, learns user's Higher Order Abstract feature, expression are as follows:
Y=f[l](w[l]*(f[l-1]*(...f[1](w[1]*x+b[1])...))+b[l])
Wherein, l is network number of plies number, and f is the activation primitive of current layer, and x is user characteristics vector, and w is Model Weight vector,
B is bigoted；
The input of depth model the last layer and linear model is stitched together by joint training linear model and depth model,
Predicted value is exported after training；
Joint training will optimize all parameters of linear model and depth model simultaneously, by the way that linear model and depth model exist
The mode that summation is weighted when training carries out.
9. potential paying customer prediction technique in smart television as described in claim 1, characterized in that
Embeded Wide&Deep model tormulation are as follows:
Wherein, y is the label of two classification, and σ () is sigmoid activation primitive, and x is user characteristics vector, and b is bigoted item；wwide
It is the weight of linear model, wdeepIt is the weight of depth model.
10. potential paying customer forecasting system in smart television, characterized in that include: that memory, processor and being stored in is deposited
The computer instruction run on reservoir and on a processor when the computer instruction is run by processor, completes aforesaid right
It is required that step described in 1-9 either method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810879462.3A CN108965938B (en) | 2018-08-03 | 2018-08-03 | Method and system for predicting potential pay users in smart television |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810879462.3A CN108965938B (en) | 2018-08-03 | 2018-08-03 | Method and system for predicting potential pay users in smart television |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108965938A true CN108965938A (en) | 2018-12-07 |
| CN108965938B CN108965938B (en) | 2020-03-20 |
Family
ID=64467577
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810879462.3A Expired - Fee Related CN108965938B (en) | 2018-08-03 | 2018-08-03 | Method and system for predicting potential pay users in smart television |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108965938B (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110166803A (en) * | 2019-06-21 | 2019-08-23 | 奚华文 | A kind of interactive approach of the network media |
| CN110599234A (en) * | 2019-08-13 | 2019-12-20 | 广东工业大学 | Product sales prediction method |
| CN111191860A (en) * | 2020-04-14 | 2020-05-22 | 北京热云科技有限公司 | Prediction method based on ensemble learning, prediction system and readable storage medium |
| CN111554118A (en) * | 2020-04-24 | 2020-08-18 | 深圳职业技术学院 | Dynamic prediction method and system for bus arrival time |
| CN111709764A (en) * | 2019-03-18 | 2020-09-25 | 腾讯科技（深圳）有限公司 | Method and device for determining related parameters of multimedia content and storage medium |
| CN116934389A (en) * | 2023-06-12 | 2023-10-24 | 嘉兴华数广电网络有限公司 | Digital television value ticket card management system based on cloud computing and cross-platform technology |
| CN117909601A (en) * | 2024-03-15 | 2024-04-19 | 厦门她趣信息技术有限公司 | Payment social matching method, device, equipment and readable storage medium |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105160008A (en) * | 2015-09-21 | 2015-12-16 | 合一网络技术(北京)有限公司 | Method and device for locating suggested users |
| CN105354725A (en) * | 2015-11-20 | 2016-02-24 | 珠海多玩信息技术有限公司 | Prediction method and system of promotion effect of application |
| CN105719164A (en) * | 2016-01-21 | 2016-06-29 | 海信集团有限公司 | Paid multimedia resource recommending method and paid multimedia resource recommending device |
| CN106168976A (en) * | 2016-07-14 | 2016-11-30 | 武汉斗鱼网络科技有限公司 | A kind of specific user's method for digging based on NB Algorithm and system |
| CN106204063A (en) * | 2016-06-30 | 2016-12-07 | 北京奇艺世纪科技有限公司 | A kind of paying customer's method for digging and device |
| CN108205766A (en) * | 2016-12-19 | 2018-06-26 | 阿里巴巴集团控股有限公司 | Information-pushing method, apparatus and system |
-
2018
- 2018-08-03 CN CN201810879462.3A patent/CN108965938B/en not_active Expired - Fee Related
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105160008A (en) * | 2015-09-21 | 2015-12-16 | 合一网络技术(北京)有限公司 | Method and device for locating suggested users |
| CN105354725A (en) * | 2015-11-20 | 2016-02-24 | 珠海多玩信息技术有限公司 | Prediction method and system of promotion effect of application |
| CN105719164A (en) * | 2016-01-21 | 2016-06-29 | 海信集团有限公司 | Paid multimedia resource recommending method and paid multimedia resource recommending device |
| CN106204063A (en) * | 2016-06-30 | 2016-12-07 | 北京奇艺世纪科技有限公司 | A kind of paying customer's method for digging and device |
| CN106168976A (en) * | 2016-07-14 | 2016-11-30 | 武汉斗鱼网络科技有限公司 | A kind of specific user's method for digging based on NB Algorithm and system |
| CN108205766A (en) * | 2016-12-19 | 2018-06-26 | 阿里巴巴集团控股有限公司 | Information-pushing method, apparatus and system |
Non-Patent Citations (1)
| Title |
|---|
| H.-T. CHENG等: "Wide & deep learning for recommender systems", 《IN PROC. WORKSHOP DEEP LEARN RECOMMENDER SYST.》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111709764A (en) * | 2019-03-18 | 2020-09-25 | 腾讯科技（深圳）有限公司 | Method and device for determining related parameters of multimedia content and storage medium |
| CN111709764B (en) * | 2019-03-18 | 2023-04-18 | 腾讯科技（深圳）有限公司 | Method and device for determining related parameters of multimedia content and storage medium |
| CN110166803A (en) * | 2019-06-21 | 2019-08-23 | 奚华文 | A kind of interactive approach of the network media |
| CN110599234A (en) * | 2019-08-13 | 2019-12-20 | 广东工业大学 | Product sales prediction method |
| CN111191860A (en) * | 2020-04-14 | 2020-05-22 | 北京热云科技有限公司 | Prediction method based on ensemble learning, prediction system and readable storage medium |
| CN111191860B (en) * | 2020-04-14 | 2020-08-21 | 北京热云科技有限公司 | Prediction method based on ensemble learning, prediction system and readable storage medium |
| CN111554118A (en) * | 2020-04-24 | 2020-08-18 | 深圳职业技术学院 | Dynamic prediction method and system for bus arrival time |
| CN116934389A (en) * | 2023-06-12 | 2023-10-24 | 嘉兴华数广电网络有限公司 | Digital television value ticket card management system based on cloud computing and cross-platform technology |
| CN116934389B (en) * | 2023-06-12 | 2024-02-23 | 嘉兴华数广电网络有限公司 | Digital television value ticket card management system based on cloud computing and cross-platform technology |
| CN117909601A (en) * | 2024-03-15 | 2024-04-19 | 厦门她趣信息技术有限公司 | Payment social matching method, device, equipment and readable storage medium |
| CN117909601B (en) * | 2024-03-15 | 2024-05-17 | 厦门她趣信息技术有限公司 | Payment social matching method, device, equipment and readable storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108965938B (en) | 2020-03-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108965938A (en) | Potential paying customer prediction technique and system in smart television | |
| CN108205766A (en) | Information-pushing method, apparatus and system | |
| JP5583696B2 (en) | Conversion confidence rating | |
| CN106055617A (en) | Data pushing method and device | |
| CN105930425A (en) | Personalized video recommendation method and apparatus | |
| CN109447364A (en) | Power customer based on label complains prediction technique | |
| CN105975472A (en) | Method and device for recommendation | |
| CN106326413A (en) | Personalized video recommending system and method | |
| CN104205158A (en) | Systems, methods, and articles of manufacture to measure online audiences | |
| CN104933622A (en) | Microblog popularity degree prediction method based on user and microblog theme and microblog popularity degree prediction system based on user and microblog theme | |
| CN107657487A (en) | A kind of advertisement placement method and device based on verification | |
| CN110428087A (en) | Business stability prediction technique, device, computer equipment and storage medium | |
| CN109359868A (en) | A kind of construction method and system of power grid user portrait | |
| CN110880127B (en) | Consumption level prediction method and device, electronic equipment and storage medium | |
| CN103647671A (en) | Gur Game based crowd sensing network management method and system | |
| CN102592236A (en) | Internet advertising crowd analysis system and analysis method | |
| CN112598438A (en) | Outdoor advertisement recommendation system and method based on large-scale user portrait | |
| CN112287238B (en) | User characteristic determination method and device, storage medium and electronic equipment | |
| CN103136280A (en) | System and method of automatically searching hot point video | |
| CN103795697A (en) | Network medium information release effect simulation method and system | |
| CN109325845A (en) | A kind of financial product intelligent recommendation method and system | |
| CN108572988A (en) | A kind of house property assessment data creation method and device | |
| CN108885628A (en) | Data analysing method candidate's determination device | |
| CN110110372A (en) | A kind of user's timing behavior automatic segmentation prediction technique | |
| CN108537635A (en) | A kind of recommendation method and device of product |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20200320 Termination date: 20210803 |