CN108521586A - The IPTV TV program personalizations for taking into account time context and implicit feedback recommend method - Google Patents
The IPTV TV program personalizations for taking into account time context and implicit feedback recommend method Download PDFInfo
- Publication number
- CN108521586A CN108521586A CN201810230242.8A CN201810230242A CN108521586A CN 108521586 A CN108521586 A CN 108521586A CN 201810230242 A CN201810230242 A CN 201810230242A CN 108521586 A CN108521586 A CN 108521586A
- Authority
- CN
- China
- Prior art keywords
- program
- tensor
- user
- iptv
- personalizations
- 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
- H04N21/251—Learning process for intelligent management, e.g. learning user preferences for recommending movies
- H04N21/252—Processing of multiple end-users' preferences to derive collaborative 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]
- H04N21/20—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
- H04N21/25891—Management of end-user data being end-user 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
- H04N21/4668—Learning process for intelligent management, e.g. learning user preferences for recommending movies for recommending content, e.g. movies
Landscapes
- Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Multimedia (AREA)
- Signal Processing (AREA)
- Computing Systems (AREA)
- Computer Graphics (AREA)
- Two-Way Televisions, Distribution Of Moving Picture Or The Like (AREA)
Abstract
The invention discloses a kind of IPTV TV program personalizations for taking into account time context and implicit feedback to recommend method, this method watches the preference pattern of program by building user, determine user in different time sections to the preference of program category, then by the way of structure tensor and tensor resolution, the potential association for excavating user, program category and period, to carry out accurately and effectively personalized recommendation.Compared with existing proposed algorithm, the method for the present invention, which can provide, more preferably recommends precision and higher recall rate.
Description
Technical field
The present invention relates to IPTV service fields, and in particular to a kind of IPTV TVs for taking into account time context and implicit feedback
Program personalized recommendation method.
Background technology
In recent years, the personalized service cry of quickly propelling with " triple play ”, diversification is gradually high.Telecommunication network,
Cable network and computer network interpenetrate, and are gradually fused into unified communication network, are protected in intelligent transportation, environment
It is paid close attention in the fields such as shield, government work, public safety.As one of most important application in " triple play ”, IPTV is
Possess a large number of users.By IPTV service, user can watch any interested TV programme at any time.To improve user couple
The satisfaction of IPTV, IPTV operators are providing various personalized viewing services, such as live telecast (Live TV), video point
It broadcasts (VOD), Catch-up TV etc..These services have provided various viewing selection to the user:Live telecast is as common broadcast
TV is the same, and user can watch the video of live broadcast;VOD is the supplement to live television programs, and user can see at any time
See any video prerecorded, typically film, TV series, variety show or cartoon etc.;And introduce recently
Catch-up TV allow user to watch any video that other day (such as 3 days) is broadcast live.
IPTV considerably increases the flexibility that user selects TV programme.But thousands of films, TV play or other
When being selected in program, the problem of user will often face information overload.If spending the long time that could find favorite TV Festival
The viewing experience of mesh, user will will be greatly reduced.Therefore, it is necessary to personalized recommendation systems to help user to be quickly found out them
The TV programme seen are liked, to promote satisfaction of the user to IPTV.
Currently, in numerous such as video stream medias (such as Netflix, YouTube), music Streaming Media (such as Spotify,
Pandora), the business such as sale service of goods (such as Amazon, Booking) and social networks (such as Facebook, Linkedin)
Field, commending system all occupy a tiny space.But the accurate recommendation of IPTV faces multiple challenges.First, it is protected for privacy
Shield, the user data that IPTV service provider usually only possesses are limited.Secondly, the TV programme that most users watch are very
It is few, cause the data being collected into very sparse.Third can not be derived not to the explicit evaluation of TV programme (such as score)
The viewing preference of user.4th, IPTV viewing data are influenced by family composition, and user journal is often mixed and disorderly and imperfect
's.
Some television program recommendations algorithms are proposed in the prior art, but these algorithms are suitable for special scenes, wherein
Such as the recommendation of Popular Items algorithms, User Popular Items algorithm, collaborative filtering and Cempetency-based education
Algorithm carries out program recommendation using the two-dimentional relation of user, TV programme.But program viewing is family's behavior, is seen
It is an important factor for recommending to be considered as to see the time also.Although proposed algorithm based on context-aware combines under varying environment
User preference, but the potential applications correlation between dimension is not considered.User is not being obtained to new program based on the proposed algorithm of SVD
Evaluation data before, it can not be recommended.
Invention content
For existing algorithm there are the problem of, the object of the present invention is to provide one kind taking into account time context and implicit feedback
IPTV TV program personalizations recommend method, with realize accurately and effectively personalized TV programs recommend.
In order to realize that above-mentioned task, the present invention use following technical scheme:
The IPTV TV program personalizations for taking into account time context and implicit feedback recommend method, include the following steps:
Step 1, structure user watches the preference pattern of program
The preference pattern includes loyalty and interest-degree of the user in certain time period to program category, wherein
The loyalty is the actual play duration of duration and the type program that user watches certain type programs in certain time period
The ratio between, the interest-degree is that user watches the duration of certain type programs in certain time period and user watches all program categories
The ratio between total duration;
Preference value of the user in different time sections to program category is determined according to preference pattern;
Step 2, structure tensor is to indicate user, program category, the relationship between the period, the wherein each single item in tensor
User is represented in a certain period of time to the preference value of certain program type;
Step 3, the tensor decomposed, reconstructed, the approximate tensor of the tensor is obtained, according to the tensor sum
Approximate tensor builds loss function, and carries out regularization to the loss function, obtains object function;Solve the target letter
Number, optimal approximation tensor is determined according to solving result;
Step 4, using the element in the optimal approximation tensor as prediction result, recommended to user according to prediction result
Program category.
Further, the determination process of the preference value includes:
It assigns different weights respectively for the loyalty, interest-degree, and the loyalty after weights, interest-degree will be assigned
The sum of as the preference value.
Further, it is that the weights that the loyalty assigns are more than the weights that the interest-degree described in being assigns.
Further, the tensor A described in step 2 is expressed as:
Wherein, i indicates that user, j indicate that program category, d indicate that period, M indicate that user's number, N indicate program category
Number, L indicate the number of period.
Further, the preparation method of the approximate tensor described in step 3 includes:
The tensor A is decomposed into the mould product of core tensor S and each factor matrix (U, G, T), to core tensor S and respectively
Factor matrix (U, G, T) is initialized, and U ∈ R are obtainedM×du, G ∈ RN×dg, T ∈ RL×dt, S ∈ Rdu×dg×dt;Wherein, du, dg and
Dt is positive integer value, and meets du<<M,dg<<N,dt<<L, R indicate set of real numbers;
According to the core tensor S and each factor matrix (U, G, T) after initialization, tensor A is reconstructed using following formula, with
Obtain approximate tensor
In above formula, S ×UU indicates that the U- moulds of tensor S and matrix U accumulate, ×GG、×TT indicates G- moulds product, T- moulds product respectively.
Further, the expression formula of the loss function described in step 3 is:
In above formula, D is mark value, each single item a in Ai,j,dPreference values of the user i in period d to program j is represented, such as
This of fruit ai,j,dD is then set to 1, D is otherwise set to 0 by non-zero.
Further, the expression formula of the object function described in step 3 is:
In above formula,Indicate that Frobenius norms, variable λ and σ are different adjusting parameters.
Further, the object function described in the solution described in step 3, including:
The matrix U that one group of items element is real number is randomly selected from U, G, T, the S obtained after initializationi,*、Gj,*、
Td,*、S*Come approximate representation U, G, T, S, local derviation is taken by each component to object function respectively, obtains following equation:
In above formula,Indicate Kronecker products;It is right after calculating partial derivative result by above equation
Ui,*、Gj,*、Td,*、S*It is updated, recycles above equation to recalculate partial derivative after update as a result, until object function
Convergence, U when object function being made to restraini,*、Gj,*、Td,*、S*As best U, G, T, S, to obtain optimal approximation tensor.
A kind of IPTV TV program personalizations recommendation taking into account time context and implicit feedback, including sequentially connected structure
Preference pattern module is built, tensor module is built, determines optimal approximation tensor module and recommending module, these modules are respectively used to
Realize that step 1 is to the process described in step 4 in claim 1.
The present invention has following technical characterstic compared with prior art:
1. the present invention analyzes the viewing preference of different times user, time dimension is increased to track the dynamic of kinsfolk
State watches behavior, constructs the preference pattern of user's viewing so that it is more accurate that user's viewing preference is mined.
2. the present invention excavates the potential association between user, TV programme and viewing time three by tensor resolution, make
Accurately and effectively personalized recommendation can be carried out by obtaining this method.
Description of the drawings
Fig. 1 is flow chart of the method for the present invention;
Fig. 2 is the flow chart of tensor resolution algorithm;
Fig. 3 is that the precision of six recommendation methods compares figure;
Fig. 4 is that the recall rate of six recommendation methods compares figure;
Fig. 5 is six and method F-score performances is recommended to compare figure.
Six in Fig. 3 to Fig. 5 recommendation methods, recommendation method respectively provided by the invention recommend method at random
Random, welcome program commending method Popular, based on the welcome program commending method UserPopular of user, base
In the proposed algorithm SVD of singular value decomposition and proposed algorithm Context-aware based on context-aware.
Specific implementation mode
The present invention provides a kind of IPTV TV program personalizations for taking into account time context and implicit feedback to recommend method,
Relationship between user-viewing period-program category is indicated using three-dimensional tensor, using Higher-order Singular value decomposition algorithm, is dug
The semantic association between user, viewing period and program category is dug, to be embodied as the program that user recommends each period
List.The detailed step of the present invention is as follows:
Step 1, structure user watches the preference pattern of program
The preference pattern includes loyalty and interest-degree of the user in certain time period to program category, wherein
The loyalty is the actual play duration of duration and the type program that user watches certain type programs in certain time period
The ratio between, the interest-degree is that user watches the duration of certain type programs in certain time period and user watches all program categories
The ratio between total duration;Specifically:
Step 1.1, user's viewing time section is divided
The rating preference of data statistics, user is as time cycle property changes, it is contemplated that the influence of duty factor is used
Family is different with the viewing behavior of festivals or holidays same period on weekdays.In order to preferably analyze viewing data at any time
Variation was divided into 7 periods by 24 hours one day, and considers two kinds of situations of working day and festivals or holidays respectively, when obtaining 14
Between section (the sum of 7 periods of workaday 7 periods and festivals or holidays);A kind of specific dividing mode of period is as follows
Shown in table 1:
Table 1:User's viewing time section divides
Time started | End time | Time segment description |
06:01 | 08:00 | Early morning |
08:01 | 12:00 | The morning |
12:01 | 13:00 | Noon |
13:01 | 18:00 | Afternoon |
18:01 | 19:30 | At dusk |
19:31 | 24:00 | Night |
00:01 | 06:00 | Morning |
Step 1.2, structure user watches the preference pattern of program
The preference pattern is:
In above formula, m, n indicate the number of program category and service type respectively.
Loyaltyi,j,dExpression user i is to the loyalty of program category j in period d, i.e., user i is in period d
The ratio between the actual play duration of the duration and the type program of the type program is watched,Indicate user i in the time
The weighting total duration of section d viewing program categories j, tj,k,dIt is that user i watches program category j in period d by service type k
Time, Tj, k be by the actual play duration of service type k program categories j,It is by all service type programs
The broadcasting total duration of type j,.Wk is the weight of service type k.
Interesti,j,dIn period d to the interest-degree of program category j, i.e. user i is seen expression user i in period d
See that the total duration of the type program watches the ratio between total duration of all program categories with it in period d,
It is the weighting total duration that user i watches all program categories in period d by all service types.
For example, in IPTV, service type k has Catch-up TV, VOD, tri- kinds of Live TV.According to data statistics, three kinds
The weights of service type meet:Catch-up TV>VOD>Live TV, thus the weight of three kinds of service types can with value be 2,
1.5 and 1.
Step 1.3, it is that the loyalty, interest-degree assign different weights respectively, and the loyalty after weights will be assigned
As the preference value, the formula of preference value is for the sum of degree, interest-degree:
ai,j,d=Loyaltyi,j,dW1+Interesti,j,dW2
In above formula, W1,W2It is the weights assigned for loyalty, interest-degree respectively.In preference pattern, it is believed that user is to certain
The loyalty of one program can more represent preference of the user to the program category.Therefore, it is that the weights that the loyalty assigns are big
In the weights for the interest-degree imparting, i.e. W1>W2;For example, W1,W2Value can be 0.8 and 0.2.
For example, calculating user i in period d (12:01-13:00) to the viewing preference value of program category j.Assuming that service
The time that Type C atch-up TV play program category j is 10:00-14:00, user i is in the period 11:30-13:30 pass through
Service type Catch-up TV viewing TV programme (weight w shared by service type Catch-up TVkFor 2), wherein 12:30-
13:30 in viewing program category j, then the user is in the period 12:01-13:The loyalty of 00 couple of program category jInterest-degreeSo user i is in period d (12:
01-13:00) it is a to the viewing preference value of program category ji,j,d=0.125 × 0.8+0.25 × 0.2=0.15.
Step 2, structure tensor is to indicate user, program category, the relationship between the period, the wherein each single item in tensor
User is represented in a certain period of time to the preference value of certain program type;
The tensor A is expressed as:
Wherein, i indicates that user, j indicate that program category, d indicate that period, M indicate that user's number, N indicate program category
Number, L indicate the number of period, in tensor A, each single item ai,j,dUser i is represented in period d to program category j
Preference value.
Step 3, the tensor decomposed, reconstructed, the approximate tensor of the tensor is obtained, according to the tensor sum
Approximate tensor builds loss function, and carries out regularization to the loss function, obtains object function;Solve the target letter
Number, optimal approximation tensor is determined according to solving result;
Step 3.1, the tensor A is decomposed:
The mould that tensor A is decomposed into core tensor S and each factor matrix (U, G, T) using HOSVD accumulates, come excavate user,
Potential contact between program category, period.
Step 3.2, each factor matrix of core tensor sum (U, G, T) is initialized, obtains U ∈ RM×du, G ∈ RN×dg, T
∈RL×dt, S ∈ Rdu×dg×dt;Wherein, R indicates that set of real numbers, M indicate that user's number, N indicate that program category number, L indicate the time
The number of section, du, dg and dt are very small positive integer value, and meet du<<M,dg<<N,dt<<L.Usually, du<
0.0005 × M, dg<0.2 × N, dt<0.5×L.
Step 3.3, according to the core tensor S and each factor matrix (U, G, T) after initialization, using following formula to tensor A into
Row reconstruct, to obtain approximate tensor
In above formula, S ×UU indicates that the U- moulds of tensor S and matrix U accumulate, ×GG、×TT indicates (tensor and matrix U) respectively
G- moulds product, (tensor is with matrix T's) T- moulds accumulate;N (U, G, T) mould of tensor and matrix product, which calculates, belongs to the prior art, herein not
It repeats.
Then approximate tensorIn each single item elementIt is expressed as:
In above formula, sdu,dg,dtIndicate the element positioned at position (du, dg, dt) of core tensor S, ui,duRepresenting matrix U
In the i-th row, du row element, gj,dgIndicate the element arranged positioned at matrix G jth row, dg, td,dtIt indicates to be located at matrix T d
The element of row, dt row.
Step 3.4, best approximate tensor in order to obtain, defines loss function, and quantify that above-mentioned tensor resolution obtains forces
Close values, to obtain optimal model parameter, i.e., each factor matrix (U, G, T) and core tensor S:
In above formula, A is the tensor that step 2 is built,For approximate tensor, whether D is mark value, lost to label information
It loses.Each single item a in Ai,j,dPreference values of the user i in period d to program j is represented, if thisi,j,dNon-zero then sets D
It is 1, D is otherwise set to 0.
Step 3.5, the program category watched in different time sections due to user is fewer, obtains data than sparse, institute
With by the data set being made of sparse tensor come training pattern to obtain optimal parameter (each factor matrix (U, G, T) and core
Heart tensor S) normally result in data overfitting.Model over-fitting in order to prevent, improve model generalization ability, using based on
The item of Frobenius norms to carry out regularization to formula, i.e., adds regular terms in the loss function to model, obtain target letter
Number C (U, G, T, S):
In above formula,Indicate that Frobenius norms, variable λ and σ are different adjusting parameters, by largely testing point
The universal experience value of analysis, adjusting parameter is { 0.1,0.2,0.01,0.001 }, can be according to the multiple adjustment of data so that target
Function is optimal.
Step 3.6, the object function is solved
In order to minimize loss function, usage factor matrix U, the stochastic gradient of G, T and core tensor S optimizes.
Stochastic gradient i.e. when data volume is too big or can not once obtain total data, the partial data that is randomly selected by one group come
It obtains " gradient ".Therefore, the U ∈ R obtained after step 3.2 initializationM×du, G ∈ RN×dg, T ∈ RL×dt, S ∈ Rdu×dg×dtIn with
Machine chooses the matrix U that one group of items element is real numberi,*、Gj,*、Td,*、S*Come approximate representation U, G, T, S, by target letter
Several each components take local derviation respectively, obtain following equation:
In above formula group, variable λ and σ are different adjusting parameters, empirical value be generally 0.1,0.2,0.01,
0.001}。Indicate Kronecker products (Kronecker product), above subscript in formula before matrix parameter indicate mould product.
Based on obtained partial derivative as a result, by the way of iteration, in each iteration, group calculates according to formula above
The partial derivative of each component of object function after calculating, updates each component by following formula group:
In above formula group, μ indicate iterations, η indicate learning rate, plus mark ' updated point of representation in components
Amount does not add the component before the update of mark ' representation in components, such as Ui′,*Indicate updated Ui,*。
Continuous iteration update, until object function convergence meets Cμ-Cμ+1>ε, wherein Cμ、Cμ+1Indicate respectively the μ times,
Object function is calculated after+1 iteration of μ, ε indicates that arbitrarily small positive integer, the smaller explanation of value and original tensor more connect
Closely, value range is generally [0.01,0.0001].U when object function is restrainedi,*、Gj,*、Td,*、S*As best U, G,
T, (i.e. the value of these parameters is U to Si,*、Gj,*、Td,*、S*Value), U, G, T, S are brought into formula 1, in the hope of best close
Like tensor
Step 4, using the element in the optimal approximation tensor as prediction result, recommended to user according to prediction result
Program category.
According to optimal approximation tensorMiddle user i recommends the preference value of various types of programs in period d.Specifically,
If we want to recommend K program category to user i in certain time period d, can by user i in period d to difference
The preference value of program category sorts by size, the K program category that right rear line recommends preference value larger.
On the basis of above-mentioned recommendation method, the present invention further provides a kind of time context and implicit feedbacks taken into account
IPTV TV program personalizations recommend, including sequentially connected structure preference pattern module, structure tensor module, determine most preferably it is close
Like tensor module and recommending module, these modules are respectively used to realize that step 1 is to the process described in step 4 in claim 1.
The performance evaluation of the method for the present invention
The data set being collected into is subjected to offline evaluation, shadow of more each proposed algorithm to various aspects of performance to recommendation method
It rings.
Step 1, data set statistics
In this experiment, the data set provided using China Unicom.The data set includes 1 day to 2017 October in 2016
The TV programme daily record and timetable daily record that 48.1 general-purpose family of Henan Province is watched during January 1.In our experiment, it only examines
The log recording of continuous viewing program 10 minutes or more is considered, the key data of filtered data collection is as shown in table 2, a total of
478,170 users and 517,277,222 viewing records.Table 3 shows the partial data of program viewing daily record.Table 4
Show that the partial data of Live TV program log daily records, the format of data are identical as program viewing daily record.
Data set is divided into one group of training set and one group of test set, wherein training set includes 1 day to 2016 October in 2016
The daily record on December 14, in (75 days, 421,821,170 entrants), test set include 15 days to 2017 years 1 December in 2016
The moon (18 days, 95,456,052 entries) on the 1st.The former recommends performance for establishing model, the latter for assessing.
2 data set of table counts
Users | 481,232 |
Time span | 93days |
views | 536,649,128 |
programs | 7,154 |
A part for the viewing daily record of 3 TV programme of table
A part for 4 Live TV program log daily records of table
Step 2, program category divide
62 channels are shared in the data set that we count, are classified as following 16 classifications:Finance, international, sport,
Military and agricultural, opera, news, children's garment, shopping, wushu, documentary film, science and education, social and law, music, variety section
Mesh, TV series and film.In order to make three kinds of services that can preferably classify to program, we are by TV play, film
These three classifications are further divided with variety show.Wherein TV series have 17 types:Family, crime,
Suspense, in ancient times, comedy, love, war, science fiction, swordsman, action, city, history, idol is military, police, criminal investigation and
Spy.Film is divided into 14 classes:Crime, terrible, action, comedy, love, science fiction, venture, suspense, drama film is terrified, and literature is gone through
History, war and cartoon.Variety show is divided into 9 classes:Reality TV show, music variety show are interviewed, life, party, tourism, talk show,
Xiucai's show and wisdom select-elite.So in conjunction with ready-portioned 13 classifications (finance, international, sport is military with agricultural, opera,
News, children's garment, shopping, wushu, documentary film, science and education, society and law, music, finance is international, sport, military and agriculture
Industry, opera, news, children's garment, shopping, wushu, documentary film, science and education, society and law, music), we obtain 53 altogether
Television program type, and program category can be obtained by finding programm name in tracking list.
Step 3 analyzes experimental result
Fig. 3 shows the precision comparison result of six commending systems.Precision be user watch the quantity of type of recommendation with it is total
The ratio of type of recommendation.From Fig. 3 intuitively from the point of view of, recommendation method precision that we use totally recommend better than other five be
System.Since the programme variety that user watches in reality is seldom, precision can be continuously increased when incipient, later with recommendation class
The increase of the quantity of type and decline.
Fig. 4 shows recall rate of six commending systems to different n values.Recall rate refers to being pushed away in the program that user watches
Recommend the ratio shared by program.Recall rate is higher to indicate that the user preference of commending system prediction is more accurate.In Fig. 4, we can be with
See that recall rate increases with the quantity of type of recommendation, and since n=7, the recall rate of our commending system is always
In optimum state.
Fig. 5 shows the F-score performance comparison results of six commending systems.F scores are the comprehensive of accuracy and recall rate
Close assessment.From fig. 5, it can be seen that F-score performances are also being continuously increased with the quantity of type of recommendation.Further, it can be seen that
Our recommendation effect at least improves 10% than other five commending systems.
Claims (9)
1. the IPTV TV program personalizations for taking into account time context and implicit feedback recommend method, which is characterized in that including with
Lower step:
Step 1, structure user watches the preference pattern of program
The preference pattern includes loyalty and interest-degree of the user in certain time period to program category, wherein described
Loyalty be user watch the ratio between the duration of certain type programs and the actual play duration of the type program in certain time period,
The interest-degree is that user watches the duration of certain type programs in certain time period and user watches the total of all program categories
The ratio between duration;
Preference value of the user in different time sections to program category is determined according to preference pattern;
Step 2, to indicate user, program category, the relationship between the period, each single item wherein in tensor represents structure tensor
User is in a certain period of time to the preference value of certain program type;
Step 3, the tensor decomposed, reconstructed, obtain the approximate tensor of the tensor, according to tensor sum approximation
Tensor builds loss function, and carries out regularization to the loss function, obtains object function;The object function is solved,
Optimal approximation tensor is determined according to solving result;
Step 4, using the element in the optimal approximation tensor as prediction result, program is recommended to user according to prediction result
Type.
2. the IPTV TV program personalizations recommendation method as described in claim 1 for taking into account time context and implicit feedback,
It is characterized in that, the determination process of the preference value includes:
It assigns different weights respectively for the loyalty, interest-degree, and the sum of the loyalty after weights, interest-degree will be assigned
As the preference value.
3. the IPTV TV program personalizations recommendation method as claimed in claim 2 for taking into account time context and implicit feedback,
It is characterized in that, being more than the weights that the interest-degree described in being assigns for the weights that the loyalty assigns.
4. the IPTV TV program personalizations recommendation method as described in claim 1 for taking into account time context and implicit feedback,
It is characterized in that, the tensor A described in step 2 is expressed as:
Wherein, i indicates that user, j indicate that program category, d indicate that period, M indicate that user's number, N indicate program category number,
L indicates the number of period.
5. the IPTV TV program personalizations recommendation method as described in claim 1 for taking into account time context and implicit feedback,
It is characterized in that, the preparation method of the approximate tensor described in step 3 includes:
The tensor A is decomposed into the mould product of core tensor S and each factor matrix (U, G, T), to core tensor S and each factor
Matrix (U, G, T) is initialized, and U ∈ R are obtainedM×du, G ∈ RN×dg, T ∈ RL×dt, S ∈ Rdu×dg×dt;Wherein, du, dg and dt are equal
For positive integer value, and meet du<<M,dg<<N,dt<<L, R indicate set of real numbers;
According to the core tensor S and each factor matrix (U, G, T) after initialization, tensor A is reconstructed using following formula, to obtain
Approximate tensor
In above formula, S ×UU indicates that the U- moulds of tensor S and matrix U accumulate, ×GG、×TT indicates G- moulds product, T- moulds product respectively.
6. the IPTV TV program personalizations recommendation method as described in claim 1 for taking into account time context and implicit feedback,
It is characterized in that, the expression formula of the loss function described in step 3 is:
In above formula, D is mark value, each single item a in Ai,j,dPreference values of the user i in period d to program j is represented, if should
Item ai,j,dD is then set to 1, D is otherwise set to 0 by non-zero.
7. the IPTV TV program personalizations recommendation method as described in claim 1 for taking into account time context and implicit feedback,
It is characterized in that, the expression formula of the object function described in step 3 is:
In above formula,Indicate that Frobenius norms, variable λ and σ are different adjusting parameters.
8. the IPTV TV program personalizations recommendation method as described in claim 1 for taking into account time context and implicit feedback,
It is characterized in that, the object function described in solution described in step 3, including:
The matrix U that one group of items element is real number is randomly selected from U, G, T, the S obtained after initializationi,*、Gj,*、Td,*、S*
Come approximate representation U, G, T, S, local derviation is taken by each component to object function respectively, obtains following equation:
In above formula,Indicate Kronecker products;After calculating partial derivative result by above equation, to Ui,*、Gj,*、
Td,*、S*It is updated, recycles above equation to recalculate partial derivative after update as a result, until object function convergence, makes mesh
U when scalar functions are restrainedi,*、Gj,*、Td,*、S*As best U, G, T, S, to obtain optimal approximation tensor.
9. a kind of IPTV TV program personalizations for taking into account time context and implicit feedback are recommended, which is characterized in that including according to
The structure preference pattern module of secondary connection builds tensor module, determines optimal approximation tensor module and recommending module, these moulds
Block is respectively used to realize that step 1 is to the process described in step 4 in claim 1.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810230242.8A CN108521586B (en) | 2018-03-20 | 2018-03-20 | IPTV television program personalized recommendation method giving consideration to time context and implicit feedback |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810230242.8A CN108521586B (en) | 2018-03-20 | 2018-03-20 | IPTV television program personalized recommendation method giving consideration to time context and implicit feedback |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108521586A true CN108521586A (en) | 2018-09-11 |
CN108521586B CN108521586B (en) | 2020-01-14 |
Family
ID=63433846
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810230242.8A Active CN108521586B (en) | 2018-03-20 | 2018-03-20 | IPTV television program personalized recommendation method giving consideration to time context and implicit feedback |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108521586B (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109921799A (en) * | 2019-02-20 | 2019-06-21 | 重庆邮电大学 | A kind of tensor compression method based on cumulative amount dictionary learning |
CN110149556A (en) * | 2019-05-20 | 2019-08-20 | 河北工业大学 | A kind of IPTV user's behavior pattern mining method based on TDC-LDA model |
CN110196946A (en) * | 2019-05-29 | 2019-09-03 | 华南理工大学 | A kind of personalized recommendation method based on deep learning |
CN111753182A (en) * | 2019-03-28 | 2020-10-09 | 北京达佳互联信息技术有限公司 | Multimedia information recommendation method and device, electronic equipment and readable storage medium |
WO2020233245A1 (en) * | 2019-05-20 | 2020-11-26 | 山东科技大学 | Method for bias tensor factorization with context feature auto-encoding based on regression tree |
CN113051463A (en) * | 2019-12-26 | 2021-06-29 | 中移物联网有限公司 | Project pushing method and system |
CN113254794A (en) * | 2021-07-15 | 2021-08-13 | 中国传媒大学 | Program data recommendation method and system based on modeling |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103260061A (en) * | 2013-05-24 | 2013-08-21 | 华东师范大学 | Context-perceptive IPTV program recommending method |
CN103297853A (en) * | 2013-06-07 | 2013-09-11 | 华东师范大学 | IPTV (internet protocol television) program recommendation method based on context recognition for multiple users |
CN104935968A (en) * | 2015-06-25 | 2015-09-23 | 广州视源电子科技股份有限公司 | Method and system for recommending program by interactive personality TV (IPTV) |
CN104994408A (en) * | 2015-06-25 | 2015-10-21 | 青岛海信电器股份有限公司 | Intelligent television program recommendation method, device and intelligent television |
CN107071578A (en) * | 2017-05-24 | 2017-08-18 | 中国科学技术大学 | IPTV program commending methods |
US20170332150A1 (en) * | 2016-05-13 | 2017-11-16 | Hulu, LLC | Personalized Content Ranking Using Content Received from Different Sources in a Video Delivery System |
-
2018
- 2018-03-20 CN CN201810230242.8A patent/CN108521586B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103260061A (en) * | 2013-05-24 | 2013-08-21 | 华东师范大学 | Context-perceptive IPTV program recommending method |
CN103297853A (en) * | 2013-06-07 | 2013-09-11 | 华东师范大学 | IPTV (internet protocol television) program recommendation method based on context recognition for multiple users |
CN104935968A (en) * | 2015-06-25 | 2015-09-23 | 广州视源电子科技股份有限公司 | Method and system for recommending program by interactive personality TV (IPTV) |
CN104994408A (en) * | 2015-06-25 | 2015-10-21 | 青岛海信电器股份有限公司 | Intelligent television program recommendation method, device and intelligent television |
US20170332150A1 (en) * | 2016-05-13 | 2017-11-16 | Hulu, LLC | Personalized Content Ranking Using Content Received from Different Sources in a Video Delivery System |
CN107071578A (en) * | 2017-05-24 | 2017-08-18 | 中国科学技术大学 | IPTV program commending methods |
Non-Patent Citations (1)
Title |
---|
朱沿熹: "基于张量理论的移动推荐系统研究", 《硕士学位论文》 * |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109921799A (en) * | 2019-02-20 | 2019-06-21 | 重庆邮电大学 | A kind of tensor compression method based on cumulative amount dictionary learning |
CN111753182A (en) * | 2019-03-28 | 2020-10-09 | 北京达佳互联信息技术有限公司 | Multimedia information recommendation method and device, electronic equipment and readable storage medium |
CN111753182B (en) * | 2019-03-28 | 2024-05-17 | 北京达佳互联信息技术有限公司 | Multimedia information recommendation method and device, electronic equipment and readable storage medium |
CN110149556A (en) * | 2019-05-20 | 2019-08-20 | 河北工业大学 | A kind of IPTV user's behavior pattern mining method based on TDC-LDA model |
WO2020233245A1 (en) * | 2019-05-20 | 2020-11-26 | 山东科技大学 | Method for bias tensor factorization with context feature auto-encoding based on regression tree |
CN110196946A (en) * | 2019-05-29 | 2019-09-03 | 华南理工大学 | A kind of personalized recommendation method based on deep learning |
CN110196946B (en) * | 2019-05-29 | 2021-03-30 | 华南理工大学 | Personalized recommendation method based on deep learning |
CN113051463A (en) * | 2019-12-26 | 2021-06-29 | 中移物联网有限公司 | Project pushing method and system |
CN113254794A (en) * | 2021-07-15 | 2021-08-13 | 中国传媒大学 | Program data recommendation method and system based on modeling |
Also Published As
Publication number | Publication date |
---|---|
CN108521586B (en) | 2020-01-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108521586A (en) | The IPTV TV program personalizations for taking into account time context and implicit feedback recommend method | |
CN108737856B (en) | Social relation perception IPTV user behavior modeling and program recommendation method | |
Ahmed et al. | Movie recommendation system using clustering and pattern recognition network | |
US20120323725A1 (en) | Systems and methods for supplementing content-based attributes with collaborative rating attributes for recommending or filtering items | |
CN102780920A (en) | Television program recommending method and system | |
CN103686382A (en) | Program recommendation method | |
CN103106208A (en) | Streaming media content recommendation method and system in mobile internet | |
CN106791964B (en) | Broadcast TV program recommender system and method | |
CN104933135A (en) | Method and device for clustering multimedia data | |
CN111858972A (en) | Movie recommendation method based on family knowledge graph | |
Liu et al. | Resource allocation and scheduling in the intelligent edge computing context | |
CN109508407A (en) | The tv product recommended method of time of fusion and Interest Similarity | |
CN115618128A (en) | Collaborative filtering recommendation system and method based on graph attention neural network | |
CN102695086A (en) | Content pushing methods and device for interactive network protocol television | |
CN104935967B (en) | The interest recognition methods of video terminal user a kind of and device | |
Garg et al. | Social network analysis of YouTube: a case study on content diversity and genre recommendation | |
CN105550312B (en) | Context information processing method and device | |
Gao et al. | Deep learning with consumer preferences for recommender system | |
CN115618127A (en) | Collaborative filtering algorithm of neural network recommendation system | |
CN115422396A (en) | Implicit video recommendation algorithm | |
Oh et al. | Comparison of techniques for time aware TV channel recommendation | |
Pattanayak et al. | Review of recommender system for OTT platform through artificial intelligence | |
Chughtai et al. | Goal-based hybrid filtering for user-to-user Personalized Recommendation | |
He | Construction of internet TV industry ecosystem based on data mining technology | |
Xiao et al. | Social4rec: Distilling user preference from social graph for video recommendation in tencent |
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 |