CN108521586B - IPTV television program personalized recommendation method giving consideration to time context and implicit feedback - Google Patents

IPTV television program personalized recommendation method giving consideration to time context and implicit feedback Download PDF

Info

Publication number
CN108521586B
CN108521586B CN201810230242.8A CN201810230242A CN108521586B CN 108521586 B CN108521586 B CN 108521586B CN 201810230242 A CN201810230242 A CN 201810230242A CN 108521586 B CN108521586 B CN 108521586B
Authority
CN
China
Prior art keywords
tensor
program
user
time period
preference
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.)
Active
Application number
CN201810230242.8A
Other languages
Chinese (zh)
Other versions
CN108521586A (en
Inventor
尹小燕
王�华
米晓倩
刘浩
徐成
许鹏飞
汤战勇
陈晓江
房鼎益
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Northwest University
Original Assignee
Northwest University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Northwest University filed Critical Northwest University
Priority to CN201810230242.8A priority Critical patent/CN108521586B/en
Publication of CN108521586A publication Critical patent/CN108521586A/en
Application granted granted Critical
Publication of CN108521586B publication Critical patent/CN108521586B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

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 an IPTV television program personalized recommendation method giving consideration to time context and implicit feedback. Compared with the existing recommendation algorithm, the method can provide better recommendation precision and higher recall rate.

Description

IPTV television program personalized recommendation method giving consideration to time context and implicit feedback
Technical Field
The invention relates to the field of IPTV service, in particular to an IPTV television program personalized recommendation method giving consideration to time context and implicit feedback.
Background
In recent years, with the rapid progress of "three networks integration", diversified personalized services have been called for more and more. Telecommunication networks, wired networks and computer networks permeate each other, are gradually integrated into a unified information communication network, and are concerned in the fields of intelligent transportation, environmental protection, government work, public safety and the like. As one of the most important applications in "three-network convergence", IPTV already has a large number of users. Through the IPTV service, the user can watch any interested television program at any time. To improve the satisfaction of the IPTV to the user, IPTV operators are providing various personalized viewing services, such as Live TV (Live TV), Video On Demand (VOD), Catch-up TV, and the like. These services provide the user with a variety of viewing options: live television is just like common broadcast television, and a user can watch live video; VOD is a supplement to live television programs, and users can watch any pre-recorded video at any time, usually movies, television series, art programs or cartoons and the like; while the newly introduced Catch-up TV allows the user to watch any video that is live a few days ago (e.g., 3 days).
IPTV greatly increases the flexibility of user selection of television programs. However, users often face the problem of information overload when selecting among thousands of movies, television shows, or other programs. If it takes too long to find the desired tv program, the user's viewing experience will be greatly reduced. Therefore, there is a need for a personalized recommendation system to help users quickly find out the tv programs they like to watch, thereby increasing the satisfaction of the users with IPTV.
Currently, recommendation systems have a place in many business fields such as video streaming (e.g., Netflix, YouTube), music streaming (e.g., Spotify, Pandora), sales product services (e.g., Amazon, bookking), and social networking (e.g., Facebook, Linkedin). However, accurate recommendation of IPTV faces multiple challenges. First, IPTV service providers are typically limited in the user data they own for privacy protection. Secondly, most users watch very few tv programs, resulting in very sparse collected data. Third, without explicit ratings (e.g., scores) for television programs, the viewing preferences of the user cannot be inferred. Fourth, IPTV viewing data is affected by the composition of the home, and the user log is often cluttered and incomplete.
Some television program recommendation algorithms are proposed in the prior art, but the algorithms are all suitable for specific scenes, wherein, for example, a popup Items algorithm, a User popup Items algorithm, a collaborative filtering algorithm and a recommendation algorithm based on content filtering all utilize the two-dimensional relationship between a User and a television program to recommend the program. However, television program viewing is a family behavior, and viewing time is also an important factor to be considered for recommendation. The context-aware based recommendation algorithm, although combining user preferences in different environments, does not consider the potential semantic relevance between dimensions. The SVD-based recommendation algorithm cannot recommend a new program before evaluation data of a user on the new program is not obtained.
Disclosure of Invention
Aiming at the problems of the existing algorithm, the invention aims to provide the IPTV television program personalized recommendation method which gives consideration to time context and implicit feedback so as to realize accurate and effective personalized television program recommendation.
In order to realize the task, the invention adopts the following technical scheme:
the IPTV television program personalized recommendation method considering both time context and implicit feedback comprises the following steps:
step 1, establishing a preference model of a user watching a program
The preference model comprises loyalty and interestingness of a user to program types in a certain time period, wherein the loyalty is the ratio of the time length of the user watching a program of a certain type in the certain time period to the actual playing time length of the program of the type, and the interestingness is the ratio of the time length of the user watching the program of the certain type in the certain time period to the total time length of the user watching all program types;
determining preference values of the user on the program types in different time periods according to the preference model;
step 2, constructing tensors to express the relation among users, program types and time periods, wherein each item in the tensors represents the preference value of the users to a certain program type in a certain time period;
step 3, decomposing and reconstructing the tensor to obtain an approximate tensor of the tensor, constructing a loss function according to the tensor and the approximate tensor, and regularizing the loss function to obtain a target function; solving the objective function, and determining the optimal approximate tensor according to the solving result;
and 4, taking the elements in the optimal approximate tensor as a prediction result, and recommending the program type to the user according to the prediction result.
Further, the process of determining the preference value includes:
and respectively endowing different weights for the loyalty and the interest degree, and taking the sum of the loyalty and the interest degree endowed with the weights as the preference value.
Further, the weight value given to the loyalty is greater than the weight value given to the interestingness.
Further, the tensor A in the step 2 is expressed as:
Figure BDA0001602362620000031
wherein i represents a user, j represents a program type, d represents a time period, M represents the number of users, N represents the number of program types, and L represents the number of time periods.
Further, the method for obtaining the approximate tensor described in step 3 includes:
decomposing the tensor A into the modulus product of the core tensor S and each factor matrix (U, G, T), initializing the core tensor S and each factor matrix (U, G, T) to obtain U belonging to RM×du，G∈RN×dg，T∈RL×dt，S∈Rdu×dg×dt(ii) a Wherein du, dg and dt are positive integer values and satisfy du<<M,dg<<N,dt<<L and R represent a real number set;
reconstructing tensor A according to the initialized core tensor S and each factor matrix (U, G, T) by using the following formula to obtain approximate tensor
Figure BDA0001602362620000032
Figure BDA0001602362620000033
In the above formula, S isUU represents the U-modulus product of tensor S and matrix UGG、×TT represents G-modulus and T-modulus, respectively.
Further, the expression of the loss function described in step 3 is:
Figure BDA0001602362620000034
in the above formula, D is a label value, and each item a in Ai,j,dRepresents the preference value of the user i for the program j in the time period d if the item ai,j,dIf not, D is set to 1, otherwise, D is set to 0.
Further, the expression of the objective function in step 3 is:
Figure BDA0001602362620000035
in the above formula, the first and second carbon atoms are,
Figure BDA0001602362620000036
representing the Frobenius norm, the variables λ and σ are different tuning parameters.
Further, the solving of the objective function in step 3 includes:
randomly selecting a group of matrixes U with real number of elements from U, G, T, S obtained after initializationi,*、Gj,*、Td,*、S*To approximate U, G, T, S, by separately taking the partial derivatives for each component of the objective function, the following equation is obtained:
Figure BDA0001602362620000041
Figure BDA0001602362620000042
Figure BDA0001602362620000043
Figure BDA0001602362620000044
in the above formula, the above formula is shown,
Figure BDA0001602362620000045
represents the Kronecker product; after calculating the partial derivative result by the above equation, for Ui,*、Gj,*、Td,*、S*Updating, and recalculating partial derivative result by using the above equation after updating until the target function is converged to make U when the target function is convergedi,*、Gj,*、Td,*、S*U, G, T, S, the best approximate tensor is obtained.
An IPTV television program personalized recommendation considering both time context and implicit feedback comprises a preference model building module, a tensor building module, an optimal approximation tensor determining module and a recommending module which are connected in sequence, wherein the modules are respectively used for realizing the processes of the steps 1 to 4 in the claim 1.
Compared with the prior art, the invention has the following technical characteristics:
1. the method analyzes the watching preference of the user in different periods, increases the time dimension to track the dynamic watching behavior of family members, and constructs a preference model for the user to watch, so that the watching preference of the user is mined more accurately.
2. The method can accurately and effectively recommend the television programs individually by decomposing and mining the potential association among the users, the television programs and the watching time through tensor.
Drawings
FIG. 1 is a flow chart of a method of the present invention;
FIG. 2 is a flow chart of a tensor resolution algorithm;
FIG. 3 is a graph of accuracy comparisons for six recommended methods;
FIG. 4 is a chart comparing recall rates for six recommended methods;
FIG. 5 is a graph comparing the performance of the six recommended methods F-score.
The six recommendation methods in fig. 3 to 5 are the recommendation method, the Random recommendation method Random, the Popular program recommendation method popup, the user Popular program recommendation method UserPopular, the recommendation algorithm SVD based on singular value decomposition, and the Context-aware recommendation algorithm Context, respectively, provided by the present invention.
Detailed Description
The invention provides an IPTV television program personalized recommendation method giving consideration to time context and implicit feedback, which adopts a three-dimensional tensor to express the relation between a user and a watching time period and a program type, and adopts a high-order singular value decomposition algorithm to mine semantic association among the user, the watching time period and the program type, thereby realizing the recommendation of a program list of each time period for the user. The detailed steps of the invention are as follows:
step 1, establishing a preference model of a user watching a program
The preference model comprises loyalty and interestingness of a user to program types in a certain time period, wherein the loyalty is the ratio of the time length of the user watching a program of a certain type in the certain time period to the actual playing time length of the program of the type, and the interestingness is the ratio of the time length of the user watching the program of the certain type in the certain time period to the total time length of the user watching all program types; specifically, the method comprises the following steps:
step 1.1, dividing the user viewing time period
And (4) counting data, wherein the viewing preference of the user is periodically changed along with time, and the viewing behaviors of the user in the same time period of a working day and a holiday are different in consideration of the influence of working factors. In order to better analyze the change of the viewing data along with the time, 24 hours a day is divided into 7 time periods, and two conditions of a working day and a holiday are respectively considered, so that 14 time periods (the sum of the 7 time periods of the working day and the 7 time periods of the holiday) are obtained; a specific time period division manner is shown in table 1 below:
table 1: user viewing time period division
Starting time End time Description of time periods
06:01 08:00 Early morning (morning)
08:01 12:00 In the morning
12:01 13:00 Noon is a Chinese traditional musical instrument
13:01 18:00 In the afternoon
18:01 19:30 In the evening
19:31 24:00 Night time
00:01 06:00 Early morning
Step 1.2, constructing a preference model of the program watched by the user
The preference model is as follows:
Figure BDA0001602362620000062
in the above formula, m and n respectively represent the number of program types and service types.
Loyaltyi,j,dIndicating the loyalty of user i to type j of the program during time period d, i.e. the ratio of the time during which user i views the type of program during time period d to the actual playing time of the type of program,
Figure BDA0001602362620000063
representing the weighted total duration, t, of the program type j viewed by the user i during the time period dj,k,dIs the time when user i watches program type j through service type k in time period d, Tj, k is the actual playing time length of program type j through service type k,
Figure BDA0001602362620000064
is the total duration of the play through all service types program type j. wk is the weight of service type k.
Interesti,j,dRepresenting the interest of the user i in the program type j in the time period d, i.e. the ratio of the total time for the user i to watch the program of the type in the time period d to the total time for the user i to watch all the program types in the time period d,
Figure BDA0001602362620000065
is the weighted total length of time that user i watches all program types through all service types during time period d.
For example, in IPTV, service types k include Catch-up TV, VOD, and Live TV. According to data statistics, the weights of the three service types meet the following conditions: catch-up TV > VOD > Live TV, so the weights for the three service types can take on values of 2, 1.5 and 1.
Step 1.3, respectively endowing different weights for the loyalty and the interest degree, and taking the sum of the loyalty and the interest degree endowed with the weights as the preference value, wherein the formula of the preference value is as follows:
ai,j,d＝Loyaltyi,j,dW1+Interesti,j,dW2
in the above formula, W1,W2The weights assigned to loyalty and interestingness, respectively. In the preference model, it is believed that the loyalty of a user to a program is more representative of the user's preference for that program type. Thus, the weight given to the loyalty is greater than the weight given to the interest, W1>W2(ii) a For example, W1,W2Values of (d) can be 0.8 and 0.2.
For example, the viewing preference value of the user i for the program type j in the time period d (12:01-13:00) is calculated. Suppose the service type Catch-up TV shows program type j at 10 times: 00-14:00, user i watches TV programs via the service type Catch-up TV in time period 11:30-13:30 (weight w occupied by the service type Catch-up TV)k2) where 12:30-13:30 is watching program type j, the user has loyalty to program type j during time period 12:01-13:00
Figure BDA0001602362620000071
Degree of interest
Figure BDA0001602362620000072
The viewing preference value of the user i for the program type j in the time period d (12:01-13:00) is ai,j,d＝0.125×0.8+0.25×0.2＝0.15。
Step 2, constructing tensors to express the relation among users, program types and time periods, wherein each item in the tensors represents the preference value of the users to a certain program type in a certain time period;
the tensor A is expressed as:
Figure BDA0001602362620000073
wherein i represents a user, j represents a program type, d represents a time period, M represents the number of users, and N represents a program typeNumber, L denotes the number of time segments, in tensor A, each term a thereofi,j,dRepresenting the preference value of user i for program type j over time period d.
Step 3, decomposing and reconstructing the tensor to obtain an approximate tensor of the tensor, constructing a loss function according to the tensor and the approximate tensor, and regularizing the loss function to obtain a target function; solving the objective function, and determining the optimal approximate tensor according to the solving result;
step 3.1, decomposing the tensor A:
the potential connection between users, program types, time periods is mined using HOSVD to decompose tensor A into the product of the core tensor S and the factor matrices (U, G, T).
Step 3.2, initializing the core tensor and each factor matrix (U, G, T) to obtain U e to RM×du，G∈RN×dg，T∈RL×dt，S∈Rdu×dg×dt(ii) a Wherein R represents a real number set, M represents a user number, N represents a program type number, L represents a time period number, du, dg, and dt are all very small positive integer values, and du is satisfied<<M,dg<<N,dt<<And L. Generally du<0.0005×M，dg<0.2×N，dt<0.5×L。
And 3.3, reconstructing the tensor A by using the following formula according to the initialized core tensor S and each factor matrix (U, G, T) to obtain an approximate tensor
Figure BDA0001602362620000081
Figure BDA0001602362620000082
In the above formula, S isUU represents the U-modulus product of tensor S and matrix UGG、×TT represents the G-modulus (of the tensor and the matrix U) and the T-modulus (of the tensor and the matrix T), respectively; the computation of the n (U, G, T) product of the tensor and matrix is well known in the art and will not be described in detail herein.
Then approximate tensor
Figure BDA0001602362620000083
Each element of
Figure BDA0001602362620000084
Expressed as:
in the above formula, sdu,dg,dtRepresenting the element, u, located at the (du, dg, dt) position of the core tensor Si,duElements representing the ith row and du column of the matrix U, gj,dgRepresenting the elements in the jth row and dg column of the matrix G, td,dtRepresenting the elements located in the d-th row and dt-th column of the matrix T.
Step 3.4, in order to obtain the optimal approximate tensor, defining a loss function, quantizing the approximated values obtained by decomposing the tensor to obtain the optimal model parameters, namely, each factor matrix (U, G, T) and the core tensor S:
Figure BDA0001602362620000086
in the above formula, A is the tensor constructed in step 2,
Figure BDA0001602362620000087
to approximate the tensor, D is a marker value to mark if information is missing. Each item a in Ai,j,dRepresents the preference value of the user i for the program j in the time period d if the item ai,j,dIf not, D is set to 1, otherwise, D is set to 0.
Step 3.5, since the types of programs watched by the user at different time periods are less, and the obtained data is sparse, training the model through the data set composed of sparse tensors to obtain the optimal parameters (each factor matrix (U, G, T) and the core tensor S) usually results in data overfitting. In order to prevent overfitting of the model and improve the generalization capability of the model, a formula is normalized by using a term based on a Frobenius norm, that is, a regularization term is added to a loss function of the model to obtain an objective function C (U, G, T, S):
Figure BDA0001602362620000091
in the above formula, the first and second carbon atoms are,the Frobenius norm is represented, the variables lambda and sigma are different adjusting parameters, general experience values of the adjusting parameters are {0.1,0.2,0.01 and 0.001} through a large amount of experimental analysis, and an objective function can be optimized according to multiple times of adjustment of data.
Step 3.6, solving the objective function
To minimize the loss function, the random gradient of the factor matrices U, G, T and the core tensor S is used for optimization. Random gradient is the "gradient" obtained by a set of randomly selected partial data when the amount of data is too large or when the entire data cannot be obtained at one time. Thus, U e R resulting from the initialization of step 3.2M×du，G∈RN×dg，T∈RL×dt，S∈Rdu×dg×dtRandomly selecting a group of matrix U with each element being real numberi,*、Gj,*、Td,*、S*To approximate U, G, T, S, by separately taking the partial derivatives for each component of the objective function, the following equation is obtained:
Figure BDA0001602362620000093
Figure BDA0001602362620000094
Figure BDA0001602362620000095
Figure BDA0001602362620000096
in the above formula set, the variables λ and σ are differentThe empirical values of the tuning parameters are typically {0.1,0.2,0.01,0.001 }.
Figure BDA0001602362620000097
Representing the Kronecker product (Kronecker product) and the subscript preceding the matrix parameter in the above formula represents the modulus product.
Based on the obtained partial derivative result, adopting an iteration mode, in each iteration, calculating the partial derivative of each component of the objective function according to the above formula group, and updating each component through the following formula group after calculation:
Figure BDA0001602362620000098
Figure BDA0001602362620000101
Figure BDA0001602362620000102
Figure BDA0001602362620000103
Figure BDA0001602362620000104
in the above formula set, μ represents the number of iterations, η represents the learning rate, the superscripted 'component represents the updated component, and the unsupplemented' component represents the pre-update component, e.g., Ui′,*Indicating updated Ui,*。
Continuously iteratively updating until the target function converges to satisfy Cμ-Cμ+1>Epsilon, wherein Cμ、Cμ+1Respectively representing the calculation of the mu and the mu +1 iteration to obtain an objective function, wherein epsilon represents any small positive integer, the smaller the value of epsilon is, the closer the value of epsilon is to the original tensor, and the value range is generally [0.01,0.0001 ]]. U at the time of convergence of the objective functioni,*、Gj,*、Td,*、S*U, G, T, S being optimal (i.e., the values of these several parameters are U)i,*、Gj,*、Td,*、S*Value of (d) U, G, T, S is taken into equation 1 to find the best approximate tensor
And 4, taking the elements in the optimal approximate tensor as a prediction result, and recommending the program type to the user according to the prediction result.
From the best approximation tensor
Figure BDA0001602362620000106
The user i recommends preference values for various types of programs in the time period d. Specifically, if we want to recommend K program types to the user i in a certain time period d, the preference values of the user i for different program types in the time period d can be sorted according to the size, and then the K program types with larger preference values are recommended to the user.
On the basis of the recommendation method, the invention further provides an IPTV television program personalized recommendation considering both time context and implicit feedback, which includes a preference model building module, a tensor building module, an optimal approximation tensor determining module and a recommendation module, which are connected in sequence, and these modules are respectively used for implementing the processes described in steps 1 to 4 in claim 1.
Performance analysis of the method of the invention
And carrying out off-line evaluation on the recommendation method by the collected data set, and comparing the influence of each recommendation algorithm on the performance of each aspect.
Step one, data set statistics
In this experiment, the data set provided by china unicom was used. The data set included a log of television programs and a schedule log for universal user viewing at 48.1 in Henan province between 2016, 10, month 1 and 2017, 1, month 1. In our experiment, only log records of more than 10 minutes of continuous viewing of a program were considered, the main data of the filtered data set, as shown in table 2, for a total of 478,170 users and 517,277,222 viewing records. Table 3 shows part of the data of the television program viewing log. Table 4 shows part of the data for the Live TV program schedule log, the data being in the same format as the program viewing log.
The data set is divided into a set of training sets comprising logs of 2016 [ 10/1/10 ] to 2016 [ 12/12 ] 14 (75, 421,821,170 contestants) and a set of test sets comprising 2016 [ 12/15 ] to 2017 [ 1/1 (18, 95,456,052 entries). The former is used to build models and the latter is used to evaluate recommendation performance.
Table 2 data set statistics
Users 481,232
Time span 93days
views 536,649,128
programs 7,154
TABLE 3 part of the viewing Log of television programs
Figure BDA0001602362620000111
TABLE 4 part of Live TV program schedule Log
Figure BDA0001602362620000112
Step two, program category division
There are 62 channels in our statistical data set, which are classified into the following 16 categories: finance, international, sports, military and agricultural, drama, news, children's wear, shopping, martial arts, documentaries, science and education, social and legal, music, art programs, television sitcoms and movies. In order to make the three services better able to classify the programs, we have further divided the three categories of drama, movie and art programs. There are 17 types of television series: home, crime, suspicion, ancient, comedy, love, war, science fiction novels, martial arts, action, city, history, idol, military, police, criminal investigation and spy. Movies are classified into 14 categories: crime, thriller, action, comedy, love, science fiction, adventure, suspicion, drama movie, horror, literature, history, war and cartoon. The comprehensive programs are divided into 9 types: the real person shows, the music is combined with the art program, the interview, the life, the evening, the travel, the talk show, the show talent show and the smart choice show. So in combination with the already divided 13 categories (finance, international, sports, military and agriculture, drama, news, children's garments, shopping, martial arts, documentaries, science and education, social and legal, music, finance, international, sports, military and agriculture, drama, news, children's garments, shopping, martial arts, documentaries, science and education, social and legal, music), we obtained 53 television program types and can obtain the program types by finding the program names in the tracking list.
Step three, analyzing the experimental result
Figure 3 shows the results of the accuracy comparison for the six recommendation systems. The accuracy is the ratio of the number of types of recommendations watched by the user to the total recommendation type. Intuitively, from fig. 3, the recommendation method we use is generally better accurate than the other five recommendation systems. Since the user views a small number of programs in reality, the accuracy is increased at the beginning and then decreases as the number of recommended genres increases.
Figure 4 shows the recall rates for different values of n for the six recommendation systems. The recall ratio refers to a ratio of recommended programs among programs viewed by the user. Higher recall indicates more accurate user preferences predicted by the recommendation system. In fig. 4 we can see that the recall increases with the number of recommendation types and that from n-7 the recall of our recommendation system is always at optimum.
Figure 5 shows the results of the F-score performance comparison for the six recommendation systems. The F-score is a comprehensive assessment of accuracy and recall. As can be seen from FIG. 5, F-score performance is also increasing with the number of recommended types. Furthermore, it can be seen that our recommendation is at least 10% better than the other five recommendation systems.

Claims (2)

1. The IPTV television program personalized recommendation method giving consideration to time context and implicit feedback is characterized by comprising the following steps of:
step 1, constructing a preference model of a user for watching programs;
the preference model comprises loyalty and interestingness of a user to program types in a certain time period, wherein the loyalty is the ratio of the time length of the user watching a program of a certain type in the certain time period to the actual playing time length of the program of the type, and the interestingness is the ratio of the time length of the user watching the program of the certain type in the certain time period to the total time length of the user watching all program types;
determining preference values of the user on the program types in different time periods according to the preference model;
step 2, constructing tensors to express the relation among users, program types and time periods, wherein each item in the tensors represents the preference value of the users to a certain program type in a certain time period;
step 3, decomposing and reconstructing the tensor to obtain an approximate tensor of the tensor, constructing a loss function according to the tensor and the approximate tensor, and regularizing the loss function to obtain a target function; solving the objective function, and determining the optimal approximate tensor according to the solving result;
step 4, taking the elements in the optimal approximate tensor as a prediction result, and recommending the program type to the user according to the prediction result;
the determination process of the preference value comprises the following steps:
respectively endowing different weights for the loyalty and the interest, and taking the sum of the loyalty and the interest endowed with the weights as the preference value;
the weight value given to the loyalty is larger than the weight value given to the interest;
the tensor A in step 2 is expressed as:
Figure FDA0002221335510000011
wherein i represents a user, j represents a program type, d represents a time period, M represents the number of users, N represents the number of program types, L represents the number of time periods, and in the tensor A, each item a thereofi,j,dA preference value representing the type j of the program for the user i in the time period d;
the method for obtaining the approximate tensor described in the step 3 comprises the following steps:
decomposing the tensor A into the modulus product of the core tensor S and each factor matrix (U, G, T), initializing the core tensor S and each factor matrix (U, G, T) to obtain U belonging to RM×du，G∈RN×dg，T∈RL×dt，S∈Rdu×dg×dt(ii) a Wherein du, dg and dt are positive integer values and satisfy du<<M,dg<<N,dt<<L, R represents a real number set, M represents the number of users, N represents the number of program types, and L represents the number of time periods;
reconstructing tensor A according to the initialized core tensor S and each factor matrix (U, G, T) by using the following formula to obtain approximate tensor
Figure FDA0002221335510000021
Figure FDA0002221335510000022
In the above formula, S isUU represents the U-modulus product of tensor S and matrix UGG、×TT represents G-modulus and T-modulus,
the expression of the loss function described in step 3 is:
Figure FDA0002221335510000023
in the above formula, D is a label value, and each item a in Ai,j,dRepresents the preference value of user i for program type j in time period d if the item ai,j,dIf the value is not zero, setting D to be 1, otherwise, setting D to be 0;
the expression of the objective function in step 3 is:
Figure FDA0002221335510000024
in the above formula, the first and second carbon atoms are,
Figure FDA0002221335510000025
expressing Frobenius norm, wherein variables lambda and sigma are different adjusting parameters;
the solving of the objective function in step 3 includes:
randomly selecting a group of matrixes U with real number of elements from U, G, T, S obtained after initializationi,*、Gj,*、Td,*、S*To approximate U, G, T, S, by separately taking the partial derivatives for each component of the objective function, the following equation is obtained:
Figure FDA0002221335510000027
Figure FDA0002221335510000028
in the above formula, the above formula is shown,
Figure FDA0002221335510000032
represents the Kronecker product; after calculating the partial derivative result by the above equation, for Ui,*、Gj,*、Td,*、S*Updating, and recalculating partial derivative result by using the above equation after updating until the target function is converged to make U when the target function is convergedi,*、Gj,*、Td,*、S*U, G, T, S, the best approximate tensor is obtained.
2. An IPTV television program personalized recommendation system giving consideration to both time context and implicit feedback is characterized by comprising a preference model building module, a tensor building module, an optimal approximation tensor determining module and a recommendation module which are connected in sequence, wherein the preference model building module, the tensor building module, the optimal approximation tensor determining module and the recommendation module are respectively used for realizing the processes in the steps 1 to 4 in the claim 1.
CN201810230242.8A 2018-03-20 2018-03-20 IPTV television program personalized recommendation method giving consideration to time context and implicit feedback Active CN108521586B (en)

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 CN108521586A (en) 2018-09-11
CN108521586B true 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)

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109921799B (en) * 2019-02-20 2023-03-31 重庆邮电大学 Tensor compression method based on energy-gathering dictionary learning
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
CN110149556B (en) * 2019-05-20 2021-05-04 河北工业大学 IPTV user behavior pattern mining method based on TDC-LDA model
CN110209933A (en) * 2019-05-20 2019-09-06 山东科技大学 A kind of biasing tensor resolution method based on regression tree contextual feature autocoding
CN110196946B (en) * 2019-05-29 2021-03-30 华南理工大学 Personalized recommendation method based on deep learning
CN113051463B (en) * 2019-12-26 2023-07-07 中移物联网有限公司 Project pushing method and system
CN113254794B (en) * 2021-07-15 2021-10-01 中国传媒大学 Program data recommendation method and system based on modeling

Citations (5)

* Cited by examiner, † Cited by third party
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

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CA3022268C (en) * 2016-05-13 2022-06-21 Hulu, LLC Personalized content ranking using content received from different sources in a video delivery system

Patent Citations (5)

* Cited by examiner, † Cited by third party
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

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
基于张量理论的移动推荐系统研究;朱沿熹;《硕士学位论文》;20140401;第6-24页 *

Also Published As

Publication number Publication date
CN108521586A (en) 2018-09-11

Similar Documents

Publication Publication Date Title
CN108521586B (en) IPTV television program personalized recommendation method giving consideration to time context and implicit feedback
US10885063B2 (en) Feedback loop content recommendation
CN108737856B (en) Social relation perception IPTV user behavior modeling and program recommendation method
CN106802956B (en) Movie recommendation method based on weighted heterogeneous information network
CN108875022B (en) Video recommendation method and device
JP5032477B2 (en) System and method for recommending items of interest to users
US20120036523A1 (en) Adaptive Blending of Recommendation Engines
US20140149424A1 (en) Time weighted content recommendation
CN104182449A (en) System and method for personalized video recommendation based on user interests modeling
US20140150005A1 (en) Content recommendation pre-filtering
CN103686382A (en) Program recommendation method
CN109246450B (en) Movie and television preferred recommendation method based on implicit information scoring
US20190082206A1 (en) Systems and Methods for Predicting Audience Measurements of a Television Program
US20140149326A1 (en) Post-processed content recommendation
CN106227881B (en) Information processing method and server
CN107368584B (en) Personalized video recommendation method and system
CN114339417A (en) Video recommendation method, terminal device and readable storage medium
Wang et al. User identification for enhancing IP-TV recommendation
Fazelnia et al. Variational user modeling with slow and fast features
CN109640128B (en) Television user watching behavior feature extraction method and system
CN113051468B (en) Movie recommendation method and system based on knowledge graph and reinforcement learning
CN109543094B (en) Privacy protection content recommendation method based on matrix decomposition
US20140074858A1 (en) Percent-consumed weighted content recommendation
CN108182600A (en) A kind of method and system that extending user is determined according to weighted calculation
CN108053260A (en) A kind of method and system that extending user is determined according to statistics interest-degree

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