CN108521586B

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

Info

Publication number: CN108521586B
Application number: CN201810230242.8A
Authority: CN
Inventors: 尹小燕; 王�华; 米晓倩; 刘浩; 徐成; 许鹏飞; 汤战勇; 陈晓江; 房鼎益
Original assignee: Northwest University
Current assignee: Northwest University
Priority date: 2018-03-20
Filing date: 2018-03-20
Publication date: 2020-01-14
Anticipated expiration: 2038-03-20
Also published as: CN108521586A

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:

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 R^M×du，G∈R^N×dg，T∈R^L×dt，S∈R^du×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

In the above formula, S is_UU represents the U-modulus product of tensor S and matrix U_GG、×_TT represents G-modulus and T-modulus, respectively.

Further, the expression of the loss function described in step 3 is:

in the above formula, D is a label value, and each item a in A_i,j,dRepresents the preference value of the user i for the program j in the time period d if the item a_i,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:

in the above formula, the first and second carbon atoms are,

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 initialization_i,*、G_j,*、T_d,*、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:

in the above formula, the above formula is shown,

represents the Kronecker product; after calculating the partial derivative result by the above equation, for U_i,*、G_j,*、T_d,*、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 converged_i,*、G_j,*、T_d,*、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:

in the above formula, m and n respectively represent the number of program types and service types.

Loyalty_i,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,

representing the weighted total duration, t, of the program type j viewed by the user i during the time period d_j,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,

is the total duration of the play through all service types program type j. wk is the weight of service type k.

Interest_i,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,

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:

a_i,j,d＝Loyalty_i,j,dW₁+Interest_i,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

Degree of interest

The viewing preference value of the user i for the program type j in the time period d (12:01-13:00) is a_i,j,d＝0.125×0.8+0.25×0.2＝0.15。

the tensor A is expressed as:

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 thereof_i,j,dRepresenting the preference value of user i for program type j over time period d.

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 R^M×du，G∈R^N×dg，T∈R^L×dt，S∈R^du×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

In the above formula, S is_UU represents the U-modulus product of tensor S and matrix U_GG、×_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

Each element of

Expressed as:

in the above formula, s_du,dg,dtRepresenting the element, u, located at the (du, dg, dt) position of the core tensor S_i,duElements representing the ith row and du column of the matrix U, g_j,dgRepresenting the elements in the jth row and dg column of the matrix G, t_d,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:

in the above formula, A is the tensor constructed in step 2,

to approximate the tensor, D is a marker value to mark if information is missing. Each item a in A_i,j,dRepresents the preference value of the user i for the program j in the time period d if the item a_i,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):

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 3.2^M×du，G∈R^N×dg，T∈R^L×dt，S∈R^du×dg×dtRandomly selecting a group of matrix U with each element being real number_i,*、G_j,*、T_d,*、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:

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 }.

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:

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., U_i′_,*Indicating updated U_i,*。

Continuously iteratively updating until the target function converges to satisfy C_μ-C_μ+1>Epsilon, wherein C_μ、C_μ+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 function_i,*、G_j,*、T_d,*、S_*U, G, T, S being optimal (i.e., the values of these several parameters are U)_i,*、G_j,*、T_d,*、S_*Value of (d) U, G, T, S is taken into equation 1 to find the best approximate tensor

From the best approximation tensor

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

TABLE 4 part of Live TV program schedule Log

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

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;

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:

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 thereof_i,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 R^M×du，G∈R^N×dg，T∈R^L×dt，S∈R^du×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;

In the above formula, S is_UU represents the U-modulus product of tensor S and matrix U_GG、×_TT represents G-modulus and T-modulus,

the expression of the loss function described in step 3 is:

in the above formula, D is a label value, and each item a in A_i,j,dRepresents the preference value of user i for program type j in time period d if the item a_i,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:

in the above formula, the first and second carbon atoms are,

expressing Frobenius norm, wherein variables lambda and sigma are different adjusting parameters;

the solving of the objective function in step 3 includes:

in the above formula, the above formula is shown,

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.