KR101372928B1 - Apparatus, System, Method and Computer Readable Recording Media Storing the Program for Automatic Recommendation of TV Program Contents based on User-Preferred Topic - Google Patents

Abstract

The present invention applies LDA (Latent Dirichlet Allocation), which is a basic model of topic modeling, to a user's viewing history, and finds and recommends user preference topics by using collaborative filtering techniques considering personal preferences and public preferences. We present technology for auto recommending TV program contents. According to an aspect of the present invention, a user grouping unit configured to set a user group by using preference viewing information of a plurality of users, and a subject that learns a topic model parameter for each user group using the preferred viewing information. A model parameter learning unit, a parameter mapping unit for mapping a hidden topic parameter among the subject model parameters to individual and group preferences, and additional processing or inference to the subject model parameters to reveal the diversity of the individual preferences. It provides a user preference topic-based TV program content automatic recommendation apparatus including a parameter processing unit for performing the. According to the present invention, the TV program recommendation is made in consideration of the user's viewing habits and popular content, thereby facilitating analysis of viewing characteristics of the user and increasing user's convenience in using the TV.

Description

TV program content recommendation system based on user preferences Apparatus, System, Method and Computer Readable Recording Media Storing the Program for Automatic Recommendation of TV Program Contents based on User-Preferred Topic}

The present invention relates to an apparatus, system, method for automatically recommending TV program content based on user preferences, and a computer-readable recording medium having recorded thereon a program for executing the method. In detail, the present invention utilizes only a predetermined upper part of the preferred viewing information as it is while learning the Dirichlet distribution, which is a prior probability of the hidden subject parameter, and leaves the remaining lower part as 0, or asymmetrics the Dirichlet distribution. asymmetric) Automatically infers the latent topic by calculating with the LDA (Latent Dirichlet Allocation) model, and automatically applies the TV program contents based on the user's preference topic applying the collaborative filtering technology using the user's preference and the group's preference to the inferred hidden topic. The present invention relates to a computer-readable recording medium having recorded thereon a recommended apparatus, a system, a method, and a program for executing the method.

Recently, with the emergence of IPTV and Smart TV due to the convergence of multi-channel, multi-media, broadcasting and communication, a large amount of TV programs are provided to viewers, which makes it difficult for viewers (users) to find and watch TV program contents. It became. According to the survey, 75% of users take more than a few minutes to find the content they want, and 34% of customers are afraid to choose that content on TVs, PCs, and mobile devices. Interested in recommendations tailored to the viewing experience, 43% of British are willing to pay for Vod content consumption if interesting content is recommended.

Research on content recommendations has been studied through online store and search companies such as Amazon, Google, Apple, and short-selling apps such as video and music recommendations, book recommendations, and news recommendations, and app recommendations on mobile phones. Its effectiveness has been proven. Recommendation is closely related to search, and related search terms are derived from keyword-based query. Product recommendation to the user has been similarly applied in the form of matching the user's consumption history to a query and recommending related contents. As a representative model representing the user's preference as a feature vector and expressing it in a hidden topic, a latent topic, many researches have been conducted based on Hofmann's Probabilistic Latent Semantic Indexing (PLSI).

First, there is a related art regarding recommendation using a viewing preference vector based on a series of categories. Based on a series of known categories (genres, channels, etc.), user preferences are constructed based on the user's preferences for feature vectors as many as the number of categories, and the feature vectors of each user and the characteristics of each content. It is a way of recommending preference items based on similarity (or correlation, correlation) between vectors. Widely used methods for similarity calculation include Vector Cosine Correlation and Euclidian distance.

Second, there is a related art related to a system for recommending a film based on topical modeling. This method excludes the recommendation of short content such as movies from the main recommendation targets, since films have low re-watching rate and are mainly short stories, so there are few films composed of several circuits under one title. In addition, since the data are largely lean based on the evaluation data of the users, they are mainly used for the purpose of predicting the user's evaluation of non-evaluation items.

1 is a flowchart of a conventional subject modeling-based user preference content recommendation technology. According to this prior art, after learning modeling-based parameters for recommendation, evaluating predictions for non-evaluation items based on parameter learning data, except for the user's viewing items in the recommendation candidates, the similarity with the model-based user ( The user's preference content recommendation is made by recommending highly recommended items by sorting them to the top.

Existing TV program recommendation system based on the prior art has a lack of meaningful analysis of the reason for viewing a specific TV program of a user who is limited to TV program genre or channel when inferring user preference in recommendation model composition. Existing automated recommendation systems that have applied hidden subject modeling and collaborative filtering techniques based on user-evaluated data on short film-oriented content, compared to TV program content, which consists of multiple episodes under one title, TV programs are primarily aimed at predicting ratings for content you haven't watched, while automatic recommendation of TV shows is primarily aimed at screening content that users can watch for the purpose of ease of use. There was a limit to the effective application of the recommended application.

In order to solve the above problems, the present invention aims to enable a semantic recommendation for a TV program by inferring and learning a user's preference topic using the actual user's viewing history data.

In addition, by using a collaborative filtering technique based on the LDA model, asymmetric optimization of the Dirichlet distribution aims to provide improved recommendation performance that reflects the diversity of personal preferences with emphasis on individual preferences versus public preferences.

In order to achieve the above object, a first aspect of the present invention provides a user grouping unit for setting a user group using preference viewing information of a plurality of users, and a subject model parameter for each user group using the preference viewing information. a subject model parameter learning unit for learning model parameters, a parameter mapping unit for mapping a hidden topic parameter among the subject model parameters to individual preferences and group preferences, and the subject to reveal diversity of individual preferences. A user preference topic based TV program content recommendation apparatus including a parameter processing unit that performs additional processing or inference on model parameters is provided.

In addition, the parameter mapping unit, the mapping (mapping) meaning the words appearing in the document in the document generation modeling, the mapping configured to view the user viewing token as a word in the document generation modeling, and to view each document to the TV program, An apparatus for automatically recommending TV program content based on user preferences that performs any one of a mapping corresponding to a viewing token to a viewing user or a mapping consisting of subject modeling and popular preferences and personal preferences is provided.

In addition, the parameter processing unit, in learning the Dirichlet distribution (Dirichlet distribution), which is a prior probability of the hidden subject parameter, utilizes only a predetermined upper part of the preferred viewing information as it is and leave the remaining lower part as 0, or the processing An apparatus for automatically recommending TV program content based on a user's preference that calculates a Dirichlet distribution into an asymmetric Dirichlet distribution.

In addition, by using the preferred viewing information provides a user preference topic-based TV program content automatic recommendation device further comprises a non-preferred viewing item filter to exclude any one or more of the end of the broadcast content or the content that the user has not subscribed to do.

In addition, the present invention provides a user preference topic-based TV program content automatic recommendation apparatus further comprising a recommendation candidate content aligning unit for arranging the recommended candidate content for each user using the subject model parameter.

In addition, the user preference topic-based TV program content using any one or more of a technique such as Gibbs Sampling (CGS), Collapsed Gibbs Sampling (CGS), Variational Bayesian Inference (CBI), or Collapsed Variational Bayesian Inference (CVBI) for the mapping or the additional inference. Provide an automatic recommendation device.

In order to achieve the above object, the second aspect of the present invention is a media streaming server for broadcasting TV content; A user TV receiving a TV content recommendation list by transmitting a viewing record while receiving TV content from the media streaming server; A TV user viewing record database for receiving and storing a viewing record from the user TV; Set a user group using the preference viewing information generated from the viewing records received from the plurality of user TVs, and learn a topic model parameter for each user group using the preferred viewing information. Mapping hidden topic parameters among model parameters to individual preferences and group preferences, and performing additional processing or inference on the subject model parameters to reveal the diversity of individual preferences, thereby providing a TV content recommendation list to the user TV. User management and recommendation server for transmitting; And a TV content database storing a TV content to be transmitted to the media streaming server and the user management and recommendation server.

In addition, the user management and recommendation server, while learning the Dirichlet distribution (Dirichlet distribution), which is the prior probability of the hidden topic parameters, utilizes only a predetermined upper part of the preferred viewing information as it is and processes the remaining lower part to 0 The Dirichlet distribution is calculated as an asymmetric Dirichlet distribution, and one or more of the end-of-air content or the content to which the user's TV is not subscribed are excluded from the recommendation candidate using the preferred viewing information, and the subject model parameter is used. It provides an automatic recommendation system for TV program contents based on user preference topics that sorts the recommended candidate content for each user.

In order to achieve the above object, a third aspect of the present invention provides a user grouping step of setting a user group using preference viewing information of a plurality of users, and a subject model parameter (topic) for each user group using the preference viewing information. a subject model parameter learning step of learning a model parameter, a parameter mapping step of mapping a hidden topic parameter among the subject model parameters to individual preferences and group preferences, and the subject to reveal the diversity of the individual preferences. A parameter processing step of performing additional processing or inference on a model parameter, a non-preferred viewing item filtering step of excluding any one or more of the broadcasting end content or the content not subscribed to by the user using the preferred viewing information, and Using the subject model parameters It provides a user-preferred topic-based TV program can automatically recommend how to include the recommended candidate content alignment step to align the per-user content recommendation candidates.

In the parameter processing step, when learning a Dirichlet distribution, which is a prior probability of the hidden subject parameter, only a predetermined upper part of the preferred viewing information is used as it is and the remaining lower part is left as 0, or processed. A method for automatically recommending TV program content based on user preferences that calculates the Dirichlet distribution into an asymmetric Dirichlet distribution.

In order to achieve the above object, a fourth aspect of the present invention provides a subject model parameter learning step of learning a topic model parameter for each user group set by using preference viewing information received from a plurality of users, and the subject. Mapping the hidden topic parameters among the model parameters to individual preferences and group preferences, but mapping means the words that appear in the document in document generation modeling, and user viewing tokens in the document generation modeling. A parameter mapping step of performing any one of a mapping configured to view a document, a mapping of each document to a TV program, a mapping of each viewing token to a viewer, or a mapping of popular and personal preferences to subject modeling, and the hidden topic parameter. In learning the Dirichlet distribution, Further, by using only a predetermined upper part of the preferred viewing information and processing the remaining lower part as 0, or processing the Dirichlet distribution to be an asymmetric Dirichlet distribution, additional inference that reveals the diversity of personal preferences is obtained. A method for automatically recommending a TV program content based on a user's preference including a parameter processing step of performing the same.

In order to achieve the above object, the fifth aspect of the present invention provides a computer-readable recording medium having recorded thereon a program for executing any one of the above methods.

According to the present invention, it is possible to easily analyze the viewing characteristics of the users by expressing the preference topics of the users with high frequency of viewing.

In addition, by presenting TV contents selected from numerous TV program contents through recommendation in consideration of the user's viewing habits and popular preference contents, the user's convenience in using the TV is increased.

1 is a flowchart of a conventional subject modeling-based user preference content recommendation technology.
FIG. 2 is a data distribution diagram showing the relationship between a Dirichlet distribution and a polynomial distribution according to symmetrical and asymmetrical distributions.
3 is a block diagram of a system for automatically recommending TV program content based on user preferences according to an embodiment of the present invention.
4 is a block diagram of a user preference topic-based TV program content automatic recommendation device (user management and recommendation server) according to an embodiment of the present invention.
5 is a simplified flowchart of a method for automatically recommending TV program content based on user preferences according to an embodiment of the present invention.
6 is a conceptual diagram of user preference based grouping.
7 is a schematic of a symmetric α, symmetric β (SS prior) hyperparameter based LDA model.
8 is a schematic of an asymmetric α, symmetric β (AS prior) hyperparameter based LDA model.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the following description of the present invention, if it is determined that a detailed description of a known function or configuration may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted. The following terms are defined in consideration of the functions of the present invention, and may be changed according to the intentions or customs of the user, the operator, and the like. Therefore, the definition should be based on the contents throughout this specification.

In this specification, a topic refers to a set of contents that reveal the reason for viewing each individual user, and a preference (interest) refers to a user's interest in TV content. Expressed as a feature vector, a term used to refer to a concept called a user profile.

The present invention proposes a TV program recommendation model through a technique of applying collaborative filtering based on Lateral Dirichlet Allocation (LDA) model to user viewing history data. The proposed recommendation model makes recommendations to individuals based on individual user ranking ranking models that consider both individual preferences and public preferences, but focuses on personal preferences versus personal preferences to improve recommendation performance. It reflects diversity. To do this, we select a method with high probability values from the multinomial distribution expressing individual preferences, and asymmetrically optimize the Dirichlet distribution, which is the prior distribution of the polynomial distribution. Was used.

In the present invention, the user's personal preference is expressed in a polynomial distribution, and the Dirichlet distribution is defined by its leading probability. The main reason for setting the Dirichlet distribution as the leading probability of the polynomial distribution is that the change of each element in the polynomial distribution can be expressed as a parameter of the Dirichlet distribution, and the Dirichlet distribution is the conjugate prior to the polynomial distribution. This is because the same distribution has the advantage of having the efficiency of computation.

FIG. 2 is a data distribution diagram showing the relationship between a Dirichlet distribution and a polynomial distribution according to symmetrical and asymmetrical distributions. The dark color here means that the data is highly coherent. In FIG. 2, the first picture of the first line shows a uniform distribution form when the parameter values are all 1. FIG. In FIG. 2, the second and third pictures of the first line show a normal distribution shape when the parameter value is increased to one or more values. In FIG. 2, the first and second plots of the second row are data distributions in which asymmetric parameter values are given rather than symmetric parameters. The data distribution of the asymmetric shape is revealed from the normal distribution. If the parameter value is less than 1, the probability value for the random variable of each dimension takes a deterministic form.

데이터에 대해, K-차원 디리클레 분포는 다음과 같은 수식으로 정의된다.

For the data, the K-dimensional Dirichlet distribution is defined by the following formula.

[Equation 1]

Equation 1 is also simply expressed as follows.

&Quot; (2) "

&Quot; (3) "

)으로 표현한 것이다.Particularly, in Equation 3, the Dirichlet distribution is a vector m (expressing a concentration parameter α and a variable distribution for each dimension).

).

3 is a block diagram of a system for automatically recommending TV program content based on user preferences according to a first embodiment of the present invention. In order to configure the TV program recommendation system, in addition to the user management and recommendation server 400 directly associated with the recommendation, the TV user viewing record database 320, the TV content database 330, and the media streaming server 340 for broadcasting TV program content broadcasts are provided. Etc. are required. This embodiment shows one of the configurations of the hidden topic inference-based TV program automatic recommendation system, and the present invention is not limited to the specific form of this embodiment.

When the user applies for the TV program recommendation service and the user TV 310 is registered in the user management and recommendation server 400, the recommendation service of the present embodiment is started (S310). When the user TV 310 receives the TV content from the media streaming server 340 at a certain training period or the like (S320), the viewing record is transmitted to the TV user viewing history database 320 for the purpose of concealing the user's preference topic. It is stored (S330). The user management and recommendation server 400 receives a viewing record from the TV user viewing record database 320 and receives a TV content list from the TV content database 330 (S340) of the user management and recommendation server 400. Specific configurations and functions are described below.

4 is a block diagram of a user preference topic-based TV program content automatic recommendation device (user management and recommendation server) according to a second embodiment of the present invention. 5 is a simplified flowchart of a method for automatically recommending TV program content based on a user preference topic according to a third embodiment of the present invention. In the following, the second embodiment and the third embodiment will be described simultaneously based on the case where the apparatus of FIG. 4 implements the method of FIG. 5, but the present invention is different from FIG. And other methods may be implemented in various embodiments, such as when the apparatus of FIG.

The user management and recommendation server 400 first extracts and processes information about a user's preference (user profile) from the TV program viewing record (S510), and groups similar-viewing users based on the user's preference (S520). In addition, the subject model parameters of the recommendation model are learned (S530), the subject model parameters of the hidden subjects are mapped to individual preferences and popular preferences, and the hidden subjects are inferred (S540), subject model parameters for the diversity of individual preferences. Processing and additional reasoning of (S550). After excluding TV program content that is not available to viewers using information on TV content (S560), the candidate program for each user is presented using the subject model parameters and the like (S570 and S580). The following describes the specific method according to the flow.

First, the preference viewing item extracting unit 410 extracts information of the user viewing preference item based on the viewing item within a training period of users (S510).

This is a process for recommending a TV program by learning the user's preferences without requiring the user to prefer the TV program. In the system of the present exemplary embodiment, the TV user viewing history database 320 may be implemented by receiving information on viewing history from the user TV 310 side.

Second, the user grouping unit 420 groups the users by using the extracted user preference viewing information (S520).

6 is a conceptual diagram of user preference based grouping. The user 110 having priority membership in group 1 is a group of users who are less related to other groups among the group 1 user 120.

According to an embodiment, in the grouping step (S520), a dimension of a viewing TV content item may be reduced to use a Principle Component Analysis (PCA) technique, or a technique of reducing a preference vector based on information on a genre and a channel may be used. This is because the computational efficiency becomes higher than constructing the vector as a dimension for all items of TV content. In some embodiments, the membership of a group may be matched to multi membership to increase the recommendation performance. This is a generation model in the use of subject modeling, so that the number of documents (users) included in the model is somewhat increased. This is because it must be secured. Typical grouping algorithms include K-means clustering, fuzzy clustering, hierarchical Dirichlet process, and the like.

Third, the parameter learner 430 learns a subject model parameter for each group by using the information of the user preference viewing item (S530).

7 is a schematic of a symmetric α, symmetric β (SS prior) hyperparameter based LDA model. 8 is a schematic of an asymmetric α, symmetric β (AS prior) hyperparameter based LDA model.

The recommendation model of this embodiment is based on Blei's Latent Dirichlet Allocation (LDA) model, which is a classic of topic modeling. It is a document generation model that formalizes the writing process of the user. In other words, the user goes through a series of processes, such as selecting a topic and selecting a word suitable for the topic, and writing a paragraph composed of several paragraphs, which are the main variables of the document, the subject, the word's subject, and the word. Correspond to the main nodes of the graphical model and construct a probability-based Bayesian model.

)를 디리클레 분포의 선행 확률분포로 정의한 실시예이다.Here, each node represents a random variable, a line between nodes represents an influence relationship between random variables, and a square represents an iteration process. In particular, colored nodes represent observed data variables, and colored nodes represent hidden variables. 8 illustrates a gamma distribution for learning a prior Dirichlet distribution of parameter θ in learning asymmetric α Dirichlet distribution based on data.

) Is defined as a prior probability distribution of the Dirichlet distribution.

는 u 사용자가 시청한 TV 프로그램 콘텐츠의 개수를 나타낸다. ?K는 전체 은닉 주제를 표현하고, V는 전체 TV 프로그램의 개수를 의미한다.In the LDA schematic model, the mathematical notation is ? U represents the number of viewing users,

U represents the number of TV program contents watched by the user. K represents the total hidden theme, and V represents the total number of TV programs.

에 맞춰 TV 프로그램 w를 시청한 정보를 기반으로 한다. 생성 모델 관점에서 도 8의 비대칭 LDA 기반 TV 프로그램 시청 모델링은 아래 [표 1]과 같이 설계할 수 있다.The colored node w representing the observed data is preferred by each user.

According to the information watched TV program w. In view of the generation model, the asymmetric LDA-based TV program viewing modeling of FIG. 8 may be designed as shown in Table 1 below.

[Table 1]

The performance of various inference techniques according to the prior art has been demonstrated to show similar performance through hyperparameter optimization. Such inference techniques include Gibbs Sampling (GS), Collapsed Gibbs Sampling (CGS), Variational Bayesian Inference (VBI), and Collapsed Variational Bayesian Inference (CVBI). Among them, GS (Gibbs Sampling) is easier to infer than other inference techniques. Especially, the CGS (Collapsed Gibbs Sampling) method calculates marginalization of major random variables, thus minimizing the independence assumption between random variables. It has been proven to be more accurate than other reasoning techniques. Therefore, in the present embodiment, the CGS technique is used as an inference method, but any embodiment using other inference techniques such as GS (Gibbs Sampling), VBI (Variational Bayesian Inference), and CVBI (Collapsed Variational Bayesian Inference) can be used.

Bayesian probabilistic models aim to maximize the probability P (w). P (w) is defined as follows.

&Quot; (4) "

&Quot; (5) "

&Quot; (6) "

[Equation 7]

The expansion of the above expression coexists with z and θ, which are designated as random variables. Therefore, reasoning techniques such as VB (Variational Bayesian Inference) and MCMC (Markov Chain Monte Carlo) are required, and in this embodiment, reasoning is performed using CGS, which is one of MCMC methods.

To develop the equation in CGS, the first term on the right side of Equation 5 is integrated for θ and the second term for ф. At this time, the post probability is used by the conjugate prior distribution (conjugate prior) relationship with the property of having a Dirichlet distribution form. The first term is expressed as θ, and the second term is expressed as ф for the calculation results of the two terms, and the Gibbs Sampling method is applied to sample the hidden subject (z in FIGS. 7 and 8) for each viewing token for each user. The parameters θ and ф are calculated by the update process as in the following equation.

&Quot; (8) "

&Quot; (9) "

는 사용자의 현재 시청 토큰 i를 주제 할당 값에서 제외하는 것을 의미한다.

는 현재 TV 프로그램 w가 k 주제에 할당된 개수이다.

는 주제에 할당된 TV 프로그램의 개수를 의미하며, 이는 모든 w에 대해

를 합한 값이다.?

는 현재 사용자 u에게 주제가 할당된 횟수를 의미한다. ?

는 사용자에게 할당된 전체 주제의 개수를 의미하며, 이는 모든 주제 k에 대해

를 합한 값이다.Where u is the user, w is the TV program (index of the entire TV program), i is the user's viewing token, and k is the hidden subject index. In equations (8) and (9)

Means excluding the user's current viewing token i from the theme assignment value.

Is the number of TV programs w currently assigned to k topics.

Means the number of TV shows assigned to the topic, which for all w

Is the sum of?

Denotes the number of times the topic is assigned to the current user u. ?

Means the total number of topics assigned to the user, which for all topics k

Is the sum of

는 각 TV 프로그램 w가 갖는 k번째 은닉 주제에 대한 확률을 나타내는 모델 파라미터

의 예측 치이고,

는 각 사용자 u의 k번째 은닉 주제에 대한 확률을 나타내는 모델 파라미터

에 대한 예측치를 나타낸다. ?α와 β는 각각 ?θ와 ф 파라미터의 선행 확률(prior)로서 디리클레 분포이고, θ와 ф는 앞서 설명된 것과 같이 다항 분포이다. 일반적으로 α와 β는 대칭(symmetric)적으로 설정된 작은 값을 사용하여 왔다.therefore

Is a model parameter representing the probability for the kth hidden subject of each TV program w.

Is the predicted value of,

Is a model parameter representing the probability for each user u's kth hidden subject

Represents a prediction for. ? α and β are Dirichlet distributions as priorities of the? θ and ф parameters, respectively, and θ and ф are polynomial distributions as described above. In general, α and β have used small values set symmetrically.

In the process of optimizing or estimating asymmetric hyperparameters, Minka's method, which solves the maximum likelihood estimation of the Dirichlet distribution with a simple iterative scheme, or the gamma distribution for Minka's fixed point iteration, is an efficient computing operation. Various methods can be used, including Wallah's technique, which has been improved to make this possible. When Wallah's method is used, the equation for the asymmetric Dirichlet distribution parameter α is shown in Equation 10 below.

&Quot; (10) "

은

의 개수가 n개인 개수를,

은

의 개수가 n개인 개수를 의미한다.here

silver

Number of ns,

silver

Means the number of n's.

According to an exemplary embodiment, in relation to the subject model parameter learning step S530, the cast sampling process may proceed as follows for efficient sampling. First, after the burn-in process, the sampling is averaged, and while the sampling is averaged, the asymmetric Dirichlet distribution is simultaneously learned based on Equation 10. Finally, the inference starts with the z distribution generated as the average sampling result. You can add a course to learn asymmetry.

Fourth, the parameter mapping unit 440 maps the subject model parameters related to the hidden subject to individual preferences and popular preferences (S540).

According to an embodiment, the TV viewing modeling may be mapped to the document generation model. As described above in the LDA schematic model, U in FIG. 7 represents the total number of users. In document generation modeling, the U represents the total number of documents. In addition, the colored node w represents a token for a TV program watched by a user. In document generation modeling, a mapping means a word represented in a document, and a user is viewed as a document. A mapping configured to look at as words may be advantageous. Conversely, it is also possible to map each document to a TV program and each viewing token to a viewer.

According to an embodiment, public preference and personal preference may be mapped to subject modeling. In this case, when using LDA, a classical model of subject modeling, the parameter θ represented by the polynomial distribution for each subject in FIG. 7 is mapped to the personal preference, and Ф represented by the polynomial distribution for the subject of the whole word is mapped to the public preference. do.

Fifth, the parameter processing unit 450 undergoes subject model parameter processing and additional reasoning for diversity of individual preferences (S550).

As a computational method that can make use of the diversity of preferences for each user, a method is obtained by learning the Dirichlet distribution, which is the preceding probability distribution of the polynomial distribution represented by individual preferences, as an asymmetric parameter. FIG. 8 illustrates an embodiment in which a gamma distribution Г (a, b) is used as a prior probability of the Dirichlet distribution for the prior Dirichlet distribution learning of the parameter θ for the asymmetric Dirichlet distribution learning. According to the embodiment, only a certain number of upper probability element values for the individual hidden subject may be used as they are, and the remaining values may be processed to 0 to be used for further inference.

Sixth, the non-favorite viewing item filter 460 excludes the non-preferred viewing item of the user from the recommended candidate (S560). The filtering function may be performed to exclude content that is not available due to a service that has already been aired or is not subscribed to by the content recommendation candidate.

Seventh, the recommended candidate content arranging unit 470 sorts the recommended candidate content for each user by using the subject model parameter (S570).

Predicting the preferences for each user's content based on a model composed of products and linear sums that can take into account both public and individual preferences. In the case of grouping of multi-membership, it may be implemented to take the result of calculating the recommendation item for each user in the group with the largest membership.

Herein, if the individual preference variable is θ and the public preference variable is Ф, the score for the TV program w for the user U may be calculated as in Equation 11 or Equation 12 below, where K is the number of hidden subjects. Means).

&Quot; (11) "

&Quot; (12) "

Finally, the recommendation item presenter 480 presents the user preference TV program content recommendation item arranged through the process (S580).

The third embodiment of the method of the present invention and its modified methods can also be embodied in the form of a computer program stored on a computer readable recording medium.

In the embodiment of the present invention, the preference of individual subjects, which are the main variables of the model, and the public preference of each subject are expressed in polynomial distribution. In addition, in the suggestion of individual recommendation items, in order to reflect the recommendation centered on individual preferences rather than public preferences, a method of limiting individual preference topics to higher interest topics was used. In addition, since viewing interests of TV programs are largely different from individual to viewer, the asymmetric dirichlet distribution is applied to reveal the user's personal preference.

As a result of experimenting with one embodiment of the present invention, the results of meaningful recommendation performance experiments for individual recommendation were confirmed through a parameter-based ranking alignment model trained based on the corresponding methods. In addition, as a result of the topic modeling, it was confirmed that not only the TV programs of the genres or channels consistent under one topic but also the list of preferred watching TV programs not defined by the consistent genres or channels are included. Assuming that the concealed topic in TV program recommendation reflects the topic of interest of the user who watched the TV programs belonging to the TV program, the concealed topic reasoning-based TV program automatic recommendation method proposed by the present invention is the existing preferred genre and preferred channel. It was confirmed that they express the diversity of TV viewing preferences of users better than recommending TV programs.

The modules, functional blocks or means of the present embodiment may be implemented by various known devices such as an electronic circuit, an integrated circuit, and an ASIC (Application Specific Integrated Circuit), and they may be implemented separately or two or more may be integrated into one .

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is clearly understood that the same is by way of illustration and example only and is not to be construed as limited to the specific embodiments set forth herein; And can be replaced, modified and replaced. For example, the technique of the present invention can be applied to pictures, images, etc., which can be displayed by a display such as other LCDs instead of characters. Accordingly, it is intended to cover in the appended claims all such changes and modifications that fall within the true spirit and scope of the invention.

[310] your TV
[320] TV content database
[330] TV user viewing history database
[340] media streaming server
[400] User Management and Recommendation Server
[610] User with Priority Membership in Group 1
[620] Group 1 user

Claims (12)

A user grouping unit configured to set a user group using the preferred viewing information of a plurality of users;
A subject model parameter learner learning a topic model parameter for each user group using the preferred viewing information;
A parameter mapping unit for mapping hidden topic parameters among the subject model parameters to individual preferences and group preferences, and
Parameter processing unit that performs additional processing or inference to the subject model parameters to reveal the diversity of individual preferences
Automatic recommendation device based on user preference topic, including TV program content.
The method of claim 1,
The parameter mapping unit,
Mapping means the words that appear in the document in document generation modeling, mappings configured to view users as documents and view user viewing tokens as words in document generation modeling, each document to a TV program, and each viewing token to a viewer One of the following mappings, or one that consists of subject modeling of public and personal preferences.
Automatic recommendation device for TV program content based on user preference.
The method of claim 1,
The parameter processing unit,
In learning the Dirichlet distribution, which is the prior probability of the hidden subject parameter,
Process only by using a predetermined upper part of the preferred viewing information as it is and leaving the remaining lower part as 0,
Calculating the Dirichlet distribution as an asymmetric Dirichlet distribution
Automatic recommendation device for TV program content based on user preference.
The method of claim 1,
A non-preferred viewing item filter for excluding any one or more of the broadcasting end content or the content not subscribed to by the user using the preferred viewing information.
Automatically recommend the user based on the topic topic-based TV program content.
The method of claim 1,
Recommended candidate content sorting unit for sorting the recommended candidate content for each user by using the subject model parameter
Automatically recommend the user based on the topic topic-based TV program content.
The method according to any one of claims 1 to 5.
The mapping or the additional inference uses any one or more of a technique such as Gibbs Sampling (GS), Collapsed Gibbs Sampling (CGS), Variational Bayesian Inference (VBI), or Collapsed Variational Bayesian Inference (CVBI).
Automatic recommendation device for TV program content based on user preference.
A media streaming server for broadcasting TV content;
A user TV receiving a TV content recommendation list by transmitting a viewing record while receiving TV content from the media streaming server;
A TV user viewing record database for receiving and storing a viewing record from the user TV;
Set a user group using the preference viewing information generated from the viewing records received from the plurality of user TVs, and learn a topic model parameter for each user group using the preferred viewing information. Mapping hidden topic parameters among model parameters to individual preferences and group preferences, and performing additional processing or inference on the subject model parameters to reveal the diversity of individual preferences, thereby providing a TV content recommendation list to the user TV. User management and recommendation server for transmitting; And
A TV content database storing TV content to be transmitted to the media streaming server and the user management and recommendation server
User recommendation topic-based TV program content automatic recommendation system comprising a.
8. The method of claim 7,
The user management and recommendation server,
In learning the Dirichlet distribution, which is a prior probability of the hidden subject parameter, only a predetermined upper part of the preferred viewing information is used as it is and the remaining lower part is left as 0, or the Dirichlet distribution is asymmetric. , And
By using the preferred viewing information, one or more of the end of the broadcast content or content that the user TV is not subscribed to is excluded from the recommended candidates,
Sorting recommendation content for each user by using the subject model parameter
Automatic recommendation system based on user preference topic.
A user grouping step of setting a user group by using the preference viewing information of a plurality of users;
A topic model parameter learning step of learning a topic model parameter for each user group using the preference viewing information;
A parameter mapping step of mapping a hidden topic parameter among the subject model parameters to individual preferences and group preferences;
A parameter processing step of performing additional processing or inference on the subject model parameters to reveal the diversity of individual preferences;
A non-preferred viewing item filtering step of excluding any one or more of the broadcasting end contents or the contents which the user does not subscribe to using the preferred viewing information, as a candidate for recommendation, and
Suggested candidate content sorting step of sorting the recommended candidate content for each user by using the subject model parameter
Auto recommendation method based on user preference topic, including a.
10. The method of claim 9,
The parameter processing step,
In learning the Dirichlet distribution, which is the prior probability of the hidden subject parameter,
Process only by using a predetermined upper part of the preferred viewing information as it is and leaving the remaining lower part as 0,
Calculating the Dirichlet distribution as an asymmetric Dirichlet distribution
How to automatically recommend TV program content based on user preference.
A subject model parameter learning step of learning a topic model parameter for each user group set using the preference viewing information received from a plurality of users;
Mapping the hidden topic parameters among the subject model parameters to individual preferences and group preferences, mapping the meanings of words appearing in the document in document generation modeling, and viewing the user as a document in document generation modeling. A parameter mapping step of performing any one of a mapping configured to view as a word, a mapping of each document to a TV program, a mapping of each viewing token to a viewer, or a mapping of popular and personal preferences to subject modeling, and
In learning the Dirichlet distribution, which is a prior probability of the hidden subject parameter, only a predetermined upper part of the preferred viewing information is used as it is and the remaining lower part is set to 0, or the Dirichlet distribution is asymmetric. Parametric processing step of performing additional inference that reveals the diversity of individual preferences by processing to compute into a distribution
Auto recommendation method based on user preference topic, including a.
A program for executing the method according to any one of claims 9 to 11 is recorded.
A computer readable recording medium.

Images