KR101450453B1 - Method and apparatus for recommending contents - Google Patents

Abstract

The present invention relates to a method and an apparatus for recommending document content suitable for each user in a digital content providing system with a recommendation function. According to an embodiment, the method for recommending document content and the apparatus for executing the same are provided, wherein the method for recommending the document content comprises the steps of: inputting information about a user and a document as input data; and generating a preference database based on the input data. The input data includes a set of documents, a set of users, a set of documents created by a user, a set of documents recommended by a user, and a set of words that appear in the whole document.

Description

TECHNICAL FIELD The present invention relates to a content recommendation method,

The present invention relates to a content recommendation method and apparatus, and more particularly, to a method and apparatus for recommending document contents suitable for each user in a digital content providing system having a recommendation function.

The digital content market is the largest market that has grown with the development of the Internet. With the growth of social networks, the market is expanding more rapidly. Typical digital contents include news, movies, and theses.

In recent years, various types of content providers providing digital contents including Internet portal companies have started offering services for recommending and providing customized contents according to the preferences of each user. One of the methods for recommending this content is a method using a stochastic model.

According to the probabilistic model, the user models the process of creating content or creating a summary or description of the content. At this time, the content created by the user is limited to the text-based content and is hereinafter referred to as "document ". Probabilistic modeling of documents is known, for example, in PLIFSI modeling presented in Thomas Hofmann's paper "Probabilistic Latent Indexing" (SIGIR 1999). The PLSI model is a method of automatically classifying documents based on a statistical latent class model. This technique is used in various fields such as information retrieval and filtering, natural language processing, and machine learning using documents.

However, when applying this PLSI model to the process of generating document contents, only the process of generating the content by the user can be modeled, and the process of voting the content created by the user can not be modeled. Therefore, if the user does not have the history of creating the content at all, if there is only recommendation to the contents of the other person, there is no way to analyze the user's preference and recommend other contents.

Therefore, there is a need for a method of calculating the preference of the user's content and recommending the appropriate content considering the modeling process from the process of creating the content to the process of the user voting the content of the other user.

According to an embodiment of the present invention, it is possible to provide a method and an apparatus for calculating a preference for a user's content and considering the content in consideration of modeling up to a process of generating content, as well as a process of voting for content of another user have.

Also, according to an embodiment of the present invention, it is possible to provide a method and apparatus for calculating a user's preference and recommending content through a history of voting for the content of another person even though the user has never created the content directly.

According to an embodiment of the present invention, there is provided a document content recommendation method comprising: inputting information about a document and a user as input data; And generating a preference database (DB) based on the input data, wherein the input data includes a set of documents, a set of users, a set of documents created by the user, a set of documents recommended by the user, And a set of words appearing in the document.

According to an embodiment of the present invention, there is provided a document content recommendation method comprising: calculating a preference score for each of a plurality of candidate documents based on a posterior probability distribution of each preference; And recommending a document to the user based on the preference score calculated for each candidate document.

According to another aspect of the present invention, there is provided a computer-readable recording medium having recorded thereon a program for causing a computer to perform the document content recommendation method.

According to the embodiment of the present invention, there is an advantage that the user can calculate the preference of the user's content and recommend the content by taking into account modeling from the process of creating content to voting on the content of another user.

Also, according to an embodiment of the present invention, there is an advantage that a user can calculate a user's preference and recommend contents through hysteresis in which the user has not directly written content but has already voted on the content of another person.

Also, according to an embodiment of the present invention, there is an advantage that users who vote on contents and contents of newly created contents can recommend to other users by using their preferences.

FIG. 1 is a flowchart illustrating a method of recommending content to a user according to an exemplary embodiment of the present invention.
2 is a flowchart showing an exemplary method of the learning algorithm in the learning step S10,
FIG. 3A illustrates exemplary subject preference probability distributions for illustrative purposes of the present invention,
FIG. 3B illustrates exemplary user preference probability distributions for illustrating the present invention,
FIG. 3C illustrates a probability distribution for a word to be selected for an exemplary subject to illustrate the present invention,
4 is a flow chart illustrating an exemplary method of recommendation algorithm in recommendation step S20,
5 is a diagram showing a document and a recommendation type according to the type of user,
6 is a block diagram illustrating an exemplary network configuration for content recommendation in accordance with one embodiment.

BRIEF DESCRIPTION OF THE DRAWINGS The above and other objects, features, and advantages of the present invention will become more readily apparent from the following description of preferred embodiments with reference to the accompanying drawings. However, the present invention is not limited to the embodiments described herein but may be embodied in other forms. Rather, the embodiments disclosed herein are provided so that the disclosure can be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

In the present specification, when an element is referred to as being on another element, it may be directly formed on another element, or a third element may be interposed therebetween. Further, in the drawings, the thickness of the components is exaggerated for an effective description of the technical content.

Where the terms first, second, etc. are used herein to describe components, these components should not be limited by such terms. These terms have only been used to distinguish one component from another. The embodiments described and exemplified herein also include their complementary embodiments.

In the present specification, the singular form includes plural forms unless otherwise specified in the specification. The terms "comprise" and / or "comprising" used in the specification do not exclude the presence or addition of one or more other elements.

Hereinafter, the present invention will be described in detail with reference to the drawings. In describing the specific embodiments below, various specific details have been set forth in order to explain the invention in greater detail and to assist in understanding it. However, it will be appreciated by those skilled in the art that the present invention may be understood by those skilled in the art without departing from such specific details. In some instances, it should be noted that portions of the invention that are well known in the description of the invention and not significantly related to the invention do not describe confusion in describing the present invention.

The present invention relates to a digital content providing system capable of identifying an individual through a user login or the like and capable of expressing a preference for a specific subject (for example, by voting or recommending a content created by a user or created by another person) To a method and apparatus for recommending content relating to a user's preferred topic in a system environment. Here, the content refers to text-based content at least partially including text, which is hereinafter also referred to as "document" or "document content". For example, the content corresponds to the content in the present embodiment, which is defined as a set of a plurality of words such as a news article, a product review, a blog post, a review on a movie or an art work, and the like.

In the content recommendation method and apparatus according to an exemplary embodiment of the present invention, each user's favorite content is analyzed based on past contents of his / her own content and other users' votes (or recommended) It is possible to recommend content that the user prefers among the new contents that each user has not yet touched. According to the present invention, even if the user has never created the content and only recommends the content of the other person, the user's preference can be calculated through the recommendation history to present the content preferred by the user.

FIG. 1 is a simplified diagram illustrating a method for recommending content to a user according to an embodiment of the present invention. Referring to FIG.

In one embodiment, the content recommendation method includes a learning step S10 for creating a preference database DB and a recommending step S20 for recommending a content to a user based on the generated preference DB.

In the illustrated embodiment, the learning step S10 may include inputting information about the document and the user as input data to the learning algorithm, and generating a preference database DB based on the input data.

The input data may include, for example, a set of documents, a set of users, a set of documents created by the user, a set of documents recommended by the user, and a set of words appearing in the entire document.

The generated preference DB may also include, for example, the subject preference of the document, the subject preference of the user, and the word preference of the subject. At this time, in the preferred embodiment, each preference is represented by a probability distribution. That is, the subject preference of a document is expressed by a probability distribution (θ) of probabilities of selecting a specific subject (z) in a specific document (d), and the user's subject preference is expressed by the probability that a particular user (u) Is represented by a probability distribution (γ), and the word preference of the subject can be expressed as a probability distribution (φ) of probability of selecting a particular word (w) for a particular subject (z).

After the preference DB, which is a set of preferences represented by the probability distributions, is generated by the learning algorithm and then the information of a specific user is input in the recommendation step S20, the recommendation algorithm determines the preference for a given plurality of documents Can be digitized based on the preference DB, and documents with high preference can be recommended to the user.

Hereinafter, a specific embodiment of the learning step S10 will be described in detail with reference to FIG. 2 and FIG. 2 is a flowchart showing an exemplary method of the learning algorithm in the learning step S10.

In step S110, a document creation probability, a document recommendation probability, and a word selection probability are calculated based on a dictionary (estimated) probability distribution of input data and preferences. The prior probability distribution of input data and preferences used here is as follows.

Input data

Here, the input data is a predetermined value determined as data necessary for generating the preference DB, and may include the following data.

D = {d ₁ , d ₂ , ... , dn}: A set of given documents

U = {u ₁ , u ₂ , ... , u _m }: a set of users

D (u _a ): set of documents created by user u _a

L (u _a ): _a set of documents read and recommended by user u _a

U (di): the set of users who recommended the document di

W (di): a list of words in the document di (bag-of-words)

Ni: Number of words in the document di

W = {w ₁ , w ₂ , ... , w _l }: a set of words appearing in the entire document (D)

Pre-probability distribution of preferences

The preference DB may include the subject preference of the document, the subject preference of the user, and the word preference of the subject, which are each defined by a probability distribution as follows.

θ (z | d) is the probability of choosing the subject (z) in document (d), which is the subject preference of document (d).

γ (z | u) is the probability that the user (u) will select the topic (z) and it means the subject preference of the user (u).

φ (w | z) is the probability of choosing the word (w) for the topic (z), and is the word preference of the topic (z).

In this regard, Figures 3a-3c illustrate an exemplary estimate (a dictionary) of each of the probability distribution (?) Of the subject preferences of the document, the probability distribution (?) Of the subject's preferences of the user, and the probability distribution ) Probability distribution.

Figure 3a is a prior probability distribution ([theta]) of the subject preferences of an exemplary document, where each document d is associated with a particular subject z when there are N documents and categorized into K subjects z ). ≪ / RTI > Referring to the table, the document (d1) is the probability to belong to the subject (z1) is 0.05. The subject probability to belong to the (z2) is 0.03 and is thus a document (d1) is the probability to belong to each subject (z1 to z _K) show And adds 1 to the probability for all subjects. Similarly, document (d2) to the document _(N d) is the probability to belong to each subject also each shown and adding a probability to belong to the entire subject for each document can be seen that each one.

FIG. 3B is a prior-art distribution of subject preferences γ of an exemplary user, where each user u has a document d when there are M users u and the subject z is categorized into K, It is a table showing the probability of choosing a specific subject (z). Referring to the table, for all users (u), a probability is assigned to each user to select each topic, and 1 is added to the probability of selecting the entire topic for each user.

FIG. 3C is a prior probability distribution (φ) in which a word is selected for an exemplary subject for explaining the present invention. When the subject (z) is classified into K and L words (w) Represents the probability that a particular word w will be included. Referring to the table, for all subjects (z), the probability that each topic will contain each word is assigned and becomes 1 when all the probabilities of including the whole word in each subject are all added.

It should be noted that the probability distribution of the preferences used in step S110 is an arbitrarily determined estimation value and not a value calculated based on actual input data. The probability distributions of the preference DB used in the step S110 are estimated probability distributions and then the posterior probability distributions of the preference DBs that best satisfy the actual input data are obtained through the following steps S120 and S130. That is, the present invention utilizes Bayes 'Theorem to derive posterior probability information using prior probability and observation values, and it is possible to use the Bayes' Theorem to derive the posterior probability information (θ, γ, φ) and observation value (input data), the actual probability distribution of the preference of the preference DB is obtained.

Referring again to FIG. 2, in step S110, the document creation probability, word selection probability, and document recommendation probability may be calculated as follows based on the prior probability distribution of input data and preference.

Document creation probability

The document creation probability Pa (u, d) means the probability that any user (u _a ) will create the document (d), and the document creation probability is the probability distribution (?) Of the topic preference of the document and the probability Is proportional to the similarity between the distributions ([gamma]). The document creation probability Pa (u, d) in one embodiment may be defined as Equation 1 as follows.

은 쿨백-라이블러 발산(Kullback-Leibler divergence) 함수로서 두 확률분포 γ(z|u)와 θ(z|d) 사이의 거리를 수치화하는 함수이다. 이 함수값이 크다면 두 확률분포의 거리가 큰 것을 의미하므로, 문서 작성확률은 이 함수에 반비례하도록 정의된다. Here, θ (z | d) is the probability distribution of the topic preference of the document, and γ (z | u) is the probability distribution of the user's topic preference. And, α is a constant, which is a constant indicating how many different kinds of documents the user will make different from his or her preference. The higher the value, the more users will create documents that are more similar to their topic preferences. ? can be determined according to a concrete implementation situation. And

Is a function that quantifies the distance between two probability distributions γ (z | u) and θ (z | d) as a Kullback-Leibler divergence function. If the value of this function is large, it means that the distance between two probability distributions is large. Therefore, the document creation probability is defined to be in inverse proportion to this function.

In the following embodiment, it is needless to say that the distance can be numerically calculated by using other functions according to the embodiment, although the Kullback-Leibler divergence function is used to quantify the distance between the two probability distributions.

Word selection probability

The word selection probability is a probability that the document d created by an arbitrary user u _a includes an arbitrary word w. For all documents, each document selects a topic (z) according to a probability distribution of θ (z | d), and the selected subject (z) , The word selection probability is proportional to the product of the probability distribution (?) Of the subject preference of the document and the probability distribution (?) Of the word preference of the subject. In one embodiment, the document creation probability is calculated for every document created by each user.

Document referral probability

The document recommendation probability Pv (u, d) means the probability that _a certain user (u _a ) recommends the document of the other person and the probability distribution (θ) of the topic preference of the document and the probability distribution (γ) And is inversely proportional to the number of referrals to the document. In one embodiment, the document recommendation probability Pv (u, d) may be defined by Equation 2 as follows.

(Z | d) is the probability distribution of the topic preference of the document, and γ (z | u) is the probability that the user's subject preference . And β is a constant and s represents the number of recommendation for the document (d). The larger the value of β / s, the more the user recommends documents that are similar to his or her topic preferences. Here, the denominator of the recommendation frequency for the document (d) is to consider that the greater the number of recommendation times, that is, the greater the number of documents exposed to a larger number of people, the higher the likelihood of recommendation even if they differ from their own topic preferences. KL is a Kullback-Leibler divergence function and quantifies the distance between two probability distributions θ (z | d) and γ (z | u).

Referring to FIG. 2, when the document creation probability, the document recommendation probability, and the word selection probability are calculated in step S110 as described above, in step S120, the probability that data is generated based on the calculated probabilities (Likelihood).

In one embodiment, given the prior probability distributions (?,?,?) Of preferences DB, the likelihood that data will be generated according to the above probability model, i.e., likelihood (L) .

Then, in step S130, posterior probability distributions (?,?,?) Of each preference for maximizing the likelihood are calculated. The probability distributions? (Z | d),? (Z | u), and? (W | z) of each preference of the preference DB such that the likelihood (L)

This process can use a known method known as the EM algorithm. EM algorithms are described in, for example, AP Dempster, NM Laird, and DB Rubin, "Maximum likelihood from incomplete data via EM algorithm" (Journal of Royal Statist. Soc., 39: 1-38, 1977), and a method of finding a probability distribution value when a likelihood value does not increase any more while performing E step and M step repeatedly is used as a solution.

In the present invention, the values to be calculated in the E step and M step according to the EM algorithm are expressed by the following equation (4).

Since the EM algorithm does not always guarantee that it finds the optimal value, it is desirable to run the EM algorithm multiple times to make the probability distribution with the greatest likelihood value the best.

The posterior probability distributions? (Z | d),? (Z | u), and? (W | z) of each preference of the preference DB can be obtained by performing steps S110 to S130, The posterior probability distribution preference DB may then be used in a step of recommending and recommending a document of a subject hypothetical to the user in the recommendation step (S20).

4 is a flow chart illustrating an exemplary method of recommendation algorithm in recommendation step S20.

In one illustrated embodiment, the recommendation algorithm allows the user to select one or more preferred documents among the recommended candidate documents (d1, d2, ... di) considered in the recommendation for a given user.

Referring to Fig. 4, the score (di (di)) for each of the recommended candidate documents d1, d2, ..., di is calculated by executing steps S210 to S250. At this time, step S230 calculates a preference score for each candidate document for a given user based on the posterior probability distribution of the preference DB. The number of recommended candidate documents is not particularly limited and can be determined according to the embodiment.

Thereafter, in step S260, the recommendation candidate documents are sorted based on the calculated preference score for each recommended candidate document, and in step S270, one or more recommended candidate documents having a high preference score are presented to the user.

The road in preference score calculating step in 4 (S230) preference score for each recommended candidate document is calculated by using the posterior probability distribution of the preference DB, a given user (u _a) and the recommended candidate document (di) of the information is preference DB If it exists, it can be calculated directly. In some cases, however, it may be necessary to recommend a document for a new user or to recommend a new document for an existing user. According to one embodiment of the present invention, as shown in FIG. 5, the recommendation type is classified into four cases according to whether the user information and the candidate document information are present in the preference DB, and a preference score is calculated for each recommendation type And how to do it.

FIG. 5 is a chart showing a recommendation type according to a document and a type of a user. Referring to FIG. 5, the following four types are classified according to the presence or absence of user information and document information.

- Warm-start recommendation: If the user to be recommended and the document to be recommended all exist in the preference DB created in the learning step (S10)

- Cold-start Recommendation type 1: If a new user who does not appear in the preference DB recommends documents existing in the preference DB

- Cold-start recommendation type 2: When a new document that does not appear in the preference DB is recommended to users existing in the preference DB

- Cold-start recommendation type 3: When a new user who does not appear in the preference DB recommends a new document not appearing in the preference DB

According to a preferred embodiment of the present invention, the calculation of the preference score according to each of these recommendation types is performed as follows.

Warm-start recommendations

Recommended target user (u _a) and documents (di) if that exists both in the preference DB generated in the learning phase (S10), the user (u _a) documents (di) Rating score is the topic of this article for the to be recommended Can be defined as a function proportional to the similarity between the probability distribution (θ) of the preferences and the probability distribution (γ) of the user's subject preferences. In one embodiment, this preference score may be expressed by Equation 5 below.

Here, score (di) is the preference score for the candidate document (di) of the user (u _a ), and γ (z | u _a ) is the probability distribution of the topic preference of the user (u _a ) di) is the probability distribution of the subject preference of the candidate document (di), and KL is the Kullback-Leibler divergence function.

Alternatively, however, it is of course also possible to use another function to quantify the distance between the probability distribution (?) Of the topic preferences of the document and the probability distribution (?) Of the user's topic preferences without using the above equation (5).

Cold-start Recommended type 1 person  Occation

As if the new user does not appear in the preference DB generated in the learning phase (S10) to recommend a document that exists in preference DB, the user (u _a) subject preference probability γ of | unknown (z u _a) not because calculates | (u _a z) user (u _a) a set of written documents (D (u _a)) and a set of a like article (L (u _a)) γ in the following formula 6 such as using a.

Where Z (u _a ) is a normalization value such that the sum of the preference distributions γ (z | u) for all subjects is 1. If y (z | u _a ) is calculated by Equation 6 as above, the preference score can be calculated by substituting this value into Equation 5.

Cold-start Recommended type 2 people  Occation

Since the user does not know the subject preference probability distribution θ (z | di) of the document di, it is possible to recommend a new document not present in the preference DB to the user existing in the preference DB generated in the learning step S10, based on the probability distributions of the words for the di (di), the probability distribution of the subject preference of the user who created the document di, and the probability distribution of the subject preference of the user who recommended the document (di) Calculate θ (z | di).

Here, Z (di) is a normalization value such that the sum of the preference distributions θ (z | d) for all subjects is 1, and u _j means a person who created the document di. When calculating? (Z | di) by Equation (7), the preference score can be calculated by substituting this value into Equation (5).

Cold-start Recommended type 3 people  Occation

As if the new user does not appear in the preference DB generated in the learning phase (S10) to recommend a new document that does not appear in the preference DB, the user (u _a) γ topic preferences probability distribution (z | u _a) and documents (di ) Is a case in which both of the subject preference probability distributions θ (z | di) are unknown. In this case, the preference score can be calculated by calculating the probability distribution? (Z | u _a ) and the probability distribution? (Z | di) using the above Equation 6 and Equation 7, have.

Through the above-described method, the preference score for each candidate document can be calculated not only when the user information and the document information are known but also when the user information and / or document information is not present in the preference DB.

6 is a block diagram illustrating an exemplary network configuration including a content recommendation device for content recommendation according to an embodiment.

Referring to FIG. 6, the content recommendation apparatus may include a server 30, an input data DB 50, and a preference DB 60, and may be connected to a plurality of user terminals 10 As shown in FIG.

The user terminal 10 may be, for example, a non-portable terminal, such as a portable mobile terminal, e.g., a smart phone, a tablet PC, a notebook computer, or a desktop computer.

The network 20 is any type of network that provides a data transmission / reception path between the portable terminal 10 and the server 30 and may include one of a LAN, a WAN, an Internet network, and / or a mobile communication network.

The server 30 may be a service server for providing contents to the user terminal 10 and may include a content recommendation application 40 capable of performing a learning algorithm and / . For this, the server 30 may include a processor, a memory, a storage unit, a communication unit, and the like. The content recommendation application 40 may be stored in a storage unit and loaded into a memory under the control of the processor.

In the illustrated embodiment, the server 30 is communicably connected to the input data DB 50 and the preference DB 60. [ In an alternative embodiment, the server 30 may include at least one of an input data DB 50 and a preference DB 60. The input data DB 50 may store input data necessary for generating the preference DB and the preference DB 60 may perform the pre-preference probability distribution before the learning step S10 and the learning step S10 The generated posterior preference probability distribution can be stored.

As described above, the content recommendation method and apparatus according to an embodiment of the present invention allow the user to analyze the user's preference by utilizing not only the content displayed in the digital content providing system but also the recommendation history of the content created by the other user, You can recommend interesting content. Therefore, even if there is no content created by the user, the preference propensity of the user can be analyzed through the history of the content recommended by the user, thereby making it possible to perform more accurate and appropriate content recommendation.

As described above, although the present invention has been described with reference to the limited embodiments and drawings, the present invention is not limited to the above embodiments. It will be understood by those skilled in the art that various changes and modifications may be made without departing from the spirit and scope of the present invention as defined by the appended claims. Therefore, the scope of the present invention should not be limited to the described embodiments, but should be determined by the equivalents of the claims, as well as the claims.

10: User terminal
20: Network
30: Server
40: Content recommendation application
50: input data DB
60: Preference DB

Claims (15)

A document content recommendation method,
The server receives information about a document and a user as input data; And
Wherein the server generates a preference database (DB) based on the input data,
The input data includes a set of documents, a set of users, a set of documents created by the user, a set of documents recommended by the user, and a set of words appearing in the entire document,
Wherein the preference DB includes a subject preference of a document, a subject preference of a user, and a word preference of a subject. delete The method according to claim 1,
The subject preference of the document is represented by a probability distribution (θ) of probability of selecting a subject (z) in the document (d), and the subject's preference of the user is a probability distribution of the probability y), and the word preference of the subject is represented by a probability distribution (?) of probabilities of selecting the word (w) for the subject (z). 4. The method of claim 3, wherein the generating the preference DB comprises:
(S110) the server calculating a document creation probability, a document recommendation probability, and a word selection probability based on the input data and the prior probability distribution of each preference;
(S120) calculating, by the server, a probability (likelihood) at which data is generated based on the calculated document creation probability, the document recommendation probability, and the word selection probability; And
(S130) of calculating a posterior probability distribution of each preference that maximizes the likelihood of the server. 5. The method of claim 4,
Wherein the document creation probability is proportional to a degree of similarity between a probability distribution (?) Of the subject preference of the document and a probability distribution (?) Of the subject preference of the user. 6. The method of claim 5,
The document creation probability is defined by the following equation,

Here, Pa (u, d) is the probability that the user u will create the document d, θ (z | d) is the probability distribution of the topic preference of the document, , A constant is a constant, and KL is a Kullback-Leibler divergence function. 5. The method of claim 4,
Wherein the document recommendation probability is in proportion to a degree of similarity between a probability distribution (?) Of the subject preference of the document and a probability distribution (?) Of the subject's preference of the user and inversely proportional to a recommended number of the document. Way. 8. The method of claim 7,
The document recommendation probability is defined by the following equation,

(Z | d) is the probability distribution of the topic preference of the document, and γ (z | u) is the probability that the user's subject preference , Β is a constant, s is a recommendation number for the document (d), and KL is a Kullback-Leibler divergence function. 5. The method of claim 4,
Wherein the word selection probability is proportional to a product of a probability distribution (?) Of the subject preferences of the document and a probability distribution (?) Of the word preferences of the subject. 5. The method of claim 4,
The server calculating a preference score for each of the plurality of candidate documents based on the posterior probability distribution of each preference; And
And recommending a document to the user based on the preference score calculated for each candidate document by the server. 11. The method of claim 10,
Wherein the preference score is proportional to a similarity between a probability distribution (?) Of the subject preferences of the document and a probability distribution (?) Of the subject preferences of the user. 12. The method of claim 11,
The preference score is defined by the following equation,

(Z | ua) is the probability distribution of the topic preference of the user (ua) that is the object of recommendation, and θ (z | di) is the candidate document (di | di), and KL is a Kullback-Leibler divergence function. 13. The method of claim 12,
Users of topics probability distribution _{(γ (z | u a)} ) of preference is the subject of the recommendation, if you do not know the user (u _a) is a collection of written documents (D (u _a)) and a set of recommended documents ( Wherein the probability distribution (? (Z | u _a )) is calculated based on L (u _a ). 13. The method of claim 12,
A probability distribution of words for the candidate document di when the probability distribution of the subject preferences di (? (Z | di)) is unknown, a subject preference of the user who created the document di (Z (di | di)) on the basis of a probability distribution of a user's recommendation and a probability distribution of a topic preference of a user who recommended the document di. A computer-readable recording medium having recorded thereon a program for causing a computer to execute the method according to any one of claims 1 to 14.

Images