KR20130083058A

KR20130083058A - System and method for recommendation service

Info

Publication number: KR20130083058A
Application number: KR1020110144775A
Authority: KR
Inventors: 최승진; 유지호
Original assignee: 포항공과대학교 산학협력단
Priority date: 2011-12-28
Filing date: 2011-12-28
Publication date: 2013-07-22

Abstract

The present invention provides an information recommendation service providing system and a method. In detail, the information recommendation service providing system includes an information input unit to which basic information about an object constituting a preference relationship is input, and a plurality of arranged in a row direction using the basic information. An input matrix generator configured to generate an input matrix that records preference information between a first object group including a first object and a second object group including a plurality of second objects arranged in a column direction, and the preference information of the input matrix is observed A component matrix extractor for extracting and generating a component matrix relating to a common intrinsic component corresponding to each population of the input matrix from a predetermined input matrix including at least one element, which is not included, and the element matrix is not observed from the component matrix It includes a preference prediction unit for calculating the preference information that does not.

Description

Information recommendation service provision system and method of providing the same {SYSTEM AND METHOD FOR RECOMMENDATION SERVICE}

The present invention relates to an information recommendation service providing system and a method of providing the same, and more particularly, to an information recommendation service providing system and a method of providing the same for preventing performance degradation of an information recommendation service in a situation where there is no preference information given in advance.

In a modern society where customers acquire, purchase or recommend information about products, the accessibility to customers' products or intangible services is greatly improved in that they can obtain various information. However, this has resulted in a new problem of finding out a myriad of information and products that customers need.

The online information recommendation system enhances the convenience and satisfaction of online transactions by automatically selecting and exposing products and services that are likely to be desired by customers in a situation in which a customer needs to select specific information from numerous information. .

Such an online recommendation system requires preference information for each customer and information about a product or service to be referred to, but recommendation information such as a new customer to the system and a newly registered product or service is required. In this case, since preference information cannot be obtained, it becomes difficult to perform the recommendation. The situation where there is no or lack of basic information for recommendation information for a new subscription or a new item is called a cold-start situation.

In the recommendation system, the information gathering process for the recommendation target is based on the user's active behavior, so in the general recommendation system where a large number of users are passive in the display of preference information, etc. Start situations frequently occur and degrade the overall recommendation performance. In addition, in the case of a product that is not frequently exposed to the user, there may also be a problem in that there is no chance of recommendation, and thus the preference remains not displayed.

Therefore, it is necessary to develop an application algorithm that can ensure the reliability of the recommendation system for goods or services by performing stable prediction accurately and operating under cold start conditions.

The present invention is to solve the above technical problem, to provide a system and method for recommending goods and services to increase the convenience and satisfaction of the purchase of goods or services in the online purchase process.

In particular, even if it is difficult to provide basic information on newly subscribed or registered users or target items or target items that have not been exposed to users for a long time, it is to provide a reliable recommendation system and method by predicting preferences stably and accurately. .

The technical objects to be achieved by the present invention are not limited to the above-mentioned technical problems, and other technical subjects which are not mentioned can be clearly understood by those skilled in the art from the description of the present invention .

According to an aspect of the present invention, there is provided an information recommendation service providing system. An information input unit to which basic information about an object constituting a preference relationship is input, and a plurality of agents arranged in a row direction using the basic information. An input matrix generator configured to generate an input matrix in which preference information is recorded between a first population including one entity and a second population including a plurality of second entities arranged in a column direction, and preference information of the input matrix is not observed A component matrix extractor for extracting and generating a component matrix about a common intrinsic component corresponding to each population of the input matrix from a predetermined input matrix including at least one element, and an unobserved element of the element from the component matrix It includes a preference prediction unit for calculating the preference information.

In this case, the component matrix arranges each population of the input matrix in a row (or column) direction, and arranges component parameters in a column (or row) direction that affect the preference result value of the input matrix. It may be a matrix representing preference information about a parameter.

According to an embodiment of the present invention, the component matrix may be a convergence value of a probability calculated by a learning modeling technique for learning a parameter for determining preference information of the input matrix using the predetermined input matrix.

In accordance with another aspect of the present invention, there is provided a system for providing an information recommendation service, including: an information input unit to which basic information about an object constituting a preference relationship is input, and an individual group arranged in rows and columns using the basic information; An input matrix generator for generating a plurality of input matrices recording preference information therebetween; a preference of the first input matrix in which preference information of at least one of a plurality of entities included in the population of the plurality of input matrices is not observed; For prediction, from the at least one second input matrix including any one of a first population and a second population constituting the first input matrix, from the plurality of input matrices to the first population or the second population; A component matrix extractor for extracting and generating a component matrix related to a corresponding common intrinsic component, and the extracted component Replacing a matrix with a component matrix about a common intrinsic component corresponding to the first or second population of the first input matrix, and calculating preference information for at least one unobserved entity of the first input matrix It includes a preference prediction unit.

Wherein the component matrix comprises a component matrix including preference information about the common intrinsic component of the first population extracted from a third input matrix including the first population and a fourth input matrix including the second population. It may include a component matrix including the preference information about the common intrinsic component of the extracted second population.

In the present invention, the component matrix may be a convergence value of a probability calculated by a learning modeling technique for learning a parameter for determining preference information of the second input matrix using the at least one second input matrix.

The apparatus may further include an information recommendation unit configured to complete a preference relationship between the first population group and the second population group using the preference information calculated by the preference prediction unit, and to generate and output recommendation information of the individual using the preference relationship.

In this case, the recommendation information may be generated by a collaborative filtering method for analyzing and recommending similar patterns based on the preference information according to the preference relationship.

The information recommendation service providing system may further include a storage including the basic information, the input matrix information, the component matrix information, and preference information calculated for the unobserved entity.

The storage unit may include an entity information storage unit in which the basic information is classified and stored for each entity, a relationship information storage unit for storing relationship information for assembling mutual relations between entities based on the basic information, the input matrix information, and the component. The matrix storage unit may store matrix information, and a result value storage unit configured to store the recommendation information generated based on the calculated preference information and the completed preference relationship based on the preference information.

The basic information may be tangible commodity information, intangible service information, cultural content information, and knowledge information, and is not particularly limited.

In accordance with an aspect of the present invention, there is provided a method for providing an information recommendation service, the method comprising: inputting basic information about an entity constituting a preference relationship, and establishing a relationship in which preference information is collected by combining the entities; Using the basic information and the setting relationship, recording preference information between a first population including a plurality of first entities arranged in a row direction and a second population including a plurality of second entities arranged in a column direction. Generating an input matrix, extracting and generating a component matrix about a common intrinsic component corresponding to each population of the input matrix from a predetermined input matrix including at least one element of the input matrix for which no preference information is observed; Calculating unobserved preference information of the element from the component matrix, and the acid By using the preference information to complete the preference relation between the predetermined input of the matrix, and includes the step of generating and outputting a recommendation of an object by using the affinity relationship.

In this case, the generating of the component matrix may include calculating an implicit component for determining the predetermined input matrix by a learning modeling technique and obtaining a probability value converged through the learning model.

Herein, the converged probability value may be calculated while updating the intrinsic component.

In accordance with another aspect of the present invention, there is provided a method of providing an information recommendation service, in which basic information about an entity constituting a preference relationship is input, and a combination of the entities establishes a relationship in which preference information is collected. Generating a plurality of input matrices in which preference information between the group of individuals arranged in rows and columns is recorded using the basic information and the set relationship; and at least one of a plurality of entities included in the population among the plurality of input matrices. At least one of the plurality of input matrices, including at least one of a first population and a second population constituting the first input matrix, for the prediction of the preference of the first input matrix in which the preference information on the individual is not observed Component matrix for common intrinsic components corresponding to the first or second population from a second input matrix Extracting and generating, replacing the extracted component matrix with a component matrix relating to a common intrinsic component corresponding to the first or second population of the first input matrix and not viewing the first input matrix. Calculating preference information for at least one entity, and completing the preference relationship of the first input matrix using the calculated preference information, and generating and outputting recommendation information of the entity using the preference relationship. Steps.

In this case, the generating of the component matrix may include calculating the implicit component for determining the second input matrix by a learning modeling technique and obtaining a probability value converged through the learning model.

Also, the converged probability value may be calculated while updating the intrinsic component.

The information recommendation service providing method of the present invention further includes storing the basic information, the input matrix information, the component matrix information, preference information calculated for the unobserved entity, and recommendation information.

The generating of the recommendation information may be generated by a collaborative filtering method of analyzing and recommending a similar pattern based on the completed preference information according to the preference relationship.

According to the present invention, it is possible to provide a system and method for recommending goods and services to increase the convenience and satisfaction of the purchase of goods or services in an online purchase process.

Particularly, according to the present invention, even in a cold start situation in which basic information on newly registered users or target items is difficult to be acquired, or in a situation where basic information on target items that are not exposed to users for a long time is difficult to be provided. A stable and meaningful recommendation result can be obtained.

1 is a block diagram of an information recommendation service providing system according to an exemplary embodiment.
FIG. 2 is a block diagram illustrating a configuration of a storage unit 111 according to the embodiment of FIG. 1.
3 is a diagram illustrating an example of matrix decomposition in a method for providing an information recommendation service according to an embodiment of the present invention.
4 is a diagram illustrating a process of analyzing a component matrix in a method for providing an information recommendation service according to an embodiment of the present invention.
FIG. 5 is a diagram illustrating a process of predicting a preference of an unobserved portion in an information recommendation service providing method according to an exemplary embodiment.
6 and 7 are diagrams illustrating a process of analyzing a component matrix in a cold start situation in an information recommendation service providing method according to an exemplary embodiment of the present invention.
8 is a diagram illustrating a process of predicting a preference in a cold start situation in a method for providing an information recommendation service according to an embodiment of the present invention.
9 is a flowchart illustrating a method of providing an information recommendation service according to an embodiment of the present invention.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings, which will be readily apparent to those skilled in the art to which the present invention pertains. The present invention may be embodied in many different forms and is not limited to the embodiments described herein.

In addition, in the various embodiments, components having the same configuration will be representatively described in the first embodiment using the same reference numerals, and in other embodiments, only the configuration different from the first embodiment will be described.

In order to clearly illustrate the present invention, parts not related to the description are omitted, and the same or similar components are denoted by the same reference numerals throughout the specification.

Throughout the specification, when a part is referred to as being "connected" to another part, it includes not only "directly connected" but also "electrically connected" with another part in between . Also, when an element is referred to as "comprising ", it means that it can include other elements as well, without departing from the other elements unless specifically stated otherwise.

1 is a block diagram of an information recommendation service providing system according to an exemplary embodiment.

The information recommendation service providing system 100 according to the embodiment of FIG. 1 includes an information input unit 101, an input matrix generator 103, a component matrix extractor 105, a preference predictor 107, and an information recommender 109. And a storage unit 111 for storing information in association with the components. 1 are not limited to these components since they are according to an exemplary embodiment, and components for performing necessary functions in the system for providing the information recommendation service of the present invention may be variously reduced or increased. .

The information input unit 101 receives basic information Idata necessary to provide a recommendation service according to the present invention. The basic information Idata is not particularly limited and may be input information about a user or a target for finding a user's preference for tangible or intangible goods or services, cultural contents, knowledge information, and the like. The basic information may be directly transmitted to a component for preference prediction, which will be introduced later, or may be formed as an information pool for preference prediction in the storage 111.

The input matrix generator 103 is a means for generating an input matrix to predict a preference required for recommending information based on the input basic information. The basic information may be directly transmitted through the information input unit 101 or information previously stored in the storage unit 111.

The input matrix is a matrix composed of a plurality of rows arranged in one direction and a plurality of columns arranged in another direction perpendicular to the one direction, and is a matrix configured to display a relationship between two elements (objects) of the basic information. . That is, the input matrix may be generated to display a variety of various relationships using the basic information.

Although the input matrix generator 103 may select two objects based on the basic information and generate an input matrix indicating a relation thereof, in some cases, the input matrix information itself is input and the storage 111 is input. The data may be stored in and extracted directly by the input matrix generator 103.

Specifically, as an example of the input matrix, an input matrix for a movie, which is one of user's preferred cultural information about a user element (object), may be used. That is, as recommendation service for recommending a movie to the user, relationship information for preference prediction of the user's movie may be disclosed as an input matrix.

After the input matrix generator 103 generates a combination of various input matrices, the corresponding input matrix information may be stored in the storage 111 again. On the other hand, the input matrices generated by the input matrix generator 103 are passed to the component matrix extractor 105 to extract components contained in the relation matrix of the corresponding input matrices as matrices as a process for preference prediction.

In other words, the component matrix extractor 105 of FIG. 1 extracts an embedded component matrix based on the input matrices transferred from the input matrix generator 103. Here, the component matrix is not data information about the input matrix itself, but extracts a latent factor included in or included in the input matrix to redisplay the relationship to the elements (objects) corresponding to the rows or columns of the input matrix. Say a matrix. That is, in the example of the movie preference of the user, based on an input matrix indicating a preference for the user (movie) of the user, that is, a parameter component such as movie style, genre, favorite director, favorite actor, etc. By extracting and reordering the user's preference relationship with the user, a matrix that rearranges the relationship with the user or the target for the extracted intrinsic component may be defined as the component matrix.

Similarly, the component matrices extracted from the input matrix by the component matrix extractor 105 may be stored in the storage 111.

On the other hand, the preference predictor 107 predicts the preference information from the component matrix with respect to the unidentified portion of the matrix of the input matrix.

The information recommendation service providing system of the present invention may use a collaborative filtering method as a technology for generating recommendation information. The collaborative filtering method uses a similar preference pattern based on preference information given to a target such as a product, a service, or a content by a user. Branches are a way of making recommendations by analyzing other users. A matrix factorization algorithm is used for such collaborative filtering, which is performed by extracting a component matrix about an underlying component and comparing and analyzing the input matrix, which records preference information of a user's product.

The preference predictor 107 may compare the component matrices extracted by the component matrix extractor 105 with each other and predict a preference value for an element corresponding to the corresponding row and column when there is a value not observed in the input matrix. .

In addition, if a value that is not observed in the input matrix exists for one user or a target, this is a cold start situation generated due to a lack of newly inputted basic information. Preference values can also be predicted for. As a technique for predicting the preference of elements of an unidentified input matrix under the input condition of general basic information or predicting the preference of the entire unidentified elements to some users or objects under a cold start situation, the preference predicting unit 107 is a method for predicting the preference of the matrix. Using a decomposition algorithm, a method of decomposing a specific matrix and a process of predicting preference of unidentified elements through the detailed matrix will be described in detail later.

The preference value predicted by the preference predictor 107 and information about the input matrix completed through the element are also transmitted to and stored in the storage 111.

Meanwhile, the information recommendation unit 109 completes an unidentified element of the input matrix using the result of the preference prediction, and then analyzes the preferences of other users having similar preference patterns based on the input matrix for the completed preferences to a predetermined target. Recommendation information may be generated. The recommendation information is provided to the users as output information Odata. In addition, the recommendation information is transmitted and stored in the storage unit 111, so that the storage unit 111 can secure an information pool for generating more accurate preference prediction and recommendation information based on gradually accumulated data. do.

2 is a block diagram illustrating a configuration of the storage 111 according to the embodiment of FIG. 1. As described in FIG. 1, the information recommendation service providing system of the present invention stores information generated by operating each component in the process of recommending information in the storage 111.

FIG. 2 includes a plurality of sub storage units classified by information stored through a process of providing an information recommendation service.

Specifically, referring to FIG. 2, the storage 111 may include an object information storage 201, a relationship information storage 203, a matrix storage 205, and a result value storage 207. The entity information storage unit 201, the relationship information storage unit 203, the matrix storage unit 205, and the result value storage unit 207 are provided with basic information for providing an information recommendation service, and preference based on the basic information. A sub-storage unit may generate a matrix in the relationship and classify and store the information output in the step of obtaining a result of preference and recommendation information for an unidentified element.

Since the sub storage unit is named for classifying storage spaces according to the stored information, the sub storage unit is not particularly limited to the sub storage unit and may be configured as various sub storage units according to the information content according to the present invention.

In particular, the object information storage unit 201 is a space in which basic information Idata, which is first inputted, is classified and stored in order to obtain recommendation information. For example, the user DB 2011, the target DB 2013, and the user information DB ( 2015), the target information DB 2017, and the like.

The user DB 2011 is a user database in which basic information about a user accessing a related site, such as online shopping, or a homepage for providing products, services, and various contents is input and stored.

The target DB 2013 is a target database in which basic information related to a target preferred by the user is input and stored. That is, it is a place where basic information about specific objects of each of various contents, products, and services for which the user wants to obtain recommendation information is stored.

The user information DB 2015 is a place in which additional information is stored so as to be able to understand the individual users inputted into the user database in detail. For example, the additional information about the user may include demographic information such as gender, age group, and residential area of the user.

The object information DB 2017 is a place where additional information which is specifically identified about an object input to the object database is stored. For example, the additional information about the object may include content information such as genre, actor, director, etc. of the product when the object product is one of cultural contents.

Additional information about a user or a target stored in the user information DB 2015 or the target information DB 2017 is extracted by an element matrix extractor 105 of FIG. 1 from an input matrix. Can be referred to as the intrinsic component when generating a component matrix. The information recommendation providing system of the present invention can stably predict the preference information by generating and learning the component matrix by using the additional information stored in the entity information storage unit 201 in a situation such as a cold start lacking the preference information. .

On the other hand, the relationship information storage unit 203 of the storage unit 111 may store the information by combining the relationship between each element (object) that can be extracted from the input basic information. That is, a user object and a target object may be combined to clean up a user's preference relationship with a target, or a user object and one of the additional information about the user may be combined to clean up a user's information relationship. The relationship information storage unit 203 may be stored.

In addition, the matrix storage unit 205 may store matrices of various relations generated by the input matrix generator 103 and the component matrix extractor 105 of FIG. 1.

The matrices stored in the matrix storage unit 205 may extract information and may be used as data for estimating the preference of the unidentified element or recommending the information in the preference predictor 107 and the information recommender 109.

The result value storage unit 207 is a space where the result value of the unidentified preference information between the unidentified entities generated by the information recommendation service providing system of the present invention is deduced and the information is stored. In addition, based on the result value of the preference information, information that can be substantially recommended from the corresponding target information is selected to the users, and the result value may also be stored.

3 is a diagram illustrating a concept of matrix decomposition in a method of providing an information recommendation service according to an embodiment of the present invention.

In FIG. 3, matrix decomposition is performed by extracting the preference information for an empty (unidentified) element when one input matrix X is formed by a relationship between two entities, and when preference information of one entity is recorded. It shows the rearrangement by matrix.

That is, the component matrix is reconstructed by using the component (U), which is the intrinsic information about one entity, in the input matrix Q as one axis, and the component (V), which is the intrinsic information about the other entity, is formed from the input matrix Q. By reconstructing the component matrix with respect to the axis, preference information of an unidentified element in the input matrix Q can be estimated.

Meanwhile, in FIG. 3, matrix co-decomposition means that when one input matrix Q is formed as a relation between two entities, and preference information about another entity of one entity is recorded, it is not empty (unidentified) element by element. When there is no preference information for one individual (ie, this case is defined as cold start), the concept for estimating a preference for an individual without such preference information is shown.

That is, when the preference for the entire element of the unidentified entity included in the input matrix Q is to be requested, the input matrices R and P respectively associated with the two entities included in the input matrix Q are used.

For example, if the object included in the input matrix Q is a predetermined user and a predetermined target, another input matrix R related to the user and another input matrix P related to the target are used.

After decomposing each of the input matrices Q, R, and P by using the inherent components at the same time, common components are extracted through these component matrices to estimate preference information related to an unidentified user or object.

That is, since the input matrix R is a matrix of a preference relationship between the user entity and the other entity (UI) with which the user is associated, when reconstructed into an implicit component matrix using the user matrix, the user component matrix included in the input matrix Q is included. Can be used to substitute. Similarly, since the input matrix P is a matrix of the affinity relationship between the target entity and other entities VI related to the target entity, when the input matrix P is reconstructed into an embedded component matrix using the target entity, the target component matrix included in the input matrix Q is included. Can be used to substitute.

This technique assumes that the component matrices for the two entities that can be extracted from the input matrix Q are substantially the same among each of the component matrices that can be extracted through the input matrices R, P. That is, in the above example, since only the intrinsic component matrix related to the user or object that can be extracted from the input matrix Q cannot be accurately estimated under the cold start situation, other input matrices R and P related to the user and the object are not included. Can be estimated from The other input matrix R may reconstruct the implicit component matrix related to the user and the implicit component matrix related to the user's additional information, and from the other input matrix P related to the implicit component matrix related to the object and the additional information of the object. It is possible to reconstruct the implicit component matrices. In this case, the intrinsic component matrix U related to the user may be used for preference estimation under the premise that the same extracted from the input matrix Q and the input matrix R, and similarly, the intrinsic component matrix V related to the object. May be used for preference estimation under the assumption that the extracted from the input matrix Q and the input matrix P are the same.

3 is a view illustrating the concept of FIG. 3 in detail.

4 illustrates analyzing a component matrix in a method of providing an information recommendation service according to an embodiment of the present invention, and FIG. 5 illustrates a process of predicting a preference of an unobserved part in FIG. 4.

4 is a diagram illustrating an analysis of a component matrix of an input matrix X that discloses preference information for each user and a subject.

Figure 4 shows each individual limited to six individuals. For the sake of understanding, the preference information corresponding to each element is shown in the dichotomous way of positive (dark) and negative (blur), and the unobserved part is left empty.

An information recommendation providing system according to an embodiment of the present disclosure receives basic information about a user entity and a target entity, and generates an input matrix X representing a relationship between the user and each entity based on the preference information. . In this case, an unobserved element exists because not all users may indicate a preference for the object. The present invention is to provide a system that can accurately output the recommended information by predicting the estimated preference value for the unobserved element.

In order to predict a preference value for an unobserved element, first, the input matrix X is extracted and reconstructed into a component matrix corresponding to each individual as shown in FIG. 4.

That is, a common component (limited to components 1 to 3 in the example of FIG. 4) inherent in a preference relationship between a user, which is an object of the input matrix X, and a target, which is another object, is extracted for each object and the common component. Reconstruct each of the component matrices U and V representing the relationship.

For example, if the target is a movie, the component matrix U and the component matrix V of the target may be extracted from the input matrix X in which the movie preference information for the user is disclosed, according to common components. Herein, the components 1 to 3 that are commonly embedded may be nested parameters that may be referred to in the preference information as information related in common to the user and the movie. For example, the information may be a parameter of a preference implied by the user's favorite movie information, that is, information such as a movie production region, a director, and an actor. In this case, the user extracts the user component matrix U by reconstructing the relationship between the preferences for each of the movie production regions, the directors, and the main actors corresponding to each of the internal components 1 to 3. In addition, the target component matrix V is also extracted for the target movie by reconstructing the relationship for each parameter of the film production region, the director, and the leading actor corresponding to the components 1 to 3. By extracting the component matrices U and V for the input matrix X in this manner, it is possible to measure the preference values for the unidentified elements present in the input matrix X.

FIG. 5 illustrates the measurement of an unidentified element of the input matrix X of FIG. 4, wherein (1, 5), (2, 2), (2, 4), ( User preference information is not described in the empty elements of 4,1), (4,6), (5,3), (6,2), and (6,3).

Referring to FIG. 5, for example, the preference information CQ of the user 2 (unidentified element (2,2)) of user 2 may be estimated in the input matrix X. That is, the combination of the components 1 through 3 recorded in the row of the user 2 in the user component matrix U and the components 1 through 3 recorded in the column of the object 2 in the target component matrix V are combined. The preference information CQ can be estimated. In the example of FIG. 5, User 2 is friendly (positive) to component 1 but negative for the remaining components 2 and 3, while subject 2 is closely positive for component 1, while for the other components 2 and 3 User 2 may be assumed to be preferred for subject 2, which is closely related to component 1 since it appears to be negative without. For example, the above-described movie preferences are described. When the user 2's preferences for movie 2 are not known, matrix decomposition is used. In other words, it can be seen that the user 2 prefers a Hollywood movie corresponding to component 1 from the user component matrix U and does not like the director A corresponding to component 2 and the actor B corresponding to component 3. In addition, the movie corresponding to the object 2 from the object component matrix V is identified positively closely to the Hollywood film of the component 1, and is related to the director A corresponding to the component 2 and the actor B corresponding to the component 3 Can be identified as negative. Then, since the target 2 movie is closely related to the Hollywood movie, it can be estimated that the user 2's preference for the Hollywood movie is positive.

Meanwhile, as another example, the preference of the user 1 of the target 5 is unconfirmed in the input matrix of FIG. 4, which may be estimated. Which intrinsic components are associated with "subject 5" is shown in the fifth column of the subject component matrix (V), which is positive for component 1 and component 3, and "subject 5" is a Hollywood movie and starring It can be seen that actor B is a movie. Thus, it is predicted that users who like Hollywood movies or movies starring actor B will have a high preference for "Target 5," and as a result, users who like Hollywood movies (user component matrix U) The first column, users 1 and 2, and the user who likes the movie starring actor B (the third column in the user component matrix (U), users 4 and 6) have a high preference for "Target 5". You can expect to have. The preference information for User 5, which is currently unidentified element, indicates that User 1 likes Hollywood movies, so it can be expected to have high preference.

In the example of FIG. 4 and FIG. 5, for convenience, dichotomous preferences of positive and negative are used for convenience. However, in the information recommendation service providing system according to another exemplary embodiment of the present invention, the preference is scored to predict the preference through a calculation process. Calculate the recommended information. 4 and 5, if the preference is expressed in the form of positive / negative, the matrix element is given in the form of OX, while in another embodiment, if the preference is given as a score, the input score is as it is. Becomes In the recommendation system according to an embodiment of the present invention, the preference prediction unit 107 converts a value into a numerical form in the form of + 1 / -1 when given in the form of OX, and if the score is given as a score, By setting the value, you can predict the preference through the calculation process.

6 and 7 are diagrams illustrating a process of analyzing a component matrix in a cold start situation in an information recommendation service providing method according to an exemplary embodiment of the present invention.

Referring to the example of FIG. 6, the input matrix X ′ in which the preference information for the user 2 and the preference information for the target 5 are not confirmed is shown.

The unidentified input of the user or the object constituting the unidentified row or column in the input matrix X 'is a new subscription or registration state and thus cannot obtain preference information or is not subject to a preference survey. Each object in the matrix causes difficulty in predicting preferences and generating recommendation information. 4 and 5, the preference values can be predicted from the extraction of the component matrix through learning for each of the unidentified elements. However, in the cold start situation, the preference information for each individual is missing, so that the user component matrix U 'is missing. ) And even to extract and reconstruct the target component matrix (V '), even if the component matrix is extracted, even in the component matrix, as shown in Figure 6, there is no preference information for the missing object, as shown in FIG. It is hard to predict.

Accordingly, an embodiment of the present invention may use the co-decomposition method of the matrix as shown in FIG. 7 by extending the matrix decomposition method disclosed in FIGS. 4 and 5.

Referring to FIG. 7, in the cold start situation in which the preference information is not confirmed for the user 2 and the target 5, the preference information for the user 2 and the target 5, which are unidentified entities in the input matrix Q, is extracted from the input matrix Q. Since it is difficult to predict only the user component matrix U and the target component matrix V, the other input matrix P and the input matrix R that can share the intrinsic components of the input matrix Q together Analyze

That is, as shown in FIG. 6, the user component matrix U and the target component matrix V extracted from the input matrix Q itself include elements of the unidentified user 2 and the target 5, and thus, are not identified. Since there is no information that can be extracted at all, and thus the preference cannot be predicted, the component matrix (U, T) extracted from the other input matrix (P) and the other component matrix (U, T) extracted from the other input matrix (R). ). In this case, the component matrix of one entity of the component matrix extracted from the other input matrix P and the component matrix extracted from the other input matrix R is extracted from the main input matrix Q to predict the preference. It may be assumed that they are the same as the user component matrix U and the target component matrix V, respectively. In other words, in the example of FIG. 7, the other input matrix P may be extracted and reconstructed into a target information component matrix W and a target component matrix V, and the other input matrix R may be a user component matrix ( U) and the user component matrix T may be extracted and reconstructed. Since the user component matrix U and the target component matrix V, which are common component matrices of the objects extracted from the input matrix Q, are the same, they can be used. Since the component matrices are extracted through a learning process with respect to a pool formed by a plurality of information collections, as the learning is repeated, the differences between the intrinsic components may gradually converge to become the same.

In this way, after obtaining the component matrices (U and V in the above example) extracted by the other input matrices (P, R), the individual component matrices are analyzed and each individual object is not observed in the original input matrix (Q). The preference information corresponding to the element of may be calculated. The method of predicting and calculating the preference value at this time has already been described above and will be omitted.

FIG. 8 is a diagram illustrating a process of predicting a preference in a cold start situation as shown in FIG. 7.

The input matrix Q is a matrix representing a preference relationship composed of a user row (a) and a target (movie in the above example) column (b), and in a cold start situation that does not include affinity information for some user rows and some target columns. Input matrix information of.

The user component matrix (U) and the target component matrix (V) cannot be used from the direct input matrix (Q) to calculate the preference prediction values for the unidentified some user rows and some target columns. The input matrix P between the user-user information, which can be commonly extracted, and the target-target information, which can commonly extract the target component matrix V, are used. That is, if the component matrix T for the user information such as the user component matrix U and the occupation is extracted and separated from the input matrix R between user-user information, the user component matrix U is an input matrix ( Since it is the same as the user component matrix U extracted from Q), it can be replaced to predict the preference information. Similarly, extracting and separating the component matrix (W) for the target information such as the target component matrix (V) and the movie genre from the input matrix (P) between the object and the target information, the target component matrix (V) is the input matrix. Since it is the same as the target component matrix V extracted from (Q), it can be replaced to predict the preference information.

However, in order to predict a preference value according to an embodiment of the present invention, first, a user component matrix and a target component matrix must be found from input matrices which are relationship data of each entity for a given preference. In the above example, for convenience of explanation, it is assumed that each implicit component has a specific meaning, but in reality, basic information is input and an input matrix is generated based on the extracted information, and extracted from these input matrix data. It is not known what the underlying ingredients become. That is, it is not possible to directly know the meaning of the n-th implicit component common to the n-th column of the user component matrix U and the n-th row of the target component matrix V. Therefore, the user component matrix and the target component matrix are found after assuming the approximate number of intrinsic components through an iterative learning process based on the input matrix data.

There are various methods for the learning, and the learning method is not particularly limited. In the simplest method, however, the observed value and the predicted value thereof are determined based on whether the observed value inside the data input matrix X is well predicted by the product of the user component matrix U and the target component matrix V. A method of transforming each component matrix in the direction of reducing the difference is used. In an embodiment of the present invention, a method of probabilistically modeling and solving such a relationship may be used.

Referring to FIG. 8, a method of finding a component matrix will be described mathematically.

The general matrix decomposition is to find two component matrices U and V for the input matrix Q (hereinafter referred to as X), where the result of multiplying two matrices (U, V) is similar to that of the input matrix (Q). The goal is to find the matrix by iteratively learning. When the position of each element of the input matrix is represented by (i, j) using the row index i and the column index j, such a relationship may be expressed by Equation 1 below.

(1)

Collaborative filtering using matrix decomposition first uses a preference matrix Q for the user's subject. Since the user only expresses preferences for a few of the objects, this matrix Q is mostly composed of unobserved values. The matrix decomposition of the preference matrix including the unobserved values may be expressed as Equation 2 when the index set of the observed values is expressed as O. FIG.

(2)

That is, when there are unobserved values in the input matrix, matrix decomposition is done using only the observed values. That is, the component matrix is found based on whether the product of the component matrices U and V becomes similar to the observed value of Q. 4 and 5 have already shown how to obtain a component matrix using only the observed values.

In the information recommendation providing system according to an embodiment of the present invention, an object of the present invention is to predict an unobserved preference value from a preference matrix used as an input matrix. Can be used to make predictions about unobserved values. Generally, unobserved values

The predicted value for can be obtained as follows.

(3)

That is, the preference prediction value for the j-th object of the user i is calculated by the product of the i-th row Ui of the user component matrix and the j-th row Vj of the target component matrix. 5 shows an example in which the preference prediction is performed in this manner.

On the other hand, co-decomposition of the matrix as in the example disclosed in FIG. 8 is a method of obtaining a decomposed matrix by using the two or more input matrices integrally. That is, if an additional user-user information matrix R (hereinafter referred to as Y in the formula) is given in addition to the user-target preference information matrix Q, the user components obtained from the user component matrices U and R obtained from Q are given. Assuming that the matrix U is the same, it can be expressed as follows.

(Equation 4)

Here, it can be seen that the matrix U corresponding to the user component matrix is used simultaneously to decompose two input matrices Q and R. When the plurality of input matrices have information corresponding to the common target "user", it is the core of the joint decomposition model to assume that the user component matrix U obtained by decomposing them is the same. As a result, in the conventional matrix decomposition, the component matrix U is associated with the input matrix Q as well as the R in the matrix decomposition, whereas the component matrix U is only associated with the input matrix Q. Therefore, the learning algorithm f can be expressed in the same manner as U = f (X) for the matrix decomposition algorithm, but in the same manner as U = f (X, Y) for the matrix decomposition algorithm of the matrix. Similarly, the component matrix V is associated with P as well as the input matrix Q, and through the joint decomposition algorithm of the matrix, the component matrix V can also be expressed as a dependent function of the input matrices Q and P. After constructing a joint decomposition model in this way and developing a learning algorithm, it is possible to perform learning by integrating information contained in various data matrices by learning a common component matrix by simultaneously utilizing a plurality of input matrices.

If user-occupational information (input matrix R) and object-genre information (input matrix P) are given in addition to user-target preference information (input matrix Q) as shown in FIG. 8, the entity here is a) user, b) object c. Jobs can be organized into four categories of d) and relationships can be organized into three categories: user-object (a, b), user-job (a, c), and object-genre (b, d). Since we have a data matrix for each relationship, we can write the input data matrix for the relationship (a, b) as X ^{(a, b)} , and the component matrix obtained by decomposing the It can be written as U ^(a) and V ^(b) respectively. As a result, if the set of all entities in a given entity-relationship model is denoted by E , and the set of all relations is denoted by R , then the co-decomposition model of the matrix can be written as

(5)

Bayesian learning may be used in an embodiment of the present invention as a method for making a learning algorithm based on the joint decomposition model of the matrix. Bayesian learning uses input values from the model

Parameter to determine the input value

When expressed in terms of parameters,

As a way of learning, we aim to calculate the posterior probability

(6)

Since it is not easy to know what form the post-probability usually has, it is necessary to set the likelihood and prior probability, which are easier to assume the form, and then use Bayes' rule. Calculate using The input value in the model

And the parameter to be

In this case, likelihood is expressed as a probability that an input value appears for a predetermined parameter value as follows.

(7)

The prior probability is a probability indicating prior knowledge of a parameter regardless of an input value and is expressed as follows.

(8)

After setting the likelihood and prior probabilities, the Bayes' law can be used to calculate the post-probabilities in the form of the following equation.

(9)

Taking a co-decomposition model of a matrix as an example, if user-target preference information Q and user-user information R are used as input matrices, and each is decomposed into the form X = UV ^T and Y = UT ^T while sharing the user component matrix U Here, the inputs are the input matrices Q (X) and R (Y), and the parameters are the component matrices U, V, and T (hereinafter expressed as W ). Therefore, likelihood, prior probability, and posterior probability can be expressed in the form of the following equation.

(10)

It is assumed here that the likelihood and prior probability are assumed to be normal distributions, and that each input value and prior probability value are distributed independently. The likelihood can be written as

(11)

here

And

Is a parameter representing the magnitude of an error that occurs when the input matrices Q (X) and R (Y) are represented by a matrix decomposition model. I, j are indices representing the rows and columns of Q (X), i, k is an index indicating a row and a column of R (Y). Since both matrices represent information about the user, the index i corresponding to the user can also be used. Also,

Means that data x follows a normal distribution with mean m and variance s. The prior probability can also be expressed as follows using a normal distribution.

(12)

In the above equation

,

Are parameters specifying the variance of each component matrix. Post-probability based on assumptions about these likelihoods and prior probabilities

Since it is impossible to calculate the integral located in the denominator in equation (8), it cannot be calculated as it is. The variational Bayesian method assumes that in order to calculate an approximate posterior probability in this situation, the posterior probability is expressed in the form as in the following equation.

(13)

Although it is impossible to calculate the overall posterior probability, it is possible to approximate the overall posterior probability by independently calculating the posterior probabilities corresponding to each component matrix based on this assumption. Deriving the algorithm based on the variable Bayesian method can be summarized as follows.

1. Initialize the target component matrices U ^(a) using random values. Also, the parameters of the probability model (

^{(a, b)} ,

^(a) is also initialized to the appropriate value. This step involves the initialization of the algorithm, where the variables to be initialized are the component matrices U ^(a) to be learned and the parameters used to define the normal distribution.

^{(a, b)} and

^(a) As the algorithm is executed, each variable is adjusted toward an appropriate value, so that initialization can be performed using random numbers. Except that

^{(a, b)} and

^(a) represents variance and must be initialized to a positive random number.

2. Compute the posterior probability q (U ^(a) ) for each component matrix with the following normal distribution:

The value of the component matrix can be obtained from the average of the normal distributions calculated as described above. That is, it is the part which calculates the posterior probability q (U ^(a) ) in ^a variable Bayesian. The posterior probability is the value of each row of the component matrix

Calculated in the form of a normal distribution for, the mean of the normal distribution

And dispersion

The posterior probabilities can be calculated by calculating <Used for calculation

> And <

^{The T} > term is the expected value for the current posterior probability, which can be found by calculating the mean and variance of the calculated normal distribution.

3. Update each parameter using the following update formula.

These steps are parameters

^{(a, b)} and

^is a part of learning ^(a) , where tr (A) is the sum of the diagonal elements of matrix A (tr (A) =

), And ddiag (A) means a matrix composed only of the diagonal elements of the matrix A. N ^{(a, b)} means the total number of observed values contained in the input matrix X ^{(a, b)} . I ^(a) means the number of objects a, and a value indicating the total number of users when a is a user.

4. Repeat steps 2 and 3 to obtain progressively appropriate convergence results.

5. Use the user component matrix and the target component matrix to calculate predictions for unobserved values. As part of performing the preference prediction, after the execution of the algorithm and calculating the posterior probabilities for the user component matrix and the target component matrix, the average value of each posterior probability is set to the value of each component matrix and then (

), A preference prediction value may be calculated in the same manner as in Equation 3 above.

9 is a flowchart illustrating a method of providing an information recommendation service according to an exemplary embodiment. In particular, FIG. 9 illustrates a process of predicting preference and outputting recommendation information in a cold start situation in which no preference information corresponding to a user entity or a target entity is observed.

First, entity and relationship information are input to the information recommendation service providing system according to an exemplary embodiment of the present invention (S1). Individual information may be obtained from basic information input by individual or integrated information between objects, and is not limited to the information acquisition method. In addition, the relationship information may be input as information representing a preference relationship in which preference information between entities is recorded, or the relationship information may be acquired by displaying preference information generated by a preference survey after the entity information is obtained first. The generation and acquisition forms of such entity information or relationship information are not particularly limited.

When the individual information is first acquired in the step S1, relationship information indicating a preference relationship may be generated as an input matrix for each relationship through a preference investigation (S2). In another embodiment, if the input matrix representing the preference relationship between individuals is acquired as the relationship information in step S1, the step S2 may be omitted.

The preference of the unobserved entity among the objects constituting the row or column among the relational input matrices obtained in the above process can be obtained by co-decomposing other input matrices which can potentially share the component matrix corresponding to the row or column. have. Therefore, the input matrix including the unobserved entity to be estimated and the other input matrix sharing the intrinsic component matrix are jointly decomposed by the components (S3). That is, in order to obtain the component matrix of the input matrix, the component matrix common to other input matrices is obtained. Such a component matrix can be obtained by calculating a posterior probability through an algorithm for learning the intrinsic component, that is, a parameter for determining the input matrix with respect to the input matrix. In order to calculate such a post probability through a learning model, a likelihood and a prior probability to easily obtain a result value are set (S4).

Next, the posterior probability for each component matrix is calculated using the set likelihood and prior probabilities (S5). The value of the component matrix can be calculated by calculating the posterior matrix as a normal distribution, as described through the above equations, and calculating the average of the calculated normal distributions.

Then update each parameter using an update formula (S6). Steps S5 and S6 may be repeated to learn.

Then, it is possible to gradually obtain an appropriate convergence value through iterative convergence, thereby calculating the component matrix of the individual (S7).

A common component matrix of the component matrix of the entity calculated in step S7, that is, the user component matrix including the entity information unobserved from the input matrix and the other input matrix that can replace the target component matrix can be obtained. Then, the unobserved preference prediction value is calculated using the common intrinsic component matrix calculated through the calculation of the other input matrix (S8).

By using the calculated preference prediction value, preference information of unobserved entity information of the original input matrix may be obtained in a cold start situation. Then, accurate recommendation information can be obtained even under a cold start situation using the completed input matrix, and can be output to the user (S9).

It is to be understood that both the foregoing general description and the following detailed description of the present invention are illustrative and explanatory only and are intended to be illustrative of the invention and are not to be construed as limiting the scope of the invention as defined by the appended claims. It is not. Therefore, those skilled in the art can readily select and substitute it. Those skilled in the art will also appreciate that some of the components described herein can be omitted without degrading performance or adding components to improve performance. In addition, those skilled in the art may change the order of the method steps described herein depending on the process environment or equipment. Therefore, the scope of the present invention should be determined by the appended claims and equivalents thereof, not by the embodiments described.

100: information recommendation service providing system 101: information input unit
103: input matrix generator 105: component matrix extractor
107: preference prediction unit 109: information recommendation unit
111: storage unit 201: object information storage unit
203: relationship information storage unit 205: matrix storage unit
207: result storage unit

Claims

An information input unit for inputting basic information about an object constituting a preference relationship,
Generating an input matrix in which preference information is recorded between a first population including a plurality of first entities arranged in a row direction and a second population including a plurality of second entities arranged in a column direction using the basic information; Input matrix generator,
A component matrix extracting unit for extracting and generating a component matrix relating to a common intrinsic component corresponding to each population group of the input matrix from a predetermined input matrix including at least one element in which no preference information is observed in the input matrix;
And a preference predicting unit for calculating unobserved preference information of the elements from the component matrix.

The method of claim 1,
The component matrix arranges each population of the input matrix in a row (or column) direction, and arranges component parameters in a column (or row) direction that affect the preference result value of the input matrix, thereby forming the component parameters of the population. Information recommendation service providing system, characterized in that the matrix representing the preference information for.

The method of claim 1,
And the component matrix is a convergence value of a probability calculated by a learning modeling technique for learning a parameter for determining preference information of the input matrix using the predetermined input matrix.

An information input unit for inputting basic information about an object constituting a preference relationship,
An input matrix generator for generating a plurality of input matrices in which preference information between the group of individuals arranged in rows and columns is recorded using the basic information;
The first input matrix is configured among the plurality of input matrices to predict the preference of the first input matrix in which preference information of at least one of the plurality of objects included in the population among the plurality of input matrices is not observed. A component matrix extracting unit configured to extract and generate a component matrix about a common intrinsic component corresponding to the first population or the second population from at least one second input matrix including any one of a first population and a second population , And
Replace the extracted component matrix with a component matrix relating to a common intrinsic component corresponding to the first population or the second population of the first input matrix and for at least one unobserved entity of the first input matrix. An information recommendation service providing system comprising a preference prediction unit for calculating preference information.

5. The method of claim 4,
The component matrix is extracted from a component matrix including preference information about the common intrinsic component of the first population extracted from a third input matrix including the first population and a fourth input matrix including the second population. And a component matrix including preference information about the common intrinsic component of the second population.

5. The method of claim 4,
The component matrix is a convergence value of a probability calculated by a learning modeling technique for learning a parameter for determining preference information of the second input matrix using the at least one second input matrix. .

The method according to claim 1 or 4,
The information recommendation unit further comprises an information recommendation unit for completing a preference relationship between the first and second population groups using the preference information calculated by the preference prediction unit and generating and outputting recommendation information of the individual using the preference relationship. Service delivery system.

8. The method of claim 7,
The recommendation information is an information recommendation service providing system, characterized in that it is generated by a collaborative filtering method for analyzing and recommending similar patterns based on the preference information according to the preference relationship.

The method according to claim 1 or 4,
And a storage unit including the basic information, the input matrix information, the component matrix information, and preference information calculated for the unobserved entity.

The method of claim 9,
Wherein,
An entity information storage unit in which the basic information is classified and stored for each entity;
A relationship information storage unit for storing relationship information for assembling mutual relationships between entities constructed based on the basic information;
A matrix storage unit for storing the input matrix information and the component matrix information, and
And a result value storage unit configured to store the recommendation information generated based on the calculated preference information and the preference relationship completed by the preference information.

The method according to claim 1 or 4,
And the basic information is tangible product information, intangible service information, cultural content information, and knowledge information.

Inputting basic information about an object constituting the affinity relationship,
Combining the entities to establish a relationship in which preference information is collected;
An input matrix in which preference information is recorded between a first population including a plurality of first entities arranged in a row direction and a second population including a plurality of second entities arranged in a column direction by using the basic information and a setting relationship Generating the;
Extracting and generating a component matrix about a common intrinsic component corresponding to each population of the input matrix from a predetermined input matrix including at least one element in which no preference information is observed in the input matrix;
Calculating unobserved preference information of the element from the component matrix, and
And using the calculated preference information to complete a preference relationship of the predetermined input matrix, and generating and outputting recommendation information of the entity using the preference relationship.

13. The method of claim 12,
Generating the component matrix,
And calculating the intrinsic component for determining the predetermined input matrix by a learning modeling technique and obtaining a probability value converged through the learning model.

The method of claim 13,
The converged probability value is calculated while updating the internal component.

Inputting basic information about an object constituting the affinity relationship,
Combining the entities to establish a relationship in which preference information is collected;
Generating a plurality of input matrices in which preference information between groups of individuals arranged in rows and columns is recorded using the basic information and setting relations;
The first input matrix is configured among the plurality of input matrices to predict the preference of the first input matrix in which preference information of at least one of the plurality of objects included in the population among the plurality of input matrices is not observed. Extracting and generating a component matrix of common intrinsic components corresponding to the first population or the second population from at least one second input matrix comprising any one of a first population and a second population;
Replace the extracted component matrix with a component matrix relating to a common intrinsic component corresponding to the first population or the second population of the first input matrix and for at least one unobserved entity of the first input matrix. Calculating preference information, and
And completing a preference relationship of the first input matrix by using the calculated preference information, and generating and outputting recommendation information of the entity using the preference relationship.

16. The method of claim 15,
Generating the component matrix,
And calculating the intrinsic component for determining the second input matrix by a learning modeling technique and obtaining a probability value converged through the learning model.

17. The method of claim 16,
The converged probability value is calculated while updating the internal component.

16. The method according to claim 12 or 15,
And storing the basic information, the input matrix information, the component matrix information, preference information calculated for the unobserved entity, and recommendation information.

16. The method according to claim 12 or 15,
Generating the recommendation information,
Method for providing an information recommendation service, characterized in that by generating a collaborative filtering method that analyzes and recommends similar patterns based on the completed preference information according to the preference relationship.