KR101441983B1 - Apparatus and method for generating user profile - Google Patents

Abstract

The present invention relates to an apparatus and method for generating a user profile, in which a user profile generation apparatus performs learning by using a plurality of pieces of learning data as one learning unit, and calculates a weight value And a text profile generating unit for generating a text profile by calculating first class information for all the sets of texts by reflecting a weight vector on all the texts generated by the user .

Description

[0001] APPARATUS AND METHOD FOR GENERATING USER PROFILE [0002]

The present invention relates to an apparatus and method for generating a user profile.

As smartphones become widespread, various information, such as sensors, services, and applications, provided by smart phones, can be obtained. For example, it is possible to acquire the current location information of the user, the visit information of the user, and the movement information through the GPS sensor, and the social network service (SNS) such as Facebook, Twitter, ) To obtain information about the user from the text created by the user.

In recent years, researches have been introduced to provide customized information to users by analyzing various information about users. Existing customized services are generally based on user-entered profiles. However, when the user does not input the attributes such as the presence of love or the political tendency in particular, it is difficult to obtain the user profile related to the attribute, and in the case of the specific service, the user profile is not collected directly to protect the privacy.

Recently, studies have been introduced to identify the profile of a user with information generated by a user in a domain, such as a blog, a phone conversation document, or a movie review, for example. Korean Patent Laid-Open Publication No. 10-2012-0050255 (May 18, 2012) discloses a method and system for providing context-aware services based on user identity modeling. Conventional user-customized services utilize automation methods to identify user profiles utilizing a large amount of content, and these automation methods are generally based on machine learning methods. That is, first, the contents generated by the users are collected, and the contents are classified by a person, and a pattern for profile identification is searched from the contents using a machine learning method. Thereafter, when a new user is input, the user's profile is identified from the content created by the new user using the pattern thus constructed.

Incidentally, among the contents generated by the user, there are contents which are helpful for the profile identification, and some contents are not. For example, some texts written by men in their 30s on the SNS include texts related to men in their 30s, such as promotion or parenting, while there may be texts in their 30s that are not related to men in their daily lives. Existing machine learning based profile identification methods use both to identify profiles. That is, in the existing method, it is assumed that all the contents created by the user are related to the user's profile, but there is also text that is not related to the user in the contents actually created by the user.

Therefore, according to the existing method, since the contents related to the user profile are included in the learning and testing for identifying the user profile, the performance of the profile identification is degraded. In order to use only contents related to the user profile, it is possible to consider how to check whether the given contents are related to the user profile. However, the operation to distinguish whether the contents are not related to the current profile is very low in terms of efficiency, Misleading results can be propagated to the following results. In addition, since it is necessary to construct learning data by human classification in order to classify the relationship with the user profile, it takes a lot of time and cost.

Another problem with existing user profile identification schemes is that they do not take into account the correlation or mutual influence between profiles having different characteristics. That is, in the conventional user profile identification method, for example, the user identifies attributes such as local information using the information moved by the user, identifies attributes such as gender using the text generated by the user, Attributes are merely added to the level of the profiles generated. Accordingly, the user profile does not reflect all the information related to the specific attribute of the user inherent in each profile, or the importance of the different profiles is not considered. do.

It is an object of the present invention to provide an apparatus and method for generating a user profile capable of identifying a user profile only by reflecting contents related to the user.

It is another object of the present invention to provide an apparatus and method for generating a user profile that generates an accurate user profile by minimizing the influence of text that is not helpful in identifying a user profile or adversely affecting the user profile.

It is another object of the present invention to provide an apparatus and method for generating a user profile by integrating profiles having different characteristics for each attribute.

The problems to be solved by the present invention are not limited to the above-mentioned problems. Other technical subjects not mentioned will be apparent to those skilled in the art from the description below.

The apparatus for generating a user profile according to an aspect of the present invention performs learning by using a plurality of pieces of learning data as one learning unit and classifies the classification information including a weight vector indicating the influence of each learning data on classification of classes for each attribute A learning unit for calculating a learning result; And a text profile generator for generating a text profile by calculating the first class information for all the sets of texts by reflecting the weight vectors on all the texts generated by the user.

In one embodiment, the learning unit may calculate a hyperplane representing a criterion for classifying the plurality of pieces of learning data by the attribute vector and a weight vector using a support vector machine.

In one embodiment, the text profile generator reflects only a vector component corresponding to learning data corresponding to a support vector closest to the hyperplane among the weight vectors, and sets the hyperplane as a reference The first class information can be calculated.

In one embodiment, the attribute may include at least one of gender, age, region, occupation, political orientation, hobby, and interest.

In one embodiment, the text may include text generated by the user via a Social Network Service.

In an exemplary embodiment, the user profile generation device may include: a mobility profile generation unit that generates a mobility profile for the user by calculating second class information for the set of all the texts based on the movement information of the user; And a profile integrating unit for merging the text profile and the mobility profile to generate a user profile for the user.

In one embodiment, the profile integration unit may include a profile generation unit that reflects a first importance degree to the first class information and a second importance degree to the second class information to generate the user profile.

In one embodiment, the profile integrating unit updates the first importance degree based on a difference between the class information of the user profile and the actual class information of the user, and updates the first importance degree based on the difference between the class information of the user profile and the actual class information of the user The importance level updating unit may update the second importance level based on the first importance level.

According to another aspect of the present invention, there is provided an information processing apparatus including: a weighting vector indicating the influence of a plurality of pieces of learning data on classification of classes for each attribute; and classification information including a hyperplane representing a criterion for classifying the plurality of pieces of learning data, A storage unit for storing the data; And a text profile generating unit for generating a text profile for the other user by reflecting the weight vector on texts transmitted from the other user and calculating first class information for the set of texts, have.

In one embodiment, the text may include a Social Network Service text generated by the other user and transmitted to the terminal.

According to another aspect of the present invention, there is provided a classifier including a weight vector indicating the influence of a plurality of pieces of learning data on classification of a class for each attribute, and a hyperplane representing a criterion for classifying the plurality of pieces of learning data by the attribute A database for storing information; And a text profile generation unit for generating a text profile for the user by calculating the first class information for the set of texts by reflecting the weight vector on the texts transmitted from the user terminal Can be provided.

In one embodiment, the user profile creation server system receives a log file indicating a history of location information of the user terminal from the user terminal, and calculates second class information on the set of text based on the log file A mobility profile generator for generating a mobility profile for the user; And a profile integration unit for integrating the text profile and the mobility profile to generate a user profile for the user.

According to another aspect of the present invention, learning is performed using a plurality of pieces of learning data as one learning unit, and classification information including a weight vector indicating the influence of each piece of learning data on classification of each class of attributes is calculated ; And generating a text profile by calculating the first class information for all the sets of texts by reflecting the weight vector on all texts generated by the user.

In one embodiment, the step of calculating the classification information may include generating a hyperplane representing a criterion for classifying the plurality of learning data by the attribute using a support vector machine, and calculating a weight vector . ≪ / RTI >

In one embodiment, the step of generating the text profile may include a step of reflecting only vector components corresponding to learning data corresponding to a support vector closest to the hyperplane among the weight vectors, And calculating the first class information based on the hyperplane.

In one embodiment, the method of generating a user profile comprises: generating a mobility profile for the user by calculating second class information for a set of all the texts based on movement information of the user; And combining the text profile and the mobility profile to generate a user profile for the user.

In one embodiment, the step of generating the user profile may include the step of reflecting the first importance degree in the first class information and the second importance degree in the second class information to generate the user profile .

In one embodiment, the step of generating the user profile may include updating the first importance degree based on a difference between the class information of the user profile and the actual class information of the user, And updating the second importance degree based on the difference of the class information.

According to another aspect of the present invention, there is provided a computer-readable recording medium having recorded thereon a program for executing the method of generating a user profile.

According to an embodiment of the present invention, a user profile can be accurately reflected by only identifying contents related to a user and minimizing the influence of text that is not helpful in identifying a user profile.

In addition, according to the embodiment of the present invention, profiles having different characteristics can be integrated for each attribute to more accurately generate a user profile.

The effects of the present invention are not limited to the effects described above. Unless stated, the effects will be apparent to those skilled in the art from the description and the accompanying drawings.

1 is a configuration diagram of a user profile creation system 10 according to a first embodiment of the present invention.
2 is a flowchart of a method of generating a user profile according to an embodiment of the present invention.
3 is a configuration diagram of a user profile generating apparatus according to a second embodiment of the present invention.

Other advantages and features of the present invention and methods of achieving them will become apparent with reference to the embodiments described below in detail with reference to the accompanying drawings. The present invention may, however, be embodied in many different forms and should not be construed as being limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the concept of the invention to those skilled in the art. Is provided to fully convey the scope of the invention to those skilled in the art, and the invention is only defined by the scope of the claims. Although not defined, all terms (including technical or scientific terms) used herein have the same meaning as commonly accepted by the generic art in the prior art to which this invention belongs. Terms defined by generic dictionaries may be interpreted to have the same meaning as in the related art and / or in the text of this application, and may be conceptualized or overly formalized, even if not expressly defined herein I will not.

The user profile creation system according to the embodiment of the present invention performs learning by using a plurality of pieces of learning data as one learning unit to classify classification information including a weight vector indicating the influence of each piece of learning data on classification of each class . According to the embodiment of the present invention, a text profile is generated by calculating the class information for all the sets of text by reflecting the weight vector on all the texts generated by the user, so that the text that does not help identifying the user profile It is possible to generate a more accurate user profile.

1 is a configuration diagram of a user profile creation system 10 according to a first embodiment of the present invention. Referring to FIG. 1, a user profile creation system 10 according to a first embodiment of the present invention includes a plurality of user terminals 100, and a user profile creation server system 200. Each user terminal 100 is connected to a user profile creation server system 200 via a wired or wireless communication network. The user terminal 100 may be a mobile communication device such as a smart phone, a notebook, a PDA (Personal Digital Assistant), or a fixed communication device such as a desktop PC. In the case of the mobile communication device, the user profile can be generated by integrating the text profile and the mobility profile as described later. If the user terminal 100 is a fixed-type communication device, a user profile including only a text profile may be generated.

The user terminal 100 includes a user interface unit 110, a display unit 120, a communication unit 130, and a storage unit 140. In this case, The user interface unit 110 may be implemented as a keyboard, a mouse, or the like. As another example, in the case of a user terminal 100 such as a smart phone, the user interface unit 110 may be implemented in a touch screen manner by a graphic user interface or the like. The display unit 120 displays various contents to the user on the screen. The communication unit 130 may be connected to the user terminal 100 either between the user terminals 100 or between the user terminal 100 and the user profile creation server system 200 or between the user terminal 100 and another server system And performs communication. The storage unit 140 may store various contents, movement information of the user terminal 100, and the like. If the user terminal 100 is a mobile phone such as a smart phone, the user terminal 100 may further include a normal phone module (not shown) for performing a call function.

The user profile creation server system 200 includes a learning unit 210, a text profile creation unit 220, a mobility profile creation unit 230, a profile integration unit 240, and databases 250, 260 and 270. In one embodiment, the user profile creation server system 200 may be provided separately from a communication server system (not shown) connecting the communication between the user terminals 100, or may be implemented in an integrated manner with the communication server system have.

When the server system 200 is provided separately from the communication server system, the user profile creation server system 200 transmits SNS (Social Network System) texts transmitted between the user terminals 100 from the communication server system And stores it in a learning data database (DB) 250. Alternatively, when the user profile creation server system 200 is implemented in a form integrated with the communication server system, the SNS text may be directly received from the user terminal 100 and stored in the learning data DB 250.

Learning data DB 250 stores learning data for users whose actual profile information is known. The learning data DB 250 may store learning data for a plurality of users for each attribute such as, for example, sex, age, area, occupation, political disposition, hobbies, or area of interest. The learning data may be collected, for example, via an on-line or off-line survey, or may be collected based on information entered directly by the user when the user subscribes to the network.

The learning unit 210 generates classification information using learning data for users who know the user profile in advance. Embodiments of the present invention generate classification information in a manner different from conventional methods. The text profile generating unit 220 generates a text profile in a manner different from the conventional method using the classification information. Before describing the embodiment of the present invention, the conventional learning and user profile identification method will be described as follows.

Traditionally, individual instance learning methods have identified user profiles for individual texts created by the user, and thus failed to obtain accurate user profiles due to texts that either do not help identify the user profile or have adverse effects. That is, in the conventional profile identification model, when new texts created by the user are given, the class is estimated for each text and the user profile is identified by the most frequently classified class among the estimated classes. For example, when there are ten texts created by the user, the profile is identified for each of the ten texts, and then the results are combined to ultimately identify the user profile.

However, there are some things that user-written text do not affect or degrade performance in estimating the profile. When these are used in model construction, incorrect model parameters are selected and finally the performance of the profile identification model can be degraded. Accordingly, embodiments of the present invention propose a profile identification model based on multi-instance learning to solve the problems of the above-described prior art. According to the embodiment of the present invention, the profile is identified by considering the entire group of texts as one, rather than individually identifying the profile for each text created by the user.

To this end, the learning unit 210 according to the embodiment of the present invention performs learning using a plurality of pieces of learning data as one learning unit, and calculates a weight vector indicating the influence of each piece of learning data on classification of each class And generates classification information including information on the hyperplane. In one embodiment, the learning unit 210 uses a support vector machine (not shown) to classify a normal vector of a hyperplane representing a criterion for classifying a plurality of pieces of learning data by attribute, and a classification including a weight vector Information can be calculated. A more detailed process of calculating classification information will be described later with reference to FIG. The classification information calculated by the learning unit 210 may be stored in the classification information DB 260. [

The text profile generating unit 220 generates a text profile from the SNS texts created through a social network service such as Facebook or Twitter using the user terminal 100, You can identify the text profile. At this time, among the texts created by users, texts are included to help identify the user profile, but there may be texts that do not help or adversely affect the identification of the user profile.

The embodiment of the present invention generates a text profile by calculating first class information on a set of all texts by reflecting a weight vector on all texts generated by a user. According to the embodiment of the present invention, all the texts created by the user are regarded as one unit for profile identification, and the text profile is identified. Therefore, among the texts created by the user, the texts not related to the profile identification affect the identification . This means that the vector component of the learning data that does not correspond to the support vector close to the hyperplane serving as the reference for the profile identification is calculated as '0' among the components of the weight vector so that the influence of the learning data which does not affect the profile identification is minimized It is because.

That is, according to the embodiment of the present invention, the text profile generator 220 reflects only the vector components corresponding to the learning data corresponding to the support vector (Support Vector) closest to the hyperplane among the weight vectors of classification information, The first class information is calculated based on the hyperplane. Since other vector components corresponding to learning data other than the support vector have a value of, for example, '0', errors in user profile identification due to texts that do not affect class classification of the attribute are prevented.

The movement information DB 270 may store movement information such as a log file indicating the history of the position information of the user terminal 100 with respect to time. The movement information of the user may be received from the user terminal 100 or from a location tracking server system (not shown) capable of tracking the location of the user terminal 100, for example. The location information of the user terminal 100 may be obtained, for example, by a Global Positioning System (GPS) module (not shown). The GPS module measures the distance and time from the satellites of the user terminal 100 and can detect the location information of the user terminal 100 according to the triangular method.

The mobility profile generation unit 230 generates a mobility profile for the user by calculating second class information for all the sets of texts of the user based on the movement information of the user. In one embodiment, the mobility profile generation unit 230 extracts movement information visited from a base point at a base point, visit frequency of each base point, visit time period per base point, stay time per base point from GPS information received using a mobile device or the like, To identify the mobility profile.

The mobility profile generation unit 230 may generate a mobility profile using a k-Nearest Neighborhood (kNN) scheme, which is known as one of the machine learning methods, in order to automate and acquire movement information of many users. The mobility profile generating unit 230 may find k users having a pattern similar to the user mobility pattern through the kNN scheme, for example, and then identify the user profile from k user profiles.

The profile integration unit 240 integrates the text profile and the mobility profile to create a user profile. The profile integrating unit 240 combines the two models having different characteristics into one to finally identify the user profile. To this end, the profile integration unit 240 reflects the first importance degree to the first class information, which is the output value of the text profile generation unit 220, and the second importance degree information, which is the output value of the mobility profile generation unit 230, To generate a user profile.

In one embodiment, the profile integrating unit 240 may include a profile generating unit 241 and an importance updating unit 242. The importance updating unit 242 updates the first importance degree and the second importance degree through learning. In other words, the importance updating unit 242 updates the first importance based on the difference between the class information of the user profile and the actual class information of the user, and updates the importance of the second profile based on the difference between the class information of the user profile and the actual class information of the user. Update the importance. At this time, the profile integrating unit 240 always keeps the sum of the first importance degree and the second importance degree at a constant value, for example, '1'. After the importance is updated, the profile generator 241 finally generates a user profile using the updated first importance and second importance.

As a modified embodiment of the first embodiment of the present invention, the user profile creation server system 200 may not include the learning unit 210 and the learning data DB 250. That is, in a modified embodiment of the present invention, the user profile creation server system 200 can generate a text profile from the texts input from the user terminal 100 using the pre-established classification information DB 260 . At this time, the classification information DB 260 may be constructed in association with a database constructed in another server system or by receiving data from another server system.

Hereinafter, an operation and an operation of the apparatus for generating a user profile according to an embodiment of the present invention, and a method for generating a user profile according to an embodiment of the present invention will be described. 2 is a flowchart of a method of generating a user profile according to an embodiment of the present invention. 1 and 2, in step S21, the learning unit 210 performs learning by using a plurality of pieces of learning data as one learning unit, and indicates the influence of each piece of learning data on the classification of classes for each attribute A weight vector and a hyperplane. The learning unit 210 learns the learning data stored in the learning data DB 250 and generates classification information serving as a reference for identifying a new user profile, in order to identify a new user profile. The learning data stored in the learning data DB 250 can be expressed in the following Equation 1.

[Formula 1]

D = {(X ₁ , Y ₁ ), (X ₂ , Y ₂ ), ..., (X _k , Y _k )

X _i = {x _i1 , x _i2 , ..., x _il }

In Equation 1, D represents a set of all the texts (learning data) created by k users, and X _I (I = 1,2, ..., k) represents a set of texts represents _{a, Y I (I = 1,2,} ..., k) are in the class represents the actual information of each I-th user _{attribute, X I (I = 1,2,} ..., k) x I1, x _I2 , ..., _xIl represent the individual texts created by the I-th user. For example, Y _I may have a value of +1 or -1. If the I-th user of a particular attribute, such as gender is male, and Y _I = +1 is stored class value, when the gender is a woman, there is a Y _I = -1 for a class value can be stored.

A user profile creation method according to an embodiment of the present invention uses all texts created by a user as one learning unit. The learning unit 210 performs learning by searching only a text that is helpful for identification in a set of texts during learning, thereby minimizing the performance degradation from unrelated or mis-tagged texts of the identification of the user profile. The learning unit 210 may calculate classification information including a hyperplane and a weight vector indicating a criterion for classifying a plurality of pieces of learning data by attributes, for example, using a support vector machine (Support Vector Machine) . The learning unit 210 may generate classification information according to the following Equation 2 using, for example, a multi-instance support vector machine (MISVM).

[Formula 2]

In Equation 2, x _i is an _i ( _i = 1, 2,..., I) existing in an I- , ..., l) represents a second instance (instance), Y _I represents the I-th user's profile (class value), ε _I represents a slack variable (slack variable), w is the normal to the hyperplane (hyperplane) denotes a vector (normal vector), ∥w∥ represents the magnitude of the normal vector of the hyperplane, <w, x _i> denotes the inner product (dot product) of the w and x _i, C is directly and learning data and model characteristics Represents a constant determined as an optimal value experimentally and not related, and b represents a shift amount from the center of the hyperplane.

The learning unit 210 can calculate classification information including the normal vector w of the hyperplane, the deviation amount b, and the slack variable? _I through the optimization of Equation (2). The learning unit 210 may calculate classification information using, for example, Mixed Integer Programming, which is known as one of optimization methods.

Next, in step S22, the text profile generation unit 220 generates a text profile, and in step S23, the mobility profile generation unit 230 generates a mobility profile. Although the embodiment of FIG. 2 shows that step S23 is performed after the execution of step S22, step S23 may be performed before or simultaneously with step S22.

In step S22, the text profile generation unit 220 generates the text profile by calculating the first class information for all the sets of text by reflecting the weight vector on all the texts generated by the user. The weight vector represents information for distinguishing whether each learning data is a support vector or not, and the weight vector can be calculated from the hyperplane generated in step S21. In one embodiment, the learning data closest to the hyperplane is determined as a support vector. In one embodiment, the text profile generator 220 reflects only the vector component corresponding to the learning data corresponding to the support vector (nearest to the hyperplane) among the components of the weight vector, The first class information can be calculated on the basis of the first class information.

가 주어지면, 텍스트 프로파일 생성부(220)는 사용자의 프로파일을 식별하기 위하여, 예를 들어 아래의 식 3에 따라 텍스트 프로파일을 식별할 수 있다. 이때, 새로운 사용자의 데이터는 n개의 텍스트들(x₁,x₂,..x_n)을 포함한다.Data for new users

The text profile creation unit 220 can identify the text profile according to the following Equation 3, for example, in order to identify the user's profile. At this time, the data of the new user includes n texts (x ₁ , x ₂ , .. x _n ).

[Formula 3]

The normal vector w of the hyperplane obtained in Eq. (2) can be expressed as Equation (3). According to Equation 3, the text profile generation unit 220 reflects the data X of the new user in the normal vector w of the hyperplane, and calculates class information in the form of "f (X) = wX + b". At this time, the text profile creation unit 220 compares all the data used for the learning with the data created by the user. The class information generated by the text profile generation unit 220 may be used to generate a user profile in a final state by being integrated with the class information generated by the mobility profile generation unit 230 as described later. Alternatively, the text profile generation unit 220 may assign a class of +1 if the output class value is greater than 0, and a class of -1 otherwise.

The text profile generating unit 220 generates a text profile using the weight vector information a _j of the normal vector w of the hyperplane. At this time, the text profile generating unit 220 generates data that does not directly affect the classification of the learning data is not the data) is calculated, the value of a _j to '0'. Therefore, only the data (learning data corresponding to the support vector in the SVM) that substantially affects the classification is compared with the text data of the new user to generate the text profile, so that the user's profile can be efficiently identified. In addition, according to the embodiment of the present invention, since the class value is calculated once and collectively for all the texts, as in Equation 3, instead of synthesizing the individual text after class estimation, You can immediately identify the text profile for the new user.

The text used to identify the user profile can be obtained by expressing the content created by the user in the quality. Since the SNS text is short in length unlike the existing document, and the emoticon, the alphabet repetition, and the like are frequently used, the qualities can be defined as shown in Table 1 below to reflect the characteristics of the SNS text.

endowment Contents Bag-of-Word Noun, unregistered word, symbol Presence of emoticons ㅡ ㅡ, ㅠㅠ, ㅜㅜ, T_T, ㅇㅁㅇ, ㅇㅅㅇ, ^^, ^ _ ^,
1 if it contains :), :(, ♡, ..., or 0 if it contains Whether the same character, symbol repeat 1 if it matches with [ㅋ | ㅎ | ㄷ |. | | | | | ~] +, Or 0

As shown in Table 1, the attributes such as the bag-of-word, the emoticon characteristic of the social text, and the repetition of characters and symbols, which are widely used in document classification, are defined and used. For example, after a given text is morphed, a back-word is extracted from the qualities using only nouns, unrecognized words and symbols, and only the extracted back-word is used as text for identifying the text profile.

In step S23, the mobility profile generation unit 230 generates the mobility profile for the user by calculating the second class information on the set of all the texts based on the movement information of the user. The mobility information that the user has moved can play an important role in identifying the user profile. For example, suppose that there is movement information of a user who frequently goes to a beauty salon and stays for a long time. Based on this movement information, it can be seen that the user's gender is female. As another example, if there is movement information of a user returning to school in the morning in the evening, school to school in the evening, and user returning from the school at night, it can be assumed that the user is 10 or 20 years old. If the user's GPS information is given by the smartphone, the mobility profile generator 230 may extract four pieces of information, for example, as shown in Table 2 below.

Information example Frequency of visit by base Nail Shop: 5 times, PC Room: 10 times, ... Visit time by base School: 1 am and 3 pm
PC room: 2 PM, 4 PM, 3 PM, ... Time spent by base Beauty salon: 200 minutes, PC room: 360 minutes, ... Number of movements between bases Home -> School: 3 times, School -> School: 3 times
Home -> School: 3 times, school -> Company: 3 times ...

The mobility profile generation unit 230 may express the user information from the GPS information into the n-dimensional vector, and then use the kNN algorithm to identify the user's mobility profile. That is, it is possible to find k users similar to the movement information of a given user, and to identify the user's mobility profile from k user profiles. In one embodiment, the mobility profile generator 230 may use a cosine similarity function to find similar users. After finding k users like this, the profile can be identified by Majority voting to finally identify the mobility profile. However, the manner in which the mobility profile is generated is not limited thereto, and a mobility profile can be generated in various manners known in the technical field of the present invention.

In step S24, the profile integration unit 240 integrates the text profile and the mobility profile to create a user profile for the user. In one embodiment, the profile integrating unit 240 may reflect the first importance degree in the first class information and reflect the second importance degree in the second class information, thereby generating a user profile.

Since the profiles generated from the movement information and the text respectively reflect different characteristics, the profile integration unit 240 integrates these results to finally identify the user profile. At this time, since the model considered to be important for each class (sex, age, teenage, twenties, etc.) is different, the profile integrating unit 240 can reflect the different weights for each model for each class. For example, when identifying gender, the mobility model and the text model are similarly reflected, whereas when identifying the user's political tendencies, the text model should be considered more important.

Accordingly, the profile integrating unit 240 can update the first importance degree and the second importance degree. In one embodiment, the importance updating unit 242 updates the first importance based on the difference between the class information of the user profile and the actual class information of the user, and based on the difference between the class information of the user profile and the actual class information of the user The second importance degree can be updated. Given the user x _i , the profile integration unit 240 may determine the user's final profile, for example, according to Equation 4 through Equation 5 below.

[Formula 4]

[Formula 5]

wTv + wMv = 1

In Equations 4 to 5, y _i ^* denotes a user profile, V denotes a set of classes, and v denotes an element or class of V. For example, if the profile to be identified is gender, V is {+1 (male), -1 (female)}. Text _v (x _i ) represents the result value from the text profile model for class v, and Mobility _v (x _i ) represents the result value from the mobility profile model for class v. w _Tv represents the importance of the text profile model for class v, and w _Mv represents the importance of the mobility profile model for class v. The sum of the importance of the mobility profile model and the importance of the text profile model is always maintained at '1' according to Equation (5).

For example, in the case of identifying a political tendency, since the user's political tendency is not well understood in the mobility, the importance value w _Mv is determined to be a relatively small value, for example, 0.1. On the other hand, since the user's political tendency is well received in the text, the importance value w _Tv is determined to be a relatively large value, for example, 0.9. For example, when outputting the attribute of the user as a complement in the mobility profile model and outputting the attribute of the user as progress in the text profile model, the weight of the text profile in the attribute of the political tendency is estimated to be higher, It can be identified as a progressive tendency.

At this time, the importance value may be arbitrarily set, but it is more preferable that the importance value is automatically estimated using learning data as described later. In one embodiment, the importance updating unit 242 can automatically learn the importance of the two models according to the perceptron among the linear models using the learning data. The importance updating unit 242 can update the importance according to the following Equations 6 to 8, for example.

[Formula 6]

[Equation 7]

[Equation 8]

In Equations 6 to 8, y _iv represents the actual class of user i for class v, y _iv ^* represents the class estimated in the model, E represents the actual class of user i for class v, Where w _Tv represents the importance of the text profile model for class v, w _Mv represents the importance of the mobility profile model for class v, and λ represents the rate of learning. , Text _v (x _i ) represents the result value derived from the text profile model for class v, and Mobility _v (x _i ) represents the result value derived from the mobility profile model for class v.

The importance updating unit 242 can update (update) the importance wTv of the text profile model and the importance wMv of the mobility profile model by performing iterative learning to minimize the error of Equation (6) according to Equation 7 and Equation (8). That is, the importance updating unit 242 can update the importance by multiplying the error of the class of the user i and the error of the class estimated for the user i in each model by the result value of the current text or profile model and the learning ratio. The importance update can be repeatedly performed for the learning user and can be repeated until there is no longer a change in the importance of the model or the difference between the existing importance and the currently updated importance becomes a specific value, have.

When the importance level is updated by the importance level update unit 242, the profile generation unit 241 may finally generate the user profile according to the above-described Equation 4, reflecting the updated importance level. The user profile generated by the user profile creation system 10 may be utilized as data for providing customized contents, advertisements, etc. to the user, for example.

The user profile creation method according to the embodiment shown in FIG. 2 can be implemented in a general-purpose digital computer that can be created as a program that can be executed by a computer and operates the program using a computer-readable recording medium have. The computer readable recording medium may be a volatile memory such as SRAM (Static RAM), DRAM (Dynamic RAM), SDRAM (Synchronous DRAM), ROM (Read Only Memory), PROM (Programmable ROM), EPROM Volatile memory such as an electrically erasable programmable read-only memory (EEPROM), a flash memory device, a phase-change RAM (PRAM), a magnetic RAM (MRAM), a resistive RAM (RRAM), a ferroelectric RAM (FRAM) But is not limited thereto. The computer-readable recording medium may be, but not limited to, a floppy disk, a hard disk, or an optical storage medium such as a CD-ROM, a DVD, or the like.

3 is a configuration diagram of a user profile generating apparatus according to a second embodiment of the present invention. Referring to FIG. 3, a user profile generation apparatus 20 according to a second embodiment of the present invention includes a user terminal 100, a counterpart user terminal 300, and a server system 400. 3, the user interface unit 110, the display unit 120, the communication unit 130, and the storage unit 140 of the user terminal 100, and the learning data DB 140 of the server system 400, The learning unit 420, and the classification information DB 430 are the same as those shown in FIG. 1, so duplicate descriptions are omitted.

3, the text profile generation unit 150 is provided in the user terminal 100, not the server system 400, and the storage unit 140 of the user terminal 100 includes a weight vector Classification information including hyperplanes is stored. As described in the first embodiment shown in Fig. 1, the weight vector indicates the influence of a plurality of pieces of learning data on classification of classes for each attribute. The hyperplane represents a criterion for classifying a plurality of pieces of learning data by attribute, and is set to be farthest from the learning data corresponding to the support vector.

In the second embodiment of the present invention, the text profile generation unit 150 of the user terminal 100 reflects the weight vector to the texts transmitted from the opposite user, calculates the first class information for the set of texts, Create a text profile for the user. In one embodiment, the text may include Social Network Service text generated by the other user terminal 300 and transmitted to the user terminal 100.

The text profile generator 150 reflects only the vector components corresponding to the learning data corresponding to the support vector closest to the hyperplane among the weight vectors, and calculates the first class information on the basis of the hyperplane for the set of texts can do. According to the second embodiment of the present invention, a user of the user terminal 100 can create a user profile for the other user to exchange SNS contents through the SNS service using his or her own user terminal 100. The text profile generated by the text profile creation unit 150 may be displayed on the display unit 120 of the user terminal 100.

Accordingly, the user of the user terminal 100 obtains an advantage of confirming and utilizing the propensity and interest information of the other user. In the second embodiment of the present invention, the user of the user terminal 100 transmits the text profile creation unit 150 to the user terminal 100 (e.g., ).

In the second embodiment of the present invention, when the user terminal 100 is allowed to receive location information from the counterpart user terminal 300, it is also possible to combine the text profile and the mobility profile to generate the user profile of the counterpart user . In this case, it is also possible that the mobility profile generating unit 230 and the profile integrating unit 240 among the configurations of the first embodiment shown in FIG. 1 are provided in the user terminal 100 of the second embodiment shown in FIG. 3 .

It is to be understood that the above-described embodiments are provided to facilitate understanding of the present invention, and do not limit the scope of the present invention, and it is to be understood that various modifications are possible within the scope of the present invention. It is to be understood that the technical scope of the present invention should be determined by the technical idea of the claims and that the technical scope of the present invention is not limited to the literary description of the claims, To the invention of the invention.

10, 20: user profile creation system 100: user terminal
110: user interface unit 120: display unit
130: communication unit 140:
150: text profile creation unit 200: user profile creation server system
210, 420: Learning part 220: Text profile creation part
230: mobility profile generation unit 240: profile integration unit
241: Profile generating unit 242:
250, 410: learning data DB 260, 430: classification information DB
270: Movement information DB 300:
400: Server system

Claims (20)

A learning method, comprising: learning a plurality of learning data for users who know a user profile in advance as one learning unit; displaying a criterion for classifying the plurality of learning data by an attribute, A learning unit that generates classification information including a normal vector of a hyperplane and a weight vector of the normal vector, the classification vector being a criterion for classifying a star class, the influence of each learning data on classification of each class of the attribute; And
And a text profile generator for generating a text profile by calculating first class information for all the sets of text by reflecting the weight vector on all texts generated by a user who wants to know the user profile,
The text profile generator reflects only the vector component corresponding to the learning data corresponding to the support vector closest to the hyperplane among the weight vectors and calculates the first class information for the set of all the texts User profile creation device. The method according to claim 1,
Wherein the learning unit calculates the hyperplane and the weight vector indicating a criterion for classifying the plurality of learning data by the attribute using a support vector machine. 3. The method of claim 2,
Wherein the text profile generating unit comprises:
A user who reflects only vector components corresponding to learning data corresponding to a support vector closest to the hyperplane among the weight vectors and calculates the first class information on the basis of the hyperplane for all the sets of texts, Profile generating device. The method according to claim 1,
Wherein the attribute includes at least one of gender, age, region, occupation, political orientation, hobby, and interest. The method according to claim 1,
Wherein the text includes text generated by a user who wants to know the user profile through a Social Network Service. 6. The method according to any one of claims 1 to 5,
A mobility profile generating unit for generating a mobility profile for a user who wants to know the user profile by calculating second class information on the set of all the texts based on the movement information of the user; And
And a profile integrating unit for merging the text profile and the mobility profile to generate a user profile for a user who wants to know the user profile. The method according to claim 6,
The profile integrating unit,
And a profile generator for reflecting the first importance information to the first class information and generating the user profile reflecting the second importance information to the second class information. 8. The method of claim 7,
The profile integrating unit,
Updating the first importance level based on a difference between the class information of the user profile and the actual class information of a user who wants to know the user profile, And an importance updating unit for updating the second importance degree based on a difference of information. A user who is provided by performing learning using a plurality of learning data for users who know a user profile in advance as one learning unit and is provided with a criterion for classifying the plurality of learning data by attributes, A hyperplane normal vector representing the influence of each learning data on classification of each class of attribute and a storage storing classification information including a weight vector of the normal vector part; And
And a text profile generator for generating a text profile for the counterpart user by calculating the first class information for the set of texts by reflecting the weight vector to texts transmitted from the counterpart user,
Wherein the text profile generator reflects only a vector component corresponding to learning data corresponding to a support vector closest to the hyperplane among the weight vectors and calculates the first class information for a set of the texts, . 10. The method of claim 9,
The text profile generator reflects only the vector component corresponding to the learning data corresponding to the support vector closest to the hyperplane among the weight vectors so as to generate the first class information . 11. The method according to claim 9 or 10,
Wherein the text includes a Social Network Service text generated by the counterpart user and transmitted to the terminal. A user who is provided by performing learning using a plurality of learning data for users who know a user profile in advance as one learning unit and is provided with a criterion for classifying the plurality of learning data by attributes, A database for storing classification information including a normal vector of a hyperplane representing the influence of each learning data on classification of each class of the attribute and a weight vector of the normal vector, ; And
Generating a text profile for the counterpart user by calculating the first class information for the set of texts by reflecting the weight vector to the texts transmitted from the terminal of the counterpart user who wants to know the user profile, A user profile creation server system comprising:
Wherein the text profile generator reflects only the vector component corresponding to the learning data corresponding to the support vector closest to the hyperplane among the weight vectors and calculates the first class information for the set of texts Profile creation server system. 13. The method of claim 12,
Receiving a log file indicating a history of location information of the terminal of the other user from the terminal of the other user and calculating second class information for the set of texts based on the log file, A mobility profile generating unit for generating a mobility profile; And
And a profile integrating unit for merging the text profile and the mobility profile to generate a user profile for the other user. A learning method, comprising: learning a plurality of learning data for users who know a user profile in advance as one learning unit; displaying a criterion for classifying the plurality of learning data by an attribute, Calculating classification information including a normal vector of a hyperplane and a weight vector of the normal vector, which is a criterion for classifying a star class, and which has an effect on classification of each class of the learning data; And
And generating a text profile by calculating first class information for all the sets of texts by reflecting the weight vectors to all texts generated by a user who wants to know the user profile,
The step of generating the text profile may include the step of reflecting only the vector component corresponding to the learning data corresponding to the support vector closest to the hyperplane among the weight vectors, / RTI &gt; 15. The method of claim 14,
Wherein the step of calculating the classification information includes the step of calculating the hyperplane and the weight vector indicating a criterion for classifying the plurality of learning data by the attribute using a support vector machine Way. 16. The method of claim 15,
Wherein the step of generating the text profile comprises:
Reflecting only the vector component corresponding to learning data corresponding to a support vector closest to the hyperplane among the weight vectors and calculating the first class information with respect to the hyperplane for all the sets of texts / RTI &gt; 15. The method of claim 14,
Generating a mobility profile for a user who wants to know the user profile by calculating second class information for the set of all the texts based on the movement information of the user; And
Further comprising merging the text profile and the mobility profile to generate a user profile for a user who wants to know the user profile. 18. The method of claim 17,
Wherein the generating the user profile comprises:
And generating the user profile by reflecting the first importance to the first class information and reflecting the second importance to the second class information. 19. The method of claim 18,
Wherein the generating the user profile comprises:
Updating the first importance level based on a difference between the class information of the user profile and the actual class information of a user who wants to know the user profile, And updating the second importance degree based on the difference of the information. A computer-readable recording medium on which a program for executing the method according to any one of claims 14 to 19 is recorded.

Images