JP5829180B2 - Class type estimation apparatus, program, and method for estimating ratio of each class type in all member objects in group - Google Patents

Description

  The present invention relates to a technique for estimating a class (attribute) type for an estimation target group including a plurality of member objects.

  In recent years, an unspecified number of third parties can actively transmit their own comment sentences (text information) via an SNS (Social Networking Service) site server. The “SNS site server” publishes posted text posted by one user to a group of a plurality of users. For example, there are facebook (registered trademark), twitter (registered trademark), google + (registered trademark), and mixi (registered trademark), which are also generally referred to as miniblog sites. Each user registers an account with the SNS site server, and the posted text is published together with the account. The profile information for each user is also disclosed. The profile information is a self-introduction sentence of the user, represents the attribute type of the user, and is freely browsed by an unspecified number of third parties.

  On the other hand, the posted text transmitted by the SNS site server is often word-of-mouth information related to products and services. In this case, according to marketing, it is desired to estimate what kind of user is doing what kind of evaluation. In general, a large number of posted sentences based on a predetermined purpose are grouped, and the tendency of user attributes is estimated based on the profile information of the member posters. In this case, it is preferable that the profile information of the member poster is as accurate as possible. Therefore, it is also conceivable to estimate profile information, that is, user attributes (age, sex, hobbies, etc.) from the contents of the posted sentences of member posters.

  Conventionally, there is a technique for estimating a user's profile information based on text information posted by the user (see, for example, Non-Patent Document 1). According to this technology, it is possible to estimate a user attribute from a word distribution in a past tweet for each user, and to estimate a ratio of user attributes of the gender of all users in the target group. Further, according to this technology, it is possible to collect one or more user attributes and TV viewing history in advance, and to present content (for example, a TV program) that matches the preference of the user group. Thereby, in the social communication place of the user group, the possibility of viewing the content is increased, and smooth communication is provided. For example, according to word-of-mouth marketing, there is a great need for acquiring user profile information using SNS.

Kazufumi Ikeda, Hajime Hattori, Kazunori Matsumoto, Tomohiro Ono, Teruo Higashino, “Twitter Poster Profile Estimation Method for Market Estimation”, DICOMO 2011, 7E-1. Chubu University, Satoshi Suzuki, “Least Squares Method”, [online], [Search May 3, 2012], Internet <URL: http://szksrv.isc.chubu.ac.jp/lms/lms1.html > Wikipedia, “Least Squares”, [online], [Search May 3, 2012], Internet <URL: http://en.wikipedia.org/wiki/%E6%9C%80%E5%B0% 8F% E4% BA% 8C% E4% B9% 97% E6% B3% 95>

  According to the field of marketing, it is generally considered that there is a need to “know only the ratio of each class type among a plurality of class (attribute) types for a group to be estimated including a plurality of member objects”. . For that purpose, it is necessary to clearly specify the profile information of the user. However, since the profile information in the SNS site is generally described by the user himself, it is often not clearly described with intention and interest. For example, there is a case where a male user dares to pretend to be a woman and posts a post. Therefore, according to the technique described in Non-Patent Document 1, a characteristic word is extracted from the past posted text of each contributor in the group, and the profile information of each contributor is estimated. Thereby, the ratio of each profile item (class) in the whole user in the group can also be estimated.

  However, according to the technique described in Non-Patent Document 1, first, profile information is clearly estimated for each user (member object), and the results are tabulated. However, in order to clearly identify the profile information for each user, there arises a problem that the number of users in the estimation target group increases and the necessary calculation amount also increases. In addition, by clearly identifying the profile information for each user, if the profile information is ambiguous, there is a problem that the estimation accuracy of the ratio of profile items deteriorates.

  Accordingly, the present invention provides a class ratio estimation device capable of calculating the ratio of each class type without accurately identifying the class (attribute) type for each object for a group to be estimated including a plurality of member objects, The purpose is to provide a program and method.

According to the present invention, for a group including a plurality of estimation target member objects i (i = 1 to N), the ratio of each class type j (j = 1 to M) in a plurality of class (attribute) types is calculated. A class ratio estimation device comprising:
The member object is composed of a member identifier i and a feature value f _ki of each element in a plurality of elements k (k = 1 to K).
Learning feature value accumulating means for accumulating the probability distribution function of the feature value of each element k as a learning feature value distribution P _kj for each class type j from all member objects i (i = 1 to L) of the learning target group When,
A target feature amount distribution calculating means for calculating a probability distribution function of the feature amount of each element k as a target feature distribution P _k from all the member objects of the estimation target group;
The distribution distance d (Σ _j w _j P _kj , P _k ) between the weighted sum Σ _j w _j P _kj of the learning feature quantity distribution P _kj for each class type j and the target feature quantity distribution P _k is expressed as w _j ≧ so as to minimize under the condition that 0, a weight determining means for determining the weights w _j to be given to learning feature distribution P _kj for each class type j,
Class weight normalizing means for normalizing the determined weight w _j of each class type j and calculating a ratio for each class type j is provided.

According to another embodiment of the class ratio estimation apparatus of the present invention,
The weight determining means calculates the weighted sum Σ _j w _j μ _kj of the average value μ _kj of the learning feature quantity distribution P _kj of each element k calculated for each class type j and the target feature quantity distribution P _k for each element _k . Inter-distribution distance d (Σ _j w _j P _kj , P _k ) = Σ _k (Σ _j w _j μ _kj −μ _k ) ² defined by the square error from the mean value μ _k, with the condition that w _j ≧ 0 It is also preferable to determine the weight w _j to be assigned to the learning feature amount distribution P _kj of each class type j so as to be minimized under.

According to another embodiment of the class ratio estimation apparatus of the present invention,
Enter the group including a plurality of member object i to be learned (i = 1 to L), from all the member objects of the group, for each class type j of feature parameters of each element k of the probability distribution function P _ki It is also preferable to further include a learning feature amount distribution calculating unit that calculates and outputs the probability distribution function as a learning feature amount distribution P _kj to the learning feature amount storage unit.

According to another embodiment of the class ratio estimation apparatus of the present invention,
The member identifier is the poster identifier,
Each member object is a user text related to the member poster corresponding to the poster identifier,
Class type is an attribute type related to user profile,
The element is also preferably a keyword included in the user text.

According to another embodiment of the class ratio estimation apparatus of the present invention,
User text acquisition means for acquiring the user text of each member contributor belonging to the estimation target group;
A keyword extracting means for extracting a keyword as an element from user text by morphological analysis;
It is also preferable that each member contributor further includes a feature amount extraction means for extracting the number of each keyword as an element as a feature amount.

According to another embodiment of the class ratio estimation apparatus of the present invention,
Preferably, the keyword extracting means further extracts a domain name in a URL (Uniform Resource Locator) as a keyword as an element.

According to another embodiment of the class ratio estimation apparatus of the present invention,
Preferably, the keyword extracting means further extracts the member identifier of each member belonging to the estimation target group in the user text obtaining means as a keyword as an element.

According to another embodiment of the class ratio estimation apparatus of the present invention,
It is also preferable that the user text is a posted message sent by a member poster and / or a profile message of a member poster in the communication server.

According to another embodiment of the class ratio estimation apparatus of the present invention,
The target feature amount distribution calculating means includes, when the member poster's profile sentence includes a keyword representing the class type j stored in the learning feature amount storage means, in the class type j of the learning feature amount storage means. It is also preferable to use the probability distribution function P _kj of the learning feature value of each corresponding element k as it is as the feature value f _ki of each element k of the member object of the member poster.

According to the present invention, the ratio of each class type j (j = 1 to M) in a plurality of class (attribute) types is calculated for a group including a plurality of member objects i (i = 1 to N) to be estimated. A class ratio estimation program for causing a computer to function as follows:
The member object is composed of a member identifier i and a feature value f _ki of each element in a plurality of elements k (k = 1 to K) .
Learning feature value accumulating means for accumulating the probability distribution function of the feature value of each element k as a learning feature value distribution P _kj for each class type j from all member objects i (i = 1 to L) of the learning target group When,
A target feature amount distribution calculating means for calculating a probability distribution function of the feature amount of each element k as a target feature distribution P _k from all the member objects of the estimation target group ;
The distribution distance d (Σ _j w _j P _kj , P _k ) between the weighted sum Σ _j w _j P _kj of the learning feature quantity distribution P _kj for each class type j and the target feature quantity distribution P _k is expressed as w _j ≧ Weight determining means for determining a weight w _j to be assigned to the learning feature amount distribution P _kj of each class type j so as to be minimized under the condition of 0 ;
Class type normalizing means for normalizing the weight w _j of each determined class type j and calculating a ratio for each class type j;
And making the computer function.

According to the present invention, the ratio of each class type j (j = 1 to M) in a plurality of class (attribute) types is calculated for a group including a plurality of member objects i (i = 1 to N) to be estimated. An apparatus class ratio estimation method comprising:
The member object is composed of a member identifier i and a feature value f _ki of each element in a plurality of elements k (k = 1 to K) .
The apparatus accumulates a probability distribution function of a feature quantity of each element k as a learning feature quantity distribution P _kj for each class type j from all member objects i (i = 1 to L) of the learning target group. A quantity storage unit,
The device
A first step of calculating a probability distribution function of a feature value of each element k as a target feature distribution P _k from all the member objects of the estimation target group ;
The distribution distance d (Σ _j w _j P _kj , P _k ) between the weighted sum Σ _j w _j P _kj of the learning feature quantity distribution P _kj for each class type j and the target feature quantity distribution P _k is expressed as w _j ≧ A second step of determining a weight w _j to be assigned to the learning feature amount distribution P _kj of each class type j so as to be minimized under the condition of 0 ;
A third step of normalizing the determined weight w _j of each class type j and calculating a ratio for each class type j;
It is characterized by performing .

  According to the class ratio estimation apparatus, program, and method of the present invention, for a group to be estimated including a plurality of member objects, the ratio of each class type is calculated without identifying the class (attribute) type for each object. Can do.

It is a system configuration diagram in the present invention. It is a functional block diagram of the class ratio estimation apparatus in this invention. A feature amount f _ki of each element k (k = 1~K) in each member identifier i, a diagram of the probability distribution density P _k of an average value mu. It is explanatory drawing showing the probability distribution function _Pkj accumulate _| stored in the learning feature-value accumulation _| storage part. A probability distribution function P _kj learning feature quantity, which is a schematic diagram of a comparison between the target characteristic quantity of the probability distribution function P _kj. It is explanatory drawing showing the processing content in a learning feature-value distribution calculation part.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  FIG. 1 is a system configuration diagram according to the present invention.

  According to FIG. 1, an unspecified number of third parties can transmit a posted message to the communication site server 2 via the Internet using the terminal 3 possessed by the third party. Hereinafter, the communication server 2 will be described as being an SNS site server, for example. Of course, although not limited to the SNS site server, it is necessary to form a group in which comment text is transmitted and subscribed among a plurality of contributors, and the posted text is disclosed to each contributor.

  According to FIG. 1, the class ratio estimation apparatus 1 of the present invention communicates with an SNS site server 2 via the Internet. The class ratio estimation apparatus 1 can acquire a poster's exchange relation list (group information) and a posted sentence for each poster via an API (Application Programming Interface) from the SNS site server 2. . Moreover, the profile sentence for every contributor can also be acquired. The API is a rule interface for using the function of the application service, and is prepared as different for each of various servers. Then, the class ratio estimation apparatus 1 has a ratio of each class type j (j = 1 to M) in a plurality of class (attribute) types for a group to be estimated including a plurality of member objects i (i = 1 to N). Is calculated.

  In the following description, it is assumed that the exclusive class types are two class types of “male” / “female”. The member object is assumed to be a user text on the Twitter site. Then, the class ratio estimation apparatus 1 estimates the class ratio (for example, gender ratio) in the entire group from user texts from a large number of member objects.

  FIG. 2 is a functional configuration diagram of the class ratio estimation apparatus according to the present invention.

  The class ratio estimation apparatus 1 includes a communication interface unit 10 and a user interface unit 11. The communication interface unit 10 communicates with the SNS server 2 via, for example, the Internet. The user interface unit 11 can clearly indicate to the user the ratio of class types in the group.

  The class ratio estimation apparatus 1 according to the present invention includes a user text acquisition unit 21, a keyword extraction unit 22, a feature amount extraction unit 23, a target feature amount distribution calculation unit 24, and a weight determination unit as <estimation function unit>. 25, a class type normalization unit 26, and a learning feature amount storage unit 30. The class ratio estimation apparatus 1 further includes a learning text acquisition unit 31 and a learning feature amount distribution calculation unit 34 as <learning function unit>. These functional components are realized by executing a program that causes a computer installed in the apparatus to function.

<Functional unit for estimation>
[User text acquisition unit 21]
The user text acquisition unit 21 acquires the member object of each member identifier belonging to the estimation target group via the communication interface unit 10. For example, from the SNS site server 2, user texts (including profile sentences; member objects) of a plurality of member posters belonging to the group for the estimation target person are acquired. The poster is recognized by the poster identifier (member identifier). The group for the estimation target person may be extracted by an exchange relation list (follow-follower relation) acquired from the SNS site server 2.

  Here, the member contributor in the SNS site server may be a publicity public contributor. Such a member contributor has a member identifier as a corporate publicity account, for example, and is followed by many users. That is, in the case of a publicity account, the ratio of the number of followers to the number of followers is large. Therefore, when the user text acquisition unit 21 also collects the user account of the user who follows the account for publicity, it becomes a factor that decreases the estimation accuracy of the resulting class ratio. Therefore, the user text acquisition unit 21 preferably removes the user from the estimation target when the ratio of the number of followers to the number of followers of the user exceeds a predetermined threshold.

  The class ratio estimation apparatus 1 may be one in which user texts are stored in advance in a database for each group. In this case, the class ratio estimation apparatus 1 does not need to include the user text acquisition unit 21. That is, it is not essential that the class ratio estimation apparatus 1 communicates with the SNS site server 2.

[Keyword extraction unit 22]
The keyword extraction unit 22 extracts keywords as elements from the user text by morphological analysis. The keyword extraction unit 22 divides the collected user text into morphemes by morphological analysis. A “morpheme” refers to the smallest meaningful unit among the constituent elements of a sentence. The keywords (morphemes) extracted by the present invention are only “nouns” that are elements of feature amounts described later. A keyword for each poster identifier (member identifier) is output to the feature amount extraction unit 23.

  As another embodiment, there is a case where a URL (Uniform Resource Locator) of a website is included in the user text. Therefore, it is preferable that the keyword extracting unit 22 further extracts the domain name in the URL as a keyword (morpheme) as an element. This is because there is a tendency that it is easy to access a domain of a specific URL depending on a user attribute. That is, by using the domain name as an element of the feature amount, it is possible to increase the elements that are the material for estimating the class ratio.

  As still another embodiment, the user text acquisition unit 21 may be able to acquire the member identifier of the follow user from the exchange relation list acquired from the SNS site server 2. Therefore, it is preferable that the keyword extraction unit 22 further extracts the member identifiers of the members belonging to the estimation target group in the user text acquisition unit 21 as keywords as elements. This is because there is a tendency that it is easy to follow a specific other user depending on the attribute of the user. In other words, by using the member identifier of the follow user as an element of the feature amount, it is possible to increase the element that is a material for estimating the class ratio.

[Feature Extraction Unit 23]
The feature amount extraction unit 23 extracts, for each member contributor, the number of each keyword as an element from the keywords included in the user text as a feature amount. Here, the “member object T _i ” includes a member identifier i (i = 1 to N) and feature amounts f _ki of each element in a plurality of elements k (k = 1 to K). For each member object, each keyword that is an element included in the user text (and profile sentence) posted in the past by the member poster is counted for each member identifier.
T _i : Object f _Ti : Feature quantity (vector) of object T _i
N: Number of all member objects in the estimation target group

[Target Feature Quantity Distribution Calculation Unit 24]
The target feature amount distribution calculation unit 24 calculates a probability distribution function P _k of each element k from the entire member objects of the estimation target group. The probability distribution density P _k may be, for example, an average value μ.
μ = 1 / N ・ Σ _Ti f _Ti
μ: Average value

FIG. 3 is an explanatory diagram showing the feature quantity f _ki of each element k (k = 1 to K) and the probability distribution density P _k of the average value μ for each member identifier i.

According to FIG. 3, the user text of each member indicates that the following keyword is included. Each keyword means each element.
The text of the member L ₁ includes “government” and “smart phone”.
The text of member L ₂ includes “laundry”.
The text of member L ₃ includes “government” and “smartphone”.
The text of the member L ₄ includes “smart phone”, “laundry”, and “cleaning”.
Further, according to FIG. 3, the average value μ is calculated as the probability distribution density P _k of each element k in the entire member object of the estimation target group.
f _k = (government, smartphone, washing, cleaning, ...)
= (0.5, 0.75, 0.5, 0.25, ...)

As another embodiment, the target feature amount distribution calculation unit 24 stores the learning feature amount when the keyword representing the class type j stored in the learning feature amount storage unit is included in the member poster's profile sentence. It is also preferable to use the probability distribution function P _kj of the learning feature value of each element k corresponding to the class type j of the means as it is as the feature value f _ki of each element k of the member object of the member poster. For example, when “male” is described in the profile sentence in the user text, the learning feature amount storage unit 30 has already accumulated the learning feature amount of the class type “male”. In this case, the learning feature amount distribution of the class type “male” in the learning feature amount storage unit 30 is directly taken in as the target feature amount distribution of the member poster. Thereby, the attribute of the user described in the profile sentence can be taken in the target feature amount distribution as it is.

[Learning feature amount storage unit 30]
The learning feature amount accumulation unit 30 accumulates the probability distribution function P _kj of the learning feature amount of each element k for each class type j.

FIG. 4 is an explanatory diagram illustrating the probability distribution function P _kj stored in the learning feature amount storage unit.

According to FIG. 4, the class type j is classified into “male” / “female”. For each class type j, the average value μ (probability distribution density P _k ) of each element k is stored in advance.
P (male) = (Government, smartphone, washing, cleaning, ...)
= (1, 1
, 0, 0, ...)
P (female) = (government, smartphone, washing, cleaning, ...)
= (0, 0.5, 1, 0.5, ...)
That is, the “class type” is an attribute type related to the user profile. The information stored in the learning feature amount storage unit 30 may be set in advance as a predetermined value, or may be output from a learning feature amount distribution calculation unit 34 described later.

[Weight determination unit 25]
The weight determination unit 25 calculates the inter-distribution distance d (Σ _j w _j P _kj , P _k ) between the weighted sum Σ _j w _j P _kj of the learning feature quantity distribution for each class type j and the target feature quantity distribution P _k. , so as to minimize under the condition that w _j ≧ 0, to determine the weights w _j to be given to learning feature distribution P _kj for each class type j.
d (Σ _j w _j P _kj , P _k ) is minimized under the condition of w _j ≧ 0 Σ _j w _j P _kj : Weighted sum of learning feature quantity distribution for each class type j P _k : Target feature quantity Distribution d (Σ _j w _j P _kj , P _k ): distance

FIG. 5 is an explanatory diagram illustrating a comparison between the probability distribution function P _kj of the learning feature quantity and the probability distribution function P _kj of the target feature quantity.

When there are two class types j (j = 1, 2), the weight w _{j given} to the learning feature amount distribution of each class type j can be calculated using the least square method (for example, Non-Patent Document 2). reference). Specifically, the weight determination unit 25 calculates the weighted sum Σ _j w _j μ _kj of the average value μ _kj of the learning feature amount distribution P _kj of each element k calculated for each class type j and the target for each element k. feature distribution P _k of an average between the distribution defined by the square errors between values mu _k distance _{d (Σ j w j P kj} , P k) a _{= Σ k (Σ j w j} μ kj -μ k) 2, The weight w _j to be assigned to the learning feature amount of each class type _j is determined so as to be minimized under the condition of w _j ≧ 0.
Σ _k (Σ _j w _j μ _kj −μ _k ) ² is minimized under the condition of w _j ≧ 0 Σ _j w _j μ _kj : Average value μ _kj of learning feature amount distribution P _kj for each class type j Weighted sum μ _k : Average value of target feature distribution P _k

Specifically, the square error S is calculated by the following equation. This square error S, to calculate the w _j to the minimum under the condition that w _j ≧ 0.
S = (w _c1 * μ _{1, c1} + w _c2 * μ _{1, c2} −μ ₁ ) ² +
(W _c1 * μ _{2, c1} + w _c2 * μ _{2, c2} −μ ₂ ) ² +
(W _c1 * μ _{3, c1} + w _c2 * μ _{3, c2} −μ ₃ ) ² +
(W _c1 * μ _{4, c1} + w _c2 * μ _{4, c2} −μ ₄ ) ²
c1: Class type "male"
c2: Class type “female”
k = 1 to 4: (Government, smartphone, washing, cleaning)
w _c1 : Weight of class type “male” w _c2 : Weight of class type “female” μ _{1, c1} : Average value of learning features of class type “male” in “government” μ _{2, c1} : “Smartphone” Mean value of learning features of class type of “male” in μ _{μc, c1} : average value of learning features of class type of “male” in “laundry” μ _{4, c1} : class of “male” in “cleaning” Average learning feature value of each type μ _{1, c2} : Average learning feature value of “Women” class type in “Government” μ _{2, c2} : Learning feature value of “Women” class type in “Smartphone” Average value μ _{3, c2} : Average learning feature value of “Women” class type in “Laundry” μ _{4, c2} : Average learning feature value of “Woman” class type in “Cleaning” μ ₁ : the average value of the target feature amount in government "μ _2: in the" smart phone " The average value of elephant feature value μ _3: The average of the object feature amount in the "washing" value μ _4: object feature amount of the average value in the "cleaning"

Next, constants A to F in the following equation are calculated from the above-described square error S.
S = A + Bw ₁ ² -2Cw ₁ -2Dw ₂ + 2Ew ₁ w ₂ + Fw ₂ ²
In this case, w ₁ and w ₂ are calculated as follows.
w ₁ = (FD-CE) / (FB-E ² ) = 0.5
w ₂ = (BC-DE) / (FB-E ² ) = 0.5

  In addition, when the number of classes is 3 or more (multidimensional), it can be calculated by the method described in Non-Patent Document 3, for example.

[Class type normalization unit 26]
The class type normalization unit 26 normalizes the determined weight w _j of each class type j and calculates a ratio for each class type j. The normalization here refers to a conversion that is proportionally converted to a value between 0 and 1 and can be compared for each class type.

<Functional part for learning>
[Learning text acquisition unit 31]
The learning text acquisition unit 31 acquires a group including a plurality of learning target member objects i (i = 1 to L) from the user via the user interface unit 11. The acquired member object is output to the learning feature amount distribution calculation unit 34.

[Learning feature amount distribution calculation unit 34]
The learning feature amount distribution calculation unit 34 calculates a probability distribution function P _ki of the learning feature amount of each element k for each class type j from all member objects of the group, and uses the probability distribution function as the probability distribution function P _kj. This is output to the learning feature amount storage unit 30.

  FIG. 6 is an explanatory diagram illustrating processing contents in the learning feature amount distribution calculation unit.

  According to FIG. 6, a plurality of learning member objects having a known class type are represented, and a plurality (two) of class types “male” and a plurality of class types “female” ( 2). Here, the average value of the learning feature amount of each element k is calculated for each class type. In general, it can be expected that the feature amount of each member object is distributed closer to the feature amount of the class type as the element k (number of dimensions) of the feature amount becomes larger than the number of classes. Therefore, the class ratio calculated by the present invention can be an approximate value sufficiently close to the true value.

According to the above-described embodiment, the class type is derived from the user text for each user with reference to one learning feature amount storage unit 30. Here, a multi-posting user who has posted a large number of user texts and a low-posting user who has posted only a few user texts are mixed. For example, the feature size of the element | f _Ti |, that is, the distribution tendency of the feature quantity is different between a small number of posting users with a past posting number of 100 or less and a multiple posting user with 100 or more postings. Therefore, estimating them collectively by one learning feature amount storage unit 30 may reduce accuracy. Therefore, in order to perform class estimation suitable for the user as much as possible, it is also preferable to separately provide a learning feature amount storage unit for multiple posting users and a learning feature amount storage unit for low posting users.

  As described above in detail, according to the class ratio estimation apparatus, program, and method of the present invention, the class (attribute) type for each object can be accurately identified for the estimation target group including a plurality of member objects. The ratio of each class type can be calculated. In particular, according to the present invention, an unknown class ratio can be efficiently estimated for a group of a large number of ambiguous member objects.

  As a service using the class ratio estimation apparatus of the present invention, it can be used for social media analysis in a group to be estimated. According to the prior art, only negative / positive estimation based on a member's post can be performed. According to the present invention, estimation according to the class type such as age, sex, occupation, etc. can be performed in real time, and the voice of the market that is currently a topic for a predetermined product service can be recognized. Such estimation results are not only made available for browsing on a website, but also provided as web tools for various applications.

  Various changes, modifications, and omissions of the above-described various embodiments of the present invention can be easily made by those skilled in the art. The above description is merely an example, and is not intended to be restrictive. The invention is limited only as defined in the following claims and the equivalents thereto.

DESCRIPTION OF SYMBOLS 1 Class ratio estimation apparatus 10 Communication interface part 11 User interface part 21 User text acquisition part 22 Keyword extraction part 23 Feature-value extraction part 24 Target feature-value distribution calculation part 25 Weight determination part 26 Class classification normalization part 30 Learning feature-value storage part 31 Learning text acquisition part 34 Learning feature-value distribution calculation part 2 Communication site server, SNS site server 3 Terminal

Claims (11)

A class ratio estimation apparatus that calculates a ratio of each class type j (j = 1 to M) in a plurality of class (attribute) types for a group including a plurality of member objects i (i = 1 to N) to be estimated. And
The member object includes a member identifier i and a feature quantity f _ki of each element in a plurality of elements k (k = 1 to K).
Learning feature value accumulating means for accumulating the probability distribution function of the feature value of each element k as a learning feature value distribution P _kj for each class type j from all member objects i (i = 1 to L) of the learning target group When,
A target feature amount distribution calculating means for calculating a probability distribution function of the feature amount of each element k as a target feature distribution P _k from all the member objects of the estimation target group;
The distribution distance d (Σ _j w _j P _kj , P _k ) between the weighted sum Σ _j w _j P _kj of the learning feature quantity distribution P _kj for each class type j and the target feature quantity distribution P _k is expressed as w _j ≧ so as to minimize under the condition that 0, a weight determining means for determining the weights w _j to be given to learning feature distribution P _kj for each class type j,
A class ratio estimating apparatus comprising class type normalizing means for normalizing the determined weight w _j of each class type j and calculating a ratio for each class type j. The weight determining means calculates the weighted sum Σ _j w _j μ _kj of the average value μ _kj of the learning feature amount distribution P _kj of each element k calculated for each class type j and the target feature amount distribution P _k for each element _k. Inter-distribution distance d (Σ _j w _j P _kj , P _k ) = Σ _k (Σ _j w _j μ _kj −μ _k ) ² defined as a square error with respect to the average value μ _{k of} w _j ≧ 0 2. The class ratio estimation apparatus according to claim 1, wherein a weight _wj to be assigned to the learning feature amount distribution _Pkj of each class type _j is determined so as to be minimized under conditions. Enter the group including a plurality of member object i to be learned (i = 1 to L), from all the member objects of the group, for each class type j of feature parameters of each element k of the probability distribution function P _ki 3. The class ratio estimation apparatus according to claim 1, further comprising a learning feature amount distribution calculating unit that calculates and outputs the probability distribution function as a learning feature amount distribution P _kj to the learning feature amount storage unit. . The member identifier is a poster identifier,
Each member object is a user text relating to a member poster corresponding to the poster identifier,
The class type is an attribute type related to the user profile,
The class ratio estimation apparatus according to claim 1, wherein the element is a keyword included in the user text. User text acquisition means for acquiring the user text of each member contributor belonging to the estimation target group;
Keyword extracting means for extracting a keyword as an element from the user text by morphological analysis;
The class ratio estimation apparatus according to claim 4, further comprising a feature amount extraction unit that extracts, as a feature amount, the number of each keyword as an element for each member contributor. 6. The class ratio estimation apparatus according to claim 5, wherein the keyword extraction unit further extracts a domain name in a URL (Uniform Resource Locator) as a keyword as an element. 7. The class ratio estimation according to claim 5, wherein the keyword extraction unit further extracts a member identifier of each member belonging to the estimation target group in the user text acquisition unit as a keyword as an element. apparatus.   The class according to any one of claims 5 to 7, wherein the user text is a posted message sent by a member poster and / or a profile message of a member poster in the communication server. Ratio estimation device. The target feature amount distribution calculating means, when the member poster's profile sentence includes a keyword representing the class type j stored in the learning feature amount storage means, the class of the learning feature amount storage means _9. The learning feature distribution P _kj of each element k corresponding to the type j is used as the feature quantity f _ki of each element k of the member object of the member poster as it is. The class ratio estimation apparatus described in 1. For a group including a plurality of member objects i (i = 1 to N) to be estimated, the computer is caused to calculate a ratio of each class type j (j = 1 to M) in a plurality of class (attribute) types. A class ratio estimation program,
The member object includes a member identifier i and a feature quantity f _ki of each element in a plurality of elements k (k = 1 to K) .
Learning feature value accumulating means for accumulating the probability distribution function of the feature value of each element k as a learning feature value distribution P _kj for each class type j from all member objects i (i = 1 to L) of the learning target group When,
A target feature amount distribution calculating means for calculating a probability distribution function of the feature amount of each element k as a target feature distribution P _k from all the member objects of the estimation target group ;
The distribution distance d (Σ _j w _j P _kj , P _k ) between the weighted sum Σ _j w _j P _kj of the learning feature quantity distribution P _kj for each class type j and the target feature quantity distribution P _k is expressed as w _j ≧ Weight determining means for determining a weight w _j to be assigned to the learning feature amount distribution P _kj of each class type j so as to be minimized under the condition of 0 ;
Class type normalizing means for normalizing the weight w _j of each determined class type j and calculating a ratio for each class type j;
A class ratio estimation program that causes a computer to function. Class ratio estimation method for a device that calculates a ratio of each class type j (j = 1 to M) in a plurality of class (attribute) types for a group including a plurality of estimation target member objects i (i = 1 to N) Because
The member object includes a member identifier i and a feature quantity f _ki of each element in a plurality of elements k (k = 1 to K) .
The apparatus learns to accumulate a probability distribution function of a feature quantity of each element k as a learned feature quantity distribution P _kj for each class type j from all member objects i (i = 1 to L) of a learning target group. It has a feature amount storage unit,
The device is
A first step of calculating a probability distribution function of a feature value of each element k as a target feature distribution P _k from all the member objects of the estimation target group ;
The distribution distance d (Σ _j w _j P _kj , P _k ) between the weighted sum Σ _j w _j P _kj of the learning feature quantity distribution P _kj for each class type j and the target feature quantity distribution P _k is expressed as w _j ≧ A second step of determining a weight w _j to be assigned to the learning feature amount distribution P _kj of each class type j so as to be minimized under the condition of 0 ;
A third step of normalizing the determined weight w _j of each class type j and calculating a ratio for each class type j;
A method for estimating the class ratio of the apparatus .

Images