JP2013242676A - User attribute estimation device, user attribute estimation method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To focus on attributes owned by propagated users to propagate user attributes, and thereby estimate an unknown user attribute with high accuracy.SOLUTION: A user attribute estimation device is configured to comprise the means for: extracting a user set having a specific attribute from a user set of a known user attribute stored in user attribute storage means; extracting a characteristic amount appearing characteristically in a user set having a specific attribute on the basis of a user set having the specific attribute and each dialogue log stored in dialogue log storage means, calculating attribute probability serving as probability that each characteristic amount belongs to each user attribute and storing the attribute probability in attribute probability storage means; calculating intimacy between an input user and each adjacent user on the basis of each dialogue log stored in dialogue log storage means; and calculating propagation probability of propagating each attribute of adjacent users to the input user on the basis of the intimacy and the attribute probability of each characteristic amount stored in the attribute probability storage means.

Description

  The present invention relates to a technique for estimating a user attribute, and more particularly to a technique for estimating an attribute of a user who does not disclose a user attribute in a social network.

  As a first conventional technique for estimating a user attribute, there is a user attribute estimation device that estimates an unknown user attribute from a geographic attribute by learning and using the relationship between the geographic attribute and the user attribute (for example, Patent Documents). 1).

  In addition, as a second conventional technique, there is a user attribute estimation device that estimates an unknown user attribute by defining proximity on a social network and propagating an attribute from a close user (see, for example, Non-Patent Document 1). ).

JP2011-238169A

Alan Mislove, Bimal Viswanath, Krishna P. Gummadi, Peter Druschel, "You are who you know: inferring user profiles in online social networks". In Proceedings of the third ACM international conference on Web search and data mining, pages 251-260, 2010.

  By using a method similar to that of the first conventional technique, it is possible to estimate user attributes based on feature quantities such as words used by the user, geographic information, and time information. However, user attributes cannot be estimated for inactive users who do not have such information.

  The second prior art makes it possible to estimate a user attribute even for an inactive user by using and propagating a user attribute close to the social network. However, since all attributes are propagated from a nearby user, there is a problem that attributes that the user on the side of propagation does not have are propagated at the same time.

  The present invention has been made in view of the above points, and it is possible to estimate an unknown user attribute with high accuracy by propagating the user attribute by focusing on the attribute of the user on the propagation side. The purpose is to provide technology.

In order to solve the above problems, the present invention provides a conversation log between users stored in the conversation log storage means and an attribute set of user sets whose user attributes stored in the user attribute storage means are known. A user attribute estimation device for estimating a user attribute of an input user whose user attribute is unknown,
User set extraction means for extracting a user set having a specific attribute from the user set stored in the user attribute storage means;
Based on the user set having the specific attribute and each conversation log stored in the conversation log storage unit, the feature quantity characteristically expressed in the user set having the specific attribute is extracted, and each feature quantity is each user. Calculating an attribute probability that is a probability belonging to the attribute, and storing the attribute probability in the attribute probability storage means;
A closeness calculating means for calculating a closeness between the input user and each neighboring user based on each conversation log stored in the conversation log storage means;
Propagation probability calculation means for calculating a propagation probability for propagating each attribute of a neighboring user to the input user based on the familiarity and the attribute probability of each feature quantity stored in the attribute probability storage means. It is comprised as a user attribute estimation apparatus characterized by.

  The user attribute estimation device is configured to select an attribute value having the highest propagation probability among a plurality of attribute values included in the attribute name based on the result calculated by the propagation probability calculation unit as an attribute in the attribute name of the input user. You may further have an output means to output as a value.

  The propagation probability calculation means, for example, calculates the propagation probability by multiplying the probability obtained by multiplying the probability that a feature amount is included in the conversation between the input user and each neighboring user and the attribute probability of the feature amount, by the intimacy. calculate.

  Moreover, this invention can also be comprised as a user attribute estimation method which the said user attribute estimation apparatus performs. The present invention can also be configured as a program for causing a computer to function as user set extraction means, attribute probability calculation means, intimacy calculation means, and propagation probability calculation means in the user attribute estimation device.

  According to the present invention, when propagating the attributes of the user on the propagation side based on the user attributes of the conversation log and the learning user set, the attributes of the user on the propagation side can be narrowed down. It is possible to solve the problem that attributes that the user on the side to be transmitted does not have at the same time and to estimate unknown user attributes with high accuracy.

It is a block diagram of the user attribute estimation apparatus 100 in embodiment of this invention. It is a flowchart of operation | movement of the learning part 110 in embodiment of this invention. It is a flowchart of operation | movement of the inference part 120 in embodiment of this invention. It is an example of the information stored in the user attribute memory | storage part 10 in embodiment of this invention. It is an example of the information stored in the conversation log memory | storage part 20 in embodiment of this invention. It is an example of the information memorize | stored in the attribute probability memory | storage part 50 in embodiment of this invention.

  Embodiments of the present invention will be described below with reference to the drawings. The embodiment described below is only an example, and the embodiment to which the present invention is applied is not limited to the following embodiment.

(Outline of the embodiment)
First, an outline of the present embodiment will be described. The technology according to the present embodiment is based on the assumption that people who are involved in a social network share some attributes, not all attributes, and the contents of the two users talking on the social network. Is based on the assumption that it is easy to include commonality between two users. Then, by analyzing the content of the conversation log, it is estimated which of the attributes of the user on the propagating side is shared with the user on the propagating side, and the propagation is limited to the attributes that are shared. Is what you do.

(Configuration of User Attribute Estimation Device 100)
FIG. 1 shows a block diagram of a user attribute estimation device 100 for calculating a probability that a user has a certain attribute based on the conversation log and the user attribute of the learning user set.

  As shown in the figure, the user attribute estimation device 100 includes a user attribute storage unit 10, a conversation log storage unit 20, a user set extraction unit 30, an attribute probability calculation unit 40, an attribute probability storage unit 50, an input unit 60, a closeness degree. A calculation unit 70, a propagation probability calculation unit 80, and an output unit 90 are included.

  Among the above functional units, the user attribute storage unit 10, the user set extraction unit 30, the conversation log storage unit 20, the attribute probability calculation unit 40, and the attribute probability storage unit 50 perform learning processing described later. The learning unit 110 is configured. The input unit 60, conversation log storage unit 20, familiarity calculation unit 70, user attribute storage unit 10, attribute probability storage unit 50, propagation probability calculation unit 80, and output unit 90 perform inference processing described later. And constitutes the inference unit 120.

  The user attribute estimation apparatus 100 according to the present embodiment includes a user set extraction unit 30, an attribute probability calculation unit 40, an input unit 60, and a closeness calculation in a computer including a storage device (memory, hard disk, etc.) serving as each storage unit. This can be realized by executing a program corresponding to the processing of the unit 70, the propagation probability calculation unit 80, and the output unit 90. The program may be stored in a storage medium such as a portable memory and distributed, installed on the computer, or downloaded from a server on a network and installed on the computer. In addition, any or all of the user attribute storage unit 10, the conversation log storage unit 20, and the attribute probability storage unit 50 are not provided in a computer that performs learning and inference processing, and can be accessed from the computer via a network. You may prepare for.

(Processing content of learning unit 110)
Hereinafter, the learning unit 110 will be described in detail. The learning unit 110 is a probability that each feature amount included in the conversation log belongs to each user attribute based on the user attribute stored in the user attribute storage unit 10 and the conversation log stored in the conversation log storage unit 20. The attribute probability is calculated and stored in the attribute probability storage unit 50.

  Below, each structure and operation | movement of the learning part 110 are demonstrated in detail. FIG. 2 is a flowchart showing the operation of the learning unit 110. In the following description, the corresponding step numbers of FIG.

<User attribute storage unit 10>
An example of data stored in the user attribute storage unit 10 is shown in FIG.

The user attribute storage unit 10 stores the attribute set A of the user set U. The attribute A is a set of an attribute name a _i and an attribute value a _ij . Attribute names include gender, age, place of residence, hometown, native language, available language, occupation, place of work, educational information (university or university, high school, etc.), group name, religion, instructions Political parties, experience sports, preferences (favorite food, favorite music, favorite books ...), hobbies ..., etc. For example, in the example illustrated in FIG. 4, the attribute value of the attribute whose attribute name is “profession” is “student” for the user whose ID is IDa.

  Further, the user attribute storage unit 10 may store the profile set P of the user set U. Profile P includes a self-introduction sentence and an introduction sentence from a friend or the like. In this case, as preprocessing, the profile is subjected to morphological analysis, and the attribute set A is extracted by matching whether the attribute name is included in the morphological analysis result.

  The user attribute storage unit 10 is not limited to a specific one as long as the information can be stored and restored. For example, a database, a database stored in a specific area of a general-purpose storage device (memory or hard disk device) provided in advance, a Web server holding a Web page, a database server including a database, or the like.

<Conversation log storage unit 20>
An example of data stored in the conversation log storage unit 20 is shown in FIG.

  The conversation log storage unit 20 stores a conversation log set L of the user set U. The conversation log L includes a post ID that is an identification number of the post itself, a reply destination post ID that is an identification number of a post to which the post has replied, a source user ID that is an identification number of the user who has posted, and the posting user's ID This is information including the posting destination user ID, which is an identification number, and the posting content. Further, the conversation log L may include information such as the posting time of the posted content, the posting location, supplementary information to the description content such as a hyperlink, and friend information. Note that the reply destination post ID and the post destination user ID may be singular, plural, or have no value. This means that posts that are not replies and posts that have not been replied are also included in the conversation log L. The posted content may be any content such as a sentence, a morphological analysis result of the sentence, an image, a video, “sympathy information”, “rating information”, and the like. “Empathy information” means “Like!” On Facebook (registered trademark). ”Indicates information that shows empathy, such as a button. “Rating information” represents information indicating an evaluation score for the posted content, such as an evaluation score for “taste log (registered trademark)”. The reply destination posting ID, the posting source user ID, the posting destination user ID, and the like may be included in the posting content in a certain format. The posting location may be latitude / longitude information or a place name. The friend information is information of friend registration such as follow of Twitter (registered trademark), and may be registration from one side or registration from both sides.

  The conversation log storage unit 20 is not limited to a specific one as long as the information is stored and can be restored. For example, a database, a database stored in a specific area of a general-purpose storage device (memory or hard disk device) provided in advance, a Web server holding a Web page, a database server including a database, or the like.

<User set extraction unit 30>
The user set extraction unit 30 receives the attribute set A of the user set U from the user attribute storage unit 10 as an input. Then, the same attribute user set U _ij having the same attribute value a _ij for each attribute name a _i is extracted and output to the attribute probability calculation unit 40 (step 21 in FIG. 2).

Further, similarly to the first prior art, as learning data post feature amount f 'and the attribute values a _ij of the attribute user set U _ij has the relationship of features and attributes included in the conversation log storage unit 20 And using this to estimate unknown user attributes, the attribute user set U _ij may be expanded (step 22 and step 23 in FIG. 2). At this time, as the feature amount to be used, words used by each user for posting, geographic information, time information, posting frequency, posting time zone, etc. can be considered.

<Attribute probability calculation unit 40>
The attribute probability calculation unit 40 receives as input the same attribute user set U _ij from the user set extraction unit 30 and the conversation log set L _Uij of the same attribute user set U _ij from the conversation log storage unit 20. Then, an attribute probability P (a _ij | f), which is a probability that the feature quantity f included in all the given feature quantity sets F belongs to each user attribute, is calculated and stored in the attribute probability storage unit 50 (FIG. 2). Step 24). The feature amount f includes words used by two users in each conversation, geographic information, time information, conversation frequency, number of characters, number of replies, time of reply, and the like. As a method for calculating the attribute probability P (a _ij | f), the following method can be considered.

ただし、C(f, a_ij)は、特徴量fと属性値a_ijを同時に持つユーザの数である。

Here, C (f, a _ij ) is the number of users having the feature quantity f and the attribute value a _ij at the same time.

This processing may be limited to a part rather than all the feature amount sets F. By performing processing such as a test on all feature quantity sets F, it is possible to extract feature quantities whose appearance frequencies are biased and appear characteristically in the same attribute user set U _ij . As a test method, for example, the formula of χ ² test is expressed as follows.

ただしここでO = 頻度の観測値、E = 帰無仮説から導かれる頻度の期待値（理論値）である。

Where O = observed frequency and E = expected frequency (theoretical value) derived from the null hypothesis.

<Attribute probability storage unit 50>
An example of data stored in the attribute probability storage unit 50 is shown in FIG. The attribute probability storage unit 50 stores the attribute probability input from the attribute probability calculation unit 40. For example, FIG. 6 shows that the probability that the feature quantity “final examination” belongs to the user attribute “student” is 0.7.

  The attribute probability storage unit 50 is not limited to a specific one as long as these pieces of information are stored and can be restored. For example, a database, a database stored in a specific area of a general-purpose storage device (memory or hard disk device) provided in advance, a Web server holding a Web page, a database server including a database, or the like.

(Processing contents of the inference unit 120)
Hereinafter, the inference unit 120 will be described in detail. The inference unit 120 propagates the user attribute based on the user attribute stored in the user attribute storage unit 10, the conversation log stored in the conversation log storage unit 20, and the attribute probability stored in the attribute probability storage unit 50. To estimate and output unknown user attributes.

Below, each structure and operation | movement of the inference part 120 are demonstrated in detail. FIG. 3 is a flowchart showing the operation of the inference unit 120. In the following description, the corresponding step numbers of FIG.
<Input unit 60>
The input unit 60 receives information about a user who is to be predicted as an input. And it converts into user ID and outputs it to the familiarity calculation part 70 (step 31 of FIG. 3).

<Intimacy calculation unit 70>
The familiarity calculation unit 70 receives the user ID from the input unit 60 and the conversation log set L from the conversation log storage unit 20 as inputs. Then, first extracts the neighbor user set S _u in a social network of input user u (step 32 in FIG. 3). Next, the intimacy w _uu ′ between the input user u and the neighboring user u ′ is calculated and output to the propagation probability calculation unit 80 (step 33 in FIG. 3).

The proximity user set S _u in a social network of input user u, may be a user who have a conversation in the conversation log may be used by extracting users with friendship friend information of the conversation log. As a method for calculating the intimacy w _{uu ′} between the two users, a method of calculating the frequency of conversation, the number of characters, the number of times a reply is continued, the time taken for a reply to be made, the number of common friends, and the like as feature quantities can be considered. For example, in the second prior art, it can be obtained by the following equation, but may be a method in which all or part of the above-described feature amount is added to the following equation. Further, it performs normalization so that the 1 The sum of w _{uu 'for} neighboring user set S _u.

Here, dist is the following function.

However, it is assumed that K is a route set that passes through the user u, the user u ′, and their common neighboring users. Strength is the following function.

Here, X ′ _ij is obtained as follows.

ただし、X_ijはユーザiとユーザjが過去に会話した回数であるとする。

Here, X _ij is the number of times user i and user j have talked in the past.

<Propagation probability calculation unit 80>
In the propagation probability calculation unit 80, the closeness degree calculation unit 70 and the familiarity w _{uu ',} the user attribute and the attribute set A _Su points user set S _u from the storage unit 10, close to the user set and the user u from the conversation log storage unit 20 and the conversation log set _L uSu of S _u, from the attribute probability storage unit 50 attribute probability P | accepts as input (a _ij f). Then, after calculating the probability P (f | u, u ′) that the feature quantity f is included in the conversation between the user u and each neighboring user u ′ based on the conversation log set L _uSu (step 34 in FIG. 3, step 34) 35) The propagation probability P (a _ij | u, u ′) is calculated (step 36 in FIG. 3). By propagating attributes from the neighboring user set _Su based on the propagation probability, the probability P (a _ij | u) that the input user u has each attribute is calculated and output to the output unit 90 (step of FIG. 3). 37).

  The following method can be considered as a method of calculating the probability P (f | u, u ′) that the feature quantity f is included in the conversation between the user u and each neighboring user u ′.

ただし、C(f,u,u')は、ユーザu,u'の会話において特徴量fが含まれた回数を表す。

Here, C (f, u, u ′) represents the number of times the feature quantity f is included in the conversation of the user u, u ′.

  The following method etc. can be considered as a method for calculating the propagation probability.

伝搬を行う際には、下式のようにユーザu'が持つ属性値a_ijについて入力ユーザuが各属性を持つ確率P(a_ij|u)を算出する方法などが考えられる。

When performing propagation, a method of calculating the probability P (a _ij | u) that the input user u has each attribute for the attribute value a _ij that the user u ′ has, as shown in the following equation, can be considered.

この際、ユーザu'が持つ全ての属性値を伝搬確率P(a_ij|u)に基づき伝搬してもよく、ユーザu'が持つ属性値のうち、最も伝搬確率P(a_ij|u)が高い属性値に絞って伝搬してもよい。

At this time, all attribute values possessed by the user u ′ may be propagated based on the propagation probability P (a _ij | u), and among the attribute values possessed by the user u ′, the propagation probability P (a _ij | u) is the highest. May be propagated with a high attribute value.

<Output unit 90>
In the output unit 90, the probability that the input user u has each attribute is received from the propagation probability calculation unit 80 as an input. Of the propagation attribute values a _ij, and outputs the most probable attribute value of the attribute values a _ij included in the attribute name a _i as attribute values a _ij in the attribute name a _i of the user u (FIG. 3 Step 38).

  Here, output is a concept including display on a display, printing on a printer, sound output, transmission to an external device, and the like. The output unit 90 may or may not include an output device such as a display or a speaker. The output unit 90 can be realized by driver software of an output device or driver software of an output device and an output device.

(Concrete example)
Hereinafter, the processing of the present embodiment will be described using a specific example. Here, a specific example under the following conditions will be described.

・ Input is input user u ・ Attribute name a _i includes occupation and residence ・ Attribute value a _{occupation j} includes students and office workers ・ Attribute value a _{residence j} includes Tokyo and Kanagawa Included-User attribute storage unit 10 stores profile-Conversation log includes morphologically analyzed post content, time information, and friend information In user set extraction unit 30, the word "student" appears in the profile The user group U _{occupation: students} whose _occupation attribute values are students is extracted, and similarly, U _{occupation: company employee} , U _{residence: Tokyo} , U _{residence: Kanagawa} are extracted. Then, the conversation log storage unit 20 is inquired, and user conversation logs included in each user set are extracted. By examining the features included in the extracted conversation logs, U _{occupation: students} ' conversation logs contain significant words such as `` final exam '' and `` club activities'' The feature amount “posting is performed” is obtained. For example, if there are 5,000 users who use the word “final examination”, and the word “final examination” is used, and there are 500 people who are occupation: students, then P (occupation: student | Final exam) can be calculated as 0.1. Then, the probability P (a |) that the user u has an attribute a by multiplying the probability P (f | u) including the feature quantity f by the probability P (a | f) that the feature quantity f belongs to the attribute a. u) can be calculated. For example, if a user uses the word “final exam” once in 100 times, 0.01 × 0.1 = 0.001. What is necessary is just to add this about all the feature-value sets F. By calculating the probability that each user belongs to U _{occupation: student} by the above process when each feature quantity is independent, even if the word “student” is not included in the profile, U _{occupation: student} is obtained. In addition, even if the profile includes the word “student”, it is possible to exclude U _{occupation: non-student} users.

Next, the attribute probability calculation unit 40 inquires of the conversation log storage unit 20 from the user set obtained in the previous step, and extracts conversation logs between users included in U _{occupation: student} . By examining the amount of features included in the conversation between users, U _profession: Conversations between _students include words such as “Cultural Festival” and emoticons significantly, as well as “small number of characters” The feature amount such as “the number of times the reply continues” is obtained. Then, similarly to the user set extraction unit 30, the probability P (a | f) that the feature quantity f belongs to the attribute a is calculated and stored in the attribute probability storage unit 50.

In closeness degree calculation unit 70, contact the conversation log, extracts friend set S _u of the user u, calculates the respective closeness with neighboring users u '. The calculation of the intimacy is performed by adding all the common friends of the two users, which are the number of conversations between the common friends divided by the number of hops. For example, the common friends are u _a and u _b , user u and user u ′ are 5 times, user u and user u _a are 3 times, user u ′ and user u _a are 1 time, user u and user u are _{If b} is 4 times and user u ′ and user u _b are talking 6 times, 5+ (3 + 1) / 2 + (4 + 6) / 2 = 12. By normalizing this, for example, it is assumed that it becomes 0.6.

The propagation probability calculation unit 80 calculates which attribute value should be propagated among the attribute values possessed by the user u ′ between the user u and the user u ′. First, the user attribute storage unit 10 is inquired to extract the attributes of the friend set _Su . Contact the conversation log storage unit 20, to extract a conversation log of user u and a friend set S _u. P (Occupation: Student | u) adds the attribute probabilities for each feature such as P (Occupation: Student | Talk at 16:00) P (Occupation: Student | Talk at 16:00) for all features. calculate. Then, among the probability values for each attribute value, only the highest value between the user u and the user u ′ is propagated from the user u ′ to the user u. For example, P (Occupation: Student | u, u ') is 0.8, P (Occupation: Company employee | u, u') is 0.2, P (Residence: Tokyo | u, u ') is 0.4, P (Residence) : Kanagawa | u, u ') is 0.6. In this case, even if only the method with the highest probability value (Profession: Student) is propagated and the next highest probability value (Residence: Kanagawa) is not propagated at all, all attributes are propagated based on probability values But you can. At the time of propagation, the intimacy calculated previously and the probability value are multiplied and added up for all users. That is, the probability that the attribute value of occupation: student is propagated from the user u ′ to the user u is 0.6 × 0.8 = 0.48. This can be added for all users, and as a result, P (occupation: student | u) is 0.7. Similarly, P (occupation: company employee | u) can be calculated as 0.3, P (residence: Tokyo | u) can be calculated as 0.6, and P (residence: Kanagawa | u) can be calculated as 0.4.

The output unit 90 outputs the attribute value a _ij of each attribute name a _i having the highest probability. For example, since the user u's occupation has a higher probability of the student attribute value than the office worker, the user u's occupation is output as a student. By the above method, even if user u lives in Tokyo but attends school in Kanagawa, even if there are many friends who live in Kanagawa, only the attribute value of school is propagated from those friends The correct attribute value of Tokyo can be estimated by propagating the attribute value of Tokyo from a friend of the school who went to the past.

  The present invention is not limited to the above-described embodiments, and various modifications and applications are possible within the scope of the claims.

10 user attribute storage unit 20 conversation log storage unit 30 user set extraction unit 40 attribute probability calculation unit 50 attribute probability storage unit 60 input unit 70 familiarity calculation unit 80 propagation probability calculation unit 90 output unit 100 user attribute estimation device 110 learning unit 120 Reasoning department

Claims (7)

Based on the conversation log between users stored in the conversation log storage means and the attribute set of the user set in which the user attributes stored in the user attribute storage means are known, the input user whose user attributes are unknown A user attribute estimation device for estimating a user attribute,
User set extraction means for extracting a user set having a specific attribute from the user set stored in the user attribute storage means;
Based on the user set having the specific attribute and each conversation log stored in the conversation log storage unit, the feature quantity characteristically expressed in the user set having the specific attribute is extracted, and each feature quantity is each user. Calculating an attribute probability that is a probability belonging to the attribute, and storing the attribute probability in the attribute probability storage means;
A closeness calculating means for calculating a closeness between the input user and each neighboring user based on each conversation log stored in the conversation log storage means;
Propagation probability calculating means for calculating a propagation probability for propagating each attribute of a neighboring user to the input user based on the attribute probability of each feature quantity stored in the familiarity and the attribute probability storage means;
The user attribute estimation apparatus characterized by having. Based on the result calculated by the propagation probability calculating means, output means for outputting the attribute value having the highest propagation probability among the plurality of attribute values included in the attribute name as the attribute value in the attribute name of the input user. The user attribute estimation device according to claim 1, further comprising: The propagation probability calculation means calculates the propagation probability by multiplying the probability obtained by multiplying the probability that the feature amount is included in the conversation between the input user and each neighboring user and the attribute probability of the feature amount by the familiarity. The user attribute estimation apparatus according to claim 1 or 2, wherein Based on the conversation log between users stored in the conversation log storage means and the attribute set of the user set in which the user attributes stored in the user attribute storage means are known, the input user whose user attributes are unknown A user attribute estimation method executed by a user attribute estimation device for estimating a user attribute,
A user set extraction step of extracting a user set having a specific attribute from the user set stored in the user attribute storage means;
Based on the user set having the specific attribute and each conversation log stored in the conversation log storage unit, the feature quantity characteristically expressed in the user set having the specific attribute is extracted, and each feature quantity is each user. Calculating an attribute probability that is a probability belonging to the attribute, and storing the attribute probability in the attribute probability storage means;
A familiarity calculating step of calculating a familiarity between the input user and each neighboring user based on each conversation log stored in the conversation log storage means;
A propagation probability calculating step of calculating a propagation probability for propagating each attribute of a neighboring user to the input user based on the attribute probability of each feature quantity stored in the familiarity and the attribute probability storage means;
A user attribute estimation method characterized by comprising: Based on the result calculated by the propagation probability calculating step, an output step of outputting the attribute value having the highest propagation probability among the plurality of attribute values included in the attribute name as the attribute value in the attribute name of the input user. 5. The user attribute estimation method according to claim 4, further comprising: In the propagation probability calculation step, the user attribute estimation device multiplies the intimacy by a probability obtained by multiplying a probability that a feature amount is included in a conversation between the input user and each neighboring user and an attribute probability of the feature amount. The user attribute estimation method according to claim 4 or 5, wherein a propagation probability is calculated by:   A program for causing a computer to function as a user set extraction unit, an attribute probability calculation unit, a closeness calculation unit, and a propagation probability calculation unit in the user attribute estimation apparatus according to any one of claims 1 to 3.

Images