JP2011238169A - User attribute estimating device, user attribute estimating method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a user attribute estimating device, a user attribute estimating method and a program capable of efficiently acquiring user attributes.SOLUTION: The user attribute estimating device estimates user attributes and comprises: a known user attribute storage unit 505 for storing user attributes which are already known; a model generating unit 500 for learning a relation between geographic attributes and the user attributes based on known user attributes stored in the known user attribute storage unit 505 and generating a learning model; and a user attribute estimating unit 600 for estimating unknown user attributes from the geographic attributes based on the learning model generated by the model generating unit 500.

Description

  The present invention relates to a user attribute estimation device, user attribute, and the like used for efficiently describing characteristics of places such as stations and towns in ICT (Information and Communication Technology) services for outdoor spaces such as advertisement distribution services for digital signage. The present invention relates to an estimation method and a program.

  In recent years, ICT services for outdoor spaces have become popular. For example, an ICT service that acquires location information and shares it with an acquaintance, and an ICT service that automatically selects recommended information according to location information and distributes it to a mobile phone as well as acquisition of location information are known. .

  In particular, the number of digital signage (digital signage) is increasing and the progress of networking is remarkable. Advertising distribution in digital signage can be optimized according to time and place. As the number of digital signage installations increases, it is difficult to optimize everything manually, so it is necessary to automatically optimize advertisement distribution to some extent. Information about the advertisement can be obtained from the advertiser, but it is difficult to obtain information about the place. Therefore, it is necessary to automatically describe and store the characteristics of places such as digital signage installation places (hereinafter referred to as “location characteristics”) so that they can be used in the system.

  A place feature is composed of a geographic attribute and a user attribute. Geographic attributes refer to information indicating geographical features such as buildings and natural environments, such as information about facilities around the place and information about weather. The user attribute is information indicating typical characteristics of a group such as basic attributes and behavioral characteristics of people who come and go around the place. User attributes have several items such as age, purpose of use, and accompanying persons. At present, it is necessary to obtain user attributes at a high cost by investigation or the like, and it is desired to automate this. Examples of conventional techniques applicable to the description of user attributes include the following.

  That is, there is known a video analysis technique that analyzes a video of a camera installed in a certain place and estimates attributes such as age and sex of people who come and go. For example, Non-Patent Document 1 discloses a technique for estimating the approximate number of people (congestion degree) from an image. A position information analysis technique for analyzing data acquired using a GPS (Global Positioning System) compatible terminal and extracting a behavior flow line is also known (see Non-Patent Document 2).

Tetsuya Tsuji et al., "Advertising Effect Measurement Technology by Image Processing", NTT Technical Journal 2009.7, P16-19 Masanobu Nishino et al., “Examination of Behavior Pattern Extraction Method from Location Information”, IPSJ SIG Technical Report, 2008-UBI-20 (10) 2008/11/13, P57-64

  However, according to the conventional video analysis technique, it is necessary to install a camera, and according to the conventional position information analysis technique, the user needs to carry the GPS terminal and disclose the position information. That is, according to the prior art, there are restrictions on equipment and the user side, so that it is difficult to obtain quantitatively sufficient data, and it is difficult to acquire user attributes for many places.

  An object of this invention is to provide the user attribute estimation apparatus, the user attribute estimation method, and program which can acquire a user attribute efficiently in view of the prior art mentioned above.

  In order to achieve the above object, an invention according to a first aspect is a user attribute estimation device for estimating a user attribute, comprising: a known user attribute accumulation unit that accumulates a known user attribute; and the known user attribute accumulation unit. A model generation unit that generates a learning model by learning the relationship between a geographical attribute and a user attribute based on the accumulated known user attribute, and an unknown user from the geographical attribute based on the learning model generated by the model generation unit The gist of the invention is that it includes a user attribute estimation unit that estimates an attribute.

  The invention according to the second aspect is the invention according to the first aspect, wherein the geographical attribute items are accumulated in a tree structure from the viewpoint of the relevance to the user attribute, and the reference numbers of the respective nodes constituting the tree structure are accumulated. A geographic attribute item structure storage unit, and a user attribute-geographic item structure relationship storage unit that stores information indicating a correspondence relationship between a user attribute item to be estimated and a reference number of the node, and the model generation unit includes: , Obtaining a reference number of the node corresponding to the user attribute item to be estimated with reference to the user attribute-geographic item structure relationship accumulation unit, and storing the geographic attribute item structure based on the obtained reference number of the node The gist is to select variables used in the learning model by selecting geographic attribute items accumulated in the section.

  The invention according to the third aspect is the invention according to the first or second aspect, wherein the user attribute is a supply variable group that indicates what the place is generally to do and a context variable group that changes with time. The main point is to depend on

  In order to achieve the above object, the invention according to the fourth aspect is a user attribute estimation method for estimating a user attribute, wherein a known user attribute accumulation step of accumulating a known user attribute in a known user attribute accumulation unit; A model generation step of generating a learning model by learning a relationship between a geographic attribute and a user attribute based on a known user attribute stored in the known user attribute storage unit, and a learning model generated in the model generation step. And a user attribute estimation step for estimating an unknown user attribute from a geographic attribute.

  The invention according to the fifth aspect is the invention according to the fourth aspect, wherein the geographical attribute items are accumulated in a tree structure from the viewpoint of the relevance with the user attribute, and the reference numbers of the respective nodes constituting the tree structure are A user who accumulates in the user attribute-geographic item structure relationship accumulation unit information indicating the correspondence between the user attribute item to be estimated and the reference number of the node, and the geographic attribute item structure accumulation step to be accumulated in the attribute item structure accumulation unit An attribute-geographic item structure relationship accumulation step, wherein the model generation step refers to the user attribute-geographic item structure relationship accumulation unit to obtain a reference number of the node corresponding to the user attribute item to be estimated By selecting a geographic attribute item stored in the geographic attribute item structure storage unit based on the acquired reference number of the node, a learning model And it is required to sift through the variables that are needed.

  The invention according to a sixth aspect is the invention according to the fourth or fifth aspect, wherein the user attribute is a supply variable group indicating what the place generally does, and a context variable group that changes with time. The main point is to depend on

  In order to achieve the above object, an invention according to a seventh aspect is a program for estimating a user attribute, a known user attribute accumulation step of accumulating a known user attribute in a known user attribute accumulation unit, A model generation step of generating a learning model by learning a relationship between a geographical attribute and a user attribute based on a known user attribute stored in the known user attribute storage unit, and based on the learning model generated in the model generation step The gist of the present invention is a program for causing a computer to execute a user attribute estimation step of estimating an unknown user attribute from a geographic attribute.

  ADVANTAGE OF THE INVENTION According to this invention, the user attribute estimation apparatus, the user attribute estimation method, and program which can acquire a user attribute efficiently can be provided.

It is a block diagram of the user attribute estimation apparatus in embodiment of this invention. It is a figure which shows the example of the location feature in embodiment of this invention. It is explanatory drawing of the user attribute in embodiment of this invention. It is a figure which shows the example of the data accumulate | stored in the geographical attribute item structure storage part in embodiment of this invention. It is a figure which shows the example of the data accumulate | stored in the user attribute-geographic item structure relationship storage part in embodiment of this invention. It is a figure which shows the example of the data accumulate | stored in the parameter | index selection rule storage part in embodiment of this invention. It is a figure which shows the example of the data accumulate | stored in the known user attribute storage part in embodiment of this invention. It is a flowchart which shows operation | movement of the user attribute estimation apparatus in embodiment of this invention. It is a flowchart which shows the step which produces | generates the learning model in embodiment of this invention in detail.

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

  First, an outline of an embodiment of the present invention will be described. In the embodiment of the present invention, the user attribute is estimated using the relationship between the geographic attribute and the user attribute. That is, the geographic attribute can be generated at a lower cost than the user attribute by referring to various DBs. Therefore, using a known user attribute, a relationship model between a geographic attribute and a user attribute is constructed using a technique such as learning, and an unknown user attribute is estimated based on the learning model.

  Here, since there are many candidates for variables (geographic attribute items) used in the learning model, it is necessary to select variables that are effective for estimation. However, it is difficult to collect samples (known user attribute data) used for learning etc. in the order of thousands to tens of thousands. Therefore, a learning model must be constructed from a small number of samples. In order to prevent overlearning or the like, it is better that the number of variables used in the learning model is small, and it is necessary to select a small number of variables by selecting from a large number of variable candidates. This is because when all combinations are examined, the amount of calculation becomes enormous.

  Therefore, in order to efficiently select variables, the geographical attribute items are described as a tree structure from the viewpoint of the relationship with the user attributes. Geographic attribute items that are considered to have similar effects on user attribute estimation are placed under the same node, and at the time of variable selection, only one or a few variables are selected from the same node. As a result, it is possible to efficiently select variables that are effective for estimation. The correspondence between the user attribute item to be estimated and the tree portion to be selected is determined in advance based on a hypothesis or verification result of the relationship between the user attribute and the geographic attribute. An example of this hypothesis will be described later.

  FIG. 1 is a configuration diagram of a user attribute estimation apparatus according to an embodiment of the present invention. This user attribute estimation device is a device that is used to efficiently describe a location feature (see FIG. 2) such as a station or a city in an outdoor space ICT service. A geographic attribute item structure storage unit 401, a user attribute-geographic item structure relationship storage unit 402, an index selection unit 404, an index selection rule storage unit 405, a model generation unit 500, and a location ID-position conversion unit 504. A known user attribute storage unit 505, an input unit 51, a user attribute estimation unit 600, an input unit 61, and an output unit 62. The model generation unit 500 includes a variable selection unit 501, an algorithm selection unit 502, and a model storage unit 503. The user attribute estimation unit 600 includes a model selection unit 601 and an estimation calculation unit 602.

  The index accumulation unit 31 accumulates various external DBs (databases) such as store information, land price information, transfer guidance information, event information, weather information, time information, day information (hereinafter referred to as “index”). The geographic attribute item structure accumulating unit 401 accumulates geographic attribute items in a tree structure from the viewpoint of the relevance with the user attributes, and accumulates reference numbers of the nodes constituting the tree structure. The user attribute-geographic item structure relationship storage unit 402 stores information indicating a correspondence relationship between a user attribute item to be estimated and a node reference number (hereinafter referred to as “node number”). The index selection unit 404 refers to the index selection table stored in the index selection rule storage unit 405, and is necessary for calculation of each variable group of the geographic attribute item structure storage unit 401 and the user attribute-geographic item structure relationship storage unit 402. An index is acquired from each DB of the index accumulation unit 31. The index selection rule accumulation unit 405 accumulates a correspondence table of index types and acquisition destination addresses necessary for calculating the supply and context variable groups of the geographic attribute item structure accumulation unit 401 and the user attribute-geographic item structure relation accumulation unit 402. . The model generation unit 500 generates a learning model by learning the relationship between the geographical attribute and the user attribute based on the known user attribute stored in the known user attribute storage unit 505. Specifically, the node number corresponding to the user attribute item to be estimated is acquired by referring to the user attribute-geographic item structure relation storage unit 402, and the geographic attribute item structure storage unit 401 is based on the acquired node number. By selecting the accumulated geographic attribute items, the variables used in the learning model are selected. As already described, the model generation unit 500 includes a variable selection unit 501, an algorithm selection unit 502, and a model storage unit 503. The variable selection unit 501 determines a combination of variable groups from the supply and context variable groups used for learning model construction with reference to the user attribute-geographic item structure relation accumulation unit 402. In addition, the index constituting the supply and context variable group is also selected. The algorithm selection unit 502 selects a function / algorithm used for learning model construction. The model storage unit 503 stores the model constructed by the variable selection unit 501 and the algorithm selection unit 502. The place ID-position conversion unit 504 converts the place ID used in the user attribute into position information (address, latitude / longitude, etc.). The known user attribute storage unit 505 stores known user attributes and their location IDs. The input unit 51 inputs user attribute items for which a learning model is to be generated. The input unit 61 inputs a place to be estimated, a user attribute item name, and a day / time. The user attribute estimation unit 600 estimates an unknown user attribute from the geographic attribute based on the learning model generated by the model generation unit 500. As already described, the user attribute estimation unit 600 includes a model selection unit 601 and an estimation calculation unit 602. The model selection unit 601 selects a learning model that matches the input from the input unit 61 from the learning models stored in the model storage unit 503. The estimation calculation unit 602 acquires necessary variables from the geographic attribute item structure storage unit 401 and the user attribute-geographic item structure relationship storage unit 402 via the model selection unit 601 and the model generation unit 500, and estimates the corresponding user attribute. Do. The output unit 62 outputs the estimation result of the estimation calculation unit 602.

  Next, the tree structure of geographic attribute items will be described. The configuration of the tree can be determined by the following hypothesis, for example.

  That is, since the user attribute is used for advertisement distribution, a business scene, leisure scene, home life scene, and the like can be assumed as the advertisement contact scene. For example, there are scenes such as “watching an advertisement while moving on work” and “watching an advertisement on the way to play on a holiday”. The difference in scene is considered to be related to what the place is generally doing and factors other than the place such as the time zone. As shown in FIG. 3, each of the supply variable group and the context variable group is shown in FIG. It is expressed by the two variable group. For example, whether or not there is a large number of offices that are likely to visit during work is represented by a supply variable, and whether or not it is a time zone with a high possibility of working is represented by a context variable.

  Here, the constituent variables of the supply variable group will be described. The supply variable group represents what the place is typically doing. As described above, for example, a leisure scene, a business scene, and a family life scene are assumed as advertisement contact scenes, and a large number of "purchase opportunities", "entertainment opportunities", "business opportunities", and "residential opportunities" are expressed. The supply variable group may be expressed using the following four types of indices.

supply variable group ← {Purchase index, entertainment index, business index, residence index}
In the following, examples of necessary indices are shown for obtaining each index. In calculating the index, a purchase index, an entertainment index, a business index, and a residence index that can be prepared are prepared, and the variable selection unit 501 selects several optimal ones. The purchase index represents the annual sales of the retail business at the location, the number of stores in the retail chain, the average land price of commercial land, and so on. The entertainment index represents the number of restaurants open at that location, the presence of large-scale entertainment facilities such as an amusement park, and the like. The business index represents the number of employees at the location, the number of large business establishments, the number of employees per establishment, and the like. The residence index represents the number of households in the place, the average number of people per household, the total land area, and the like. These indicators can be obtained from commercially available DBs, API (Application Program Interface) services, administrative-issued statistical materials, and published land prices.

  Next, configuration variables of the context variable group will be described. The context variable group is composed of variables that change with time, which cannot be expressed by supply variables. As described above, for example, a leisure scene, a business scene, or a family life scene is assumed as an advertisement contact scene, and “day of the week” or “time zone”, which is effective for separating the leisure scene from the business scene, what kind of leisure scene The context variable group may be composed of time-variable variables such as “meteorological information” and “event information” that are effective for determining whether the scene is a leisure scene (such as purchase or entertainment).

Context variable group ← {Day of the week, time zone, weather information, event information}
In the following, examples of necessary indices are shown for obtaining each index. In calculating the index, the day of the week, the time zone, the weather information, and the event information that can be prepared are prepared, and the variable selection unit 501 selects several optimal ones. The day of the week represents a weekday or holiday. The time zone represents morning (6: 00-11: 00: 00), daytime (11: 00-17: 00), night (17: 00-24: 00), and the like. The event information represents the event holding time zone, end time zone, event genre, etc. in the vicinity. The weather information represents the expected weather and temperature in each time zone. Among these indicators, the day of the week and the time zone are calculated from the observation time of the user attribute, and the others can be acquired by referring to the weather information providing service, and the event information of each event hosting organization and event venue.

  FIG. 4 is a diagram illustrating an example of data stored in the geographic attribute item structure storage unit 401. As shown in this figure, each node constituting the tree structure is given a node number. Here, node number 1 for the supply variable group, node number 2 for the context variable group, node number 11 for the purchase index, node number 12 for the entertainment index, node number 13 for the business index, and node number for the residence index It is assumed that node number 21, node number 21 for day of the week, and time zone, node number 22 for weather information, and 23 for event information, respectively.

  FIG. 5 is a diagram illustrating an example of data stored in the user attribute-geographic item structure relationship storage unit 402. As shown in this figure, a user attribute-geographic item structure relationship storage unit 402 stores information indicating a correspondence relationship between a user attribute item to be estimated and a node number. The node number means a combination of variables used by the model generation unit 500 for variable selection. That is, the model generation unit 500 refers to the tree structure of the geographic attribute items in the geographic attribute item structure storage unit 401, and sets the geographic attribute item of the node specified by the node number as a variable selection target.

  FIG. 6 is a diagram illustrating an example of data stored in the index selection rule storage unit 405. As shown in this figure, the index selection rule accumulating unit 405 is information indicating a correspondence relationship between a user attribute item, a combination of indices necessary for calculating a supply and context variable group, and an address of an acquisition destination database of these indices. Has accumulated. The correspondence between the index and the acquisition destination is determined in advance based on the index provided by the acquisition destination. As an acquisition destination, you may use what is disclosed and sold as data, or what is provided as an API. Specifically, national census data, public land price data, house map API, or the like can be used. However, regarding the day of the week and the time zone of the context variable, the observation time zone of data provided when data is acquired from the known user attribute storage unit 505 is used.

  FIG. 7 is a diagram illustrating an example of data stored in the known user attribute storage unit 505. As shown in this figure, the known user attribute storage unit 505 stores information indicating a correspondence relationship between a place ID, a place name, an observation period (day of the week / time zone), and the user attribute. The user attribute may describe one typical value of the corresponding user attribute item for people who visit the place during the observation period.

  FIG. 8 is a flowchart showing the operation of the user attribute estimation apparatus according to the embodiment of the present invention. Hereinafter, the operation of the user attribute estimation apparatus will be described with reference to FIG.

  First, necessary information is accumulated respectively in the index accumulation unit 31, the geographic attribute item structure accumulation unit 401, the user attribute-geographic item structure relation accumulation unit 402, the index selection rule accumulation unit 405, and the known user attribute accumulation unit 505 ( S1). Next, after the model generation unit 500 generates and accumulates a learning model by a method to be described later (S2), when the input unit 61 inputs an estimation target (S3), a model is selected that matches the input from the input unit 61. The unit 601 selects from the model storage unit 503 (S4). Accordingly, when the estimation calculation unit 602 estimates the user attribute based on the selected learning model (S5), the output unit 62 outputs the estimation result (S6).

  FIG. 9 is a flowchart showing in detail the step of generating a learning model (S2 in FIG. 8). Hereinafter, a case where a learning model for estimating the user attribute item “purpose of use” is generated will be described with reference to FIG. 9.

  First, when the model generation unit 500 receives designation of “use purpose” from the user attribute items from the input unit 51 (S11), the model generation unit 500 stores the location ID of the data in which the value of “use purpose” is written and the “use” The value of “purpose” and the observation time of the data are acquired from the known user attribute storage unit 505 (S12). The value of “use purpose” may be expressed as a discrete value (label) such as “work” or “shopping”, or the ratio of people who use for the purpose of work is defined as “work level”. Then, it may be expressed as a continuous amount.

  Next, the model generation unit 500 passes the location ID to the location ID-position conversion unit 504, and acquires location information from the location ID-position conversion unit 504 (S13). The position information is matched to a system corresponding to various indexes of the index storage unit 31, such as an address and latitude / longitude.

  Next, the variable selection unit 501 refers to the user attribute-geographic item structure relation accumulation unit 402, and in the geographic attribute item tree structure, the node number “11-14, 21, 23” in the case of estimating “use purpose” Is acquired (S14). In estimating the “purpose of use”, the geographic attribute items that are candidates for variables used in the learning model are the geographic attribute items of the portion specified by the node number.

  Next, the variable selection unit 501 refers to the geographic attribute item structure storage unit 401 and determines from the nodes corresponding to the node numbers 11-14, 21 and 23 of the geographic attribute item tree structure to “Retail annual sales” and “Number of restaurants open. ”Or the like is selected as a variable candidate (S15). Then, the index selection unit 404 is notified of the geographic attribute item name selected in step S15 and the position information acquired in step S13 (S16).

  Next, the index selection unit 404 refers to the index selection rule accumulation unit 405, selects the necessary index from various DBs of the index accumulation unit 31 using the acquisition destination and position information of each geographic attribute item, and the variable selection unit It is passed to 501 (S17). However, for the node 21, the day of the week and the time zone are calculated from the observation time of each data already acquired in step S12.

  Next, the variable selection unit 501 notifies the algorithm selection unit 502 of the value of the geographical attribute item for each location obtained above (S18).

  Next, the algorithm selection unit 502 uses the part of the geographic attribute item obtained in step S18 and the value of the user attribute item “purpose of use” acquired from the known user attribute storage unit 505 to perform several algorithms. And the algorithm with the best estimation accuracy is selected (S19). Alternatively, if there is a history of model generation using the same kind of data, the same algorithm as that may be used from the beginning. On the other hand, when there is no history, if the value of “use purpose” is a discrete value such as label data, an identification algorithm such as a support vector machine may be tried, or the value of “use purpose” may be a continuous value. If there is, a highly accurate model may be obtained from multivariate analysis methods such as a non-linear regression model. Various known methods can be employed as the multivariate analysis method (Meiji Mizuno (Author), “Multivariate Data Analysis Lecture”, Asakura Shoten).

  Next, the variable selection unit 501 receives the result of step S19, refers to the geographic attribute item structure storage unit 401, selects one or several of the geographic attribute items corresponding to the node number 11, and determines the remaining nodes. Similarly, a variable is selected (S20). Then, using the algorithm / model obtained in step S19, a part of the known user attribute data stored in the known user attribute storage unit 505 is set as test data, the rest is set as training data, and a learning model is determined from the training data. Create and evaluate the performance of the corresponding learning model with test data.

  Next, the variable selection unit 501 repeats Steps S17 to S20 by changing the combination of geographic attribute items selected from each node, and determines the combination of indices having the best performance (S21). If there is a margin in the calculation amount, all combinations may be tried to obtain the most accurate combination. As a result, for example, the number of retail stores is used for the purchase index of node number 11, “there is a large-scale entertainment facility” for the entertainment index of node number 12, and “employee” is the business index of node number 13. The combination of indicators can be determined such as “number”, “number of households” for the residence index of the node number 14, and so on. As described above, the learning variable is obtained by applying the six variables such as the number of shops, the presence / absence of large-scale entertainment facilities, the number of employees, the number of households,... To the algorithm / model obtained in step S19 (S22).

  The step of obtaining the learning model is performed for all user attribute items (S23), and the learning model created for each user attribute item is accumulated in the model accumulation unit 503 (S24).

  As described above, by defining the tree structure of geographic attribute items based on the relationship with user attribute items, it is possible to efficiently and optimally select a small number of geographic attribute items from among a large number of geographic attribute items without examining combinations of all geographic attribute items. Combinations can be found.

  Next, an example in the case of estimating the user attribute of a place where the user attribute is unknown using the learning model stored in the model storage unit 503 will be described. Here, it is assumed that the purpose of use in “Jiyugaoka” is to be obtained. In this case, the spot name “Jiyugaoka” whose user attribute is unknown and the user attribute item “purpose of use” are input to the input unit 61. Thereby, an unknown user attribute is referred to from the model storage unit 503, and a model is selected and estimated according to the user attribute item. As a result, the “use purpose” of “Jiyugaoka” is estimated as “shopping” or “work rate: 21%”.

  As described above, according to the user attribute estimation device in the embodiment of the present invention, user attributes can be acquired efficiently. That is, if the user attribute can be acquired for a part of the location, the user attribute is estimated from the data called the geographic attribute that can be acquired at a lower cost than the user attribute. As a result, it is not necessary to investigate user attributes for all locations, and cost reduction can be expected.

  More specifically, create a tree structure of geographic attribute items based on the relationship between geographic attributes and user attributes, specify the user attribute items and related tree parts, and specify the combination of geographic attribute items used in the learning model. By selecting, even when the number of known user attribute data is small and the number of geographic attribute items is large, it is possible to generate a learning model by efficiently selecting geographic attribute items that are effective in estimating user attributes. it can. As a result, user attributes can be acquired for many places, and advertisement distribution optimized for the place of installation can be realized at low cost.

  Note that the present invention can be realized not only as a user attribute estimation device, but also as a user attribute estimation method using a characteristic processing unit included in such a user attribute estimation device as a step. It can also be realized as a program executed by a computer. It goes without saying that such a program can be distributed via a recording medium such as a CD-ROM or a transmission medium such as the Internet.

  Although not particularly mentioned in the above description, a known machine learning method or feature selection method can be adopted in the present invention. For example, well-known machine learning techniques include “N. Cristianini and J. Shawe-Taylor, Takeshi Ohkita,“ Introduction to Support Vector Machine ”, pp.128-149, Kyoritsu Publishing, Tokyo, 2005.” and “ The method disclosed in Keiji Yanai, "Support Vector Machine Related Tools", The Journal of the Institute of Image Information and Television Engineers, Vol. 63, No. 12, can be used. Known feature selection methods are disclosed in “Guyon, I. and Eliseeff, A.“ An introduction to variable and feature selection, ”Journal of Machine Learning Research, 3, pp1157-1182., 2003.” Can be used.

505: Known user attribute storage unit 500 ... Model generation unit 600 ... User attribute estimation unit 401 ... Geographic attribute item structure storage unit 402 ... User attribute-geographic item structure relation storage unit

Claims (7)

A user attribute estimation device for estimating a user attribute,
A known user attribute storage unit for storing known user attributes;
A model generation unit that learns a relationship between a geographical attribute and a user attribute based on a known user attribute stored in the known user attribute storage unit and generates a learning model;
A user attribute estimation unit that estimates an unknown user attribute from a geographic attribute based on the learning model generated by the model generation unit;
A user attribute estimation device comprising: A geographic attribute item structure storage unit that stores the geographical attribute items in a tree structure from the viewpoint of the relevance to the user attributes and stores the reference numbers of the respective nodes constituting the tree structure;
A user attribute-geographic item structure relationship accumulation unit that accumulates information indicating a correspondence relationship between the user attribute item to be estimated and the reference number of the node
The model generation unit acquires a reference number of the node corresponding to the user attribute item to be estimated with reference to the user attribute-geographic item structure relation storage unit, and based on the acquired reference number of the node The user attribute estimation apparatus according to claim 1, wherein a variable used in the learning model is selected by selecting a geographic attribute item stored in the geographic attribute item structure storage unit. The user attribute estimation apparatus according to claim 1 or 2, wherein the user attribute depends on a supply variable group indicating what the place is generally a place to perform and a context variable group that changes with time. . A user attribute estimation method for estimating user attributes, comprising:
A known user attribute accumulation step of accumulating known user attributes in a known user attribute accumulation unit;
A model generation step of generating a learning model by learning a relation between a geographical attribute and a user attribute based on a known user attribute stored in the known user attribute storage unit;
A user attribute estimation step for estimating an unknown user attribute from a geographic attribute based on the learning model generated in the model generation step;
A user attribute estimation method comprising: Geographic attribute item structure accumulation step for accumulating geographic attribute items in a tree structure from the viewpoint of relevance with user attributes and accumulating reference numbers of respective nodes constituting the tree structure in the geographic attribute item structure accumulation unit;
A user attribute-geographic item structure relation accumulation step for accumulating information indicating the correspondence between the user attribute item to be estimated and the reference number of the node in the user attribute-geographic item structure relation accumulation unit
In the model generation step, the node reference number corresponding to the user attribute item to be estimated is acquired with reference to the user attribute-geographic item structure relation storage unit, and the reference number of the node is acquired based on the acquired reference number of the node. 5. The user attribute estimation method according to claim 4, wherein a variable used in the learning model is selected by selecting a geographic attribute item stored in the geographic attribute item structure storage unit. 6. The user attribute estimation method according to claim 4 or 5, wherein the user attribute depends on a supply variable group that represents what the place generally does, and a context variable group that changes with time. . A program for estimating user attributes,
A known user attribute accumulation step of accumulating known user attributes in a known user attribute accumulation unit;
A model generation step of generating a learning model by learning a relation between a geographical attribute and a user attribute based on a known user attribute stored in the known user attribute storage unit;
A user attribute estimation step for estimating an unknown user attribute from a geographic attribute based on the learning model generated in the model generation step;
A program that causes a computer to execute.

Images