JP5801242B2 - Estimated interest score database generation apparatus, method, and program - Google Patents

Description

  The present invention relates to an estimated interest score database generation apparatus, method, and program in the field of information search, and in particular, a searcher can search a geographic range by inputting a map display range and latitude / longitude information. In a document search service, the present invention relates to an estimated interest score database generation apparatus, method, and program having a function of supporting a searcher's search by displaying characteristic keywords under the conditions before the search.

  Conventionally, there is a method of identifying place name information by analyzing text in a document (see, for example, Non-Patent Document 1). By using the result of the method, it is possible to analyze which region each document relates to. By using this, it is possible to analyze in which document a keyword set prepared in advance appears in a document related to it, and it is considered possible to extract a recommended keyword in a specific region. The specific method is to divide the region by fixed values such as 200m each in east / west / north / south and every 8 seconds in latitude / longitude (hereinafter referred to as “mesh”). Analyze frequency. In a certain area including a plurality of meshes, if the frequency of a certain keyword is characteristically high in the overall frequency distribution, it can be determined that the keyword is a keyword to be recommended in the area. In the same way as described above, there is a method for determining whether a characteristic is characteristic when the appearance frequency in the mesh is higher than the average appearance frequency in all meshes by 3σ (σ is a standard deviation) or more. Alternatively, Poisson probability can be used.

  The method based on the document analysis described above does not consider information on how much the user is interested in the keywords displayed in the area. If a history that a user has selected a keyword presented in a certain area is available, consider using this to perform keyword recommendation reflecting the degree of interest of the user. For example, as shown in FIG. 1, a large number of keyword selection logs can be used as a two-dimensional coordinate histogram by regarding a click log for a keyword as an interest count for the viewing range of the keyword. By using this histogram information, it is possible to realize keyword recommendation reflecting the past user's interest using the user history for the current viewing range of the user.

Toru Hirano, Yoshihiro Matsuo, Genichiro Kikui, "Resolving ambiguity of place names using geographical distance", 70th National Convention of Information Processing Society, 2008.

  However, as shown in FIG. 2 and FIG. 3, the user browses at various scales when using the system, and selects keywords, so that the viewing range with different scales is handled as it is. Cannot be appropriately extracted from the history (Problem 1).

  In addition, when the keyword selection history of the user is small in a certain area, as shown in FIG. 4, a portion where the histogram frequency information becomes discontinuous occurs, and the area in which the user is interested cannot be estimated appropriately ( Problem 2).

  In addition to this, it is conceivable that there are a plurality of regions in which the user is interested in the keyword. However, since this number varies depending on the keyword, an appropriate number of interest distributions cannot be obtained (Problem 3).

  Due to the above three problems, there is a problem that when a user's selection history for a keyword is used, the degree of interest is appropriately reflected to the user and keyword recommendation cannot be performed.

  The present invention has been made in view of the above points, and an estimated interest score database generating device capable of generating an estimated interest score database in consideration of the influence of a click log according to a user's viewing range, and An object is to provide a method and a program.

In order to solve the above problems, the present invention (Claim 1) provides an estimated interest degree for estimating the degree of interest of a user in a document search service for searching a geographical range based on a map display range and latitude / longitude information. A score database generation device,
For each keyword, a mesh ID for a rectangle obtained by dividing the map, and a click frequency histogram storage means for storing a click frequency for the mesh ID;
By obtaining the mesh ID and click frequency of the keyword from the click frequency histogram storage means and estimating the two-dimensional normal distribution, the two-dimensional normal distribution parameter for the keyword and the mixture ratio of the two-dimensional normal distribution in the keyword Interest distribution calculation means for estimating and storing the interest distribution storage means,
Based on the two-dimensional normal distribution parameter acquired from the interest distribution storage means, the integrated value of the probability density corresponding to each mesh ID is calculated, and the keyword record corresponding to the mesh ID in the interest score database is calculated. And an interest score database generating means for obtaining an interest distribution by adding.

Further, the present invention (Claim 2) includes a keyword click log storage means for storing a keyword and a geographic range in which the keyword is clicked,
The keyword and the geographic range are acquired from the keyword click log storage means, a score is calculated according to the size of the geographic range, a set of mesh IDs corresponding to the geographic range is obtained , and the set of mesh IDs Click frequency histogram generating means for adding the score to the click frequency column of the record that matches the keyword and mesh ID in the click frequency histogram storage means for each mesh ID included in
It has further.

  In the present invention (Claim 3), in the interest distribution calculation means, the mesh ID read from the click frequency histogram storage means is converted into point information on two-dimensional coordinates, and the point information is converted into a data point set. Add two-dimensional normal distribution parameters in the two-dimensional space based on the point set on the two-dimensional coordinates composed of the data point set, and obtain the mixture number of the two-dimensional normal distribution and the mixture ratio of each distribution And means for storing the two-dimensional normal distribution parameter corresponding to the number of mixtures and the mixture ratio in the interest distribution storage means.

  As described above, according to the present invention, it is possible to consider the influence of the click log according to the viewing range of the user, and it is possible to reduce the adverse effects of the user's click in a wide range. In addition, by estimating a two-dimensional normal distribution from data, it is possible to eliminate discontinuous points that occur when the number of click logs for a keyword is small. In addition, the number of user interest distributions can be appropriately estimated from the data for the keyword. This makes it possible to estimate the degree of interest of the user with high accuracy using the user's selection history for the keyword.

It is the click frequency of the keyword in the browsing range. It is an example (the 1) of the keyword selection in the browsing range from which a scale differs. It is an example (the 2) of the keyword selection in the browsing range from which a scale differs. It is an example (non-continuous) when there are few keyword selection histories. It is a structural example of the presumed interest score DB production | generation apparatus in one embodiment of this invention. It is an example of data of keyword click log DB in one embodiment of the present invention. It is an example of data of click frequency histogram DB in one embodiment of the present invention. It is an example of addition to mesh ID by the click in one embodiment of this invention. It is a flowchart of the process of the click frequency histogram production | generation part in one embodiment of this invention. It is a flowchart of the process of the interest distribution estimation part in one embodiment of this invention. It is an example of data of interest distribution DB in one embodiment of this invention. It is an example of interest score DB in one embodiment of this invention. It is a flowchart of the process of the interest score generation part in one embodiment of this invention.

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

  FIG. 5 shows a configuration example of the estimated interest score DB generation apparatus according to the embodiment of the present invention.

  The estimated interest score DB generation device shown in FIG. 1 includes a keyword click log DB 10, a click frequency histogram generation unit 20, a click frequency histogram DB 30, an interest score D positive generation processing unit 100, and an interest score DB 70.

  As shown in FIG. 6, the keyword click log DB 10 holds information on what geographic range the keyword was clicked on, and is composed of a keyword column and a geographic range column. Here, as the geographic range, for example, a rectangular range of a screen on which the user has browsed the keyword can be used, and in the case of the rectangular range, coordinates of two points on the lower left and upper right can be expressed.

  The click frequency histogram generation unit 20 reads the keyword click log DB 10 and outputs a click frequency histogram DB 30.

  The click frequency histogram DB 30 handles geographic information in units of meshes. Here, for example, the mesh is obtained by dividing a map into 200 m squares. It is assumed that a method for converting to a mesh ID (set) corresponding to given geographical range information is given. An example of data in the click frequency histogram DB 30 is shown in FIG. Here, it represents how many click frequencies a keyword with one record has acquired in a certain mesh ID. Specifically, as shown in FIG. 8, when a keyword on a certain mesh is clicked, the number of clicks is counted for each mesh ID of the mesh.

  FIG. 9 shows a processing flow of the click frequency histogram generation unit in one embodiment of the present invention.

  Step 101) The click frequency histogram generation unit 20 acquires an unprocessed record from the keyword click log DB 10, and sets the keyword to w and the geographic range to b.

Step 102) The addition score s is calculated according to the size of the geographic range b. For example, the reciprocal of the area of the geographic range b is used for the calculation of s. When _{b = (lon 1, lat 1} , lon 2, lat 2) that two sets of latitude and longitude coordinates (in seconds) is composed of, for example, in Tokyo Datum, since one second is converted to 25 m, The score based on area is

で計算することができる。ここで、mesh_sideはメッシュの一辺の大きさ（m単位）を表し、最小範囲が1になる正規化項の役目をしている。

Can be calculated with Here, mesh_side represents the size (m unit) of one side of the mesh, and serves as a normalization term with a minimum range of 1.

  Step 103) Acquire a set of mesh IDs corresponding to the geographical range b and set it to M.

  Step 104) For each m included in M, the score s calculated in Step 102 is added to the click frequency column of the record corresponding to the keyword w in the click frequency histogram DB 30 and the mesh ID m.

  Step 105) If there is an unprocessed record in the keyword click log DB 10, the process returns to Step 101, and if not, the process ends.

  The interest score DB generation processing unit 100 reads the click frequency histogram DB 30 and outputs the interest score to the interest score DB 70. The interest score DB generation processing unit 100 includes an interest distribution calculation unit 40, an interest distribution DB 50, and an interest score DB generation unit 60.

  The interest distribution calculation unit 40 reads the click frequency histogram DB 30 and outputs the interest distribution to the interest distribution DB 50. An interest distribution that can be added to a certain keyword is composed of a plurality of normal distributions and a mixture ratio thereof. The interest distribution DB 50 stores a distribution ID, information (parameters) of each distribution, and a mixing ratio for each keyword. The distribution ID represents a mixture distribution number, the mixture ratio represents the mixture ratio of the distribution in the keyword, and the parameter represents the normal distribution parameter (average, variance covariance).

  FIG. 10 is a flowchart of processing of the interest distribution calculation unit in the position embodiment of the present invention.

  Step 201) The interest distribution calculation unit 40 acquires an unprocessed record from the click frequency histogram DB 30 and sets it as w.

  Step 202) The unprocessed mesh ID and the click frequency are acquired from the record corresponding to the keyword w from the click frequency histogram DB 30, and the unprocessed mesh ID and the click frequency are acquired from the record corresponding to the keyword w. Are m and c.

Step 203) m is converted into point information on two-dimensional coordinates, and the coordinate values are set to (x ₁ , x ₂ ).

Step 204) The calculated coordinate values (x ₁ , x ₂ ) are added to the data set X.

  Step 205) If there is an unprocessed mesh, return to Step 202. Otherwise, go to step 206.

Step 206) Based on the information of the point set on the two-dimensional coordinates composed of the data point set, the number of parameters of the two-dimensional normal distribution and the mixing ratio thereof are estimated in accordance with the data in the two-dimensional space. For this estimation, for example, the method of the document “CE Rasmussen :,“ The Infinite Gaussian Mixture Model ”. In Proceedings of Advances in Neural Information Processing Systems 12 (NIPS 1999), pp. 554-560, 1999.” it can. If this method is used, the number k of binary normal distributions corresponding to the data set X, the mixture ratio πi of each distribution, and the respective two-dimensional normal distribution parameters (mean μ ₁ , μ ₂ , variance covariance) σ ₁₁ , σ ₁₂ , σ ₁₁ ) (when σ ₂₁ is a normal distribution, the covariance is symmetric and is not necessary).

  Step 207) Information about each distribution obtained in Step 206 is output to the interest distribution DB 50. Here, since the number of mixtures varies depending on the keyword, the distribution information is output to the interest distribution DB 50 by the number k of the mixtures obtained in step 206. An example of output to the interest distribution DB 50 is shown in FIG.

  Step 208) If there is an unprocessed keyword in the click frequency histogram 30, the process returns to Step 201. If not, the process ends.

  The interest score DB generation unit 60 reads the interest distribution DB 50 and outputs the interest score to the interest score DB 70.

  As shown in FIG. 12, the interest score DB 70 stores information on how much interest score a certain keyword has in a mesh as in the click frequency histogram DB 30.

  FIG. 13 is a flowchart of processing of the interest score generation unit according to the embodiment of the present invention.

  Step 301) The interest distribution calculation unit 40 acquires an unprocessed keyword from the interest distribution DB 50 and sets it as w.

Step 302) Unprocessed distribution information corresponding to the keyword is acquired, and the mixture ratio π and the parameters of the normal distribution are set to μ ₁ , μ ₂ , σ ₁₁ , σ ₁₂ , and σ ₁₂ . For example, in the example of FIG. 11, in the case of distribution ID1, parameters μ ₁ = 1112773, μ ₂ = 112353, σ ₁₁ = 1233, σ ₁₂ = 12453, and σ ₂₂ = 10224 are acquired.

  Step 303) In the normal distribution expressed by the parameter, a mesh ID set included inside q percentage points on both sides is acquired, and an integrated value of probability density corresponding to each mesh ID is calculated. Here, both-side q percent represents a point where the integrated value of probability density is q percent of the whole, and q is a preset value.

  Step 304) For each mesh ID included in the acquired mesh ID set, the integration value calculated in step 303 is added to the record in which the keyword of the interest score DB 70 corresponds to w.

  Step 305) If there is an unprocessed distribution in the record with the keyword w in the interest distribution DB 50, the process returns to Step 302. Otherwise, the process proceeds to step 306.

  Step 306) If there is an unprocessed keyword in the interest distribution DB 50, the process returns to Step 301. Otherwise, the process is terminated.

  As described above, click information of area tag (keyword) is converted into 2D histogram based on viewing range, and infinite mixed normal distribution estimation method on 2D coordinates (actually, it is not appropriate but suitable for data instead of infinity) By using (convergence to a large number), it becomes possible to estimate the distribution of the degree of interest that matches the data distribution, and it is possible to generate an interest score DB for the keyword.

  Note that the operation of each component of the estimated interest score DB generation device shown in FIG. 5 is constructed as a program and installed in a computer used as the estimated interest score DB generation device to be executed, or the network is It is possible to circulate through.

  The present invention is not limited to the above-described embodiment, and various modifications and applications can be made within the scope of the claims.

10 Keyword click log DB (database)
20 Click frequency histogram generator 30 Click frequency histogram DB (database)
40 Interest distribution calculator 50 Interest distribution DB (database)
60 Interest score DB generator 70 Interest score DB (database)
100 Interest score DB generation processing unit

Claims (7)

In a document search service for searching a geographical range based on a map display range and latitude / longitude information, an estimated interest score database generation device for estimating the degree of interest of a user,
For each keyword, a mesh ID for a rectangle obtained by dividing the map, and a click frequency histogram storage means for storing a click frequency for the mesh ID;
By obtaining the mesh ID and click frequency of the keyword from the click frequency histogram storage means and estimating the two-dimensional normal distribution, the two-dimensional normal distribution parameter for the keyword and the mixture ratio of the two-dimensional normal distribution in the keyword Interest distribution calculation means for estimating and storing the interest distribution storage means,
Based on the two-dimensional normal distribution parameter acquired from the interest distribution storage means, the integrated value of the probability density corresponding to each mesh ID is calculated, and the keyword record corresponding to the mesh ID in the interest score database is calculated. An interest score database generating means for obtaining an interest distribution by adding,
An estimated interest score database generation device characterized by comprising: A keyword click log storage means for storing a keyword and a geographical range where the keyword is clicked;
The keyword and the geographic range are acquired from the keyword click log storage means, a score is calculated according to the size of the geographic range, a set of mesh IDs corresponding to the geographic range is obtained , and the set of mesh IDs Click frequency histogram generating means for adding the score to the click frequency column of the record that matches the keyword and mesh ID in the click frequency histogram storage means for each mesh ID included in
The estimated interest score database generation device according to claim 1, further comprising: The interest distribution calculating means includes:
The mesh ID read from the click frequency histogram storage means is converted into point information on two-dimensional coordinates, the point information is added to a data point set, and the point set on the two-dimensional coordinates configured from the data point set On the basis of the two-dimensional normal distribution parameter in the two-dimensional space, the mixture number of the two-dimensional normal distribution and the mixture ratio of each distribution, the two-dimensional normal distribution parameter and the mixture ratio according to the mixture number, The estimated interest score database generation apparatus according to claim 1, further comprising means for storing in the interest distribution storage means. In a document search service for searching a geographical range based on a map display range and latitude / longitude information, an estimated interest score database generation method for estimating the degree of interest of a user,
The interest distribution calculation means obtains the mesh ID and click frequency of the keyword from the mesh ID for the rectangle obtained by dividing the map for each keyword and the click frequency histogram storage means for storing the click frequency for the mesh ID. An interest distribution calculating step for estimating a two-dimensional normal distribution parameter for the keyword and a mixture ratio of the two-dimensional normal distribution in the keyword by storing the distribution;
Based on the two-dimensional normal distribution parameter acquired from the interest distribution storage unit, the interest score database generation unit calculates an integrated value of probability density corresponding to each mesh ID, and the mesh of the interest score database An interest score database generation step for obtaining an interest distribution by adding to a record of a keyword corresponding to the ID;
An estimated interest score database generation method characterized by: Click frequency histogram generation means
The keyword and the geographic range are acquired from the keyword click log storage means storing the keyword and the geographic range where the keyword is clicked, and a score is calculated according to the size of the geographic range, and the mesh corresponding to the geographic range Click frequency histogram generation step of acquiring a set of IDs and adding the score to the click frequency column of a record having the same keyword and mesh ID in the click frequency histogram storage means for each mesh ID included in the set of mesh IDs The
The estimated interest score database generation method according to claim 4, further comprising: In the interest distribution calculation step,
The mesh ID read from the click frequency histogram storage means is converted into point information on two-dimensional coordinates, the point information is added to a data point set, and the point set on the two-dimensional coordinates configured from the data point set On the basis of the two-dimensional normal distribution parameter in the two-dimensional space, the mixture number of the two-dimensional normal distribution and the mixture ratio of each distribution, the two-dimensional normal distribution parameter and the mixture ratio according to the mixture number, The estimated interest score database generation method according to claim 4, which is stored in the interest distribution storage means. Computer
The estimated interest score database production | generation program for functioning as each means of the estimated interest score database production | generation apparatus of any one of Claims 1 thru | or 3.

Images