JPWO2012160693A1 - Information search method, information search device, and storage medium - Google Patents

Abstract

  The present invention relates to a computer system that searches for a product by inputting a keyword, and in particular, a product group that another user has made a purchase from the entered keyword as a starting point, and an intermediate between these product group and the keyword. It is a computer system that presents a keyword string. As a result, it is possible to link keywords that could not be easily linked by a single user to database search information (product).

Description

  The present invention relates to a computer system that searches for a product by inputting a keyword, and in particular, a product group that another user has made a purchase from the entered keyword as a starting point, and an intermediate between these product group and the keyword. The present invention relates to a technique for presenting a keyword string.

  In recent years, electronic commerce has been developed with the spread of high-speed communication lines, mobile terminals, and PCs. A consumer can search for and purchase a desired product from a favorite place at a favorite time by connecting to a site for electronic commerce. Since e-commerce is convenient for consumers, the number of participating consumers, the number of dealers and the number of products provided are enormous. Due to the increase in the number of products and the number of merchants, consumers who have participated in an e-commerce site cannot find the products they originally wanted, or take a long time to find them. For example, when a consumer searches for a part of the name of a product as a keyword, a product with a similar name is also presented as a list, and a situation occurs in which many products different from the product desired by the consumer are listed.

  In order to alleviate this situation, collaborative filtering has been proposed and put into practical use as a computer system that recommends information to users of electronic commerce. Collaborative filtering, for example, presents recommended products such as "If you bought this product, you also bought this product" when you display the searched product like Amazon.com (registered trademark) Yes. Thereby, the user of electronic commerce is more likely to select a product that better meets the needs than the product first searched.

  However, in the collaborative filtering that is the information recommendation algorithm, there is a problem that a product with a small number of purchased and evaluated users is not recommended. This is generally referred to as a sparsity problem.

  In addition, the collaborative filtering has a problem that it is difficult to recommend a product if the product has just been put on the market. That is, immediately after the sale of a product, there are many cases where there are few purchased users or there are no evaluated users, and they are not recommended targets. This is said to be a cold start problem.

  Various methods have been proposed to solve the above sparsity problem and cold start problem. For example, as in Non-Patent Document 1, a method is known that uses evaluation of products with similar properties, called content-based filtering. According to Non-Patent Document 1, it is possible to estimate the evaluation of a new product that has no evaluation from the evaluation of a product with similar properties.

JP 2010-067032 A

Mustansar Ali Ghazanfar, et al., An Improved Switching Hybrid Recommender System Using Naive Bayes Classifier and Collaborative Filtering, IMECS 2010, March, 2010, Hong Kong.

  However, in the content-based filtering of Non-Patent Document 1 described above, it is necessary to prepare the properties of the product in detail, so it takes a lot of time and the properties held by the products are not uniform. There is a problem that variations occur.

  For this reason, the non-patent document 1 is difficult to use on an electronic commerce site that provides a large number of products, and the above problems have not been solved.

  Also, a method is known in which user's click history using an electronic commerce site is utilized to estimate the user's preference from the clicked URL keyword, and to identify and recommend a product from the estimated preference. (For example, Patent Document 1). In the method of Patent Document 1, it is not necessary to use the evaluation of products with similar properties in the above content-based filtering method. Therefore, it is possible to recommend even if there is no similar product.

  However, in the above-mentioned Patent Document 1, there is a problem that URL information is not valid when the URL is not linked to the product desired by the user. In other words, if there is a gap between the information on the user side such as the URL and the information on the product item, the product desired by the user cannot be reached. This tendency is particularly noticeable for products that have just entered the market. This is due to the fact that the number of users who browse (click on) the product immediately after the sale is small, as in the conventional example.

  The present invention has been made in view of the above problems, and an object of the present invention is to link keywords and merchandise information that were not linked by one user by utilizing search histories of a large number of users.

  The present invention relates to an information search method for performing a search from a predetermined database using a received search keyword in a computer including a processor and a storage device, wherein the search keyword history received by the computer and the information in the database are stored. A first step of generating a keyword string from an access history; a second step of receiving the search keyword by the computer; and searching for the keyword string including a keyword that matches or is similar to the search keyword. A third step of extracting one or more keyword strings, and a fourth step of outputting a keyword from the keyword matching or similar to the search keyword to the information in the database among the extracted keyword strings Steps.

  According to the present invention, by providing a keyword string that reaches database information from a search keyword, a database user can obtain desired information much more easily than the conventional example. In particular, by creating a recommended keyword table using the search histories of a large number of users, it becomes possible to link keywords and database information that were not easily linked by a single user.

It is a block diagram which shows embodiment of this invention and shows an example of a computer system. It is a flowchart which shows embodiment of this invention and shows the whole process performed with a computer system. It is a screen image which shows embodiment of this invention and shows an example of a search screen when searching for merchandise information with a client terminal. It is a flowchart which shows embodiment of this invention and shows an example of the process which produces | generates a recommendation keyword. It is a figure which shows embodiment of this invention and shows an example of a user access history table. It is a flowchart which shows embodiment of this invention and shows the detail of the production | generation process of the related graph performed in FIG. It is a figure which shows embodiment of this invention and shows a mode that the log | history of the same session is extracted from a user access history table. It is a figure which shows embodiment of this invention and shows a mode that a partial graph is extracted from the session log 420A of the same session. It is a figure which shows embodiment of this invention and shows a mode that a partial graph is extracted from the session log 420B of the same session. It is a figure which shows embodiment of this invention and shows a mode that a subgraph is extracted from the session log 420C of the same session. It is a figure which shows embodiment of this invention and shows a mode that a subgraph is extracted from session log 420D of the same session. It is a figure which shows embodiment of this invention and shows a mode that a partial graph is extracted from the session log 420E of the same session. It is a figure which shows embodiment of this invention and shows a mode that a subgraph is extracted from the session log 420F of the same session. It is a figure which shows embodiment of this invention and shows an example of a related graph. It is a figure which shows embodiment of this invention and shows a mode that the probability transition matrix 250A is produced | generated from a related graph. It is a figure which shows embodiment of this invention and shows a mode that the probability transition matrix B is produced | generated from the probability transition matrix 250A. It is a figure which shows embodiment of this invention and shows a mode that the probability transition matrix C is produced | generated from the probability transition matrix 250B. It is a figure which shows embodiment of this invention and shows the example which calculates | requires an eigenvector from the probability transition matrix 250C. It is a figure which shows embodiment of this invention and shows a 1st eigenvector and a related graph. It is a figure which shows embodiment of this invention and shows a 2nd eigenvector and a related graph. It is a figure which shows embodiment of this invention and shows an example of a recommendation keyword table. It is a figure which shows embodiment of this invention and shows the relationship between a search keyword and a recommendation keyword. It is a figure which shows embodiment of this invention and shows the relationship between another search keyword and a recommendation keyword. It is a figure which shows embodiment of this invention and shows the relationship of another search keyword and a recommendation keyword. It is a flowchart of the process which shows embodiment of this invention, shows the detail of step S4 of FIG. 2, and outputs a recommendation keyword. It is a flowchart of the process which shows the modification of this invention, shows the modification of step S4 of FIG. 2, and outputs a recommendation keyword.

  A mode for carrying out the present invention (hereinafter referred to as “embodiment”) will be described in detail with reference to the drawings as appropriate.

  FIG. 1 is a block diagram showing an embodiment of the present invention and showing an example of the configuration of a computer system. In FIG. 1, a WEB server 100, a database server 200, a storage device 300, and a log collection server 400 constitute an electronic commerce site 2. The electronic commerce site 2 is electronically connected to a plurality of client terminals 500 connected via the Internet 1 or WAN. Providing WEB services for commercial transactions. In FIG. 1, a WEB server 100, a database server 200, a storage device 300, and a log collection server 400 connected to the LAN 3 constitute an electronic commerce site 2. The LAN 3 of the electronic commerce site 2 is connected to the client terminal 500 via the Internet 1.

  A user who uses the electronic commerce site 2 connects to the WEB server 100 via the browser of the client terminal 500, and the WEB server 100 provides an example of providing services such as provision of product information and purchase of products.

  <Client terminal 500>

  Each client terminal 500 includes a processor 501 that performs arithmetic processing, a memory 502 that stores data and programs, a storage device 503 that stores data and programs, a display device 504 as an output device, and inputs such as a mouse and a keyboard. A device 505 is provided.

  A web browser is loaded into the memory 502 of the client terminal 500 and is executed by the processor 501. Then, the client terminal 500 receives a service from the WEB server 100 of the electronic commerce site 2 by the WEB browser. For example, the client terminal 500 transmits a search keyword input by the user from the input device 505 to the WEB server 100 and requests a search for product information. The WEB server 100 acquires information on the corresponding product from the received keyword from the database server 200 and transmits it to the client terminal 500. The client terminal 500 displays product information as a search result received from the WEB server 100 on the display device 504.

  <Electronic commerce site 2>

  The e-commerce site 2 manages the WEB server 100 that receives requests from a plurality of client terminals 500 and provides services, and the product database 310 stored in the storage device 503, and the WEB server 100 receives the requests from the client terminals 500. A database server 200 that searches for product information by keyword and a log collection server 400 that acquires and stores the access history of the client terminal 500 are connected to the LAN 3 and can communicate with each other. In the present invention, when the WEB server 100 searches the product database 310 in the database server 200 with the search keyword received from the client terminal 500, in addition to the searched product information, the recommended keyword table 220 created in advance is stored. A keyword column is also provided. Hereinafter, the structure of each server which comprises the electronic commerce site 2 is each demonstrated.

  <WEB server 100>

  The WEB server 100 includes a processor 101 that performs arithmetic processing and a memory 102 that stores information, and is connected to the LAN 3 via a network interface (not shown). The memory 102 is loaded with a WEB service unit 110 and executed by the processor 101.

  The WEB service unit 110 acquires data from the database server 200 in response to a Web server function that accepts an HTTP request transmitted by the WEB browser of the client terminal 500 and a processing request accepted by the Web server function. It is assumed that a function of an application server that performs a predetermined processing and responds to the Web server function is included. Note that the web server function and the application server function of the WEB service unit 110 can be made independent to serve as servers. The WEB service unit 110 transmits data acquired from the database server 200 to the client terminal 500.

  <Database server 200>

  The database server 200 includes a processor 201 that performs arithmetic processing and a memory 202 that stores information, and is connected to the LAN 3 via a network interface (not shown). A recommended keyword generation unit 230 that generates a database management system (hereinafter referred to as DBMS) 210, a recommended keyword table 220, a related graph 240, a probability transition matrix 250, and an eigenvector 260 is loaded into the memory 102 and executed. The DBMS 210 manages the product database 310 of the storage apparatus 300. When the DBMS 210 receives a search keyword from the WEB server 100, the DBMS 210 searches the product database 310 of the storage apparatus 300 and acquires product information that matches the search keyword. At this time, the DBMS 210 acquires a keyword string from the recommended keyword table 220 that matches the search keyword, and responds to the WEB server 100 including the product information as the search result at the end of the keyword string. In the present embodiment, the database server 200 functions as an information search device that searches the product database 310 and the recommended keyword table 220, and the DBMS 210 functions as a search unit.

  The recommended keyword generation unit 230 refers to the user access history table 420 collected by the log collection server 400 at a predetermined timing, generates a related graph 240 and a recommended keyword table 220 as described later, and stores them in the memory 202. The recommended keyword generation unit 230 may store the generated recommended keyword table 220, the eigenvector 260, and the like in the storage device 300. When generating the recommended keyword table 220, the recommended keyword generation unit 230 generates a relation graph 240, a probability transition matrix 250, and an eigenvector 260. Details of these will be described later.

  In the present embodiment, an example in which the recommended keyword generation unit 230 is executed by the database server 200 has been described, but the present invention is not limited to this. In the present invention, the recommended keyword table 220 is generated by referring to the user access history table 420, and the DBMS 210 may refer to the recommended keyword table 220 and respond to the keyword string and the product information to the WEB server 100. Therefore, the recommended keyword generation unit 230 can be executed by the log collection server 400 or the WEB server 100 other than the database server 200.

  In the present invention, since the WEB server 100 that has received the search keyword responds to the client terminal 500 with the keyword string and the product information, the WEB server 100 may hold the recommended keyword table 220, the eigenvector 260, and the like.

  <Log collection server 400>

  The log collection server 400 includes a processor 401 that performs arithmetic processing and a memory 402 that stores information, and is connected to the LAN 3 via a network interface (not shown). A log collection unit 410 is loaded into the memory 402 and executed by the processor 401.

  The log collection unit 410 acquires the search keyword and product information requested by the client terminal 500 and stores them in the user access history table 420. In this embodiment, an example in which the user access history table 420 is stored in the memory 402 is shown. However, the user access history table 420 is stored in the storage device 300 or a local storage device (not shown) of the log collection server 400. Anyway.

  When the client terminal 500 requests information on the product database 310 from the WEB server 100, the log collection unit 410 adds a new entry to the user access history table 420 shown in FIG. Accumulate information.

  FIG. 5 shows a user access history table 420. The user access history table 420 includes a date and time 421 that stores time information when the client terminal 500 requests information in the product database 310, and a user ID 422 that stores an identifier of the client terminal 500 or an identifier of a user who makes a search request. A session ID 422 for storing an identifier of a session established between the client terminal 500 and the electronic commerce site 2, an action 424 for storing the type of operation performed by the client terminal 500, and a search input or selected by the client terminal 500 One entry is composed of a keyword 425 for storing the keyword and a product information URL 426 for storing the location of the product information requested by the client terminal 500. In the action 424, “input” is stored in the action 424 when the client terminal 500 inputs a search keyword on the product information search screen described later, and when the client terminal 500 selects the product information, “ “select” is stored in action 424.

  Here, the log collection server 400 may be disposed so as to be interposed between the client terminal 500 and the WEB server 100 in order to accumulate the contents requested by the client terminal 500 from the WEB server 100 as an access history. For example, the log collection server 400 can function as a proxy that transfers packets between the client terminal 500 and the WEB server 100.

  In the present invention, when the client terminal 500 requests product information in the product database 310 from the WEB server 100, it is only necessary to add a new entry to the user access history table 420. Absent. For example, the log of the client terminal 500 may be generated by the WEB service unit 110 of the WEB server 100 and the user access history table 420 held in the WEB server 100 may be updated.

  In addition, the WEB service unit 100 of the WEB server 100, the DBMS 210 of the database server 200, the recommended keyword generation unit 230, and the log collection unit 410 of the log collection server 400 are each configured by a program. These programs are stored in the storage apparatus 300 as a non-transitory computer-readable storage medium, and loaded into the memory when each server is started.

  <Overall processing>

  FIG. 2 is a flowchart of overall processing performed by the computer system of FIG. First, in step S1, a plurality of client terminals 500 access the WEB server 100 and request product information using a search keyword.

  In step S <b> 2, the log collection server 400 stores the search keyword requested by each client terminal 500 and the product information accessed by each client terminal 500 in the user access history table 420.

  On the other hand, in step S3, the recommended keyword generation unit 230 of the database server 200 generates the recommended keyword table 220, the eigenvector 260, and the like with reference to the user access history table 420 at predetermined time intervals (for example, every hour). .

  In step S4, the DBMS 210 of the database server 200 extracts a keyword string (recommended keyword string) including the search keyword received in step S1 from the recommended keyword table 220. Then, the DBMS 210 of the database server 200 acquires the extracted keyword string and the merchandise information retrieved from the keyword string from the merchandise database 310 and responds to the WEB server 100. The WEB server 100 returns the information received from the database server 200 to the client terminal 500 that is the transmission source of the search keyword.

  As a result, a search screen 540 as shown in FIG. 3 is displayed on the display device 504 of the client terminal 500. FIG. 3 shows a search screen 540 displayed on the display device 504 when the client terminal 500 accesses the WEB server 100 and requests product information.

  The search screen 540 is divided into an area 541 for inputting a search keyword (input keyword in the figure) and an area 542 (search result display area) for displaying a search result for the search keyword from the WEB server 100. An area 542 for displaying the search result includes an area 543 for displaying a related graph (tree structure keyword string) 2400 including a search keyword and a recommended keyword in the upper half, and an end of the related graph (leaf portion of the tree structure = leaf). The node) includes an area 544 for displaying product information as a search result. The user of the client terminal 500 displays the No. displayed in the product information area 541. When the product information frame of 1 to N0.4 is selected, the WEB server 100 acquires product information from the product database 310 and displays the product information on the display device 504. As will be described later, the relation graph 2400 in the area 542 for displaying the search result is obtained by inputting the search keyword and connecting the keyword in the keyword string extracted from the recommended keyword table 220 in the process of FIG. 25 or FIG. is there.

  The log collection server 400 adds a new history to the user access history table 420 when the client terminal 500 requests or searches for product information from the WEB server 100 on the search screen 540 shown in FIG.

  That is, when the client terminal 500 inputs a search keyword in the area 541 of the search screen 540 and transmits the search keyword to the WEB server 100, the log collection server 400 adds the entry to the new entry user access history table 42, and Action 424 is set to “input”. On the other hand, when product information is selected (or clicked) in the area 544 of the search result 542 and the product information is requested from the WEB server 100, the log collection server 400 adds a new entry, and the action 424 of the entry is “select”. Set to

  <Outline of Generation of Recommended Keyword Table 220>

  FIG. 4 is a flowchart illustrating an example of processing performed by the recommended keyword generation unit 230 of the database server 200. This process is a process performed in step S3 of FIG.

  In step S <b> 31, the recommended keyword generation unit 230 refers to the access history table 420 of the log collection server 400 and acquires the search keyword input history. In step S <b> 32, the recommended keyword generation unit 230 refers to the access history table 420 of the log collection server 400 and acquires the access history of product information.

  In step S <b> 33, the recommended keyword generation unit 230 executes the process of FIG. 6, extracts a keyword string from the search keyword to the product information from the user access history table 420, and generates a related graph 240. As shown in FIGS. 7 to 13, the related graph 240 extracts partial graphs 240 </ b> A to 240 </ b> F constituting a part of the related graph 240 for each session ID 423 of the user access history table 420. Then, the recommended keyword generation unit 230 generates a keyword tree structure including a keyword string from the partial graphs 240A to 240F to the product information from the search keyword as a relation graph 240 shown in FIG.

  That is, in the related graph 240, the recommended keyword generation unit 230 combines keywords from a history of keywords searched by any of the users of the electronic commerce site 2 and a history of the product information that has been browsed to reach the product information. It is a graph.

  Next, in step S34, the recommended keyword generator 230 obtains the transition probabilities of the starting point and the end point of the search keyword (hereinafter, parent node to child node) from the related graph 240 obtained in step S33, and the probability transition matrix Z (250) is obtained. The calculation of the probability transition matrix Z is performed as shown in FIGS.

  In step S35, the recommended keyword generation unit 230 calculates K eigenvectors 260 from the probability transition matrix Z obtained in step S34. Here, K is a preset value (for example, “2”), which is a value set by an administrator or the like. By this calculation, K keyword strings are selected as shown in FIGS.

  In step S36, the recommended keyword generation unit 230 generates a recommended keyword table 220 that reaches the product information from the search keyword as a base point from the K eigenvectors 260 obtained in step S35 as shown in FIG.

  Hereafter, each of said step S33-S36 is explained in full detail below.

  <Generation of related graph>

  FIG. 6 is a flowchart showing details of generation of the relation graph 240. In step S331, the recommended keyword generation unit 230 extracts an access log within a preset period from the user access history table 420 for each session. As the preset period, a period such as within 6 months can be set. Hereinafter, the access log indicates an access log within a predetermined period extracted from the user access history table 420 and a record of access to product information.

  Next, in step S332, the recommended keyword generation unit 230 extracts a record having the same session ID 423 for the extracted access log. In step S333, the recommended keyword generation unit 230 sorts records having the same session ID 423 in ascending order of the date and time 421. Then, the recommended keyword generation unit 230 sequentially acquires the search keywords stored in the keyword 425 from the sorted search history and connects them. The recommended keyword generation unit 230 holds the linked search keyword and product information as a relation graph 240.

  The recommended keyword generation unit 230 extracts the partial graphs 240A to 240F for all the session IDs 423 within a predetermined period, and generates the related graph 240.

  For example, as illustrated in FIG. 7, the recommended keyword generation unit 230 of the database server 200 acquires an access log whose date 421 is within a predetermined period from the user access history table 420 of the log collection server 400. Then, the recommended keyword generation unit 230 extracts the records with the session ID 423 = 91119 in the illustrated example as the records with the same session ID 423 from the acquired user access history table 420, and sorts them in ascending order of the date and time 421. The session log is 420A.

  Next, as shown in FIG. 8, the recommended keyword generation unit 230 sequentially extracts the contents of the keyword 425 from the oldest record of the date and time 421 from the session log 420A, and sets the data including the extracted search keywords as a partial graph 240A. Generate.

  In the example of FIG. 8, the recommended keyword generation unit 230 extracts a plurality of records equal to the session ID 423 = 91119 as the session log 420A from the user access history table 420 and sorts them in ascending order of the date 421. The recommended keyword generation unit 230 extracts the contents of the keyword 425 from the sorted session log 420A in ascending order of the date and time 421. As a result, the search keywords “metal fitting”, “hook”, and “S-shaped hook” are chronologically searched. A connected graph is generated as a subgraph 240A. That is, in session ID 423 = 91119, after inputting the search keyword “metal fitting” from the client terminal 500, the search keyword “hook” is input, and the search keyword is input in the order of “S-shaped hook” at the end of the session. Show the history. That is, the partial graph 240A is a graph that expresses the time-series relationship of the search keywords in the connection order of the search keywords.

  FIG. 9 is the same as FIG. 8 and shows an example in which the recommended keyword generation unit 230 extracts a plurality of records equal to session ID 423 = 91121 from the user access history table 420 of FIG. 7 as the session log 420B. The recommended keyword generation unit 230 sorts the session log 420B in ascending order of the date and time 421. Then, the recommended keyword generation unit 230 extracts the contents of the keyword 425 from the sorted session log 420B in ascending order of the date and time 421. As a result, the search keyword “hook” and “S-shaped hook” are connected in time series. Generated as a subgraph 240B.

  10 is the same as FIG. 8, and the recommended keyword generation unit 230 extracts a plurality of records equal to the session ID 423 = 92111 from the user access history table 420 of FIG. 7 as the session log 420C from the user access history table 420 of FIG. An example is shown. The recommended keyword generation unit 230 sorts the session log 420C in ascending order of the date and time 421. The recommended keyword generation unit 230 extracts the contents of the keyword 425 from the sorted session log 420C in ascending order of the date and time 421. As a result, the search keyword “hook” and “S tube” are connected in time series. Generated as graph 240C.

  FIG. 11 is the same as FIG. 8, and the recommended keyword generation unit 230 extracts a plurality of records equal to the session ID 423 = 92911 from the user access history table 420 of FIG. 7 as the session log 420 D from the user access history table 420 of FIG. An example is shown. The recommended keyword generation unit 230 sorts the session log 420D in ascending order of the date and time 421. Then, the recommended keyword generation unit 230 extracts the contents of the keyword 425 from the sorted session log 420D in ascending order of the date and time 421. As a result, the search keyword “metal fitting” and “attachment metal fitting” are connected in time series. Generated as a graph 240D.

  12 is the same as FIG. 8, and the recommended keyword generation unit 230 extracts a plurality of records equal to session ID 423 = 92923 from the user access history table 420 of FIG. 7 as the session log 420E from the user access history table 420 of FIG. An example is shown. The recommended keyword generation unit 230 sorts the session log 420E in ascending order of the date and time 421. Then, the recommended keyword generation unit 230 extracts the contents of the keyword 425 from the sorted session log 420E in ascending order of the date and time 421. As a result, the search keywords “S tube” and “Unit S tube XXX” are connected in time series. A thing is generated as a subgraph 240E. Here, the recommended keyword generation unit 230 sets the “unit S tube XXX” of the generated partial graph 240E as product information to be a leaf node of the related graph 240 because there is data in the product information URL 426.

  FIG. 13 is the same as FIG. 8, and the recommended keyword generation unit 230 extracts a plurality of records equal to session ID 423 = 93211 from the user access history table 420 of FIG. 7 as the session log 420F from the user access history table 420 of FIG. An example is shown. The recommended keyword generation unit 230 sorts the session log 420F in ascending order of the date and time 421. The recommended keyword generation unit 230 extracts the contents of the keyword 425 from the sorted session log 420F in ascending order of the date and time 421. As a result, the search keywords “S tube” and “Unit S tube YYY” are connected in time series. A thing is generated as a subgraph 240F. Here, the recommended keyword generation unit 230 sets the “unit S tube YYY” of the generated partial graph 240F as product information to be a leaf node of the related graph 240 because there is data in the product information URL 426.

  As described above, the recommended keyword generation unit 230 generates, as the partial graphs 240A to 240F, the contents of the keyword 425 extracted in time series for each session ID 423 by the processing of FIG. A set of these partial graphs 240A to 240F becomes a related graph 240 in the memory 202 of the database server 200 shown in FIG. Then, the recommended keyword generation unit 230 generates a relation graph 240 shown in FIG. 14 by following the connection relations in order from the search keyword that becomes the leaf node from the generated partial graphs 240A to 240F.

  In the generation of the relation graph 240, the recommended keyword generation unit 230 converts the relation graph having the same search keyword among the old search keywords in time series between the leaf nodes associated with the product information to the same node parent node. Extract as a subgraph with. For example, since the subgraphs 240E and 240F having leaf nodes have the same “S tube” as a parent node, the recommended keyword generation unit 230 generates a subgraph having “S tube” as a parent node and a leaf node as a child node. .

  Then, the recommended keyword generation unit 230 also compares each of the partial graphs 240A to 240D from the leaf node to the parent node, and summarizes the partial graphs 240A to 240F with the same node as the parent node. A graph 240 is generated.

  4, the recommended keyword generation unit 230 extracts search keywords as subgraphs 240A to 240F from the session logs 423A to 423F. Then, the recommended keyword generation unit 230 traces the search keyword that is the parent node from the leaf node associated with the product information, and selects the keyword from the search keyword to the product information as shown in FIG. The relation graph 240 expressed as follows is generated.

  FIG. 14 is a relation graph 240 showing the number of transitions of the keyword string from the search keyword to the product information. The relation graph 240 is obtained by connecting transition relationships between search keywords with a line and displaying the number of transitions beside the connection line. For example, the transition from the “hook” as the parent node to the “S-shaped hook” as the child node has a total of two transitions of the partial graph 240A in FIG. 8 and the partial graph 240B in FIG. The number of transitions from “hook” to “S-shaped hook” in the graph 240 is two.

  <Generation of stochastic transition matrix>

  Next, the process of generating the probability transition matrix 250 of each search keyword from the relation graph 240 performed in step S34 of FIG. 4 will be described in detail.

  FIG. 15 shows an example in which the probability transition matrix 250 </ b> A is obtained from the keyword string constituting the relation graph 240.

  First, the recommended keyword generation unit 230 acquires a keyword string that configures the related graph 240, stores each search keyword in a column 251 that represents a parent node, and stores each search keyword in a row 252 that represents a child node. And the recommendation keyword production | generation part 230 sets the frequency | count of having changed from the parent node to the child node about each search keyword from the related graph 240, and produces | generates it as the probability transition matrix 250A.

  Next, the recommended keyword generation unit 230 obtains the total number of transitions to the child node for each row of the probability transition matrix 250A and sets it as a population parameter. Then, the recommended keyword generation unit 230 obtains a probability obtained by dividing the number of transitions to the child node by the parameter of each row, and generates the probability transition matrix 250B shown in FIG. For example, in the row where the parent node of the probability transition matrix 250A is “hook”, the transition to “S tube” is 1 time and the transition to “S hook” is 2, so the total number of row transitions is 3 times. It becomes. As shown in the probability transition matrix 250B of FIG. 16, the recommended keyword generation unit 230 has a transition probability to “S tube” of 1/3 and a transition probability to “S-shaped hook” in the row where the parent node is “hook”. Is set to 1/3. The recommended keyword generation unit 230 similarly determines the transition probability from the parent node to the child node in units of other rows, and generates a probability transition matrix 250B.

  Next, the recommended keyword generation unit 230 sets the transition probability of the parent node having no child node in the related graph 240 of FIG. 14 in the probability transition matrix 250B as the reciprocal of the total number of components (child nodes) of each row, It is generated as a probability transition matrix 250C shown in FIG. In the example of FIG. 17, even after reaching the parent node having no child nodes = “mounting bracket”, “S-shaped hook”, “unit S tube XXX”, “S tube YYY” in the probability transition matrix 250B in the search process. In some cases, a search keyword is input. For this reason, an equal probability is set in a parent node having no child node on the assumption that a transition is made to any other child node. That is, in the probability transition matrix 250C, since the total number of child nodes (in the figure, “metal fittings” to “S tube YYY” in the row direction) is “7”, the probability of transition from each parent node to the child node is 1 / Set to 7.

  Next, as shown in FIG. 18, the recommended keyword generation unit 230 generates a probability transition matrix 250D as a unit matrix, and generates a probability transition matrix Z from the probability transition matrix 250C generated in FIG. Here, the probability transition matrix 250D as a unit matrix is a matrix in which the transition probabilities from the parent node to the child node are all equal, and in the illustrated example, the total number of child nodes is 7, so the unit matrix is 1 / A transition probability of 7 is set.

Next, the recommended keyword generation unit 230 obtains the probability transition matrix 250 from the following equation using the probability transition matrix 250C obtained in FIG. 17 as the matrix H and the unit matrix of the probability transition matrix 250D set in FIG. . Hereinafter, the probability transition matrix 250 is represented by a matrix Z.
Z = αH + (1−α) × I (1)
However, (alpha) is a stabilization coefficient and is a preset value (for example, 0.8-0.9 etc.).

  Thus, the probability transition matrix Z is obtained by the recommended keyword generation unit 230.

  <Eigenvector calculation>

  Next, the calculation of the eigenvector 260 performed in step S35 of FIG. 4 will be described. The recommended keyword generation unit 230 calculates K eigenvectors set in advance for the probability transition matrix Z obtained by the equation (1). However, the eigenvectors are calculated in the order of K eigenvectors in descending order of eigenvalues. For example, in the case of K = 2, as shown in FIGS. 19 and 20, the recommended keyword generation unit 230 makes the first eigenvector shown in FIG. 19 and the second eigenvector shown in FIG. 20 from the sequence from the search keyword to the product information. For each.

  FIG. 19 shows an example in which the eigenvector 260 and the value of the first eigenvector are set in the relation graph 240. FIG. 20 shows an example in which the values of the eigenvector 260 and the second eigenvector are set in the relation graph 240.

  In the related graph 240 in which the first eigenvector of FIG. 19 is set, “S tube” = 0.04 having the smallest value of the eigenvector 260 in the keyword string from “metal fitting” to “unit S tube XXX” which is product information. This is the most difficult search keyword. In other words, in the session log of the user access history table 420, there are many users who have obtained the search result of “hook” from the search keyword “metal fitting”, but the “Unit S tube” of the product information from the search keyword “S tube”. The number of users who have obtained the search result “XXX” is very small. In the present invention, although it is difficult to come up with, it is possible to present a recommended keyword with a small eigenvector that can reach product information.

  Note that the value of the first eigenvector may be displayed in the related graph area 543 of the search result display area 542 in FIG. In particular, the value of the eigenvector having the smallest value among the values of the first eigenvector may be displayed together with the recommended keyword. Or you may make it specify that an eigenvector is the minimum, such as displaying the recommendation keyword of the eigenvector with the smallest value among the values of a 1st eigenvector in cooperation. Thereby, it is possible to clearly indicate which recommended keyword is difficult to come up with to reach the product information.

  <Generation of recommended keywords>

  Next, generation of the recommended keyword table 220 performed in step S36 of FIG. 4 will be described. The recommended keyword generation unit 230 obtains the first eigenvector and the second eigenvector as shown in FIGS. 19 and 20 when K = 2 in step S35. Then, in the first eigenvector shown in FIG. 19, the recommended keyword generation unit 230 has the search keyword from “metal fitting” to “hook” and “S tube” to product information = “unit S tube XXX” which is a leaf node. Get keyword sequence. This keyword string is given the highest priority = 1. In the second eigenvector shown in FIG. 20, a keyword string is obtained from which the search keyword passes from “metal fitting” to “hook” and “S tube” to the leaf node product information = “S tube YYY”. The second priority = 2 is given to the keyword string.

  Then, the recommended keyword generation unit 230 sets the search keywords from the first word to the Nth word (N = 4 in the present embodiment) as recommended keywords according to the length of the keyword string, and sets them in the order of the priorities. Are generated as the recommended keyword table 220. As a result, in the examples of FIGS. 19 and 20, as shown in FIG. 21, the keyword search keyword is set from the first word 222 to the fourth word 225 according to the priority 221. For example, the recommended keyword table 220 with the priority 221 = 1 includes a keyword string from the first word = “metal fitting” to the fourth word = “unit S tube XXX”. Further, the recommended keyword table 220 with the priority 221 = 2 includes a keyword string from the first word = “metal fitting” to the fourth word = “S tube YYY”.

  In FIG. 21, the entries with the priorities 221 = 3, 4 show an example when four eigenvectors 260 are obtained. In addition, it is desirable to add a field that stores information indicating in which field the product information is stored in the first word 222 to the Nth word in the recommended keyword table 220. Alternatively, each entry may be provided with a flag that sets “0” when the end of the keyword string in the recommended keyword table 220 is product information and sets “1” when the end of the keyword string is a keyword.

  <Presentation of recommended keywords>

  Next, the details of the process of presenting the recommended keyword stored in the recommended keyword table 220 for the search keyword shown in step S4 of FIG. 2 will be described below.

  FIG. 25 is a flowchart showing details of the process performed in step S4 of FIG. This flowchart is executed by the DBMS 210 when the database server 200 receives the search keyword input by the client terminal 500 via the WEB server 100.

  In step S <b> 41, the DBMS 210 receives a search keyword from the WEB server 100. This search keyword is a keyword input by the client terminal 500 on the search screen 540 of FIG.

  In step S42, the DBMS 210 searches the recommended keyword table 220 and identifies a recommended keyword X that matches the accepted search keyword. When a search keyword that matches the recommended keyword is not found in the recommended keyword table 220, the DBMS 210 may specify the recommended keyword 220 closest to the search keyword as a recommended keyword X from a synonym dictionary or the like. As a method for searching the synonym closest to the search keyword from the recommended keyword table 220, a known or well-known method may be used.

  In step S43, the DBMS 210 searches the recommended keyword table 220 and extracts L keyword strings including the specified recommended keyword X. “L” is a preset value, which is the maximum number of keyword strings extracted from the keyword table 220.

  In step S44, the DBMS 210 extracts a path (keyword string) from the recommended keyword X to the product information among the one or more keyword strings extracted in step S43. In step S45, the DBMS 210 generates data for outputting the path from the recommended keyword X to the product information to the area 543 for displaying the related graph in the search result display area 542 shown in FIG.

  In step S46, the DBMS 210 acquires the product information of the path extracted in step S44 from the product database 310, and generates data to be output to the area 544 for displaying the product information in the search result display area 542 shown in FIG. It responds to the WEB server 100 together with the area 543 of the related graph generated in step S45. The WEB server 100 transmits a response from the DBMS 210 to the client terminal 500.

  When the client terminal 500 transmits a search keyword to the WEB server 100 by the above processing, the WEB server 100 transfers the search keyword to the database server 200 and requests a search for the search keyword. The database server 200 generates a recommended keyword table 220 in advance, and searches the recommended keyword table 220 using the search keyword received from the WEB server 100. The database server 200 identifies a recommended keyword that matches or is similar to the search keyword and sets it as the recommended keyword X. The DBMS 210 extracts a plurality of keyword strings related to the recommended keyword X from the recommended keyword table 220 and outputs a keyword string from the recommended keyword X to the product information. The end of the keyword string is product information as a search result, and the keyword string from the recommended keyword X to the product information is output to the search result display area 542 of the client terminal 500 shown in FIG. Is done.

  For example, as shown in FIG. 22, when the search keyword 226 has a maximum value L = 4 extracted by the search keyword 226 as “metal fitting”, the first word is recommended with priority = 1 to 4 that matches “metal fitting”. A plurality of keyword strings are acquired from the keyword table 220. Then, the DBMS 210 responds to the client terminal 500 via the WEB server 100 as data for displaying the acquired keyword string in the search result display area 542 shown in FIG. That is, a plurality of paths from “metal” to product information are output to the display 504 of the client terminal 500. Note that 220 ′ in FIG. 22 is a table for searching the recommended keyword table 220 and can be temporarily generated by the DBMS 210.

  When the search keyword is “hook”, as shown in FIG. 23, the second word matches the search keyword in the entries of the priority levels 1 to 3 in the recommended keyword table 220. The DBMS 210 extracts three or more words from the recommended keyword table 220 as the keyword sequence in the recommended keyword table 220 as indicated by 220 ′ in the figure, and responds to the client terminal 500 via the WEB server 100.

  When the search keyword is “S tube”, as shown in FIG. 24, the third word matches the search keyword in the entries having the priorities of 1 to 2 in the recommended keyword table 220. The DBMS 210 extracts two keywords from the recommended keyword table 220 as the keyword string in the recommended keyword table 220 as indicated by 220 ′, and responds to the client terminal 500 via the WEB server 100.

  As described above, when the user of the electronic commerce site 2 who operates the client terminal 500 requests the search screen 540 from the WEB server 100 and inputs a search keyword, a plurality of keyword strings from recommended keywords close to the search keyword to product information are displayed. It can be obtained as a relation graph 240. The recommended keyword table 220 is generated from a user access history table 420 that stores a history of searches and browsing (access to search results) performed by a plurality of users on the electronic commerce site 2.

  This eliminates the need for the user or the like to give an evaluation to the product in the product database 310 as in the conventional example. In the present invention, the relationship between the search keyword and the product information can be provided to the user of the client terminal 500 from the structure of the keyword string (relation graph 2400). Further, by providing the relation graph 2400 having a tree structure that reaches the product information from the search keyword, the user of the client terminal 500 can obtain desired product information very easily as compared with the conventional example. Further, in the present invention, it is not necessary to prepare product information immediately after the sale as in the conventional example, so that the labor for operating the electronic commerce site 2 can be greatly reduced.

  There are many search keywords that can be easily forgotten by the user, such as the product name itself and the model number of the product, for the search keyword that reliably reaches the target product information when searching for the product.

  On the other hand, there is a problem that a search keyword that a user can immediately come up with has a high level of abstraction, so that the range of hit products becomes very wide. For this reason, in the case of a new product that is unfamiliar with the name of the product, there is a problem that it is an essential problem of the cold start problem described in the conventional example, and the target product cannot be easily reached from the search keyword.

  On the other hand, in the present invention, a keyword sequence from an abstract keyword to which an unfamiliar product belongs to a more specific keyword is generated from a history of searched keywords and a history of referenced product information. Became possible. In particular, by creating the recommended keyword table 220 using the search histories of a large number of users, it is possible to link keywords and product information that were not easily linked by one user.

  According to the present invention, even if it is a new product, for example, it is possible to return to an abstract keyword once from a product that has already been sold and to start from there, and to have a function that the user originally wanted among new products Makes it easy to search.

  In the above embodiment, an example in which the present invention is applied to the product database 310 has been described. However, a database search system or search that stores information that a user wants to refer to other users' evaluations regardless of whether or not they are sold. It can be applied to the device.

  <Modification>

  FIG. 26 is a flowchart showing a modification of the process performed in step S4 of FIG. In this modified example, the recommended keyword higher in the recommended keyword table 220 that matches or is similar to the search keyword (higher on the search origin side) is also output to the search result display area 542.

  This flowchart is executed by the DBMS 210 when the database server 200 receives the search keyword input by the client terminal 500 via the WEB server 100.

  In step S51, the DBMS 210 receives a search keyword from the WEB server 100 as in step S41 of FIG. This search keyword is a keyword input by the client terminal 500 on the search screen 540 of FIG.

  In step S52, the DBMS 210 searches the recommended keyword table 220 and identifies a recommended keyword X that matches the received search keyword. When a search keyword that matches the recommended keyword table 220 is not found, the DBMS 210 may specify the recommended keyword closest to the search keyword as the recommended keyword X from a synonym dictionary or the like.

  In step S53, the DBMS 210 searches the recommended keyword table 220 and extracts M keyword strings including the specified recommended keyword X. “M” indicates the maximum value of the number of keyword strings to be extracted with a preset value.

  In step S54, the DBMS 210 extracts a recommended keyword Y higher than the recommended keyword X (from the search start side) in the keyword string extracted in step S53.

  In step S55, the DBMS 210 searches the recommended keyword table 220, and extracts L keyword strings (paths) from the higher-order recommended keyword Y to the product information. Then, the DBMS 210 generates data for outputting the path from the recommended keyword Y to the product information to the area 543 displaying the related graph in the search result display area 542 shown in FIG. In step S56, the DBMS 210 acquires the product information of the path extracted in step S54 from the product database 310, and generates data to be output to the area 544 for displaying the product information in the search result display area 542 shown in FIG. It responds to the WEB server 100 together with the area 543 of the related graph generated in step S55. The WEB server 100 transmits a response from the DBMS 210 to the client terminal 500.

  With the above processing, the database server 200 searches the recommended keyword table 220 generated in advance with the search keyword received from the WEB server 100. The database server 200 identifies a recommended keyword X that matches or is similar to the search keyword, and further acquires a keyword string from the recommended keyword Y higher than the recommended keyword X to the product information from the search origin side. Data to be output to the search result display area 542 shown in FIG.

  According to this modification, even when a new product is added to the product database 310, a new item can be renewed relatively early by following a path from a recommended keyword of higher order (from the search origin side or above in the related graph 2400). It is possible to provide a solution to the cold start problem of the conventional example in which it is easy to recommend a new product from the recommendation keyword 220 and it takes time until it is noticed. In other words, the electronic commerce site 2 of the present invention can provide a keyword string and product information in a wider range than the search keyword that the user of the client terminal 500 came up with, thereby increasing the possibility of reaching a new product immediately after sale. it can.

  For example, when the DBMS 210 receives “hook” as a search keyword, the “hook” of the second word 223 having a priority 221 of 1 to 3 is specified as the recommended keyword X as shown in FIG. Then, since the recommended keyword 220 higher than the specified second word 223 (starting side of the search) has “metal” in the first word 222, the DBMS 210 specifies this “metal” as the recommended keyword Y. . Then, as shown in FIG. 3, the DBMS 210 transmits a keyword string of a relation graph 2400 having “metal” as a vertex to the WEB server 100. The client terminal 500 traces “metal fitting” which is a high-order recommended keyword from the search keyword = “hook” from the related graph 2400 displayed in the search result display area 542 of the search screen 540 of FIG. Product information can also be reached via the “mounting bracket”. The range of the recommended keyword Y higher than the recommended keyword X specified by the DBMS 210 can be specified within a preset range. For example, when the Nth word of the recommended keyword 220 is specified as the recommended keyword X, the (N−β) th word is specified as the recommended keyword Y. As a result, it is possible to obtain a keyword string from the recommended keyword Y that is higher than the recommended keyword X identified from the search keyword by the β level. Thereby, the electronic commerce site 2 can provide the user of the client terminal 500 with a broader keyword string than the input search keyword.

  In the above embodiment, the DBMS 210 of the database server 200 searches for the recommended keyword 220. However, the WEB server 100 may use the recommended keyword 220. For example, the recommended keyword generation unit 230 is executed by the WEB server 100 and the recommended keyword 220 is stored in the memory 102. When the WEB server 100 transfers the search keyword to the database server 200 and receives the search result from the database server 200, the WEB server 100 performs the same processing as steps S42 to S46 shown in FIG. Thereby, the WEB server 100 can add the recommended keyword 220 to the search result (product information) of the database server 200 and provide the related graph 2400 to the client terminal 500.

  In addition, the WEB server 100, the log collection server 400, the database server 200, and the storage unit 300 of the above embodiment may be provided together as an information search device.

  The present invention can be applied to a computer system that provides recommended information related to information to be searched. In particular, the present invention is applied to a database system that searches for product information using a search keyword received from a client terminal and also provides recommended information. Is preferred. The present invention can also be applied to an information recommendation device or a search device that provides recommended product information together with product information to be searched.

Claims (15)

An information search method for searching from a predetermined database using a received search keyword by a computer including a processor and a storage device,
A first step of generating a keyword string from a history of search keywords received by the computer and a history of access to information in the database;
A second step in which the calculator receives the search keyword;
A third step in which the calculator searches the keyword string including keywords that match or are similar to the search keyword, and extracts one or more keyword strings;
A fourth step of outputting a keyword from the keyword matching or similar to the search keyword to the database information in the extracted keyword string by the computer;
An information search method characterized by including: The information search method according to claim 1,
The first step includes
A fifth step of storing the search keyword and access to the database information in access history information each time the computer receives the search keyword or accesses the database information;
A sixth step in which the computer reads a search keyword of the access history information and the database information at a predetermined timing, and generates a keyword string connecting keywords from the search keyword to the database information as a related graph; ,
A seventh step in which the computer generates a probability transition matrix from the association graph to each keyword;
An eighth step in which the calculator calculates an eigenvector of the keyword from the probability transition matrix;
A ninth step in which the computer generates a plurality of keyword strings from the eigenvector and stores the keyword strings in a recommended keyword table;
An information search method characterized by including: The information search method according to claim 2,
The sixth step includes
Including database information at the end of the keyword string;
The ninth step includes
An information search method comprising database information at the end of the keyword string in the recommended keyword table. The information search method according to claim 2,
The eighth step includes
A method for retrieving information, comprising: calculating a predetermined number of eigenvectors in descending order of eigenvalues of the probability transition matrix. The information search method according to claim 2,
The fourth step includes
An information search method, wherein the keyword having the smallest value of the eigenvector is specified. The information search method according to claim 1,
The fourth step includes
An information search method, wherein the computer outputs a keyword from a keyword that is a predetermined number higher than a keyword that matches or is similar to the search keyword in the extracted keyword string to information in the database. An information search device comprising a processor and a storage device, and performing a search from a predetermined database with a received search keyword,
A recommended keyword generating unit that generates a keyword string from the history of the received search keyword and the history of access to the database information;
The search keyword is received, a keyword matching or similar to the search keyword is searched from the generated keyword string, one or more keyword strings are extracted, and the search keyword is extracted from the extracted keyword strings. A search unit that outputs keywords from matching or similar keywords to information in the database;
An information retrieval apparatus comprising: The information search device according to claim 7,
The recommended keyword generation unit includes:
A history collection unit that accumulates access to the search keyword and database information in access history information each time the search keyword is received or accessed to the database information;
The recommended keyword generation unit includes:
At a predetermined timing, the search keyword and database information of the access history information are read, a keyword string connecting keywords from the search keyword to the database information is generated as a related graph, and each keyword is generated from the related graph. Generating a probability transition matrix to the computer, the computer calculating an eigenvector of the keyword from the probability transition matrix, the computer generating a plurality of keyword strings from the eigenvector, and storing the keyword string in a recommended keyword table An information retrieval apparatus characterized by The information search device according to claim 8,
The recommended keyword generation unit includes:
An information retrieval apparatus comprising database information at the end of the keyword string. The information search device according to claim 8,
The recommended keyword generation unit includes:
An information retrieval apparatus that calculates a predetermined number of eigenvectors in descending order of eigenvalues of the probability transition matrix. The information search device according to claim 8,
The search unit
An information retrieval apparatus that identifies the keyword having the smallest value of the eigenvector. The information search device according to claim 7,
The search unit
An information search apparatus that outputs a keyword from a keyword that is a predetermined number higher than a keyword that matches or is similar to the search keyword in the extracted keyword string to information in the database. A computer that includes a processor and a storage device and performs a search from a predetermined database using a received search keyword,
A first procedure for generating a keyword string from the history of the received search keyword and the history of access to the database information;
A second procedure for receiving the search keyword;
A third procedure of searching for a keyword string including keywords that match or similar to the search keyword and extracting one or more keyword strings;
A fourth procedure for outputting a keyword from the extracted keyword string that matches or is similar to the search keyword to the information in the database;
A non-transitory computer-readable storage medium storing a program for causing the computer to execute. The storage medium according to claim 13,
The first procedure includes:
A fifth procedure for storing the search keyword and the database information in access history information each time the search keyword is received or the database information is accessed;
A sixth procedure of reading a search keyword of the access history information and the database information at the predetermined timing and generating a keyword string connecting keywords from the search keyword to the database information as a related graph;
A seventh procedure for generating a probability transition matrix from the association graph to each keyword;
An eighth procedure for calculating the eigenvector of the keyword from the probability transition matrix;
A ninth procedure for generating a plurality of keyword strings from the eigenvectors and storing the keyword strings in a recommended keyword table;
A storage medium comprising: The storage medium according to claim 14,
The sixth procedure includes:
Including database information at the end of the keyword string;
The ninth procedure includes:
A storage medium comprising database information at the end of the keyword string of the recommended keyword table.

Images