JP6856466B2 - Information processing systems, information processing methods, and programs - Google Patents

Description

The present invention relates to information processing systems, information processing methods, and programs.

A learning component that analyzes stored information retrieval data to identify relevance patterns from past information retrieval behavior, and a search component that identifies a subset of current search results based on at least part of the relevance pattern. An information retrieval system equipped with the above has been proposed (see Patent Document 1).

Japanese Unexamined Patent Publication No. 2006-285982

By the way, it may be preferable to be able to provide more useful information related to the user's search behavior.

The present invention has been made in consideration of such circumstances, and provides an information processing system, an information processing method, and a program capable of providing more useful information related to a user's search behavior. It is one of the purposes.

One aspect of the present invention includes a collecting unit that collects category paths given to each of a plurality of products or services selected by the user group on the sales site for each user group that inputs the same query. By integrating the same categories included in the plurality of category paths for each of the user groups collected by the collection unit, the generation unit that generates a category tree for each query and the similarity between the plurality of queries are determined. It is an information processing system including a derivation unit for deriving each of the queries of the above based on the similarity between the category trees generated by the generation unit.

According to one aspect of the present invention, it is possible to provide more useful information related to the user's search behavior.

It is a figure which shows the use environment of the server apparatus 10 of an embodiment. It is a figure for demonstrating the category path P of embodiment. It is a figure which shows an example of the content of the product information of an embodiment. It is a figure which shows an example of the content of the user behavior information of an embodiment. It is a block diagram which shows the structure of the similarity analysis part 300 of embodiment. It is a figure which shows an example of the set of the category path P collected by the category path collecting unit 310 with respect to the user group which input a query with an embodiment. It is a figure which shows an example of the category tree T generated by the category path integration part 321 of an embodiment. It is a figure which shows another example of the category tree T generated by the category path integration part 321 of an embodiment. It is a figure which shows an example of how to obtain the degree of similarity between category trees T of embodiment. It is a figure which shows an example of the session division by the session division part 400 of an embodiment. It is a flowchart which shows an example of the process flow of the derivation stage of the category tree T of embodiment. It is a flowchart which shows an example of the process flow of the session division and the derivation stage of the conversion prediction value using the similarity between the queries of an embodiment. It is a figure which shows the experimental result by the simulation of the conversion predicted value using the server apparatus 10 of this embodiment.

Hereinafter, an information processing system, an information processing method, and an embodiment of a program will be described with reference to the drawings. In the present embodiment, it is assumed that the information processing system is applied to the server device that provides the sales site. The sales site is not limited to the one played by the browser, but also includes the one played by the application program. The server device is communicably connected to the user terminal device via a network such as the Internet, and acquires information indicating the user's search behavior for the sales site. The server device derives the similarity between a plurality of queries input from the user, and divides the session based on the derived similarity. Further, the server device derives a predicted value of conversion of the user (probability that the user takes a predetermined action) based on the information about the divided session. Hereinafter, embodiments will be described.

FIG. 1 is a diagram showing a usage environment of the server device (information processing system) 10 of the embodiment. The server device 10 is communicably connected to the user terminal device UD and the client terminal device CD via the network NW. The network NW includes the Internet, WAN (Wide Area Network), LAN (Local Area Network) and the like. Further, the number of the user terminal device UD and the client terminal device CD that are communicably connected to the server device 10 is, for example, a plurality of each.

The user terminal device UD is an information processing device used by the user. The user terminal device UD has, for example, a browser or an application program for browsing a sales site provided by the server device 10 for selling a product or service (hereinafter referred to as "product or the like"). The user terminal device UD is used for a user to search for a product or the like introduced on a sales site, or to purchase or contract (including a reservation) (hereinafter, simply referred to as "purchase"). A product or the like is an example of an "object".

The client terminal device CD is an information processing device used by the client. The client terminal device CD is used by the client, for example, to register information such as a product to be sold (hereinafter referred to as "product information") in the server device 10.

The server device 10 provides a sales site for products and the like. When the server device 10 stores the product information registered by the client terminal device CD and receives the input of the user's search behavior through the user terminal device UD, the server device 10 publishes the product information narrowed down according to the user's search behavior. Distribute the site to the user terminal device UD.

The server device 10 of the present embodiment includes, for example, a product information acquisition unit 100, an action information acquisition unit 200, a similarity analysis unit 300, a session division unit 400, a calculation model learning unit 500, and a CV prediction unit 600. The information output unit 700 and the storage unit 800 are provided.

The product information acquisition unit 100, the behavior information acquisition unit 200, the similarity analysis unit 300, the session division unit 400, the calculation model learning unit 500, the CV prediction unit 600, and the information output unit 700 are all or part of, for example, a CPU ( It is realized by executing a program (software) by a hardware processor such as the Central Processing Unit. In addition, some or all of these components are hardware (circuits) such as LSI (Large Scale Integration), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), and GPU (Graphics Processing Unit). It may be realized by the part; including circuitry), or it may be realized by the cooperation of software and hardware.

The storage unit 800 is realized by, for example, a RAM (Random Access Memory), a ROM (Read Only Memory), an HDD (Hard Disk Drive), a flash memory, or a hybrid storage device in which a plurality of these is combined. All or part of the storage unit 800 may be realized by an external device such as NAS (Network Attached Storage) or an external storage server that can be accessed by the processor of the server device 10. The storage unit 800 stores a product information database DB1, an action information database (past log) DB2, a category tree information database DB3, a calculation model information database DB4, and a CV predicted value information database DB5.

Next, each functional unit of the server device 10 will be described in detail.

First, the product information acquisition unit 100 will be described. The product information acquisition unit 100 acquires the product information input from the client terminal device CD through the network NW. "Product information" is information about products, etc. introduced on the sales site, such as the name, contents, producer (provider), price, shipping fee, inventory quantity, estimated delivery date, etc. of the product, etc. .. The "product information" includes category information indicating a category such as a product. The category information specifies the hierarchical category to which the product or the like belongs, and also includes the "category path P" shown below.

FIG. 2 is a diagram for explaining the category path P. As shown in FIG. 2, the hierarchical category includes the first layer, the second layer, ..., The Nth layer (N is a natural number of 3 or more) in order from the broad concept. Here, for convenience of explanation, a four-layer hierarchical category will be described, but the category hierarchy may be three or less layers or five or more layers.

The “category path P” means one virtual path from the top layer category to the lowest layer category via the middle layer category in the hierarchical category. For example, in the example shown in FIG. 2, the category path P of a certain product belonging to the category "Western food" of the lowest layer (fourth layer) is "book (first layer)> childcare (second layer)> cooking (first layer). 3rd floor)> Western food (4th floor) ". There are as many category paths P as there are categories in the lowest layer in each hierarchical category. Such a category path P is given to each product or the like when the product information is registered in the server device 10 through the client terminal device CD. For example, in the hierarchical category of the sales site, by designating the category of the lowest layer to which the newly registered product or the like belongs, the category path P is given to the registered product or the like.

FIG. 3 is a diagram showing an example of the contents of the product information of the present embodiment. As shown in FIG. 3, for example, the product information registered for a product or the like includes information indicating which category the product or the like belongs to in each category of a plurality of layers. As described above, the fact that the information indicating which category the product or the like belongs to is registered in each of the plurality of layers corresponds to an example of "the category pass P is given to the product or the like". Instead of this, information indicating which category of the lowest layer the product belongs to is registered for the product, etc., and the correspondence between each category of the lowest layer and the category higher than that is registered. Even if the information indicating the above can be referred to, it corresponds to an example of "a category pass P is given to a product or the like".

Next, the behavior information acquisition unit 200 will be described. The behavior information acquisition unit 200 acquires the user's behavior information through the user terminal device UD for the sales site. The behavior information acquisition unit 200 acquires the user's behavior information each time it receives a new behavior of the user. The behavior information acquisition unit 200 registers the acquired behavior information of the user in the behavior information database DB2 and outputs it to the similarity analysis unit 300 and the CV prediction unit 600. The user behavior information includes information on the user's search behavior and information on the user's specific behavior (described later).

The "search action" is, for example, an action for searching for or narrowing down a product or the like. For example, "search behavior" means inputting a query (search query), clicking on a button for narrowing down the products set on the sales site, and changing the display order of the products set on the sales site. Includes clicks on buttons etc. to do. The "information about the search behavior" includes the content of the search behavior by the user, the time when the search behavior is performed, and the like. Further, the "button" referred to in the present application may be a virtual one displayed on the screen, and includes a button that can be equated with a button such as a radio box.

The "specific action" is, for example, an action for an individual product or the like narrowed down by a search action. For example, the "specific action" is an action of selecting one or more products or the like from a plurality of products or the like displayed on the sales site. For example, the "specific action" includes a click on a hyperlink to a detailed introduction page of a product or the like, a click on a purchase button for purchasing a product or the like, and the like. The "information about the specific action" includes the content of the specific action by the user, the time when the specific action is performed, and the like.

FIG. 4 is a diagram showing an example of the content of the user behavior information of the present embodiment. As shown in FIG. 4, in the user behavior information, each behavior of the user (for example, a search behavior and a specific behavior) is managed in association with the time when the behavior is performed on the sales site. In addition, user behavior information is managed for each user.

Next, the similarity analysis unit 300 will be described. The similarity analysis unit 300 derives the similarity between the plurality of search behaviors based on the contents of the plurality of search behaviors of the user with respect to the sales site. In the present embodiment, the similarity between the plurality of queries is derived as the similarity between the plurality of search actions based on the contents of the plurality of clicks input from the user to the sales site.

Here, the "query" referred to in the present application is defined. A query is a written representation of a request for a website in a predetermined format. A query is, for example, a character entered directly into a website. However, in the present embodiment, the similarity analysis unit 300 is not limited to the directly input query, and even when a search action that can be equated with the directly input query is performed, the directly input query is used. Treat search behaviors that can be equated as query input. "Search behavior that can be equated with a directly entered query" is a click on a hyperlink to a page related to a certain item (for example, a certain product).

Further, in the present embodiment, when a plurality of terms (keywords) are input at one time in one search box, the input plurality of terms are collectively treated as one query. For example, if an input such as "this novel" is entered in the search box, "this novel" becomes one query. In other words, the query entered as "book" and the query entered as "book" are treated as different queries.

FIG. 5 is a block diagram showing the configuration of the similarity analysis unit 300 of the present embodiment. The similarity analysis unit 300 of the present embodiment collects the category paths given to each of the plurality of products selected by the user group for each user group in which the same query is input. Then, the similarity analysis unit 300 generates a category tree for each query based on the collected plurality of category paths, and derives the similarity between the plurality of queries based on the generated category tree for each query. This content will be described in detail below. As shown in FIG. 5, the similarity analysis unit 300 includes, for example, a category path collection unit 310, a category tree generation unit 320, and a similarity derivation unit 330.

The category path collecting unit 310 collects the category path P given to each of the plurality of products or the like selected by the user group on the sales site for each user group in which the same query is input. More specifically, the category path collecting unit 310 identifies a user who has input the same query (hereinafter, referred to as "specific query") among a plurality of users who have performed a search action on the sales site. For example, the category path collecting unit 310 extracts the user who has input a specific query by referring to the user behavior information registered in the behavior information database DB2.

In addition, the category path collecting unit 310 identifies the product or the like selected immediately after the user who entered the specific query inputs the specific query. In addition, "immediately after inputting a specific query" means that there is no input of another query between the input of the specific query and the selection of the product or the like. “Immediately after inputting a specific query” means that another action such as changing the display order of products or the like may exist between the input of the specific query and the selection of the product or the like. Further, "selecting (a product or the like)" means that, for example, the above-mentioned specific action was performed on the product or the like. For example, in the present embodiment, the category path collecting unit 310 considers that a button or hyperlink related to an individual product or the like is clicked (accessed) as "a product or the like is selected". .. Instead of this, the category pass collecting unit 310 may consider that "the product or the like has been selected" only when the product or the like is actually purchased. The category path collecting unit 310 refers to the user behavior information registered in the behavior information database DB2 and identifies the product or the like selected immediately after the user inputs a specific query by referring to the time information included in the user behavior information. To do. The category path collecting unit 310 executes the above process for all users who have input the same query.

Further, the category path collecting unit 310 refers to the product information registered in the product information database DB1, and the category path P given to the product or the like selected immediately after the user who input the specific query inputs the specific query. To collect. Here, when a plurality of users who have input the same query select different products and the like, a plurality of category paths P corresponding to each of the plurality of products and the like are collected. The category path collecting unit 310 executes the above processing for all products and the like selected by the user group who input the same query.

FIG. 6 is a diagram showing an example of a collection of category paths P collected by the category path collection unit 310 for a group of users who input a certain query. FIG. 6 shows a set of category paths P such as products selected (for example, by clicking a hyperlink) immediately after inputting a specific query by a user who input "book" as a specific query.

As described above, the category path collecting unit 310 collects the category paths P given to each of the plurality of products and the like selected by the user group on the sales site for each user group in which the same query is input. The category path collection unit 310 outputs a collection of category paths P collected for each user group to the category tree generation unit 320. The category path collecting unit 310 performs the above processing for each of all the queries input to the sales site during the predetermined period.

Returning to FIG. 5, the category tree generation unit 320 is based on a plurality of category paths P collected for each user group collected by the category path collection unit 310, and the same category included in the plurality of category paths P. By integrating, a category tree T for each query is generated. The category tree generation unit 320 has, for example, a category path integration unit 321 and a weighting unit 322.

The category path integration unit 321 integrates the same categories included in a plurality of category paths P collected for each user group collected by the category path collection unit 310. That is, the category path integration unit 321 determines whether or not there are two or more category paths P having the same category in the same hierarchy among the plurality of category paths P. Then, when there are two or more category paths P having the same category in the same hierarchy, the category path integration unit 321 integrates the category and a category higher than the category, so that the two or more categories are integrated. Convert the path P into one category tree T. The category path integration unit 321 integrates the categories as described above in each layer of the hierarchical categories.

FIG. 7 is a diagram showing an example of the category tree T generated by the category path integration unit 321. The category tree T shown in FIG. 7 is a category tree generated based on the set of category paths P shown in FIG. As shown in FIG. 7, when two or more category paths P having the same category exist in the same hierarchy, the category and the categories higher than the category are integrated to generate one category tree T. There is.

Further, the category path integration unit 321 sets the root node RN at a higher level than the category of the first layer in the generated category tree T. The root node RN is treated as a virtual category set in a higher layer (0th layer) of the first layer. The root node RN has the same contents in the category tree T of all queries. The role of the root node RN will be described later.

The weighting unit 322 assigns weights to each category included in the category tree T generated by the category path integration unit 321. More specifically, the weighting unit 322 assigns weights to the categories integrated in the process of generating the category tree T. For example, the weighting unit 322 derives the magnitude of the weight given to the category according to the number of the same categories integrated in the process of generating the category tree T.

In the present embodiment, the weighting unit 322 first assigns a score to the category according to the number of the same categories integrated in the process of generating the category tree T. For example, in the example in which the category tree T shown in FIG. 7 is generated based on the plurality of category paths P shown in FIG. 6, the category “book” is set to “3” and the category “childcare” is set to. A score of "2" is given to the category "cooking", and a score of "1" is given to other categories.

Then, the weighting unit 322 derives the weight value given to each category by normalizing the scores given to all the categories included in one category tree T in the category tree T unit. In the example shown in FIG. 7, the weight value given to each category is derived by dividing the score given to each category by "12" which is the total value of the scores given to all categories. To. For example, "0.25" for the category "book", "0.167" for the category "childcare", "0.167" for the category "cooking", and "0.67" for the other categories. The weight of "083" is given.

The category tree generation unit 320 performs the above processing on all the queries input to the sales site during the predetermined period, and generates a weighted category tree T for each query. The weighted category tree T for each query generated by the category tree generation unit 320 is registered in the category tree information database DB3.

Next, another example of the category tree T generated by the category path integration unit 321 will be described with reference to FIG. The example shown in FIG. 8 is an example in which a certain query has a plurality of meanings. For example, a term included in one query may refer to a certain "fishery origin" and a "fashion brand" that has nothing to do with the marine product origin. In this case, the products, etc. that the user who entered the query clicked immediately after entering the query are the products, etc. that belong to the category related to the production area of the marine products, and the products, etc. that belong to the category related to the fashion brand. May be. In this case, since there are two different hierarchical categories, the category path P of each product or the like cannot be integrated into one category tree T as it is. Therefore, in the present embodiment, as shown in FIG. 8, the root node RN is set higher than the category of the first layer. Then, the two category paths P extracted from the two different hierarchical categories are combined by the root node RN. Thereby, even if one query has a plurality of meanings, one category tree T corresponding to the query can be generated.

Next, returning to FIG. 5, the similarity deriving unit 330 will be described. The similarity derivation unit 330 derives the similarity between the plurality of search actions based on the contents of the plurality of search actions of the user. In the present embodiment, the similarity between a plurality of queries is derived as the similarity between the plurality of search actions based on the content of the query input by the user. The similarity derivation unit 330 derives the similarity between the plurality of queries to be compared based on the similarity between the category trees T generated by the category tree generation unit 320 for each of the plurality of queries. In the present embodiment, the similarity deriving unit 330 obtains the similarity between the two weighted category trees T by using the intertree edit distance. The similarity derivation unit 330 is an example of the “first derivation unit”.

FIG. 9 is a diagram showing an example of how to obtain the similarity between two weighted category trees T. As shown in FIG. 9, the similarity derivation unit 330 derives the similarity between the plurality of queries based on the degree of matching between the plurality of categories included in the category tree T corresponding to each of the plurality of queries. For example, the similarity derivation unit 330 determines whether or not the categories included in the category tree T corresponding to each of the plurality of queries match each other in order from the highest category (for example, the root node RN), and determines whether or not the categories are the same. Derivation of similarity between multiple queries is based on how many categories match in a row from. Here, the higher the number of consecutive matching categories from the highest category, the higher the similarity is derived.

In the example shown in FIG. 9, the category tree T on the left side in FIG. 9 is the category tree corresponding to the query "book", and the category tree T on the right side in FIG. 9 corresponds to the category different from the "book". It is a tree. In the example shown in FIG. 9, the category "book", the category "childcare", and the category "cooking" are the same categories in the two category trees T when viewed from the highest category (root node RN).

In the present embodiment, the similarity deriving unit 330 has a plurality of similarity deriving units 330 based on the degree of matching between a plurality of categories included in the category tree T corresponding to each of the plurality of queries and the weight given to the matching category. Derive the similarity between queries. For example, the similarity derivation unit 330 determines whether or not the categories included in the category tree T corresponding to each of the plurality of queries match each other in order from the highest category, and the weights given to the matching categories. By adding the values of, the similarity between multiple queries is derived. As the weight given to the matching category, for example, the value of the smaller weight in the two category trees T is adopted.

In the example shown in FIG. 9, the category "book", the category "childcare", and the category "cooking" are the same categories in the two category trees T when viewed from the top category. The weight given to the category "book" of the category tree T on the left side is "0.25", and the weight given to the tegori "book" of the category tree T on the right side is "0". .20 ". In this case, "0.20", which is the smaller of "0.25" and "0.20", is the value of the weight to be added. Similarly, the smaller weight given to the category "childcare" in the two category trees T is "0.15", and the weight given to the category "cooking" in the two category trees T. The smaller "0.15" is added in order. Then, "0.5", which is the total value obtained by adding these "0.20", "0.15", and "0.15", is derived as a value indicating the similarity between the two category trees T. The larger the derived similarity value, the higher the similarity between the two queries.

Further, the similarity derivation unit 330 monitors a query input to the sales site (that is, a query accepted by the behavior information acquisition unit 200) at the stage of deriving the conversion prediction value described later. Then, the similarity derivation unit 330 acquires the category tree T corresponding to the newly input query from the category tree information database DB3 every time a query is newly input by the user. Then, the similarity derivation unit 330 compares the newly input query with the category tree T corresponding to the newly input query and the category tree T corresponding to the previously input query, so that the newly input query and 1 Derives the similarity with the previously entered query. In the similarity derivation unit 330, when the similarity between the newly input query and the previously input query is less than or equal to the threshold value, the similarity between the newly input query and the previously input query is the same. A signal indicating that the value is equal to or less than the threshold value is output to the session division unit 400.

Next, returning to FIG. 1, the session dividing unit 400 will be described. The session dividing unit 400 divides a session among the plurality of search actions based on the similarity between the plurality of search actions derived by the similarity deriving unit 330. In the present embodiment, the session dividing unit 400 receives a plurality of queries based on the similarity between the plurality of queries derived by the similarity deriving unit 330 with respect to the plurality of queries input from the user. Split the session between. The "session" is, for example, a valid period of a state management function such as a cookie. For example, the period from accessing a certain web page on a website to the elapse of a predetermined time (timeout) is treated as one session. In addition, the session may be a period from accessing a certain web page in the website to switching to another web page in the website or a web page in another website. In addition, the session may be a period from accessing a certain web page in the website to closing the web browser that displays the web page.

FIG. 10 is a diagram showing an example of session division by the session division unit 400. In the present embodiment, when the session division unit 400 receives a signal from the similarity analysis unit 300 indicating that the similarity between the newly input query and the previously input query is equal to or less than the threshold value, the session division unit 400 newly inputs. The session is divided between the time when the query entered is received and the time when the query entered immediately before is received.

In the example shown in FIG. 10, the session is divided between the time when the query "this novel" is received and the time when the query "clothes" is received. According to the session division unit 400 of the present embodiment, even if the elapsed time from the time when the query "this novel" is received to the time when the query "clothes" is received is short, the similarity between the two queries is equal to or less than the threshold value. If so, the session is split between the times when the two queries are accepted. If there is another action of the user between the input of two queries (for example, clicking on the button to narrow down the product etc.), the session is divided immediately before the time when the newly input query is accepted. .. When the session is divided, the session division unit 400 sets the start time of the divided session and the session newly started by dividing the session (hereinafter referred to as "divided session") by the calculation model learning unit 500 and the CV. Output to the prediction unit 600.

Next, returning to FIG. 1, the calculation model learning unit 500 will be described. The calculation model learning unit 500 trains a calculation model that derives a user's conversion prediction value based on the user's search behavior. In the present embodiment, the calculation model learned by the calculation model learning unit 500 determines the user's conversion prediction value for the product or the like presented to the user during the division session based on the user's search behavior in the division session. It is a calculation model to be derived.

More specifically, the input to the above calculation model includes request feature information and session feature information. On the other hand, the output from the above calculation model is the conversion prediction value of the user.

The request feature information is, for example, information indicating the content of the user's search action and the time when the search action was performed. The content of the search behavior is, for example, the content of the query input by the user, the content of the user's request regarding the change of the display order (order) of the products, etc. on the sales site, and the user regarding the narrowing conditions of the products, etc. on the sales site. The content of the request. The content of the query input by the user is converted into, for example, a vector representation and input to the above calculation model. The vector representation is, for example, a dense vector in which the contents of the query are expressed in a distributed representation, but may be a sparse vector in which the contents of the query are represented in a local representation. The display order of the products and the like on the sales site is whether to display the products or the like in the order of popularity or the order of price. The conditions for narrowing down the products on the sales site are whether or not the products are in stock, the scheduled delivery date of the products is within a predetermined number of days, and the shipping fee is free. The request feature information can be obtained based on the user behavior information acquired by the behavior information acquisition unit 200.

The session feature information includes time information related to the split session and the average value of request features in the split session. The time information regarding the split session is, for example, the elapsed time from the start of the split session. The "elapsed time from the start of the split session" is, for example, the elapsed time from the start of the split session to the time when the user receives a search action (query input, change of display order of products, request for narrowing down, etc.). For example, when a user performs a plurality of search actions in one split session, the "elapsed time from the start of the split session" is, for example, each elapsed time from the start of the split session to each time point when each search action by the user is accepted. It's time. The "mean value of request features" is, for example, a vector of "request feature information" obtained for each request when the divided session is regarded as a set of user requests (for example, query input) (contents of the above-mentioned query). It is the average value of the vector represented by the distributed representation or the local representation). The session identity information is obtained based on the information derived or acquired by the session division unit 400 and the action information acquisition unit 200.

The conversion prediction value is the probability that the user takes a predetermined action for the product or the like presented to the user during the divided session. Predicted conversion values can be defined differently for different purposes. An example of the conversion prediction value is the probability that the user purchases a product or the like presented to the user during the split session. However, the conversion prediction value is not limited to the above example, and may be the probability that the user clicks a hyperlink such as a product presented to the user during the split session, or may be based on another definition.

In the present embodiment, the calculation model learning unit 500 trains the learning model using the user's behavior information and session information acquired or derived by the behavior information acquisition unit 200 and the session division unit 400 as teacher data in the past fixed period. .. The user's behavior information acquired by the behavior information acquisition unit 200 includes the request feature information and information indicating whether or not the user has actually performed the predetermined action (for example, purchase of a product or the like). The information (for example, conversion rate) indicating whether or not the predetermined action is actually performed for the plurality of users is the correct answer data for the conversion prediction value. The session information acquired by the session division unit 400 contains time information (for example, the elapsed time from the start of the division session to the time when the predicted conversion value is derived) regarding the division session divided based on the similarity between a plurality of queries. Including.

The calculation model learning unit 500 learns a calculation model showing the relationship between the user's behavior information and session information and the conversion prediction value by machine learning using the above information as teacher data. Machine learning methods are, for example, Support Vector Machine (SVM) and deep running, but are not limited thereto. The calculation model learning unit 500 registers the learned calculation model in the calculation model information database DB4.

The CV prediction unit 600 derives the conversion prediction value of the target user based on the search behavior of the user (hereinafter referred to as "target user") who wants to obtain the conversion prediction value. The CV prediction unit 600 derives the conversion prediction value of the target user for the product or the like presented during the division session based on the search behavior of the target user in the division session divided using the similarity between queries. The CV prediction unit 600 derives a conversion prediction value of the target user based on the content of the search behavior of the target user in the split session and the time information related to the split session. For example, the CV prediction unit 600 derives a conversion prediction value of the target user based on the content of the search behavior of the target user in the split session and the elapsed time from the start of the split session. The CV prediction unit 600 is an example of a “second out-licensing unit”.

More specifically, the CV prediction unit 600 receives the user's behavior information acquired by the behavior information acquisition unit 200. Further, the CV prediction unit 600 receives the session information derived by the session division unit 400 (for example, the elapsed time from the start of the division session). Further, the CV prediction unit 600 reads out the calculation model learned by the calculation model learning unit 500 by referring to the calculation model information database DB4. Then, the CV prediction unit 600 derives a conversion prediction value based on, for example, the content of each search behavior of the user, the elapsed time from the start of the split session, and the above calculation model. The CV prediction unit 600 registers the information indicating the derived conversion prediction value in the CV prediction value information database DB5.

Instead of the above, the CV prediction unit 600 searches for each user from the start of the divided session based on the user's behavior information acquired by the behavior information acquisition unit 200 and the session information derived by the session division unit 400. Each elapsed time up to each point in time when the action was taken may be derived. Then, the CV prediction unit 600 determines the conversion prediction value based on the content of each search action of the user, each elapsed time from the start of the split session to each time point when each search action of the user is performed, and the above calculation model. May be derived.

Finally, the information output unit 700 will be described. The information output unit 700 reads out the conversion predicted value registered in the CV predicted value information database DB 5 and outputs it to the outside according to a predetermined cycle or when a request from the outside is received.

Next, an example of the processing flow by the server device 10 will be described. The processing flow by the server device 10 is roughly divided into (1) a category tree derivation stage and (2) a session division using the similarity between queries and a conversion prediction value derivation stage. Hereinafter, these contents will be described.

FIG. 11 is a flowchart showing an example of the processing flow at the derivation stage of the category tree. As shown in FIG. 11, first, the category path collecting unit 310 acquires the category path P previously assigned to the product group selected by the user group for each user group inputting the same query (S101). ..

Next, the category tree generation unit 320 integrates the same categories included in a plurality of category paths of the group of category paths P for each user group collected by the category path collection unit 310. , Generate a category tree T (S102). Next, the category tree generation unit 320 assigns weights to the same category integrated in the category tree generation process (S103).

The server device 10 performs the processes S101 to S103 for each query to generate a category tree T for each query. This completes the derivation stage of the category tree T. This makes it possible to derive the inter-query similarity using the category tree T and to divide the session using the inter-query similarity. After that, a calculation model for deriving a conversion prediction value using the user's behavior information in the past fixed period and the session information obtained for the fixed period, which is divided into sessions according to the similarity between queries) as teacher data. Is learned.

FIG. 12 is a flowchart showing an example of the processing flow of the session division and the derivation stage of the conversion prediction value using the similarity between queries. As shown in FIG. 12, first, the behavior information acquisition unit 200 receives a new query input by the user each time the user inputs a new query (S201). Next, the similarity derivation unit 330 reads the category tree T corresponding to the newly input query from the category tree information database DB3 every time a new query is input from the user. Then, the similarity derivation unit 330 compares the category tree T corresponding to the newly input query with the category tree T corresponding to the previously input query (S202). As a result, the similarity derivation unit 330 derives the similarity between the newly input query and the previously input query (S203). The category tree T corresponding to the previously input query may be read from the category tree information database DB3 in the same manner as the category tree T corresponding to the newly input query, or may be read from the category tree information database DB3, such as RAM. It may be held on the volatile memory.

Next, the similarity derivation unit 330 determines whether or not the similarity between the newly input query and the previously input query is equal to or less than the threshold value (S204). When the similarity is higher than the threshold value (S204: NO), the session dividing unit 400 does not divide the session. In this case, the process proceeds to S206.

On the other hand, when the similarity is equal to or less than the threshold value (S204: YES), the session dividing unit 400 sets the time when the newly input query is received and the time when the previously input query is received. The session is divided between (S205). The session division unit 400 outputs information indicating that the session has been divided and the start time of the newly started divided session to the CV prediction unit 600.

The CV prediction unit 600 derives a conversion prediction value based on the content of the user's search behavior acquired by the behavior information acquisition unit 200, the session information derived by the session division unit 400, and the above calculation model (S206). ). The information output unit 700 outputs information indicating a conversion prediction value derived by the CV prediction unit 600. The server device 10 repeats the processes S201 to S206 every time a new query is input by the user, for example.

FIG. 13 is a diagram showing experimental results by simulating conversion predicted values using the server device 10 of the present embodiment. The "user" according to the session division method in FIG. 13 indicates a model in which a session is divided for each user, and the "time (30 minutes)" is the elapsed time from the input of the previous query to the input of a new query. A model in which a session is divided when the time is 30 minutes or more is shown, "w2v" shows a model using word2vec, and "category tree" shows a model described in the above embodiment. As shown in FIG. 13, it was confirmed that the session division using the category tree T as in the present embodiment improves the derivation accuracy of the conversion prediction value as compared with other methods. In particular, according to the method for deriving the similarity between queries using the category tree T as in the present embodiment, the similarity between the "gift certificate" and the "book ticket" or the katakana notation and the English notation of the product name. Was also confirmed to be able to be derived higher than other methods.

According to the server device 10 described above, more useful information related to the user's search behavior can be provided. That is, in the present embodiment, the server device 10 has a category path collection unit 310, a category tree generation unit 320, and a similarity derivation unit 330. The category path collecting unit 310 collects the category path P given to each of the plurality of products or the like selected by the user group on the sales site for each user group in which the same query is input. The category tree generation unit 320 generates a category tree T for each query by integrating the same categories included in a plurality of category paths P for each user group collected by the category path collection unit 310. The similarity derivation unit 330 derives the similarity between the plurality of queries based on the similarity between the category trees T generated by the category tree generation unit 320 for each of the plurality of queries. According to such a configuration, the similarity between a plurality of queries can be easily and accurately derived. This makes it possible to provide more useful information related to the user's search behavior.

In the present embodiment, the similarity derivation unit 330 derives the similarity between a plurality of queries based on the degree of matching between the plurality of categories included in the category tree T corresponding to each of the plurality of queries. According to such a configuration, the similarity between a plurality of queries can be derived by a relatively simple process.

In the present embodiment, the similarity deriving unit 330 determines whether or not the categories included in the category tree T corresponding to each of the plurality of queries match each other in order from the highest category, and continues from the highest category. Then, based on how many categories match, the similarity between multiple queries is derived. According to such a configuration, the similarity between a plurality of queries can be derived by a relatively simple process.

In the present embodiment, the category tree generation unit 320 gives weights to the categories integrated in the process of generating the category tree T. The similarity derivation unit 330 is based on the degree of matching between a plurality of categories included in the category tree T corresponding to each of the plurality of queries and the weight given to the matching category, and the degree of similarity between the plurality of queries. Is derived. According to such a configuration, the similarity of the plurality of category trees T can be determined more accurately by giving weights to the integrated categories. As a result, the accuracy of deriving the similarity between a plurality of queries can be improved.

In the present embodiment, the category tree generation unit 320 derives the weight to be given to the category according to the number of the same categories integrated in the generation process of the category tree T. According to such a configuration, the degree of integration of categories can be reflected, and the similarity of a plurality of category trees T can be derived with higher accuracy.

In the present embodiment, the category tree generation unit 320 assigns a score to the category according to the number of the same categories integrated in the generation process of the category tree T, and assigns a score to all the categories included in one category tree. The weight is derived by normalizing the score obtained in units of the category tree. According to such a configuration, even if the scale of the category tree T is significantly different for each query, the similarity of the plurality of category trees T can be accurately determined.

In the present embodiment, the similarity derivation unit 330 determines whether or not the categories included in the category tree T corresponding to each of the plurality of queries match each other in order from the highest category, and assigns the matching categories to the matching categories. By adding the values of the weights, the similarity between a plurality of queries is derived. According to such a configuration, the accuracy of deriving the similarity between a plurality of queries can be further improved.

In the present embodiment, the server device 10 has a similarity derivation unit 330 and a session division unit 400. The similarity derivation unit 330 derives the similarity between the plurality of search behaviors based on the contents of the plurality of search behaviors of the user with respect to the sales site. The session dividing unit 400 divides a session among a plurality of search actions based on the similarity derived by the similarity deriving unit 330. With such a configuration, it is possible to divide the session in consideration of the user's intention. This makes it possible to provide more useful information related to the user's search behavior.

In the present embodiment, the similarity deriving unit 330 derives the similarity between the plurality of queries as the similarity between the plurality of search actions based on the contents of the plurality of queries input by the user to the sales site. The session division unit 400 divides the session between the times when the plurality of queries are input, based on the similarity between the plurality of queries. With such a configuration, the session can be divided based on the contents of the query. This makes it possible to provide more useful information related to the user's search behavior.

In the present embodiment, the similarity deriving unit 330 queries not only the directly input query but also the search behavior that can be equated with the directly input query even when the search behavior can be equated with the directly input query. Treat as input of. With such a configuration, it may be possible to divide the session with higher accuracy than when the session is divided using only the query.

In the present embodiment, the server device 10 derives the user's conversion prediction value for the product or the like presented during the division session based on the user's search behavior in the division session obtained by dividing the session. A unit 600 is provided. With such a configuration, it is possible to derive a conversion prediction value that reflects the session division in consideration of the user's intention. Therefore, the accuracy of deriving the predicted conversion value can be improved.

In the present embodiment, the CV prediction unit 600 derives the conversion prediction value of the user based on the time information regarding the split session. According to such a configuration, since the conversion prediction value is derived based on the time information of the division session divided in consideration of the user's intention, it is possible to improve the derivation accuracy of the conversion prediction value.

In the present embodiment, the CV prediction unit 600 derives the conversion prediction value of the user based on the elapsed time from the start of the split session. According to such a configuration, the conversion prediction value is derived based on the elapsed time from the start time of the division session divided in consideration of the user's intention, so that the derivation accuracy of the conversion prediction value can be improved. ..

Although the server device 10 of the embodiment has been described above, the embodiment is not limited to the above example. Further, in the present application, "based on XX" means "based on at least XX", and includes a case where it is based on another element in addition to XX. Further, "based on XX" is not limited to the case where XX is directly used, but also includes the case where it is based on the case where calculation or processing is performed on XX. "XX" is an arbitrary element.

Although the embodiments for carrying out the present invention have been described above using the embodiments, the present invention is not limited to these embodiments, and various modifications and substitutions are made without departing from the gist of the present invention. Can be added.

10 ... Server device (information processing system), 300 ... Similarity analysis unit, 310 ... Category path collection unit, 320 ... Category tree generation unit, 330 ... Similarity derivation unit, 400 ... Session division unit, 600 ... CV prediction unit.

Claims (10)

A collection unit that collects category paths given to each of a plurality of products or services selected on the sales site by the user group for each user group that has entered the same query.
A generation unit that generates a category tree for each query by integrating the same categories included in a plurality of category paths for each user group collected by the collection unit.
A derivation unit that derives the similarity between a plurality of queries based on the similarity between the category trees generated by the generation unit for each of the plurality of queries.
Information processing system equipped with. The derivation unit derives the similarity between the plurality of queries based on the degree of matching between the plurality of categories included in the category tree corresponding to each of the plurality of queries.
The information processing system according to claim 1. The derivation unit determines whether or not the categories included in the category tree corresponding to each of the plurality of queries match each other in order from the highest category, and how many consecutively from the highest category. Derivation of similarity between the plurality of queries based on whether the categories of
The information processing system according to claim 1 or 2. The generator assigns weights to the categories integrated in the process of generating the category tree.
The derivation unit has a degree of similarity between the plurality of queries based on the degree of matching between the plurality of categories included in the category tree corresponding to each of the plurality of queries and the weight given to the matching category. To derive,
The information processing system according to any one of claims 1 to 3. The generation unit derives the weight to be given to the category according to the number of the same categories integrated in the generation process of the category tree.
The information processing system according to claim 4. The generation unit assigns a score to the category according to the number of the same categories integrated in the process of generating the category tree, and assigns a score to all the categories included in one category tree to the category tree. The weight is derived by normalizing in units.
The information processing system according to claim 4 or 5. The derivation unit determines whether or not the categories included in the category tree corresponding to each of the plurality of queries match each other in order from the highest category, and determines the value of the weight given to the matching category. By adding, the similarity between the plurality of queries is derived.
The information processing system according to claim 6. A session division unit that divides a session between the times when the plurality of queries are received is further provided based on the similarity between the plurality of queries derived by the derivation unit for a plurality of queries input from the user. Prepared
The information processing system according to any one of claims 1 to 7. The computer
For each user group who entered the same query, the category paths given to each of the plurality of products or services selected on the sales site by the user group are collected.
By integrating the same categories included in the collected multiple category paths for each user group, a category tree for each query is generated.
The similarity between a plurality of queries is derived based on the similarity between the category trees generated for each of the plurality of queries.
Information processing method. On the computer
For each user group who entered the same query, the category paths given to each of the plurality of products or services selected on the sales site by the user group are collected.
By integrating the same categories included in the collected multiple category paths for each user group, a category tree for each query can be generated.
The similarity between a plurality of queries is derived based on the similarity between the category trees generated for each of the plurality of queries.
program.

Images