JP5508232B2 - Query recommendation device, query recommendation method, query recommendation program - Google Patents

Description

  The present invention relates to a technique for providing a recommendation query to a user based on a search log of a search engine.

  In recent years, information on the Internet has increased rapidly, and the importance of search engines for searching for information required by users has increased. Since this search engine stores a log file indicating a user's search status, an attempt is made to analyze the log file and use it to improve the search service.

  Table 1 shows a typical log file. Here, information relating to items of “date”, “user ID”, “query name”, “click URL”, and “ranking order” is recorded. Specifically, the date when the user clicked the URL of the search result is recorded in the item “date”, and the ID owned by the user who searched using the search query is recorded in the item “user ID”. The query name used by the user for the search is recorded in the “query name” item, the URL clicked by the user in the search result is recorded in the “click URL” item, and the click in the “ranking order” item. The URL search ranking is recorded.

Search engines use such log files to improve the efficiency of user search behavior, for example, to present recommended queries. As shown in FIG. 1, this recommendation query is defined using a bipartite graph in which a query input by a user to a search engine and a URL where a search result is clicked are paired. In FIG. 1, q _{1 to} q ₃ on the left side indicate user input query names, and u _{1 to} u ₅ on the right side indicate URLs clicked by the user from the search results, and the two are connected by an edge. Such a bipartite graph is analyzed, and a strongly connected query is used as a recommendation query.

  However, the recommendation query based on the bipartite graph is effective because the search log in the log file is effective for the frequently used query, but the query is not frequently used because the search log is small. There is a problem that queries that can be recommended are rare.

Therefore, Non-Patent Document 1 proposes a query recommendation technique for queries that are less frequently used. Here, a query is recommended using a click graph (bipartite graph) that links the user-input query and the actually clicked URL, and a skip graph that links the user-input query and the URL that is not actually clicked. doing. FIG. 2 shows an example of a skip graph obtained from the click graph of FIG. 1. According to the skip graph, queries not connected in the click graph are connected (for example, q ₁ .q ₂ .q ₃ are connected). with connection via a u _2), it is possible to increase the number of queries to be recommended to the user.

  Here, the feature amount obtained from the click graph is treated as an explicit degree of association between the queries connected by the clicked URL. In addition, the feature amount obtained from the skip graph is treated as an implicit degree of association between queries connected via a URL that the user has not clicked. A query recommended to the user is extracted using the sum of the feature quantities between queries obtained from such a click graph and the feature quantities between queries obtained from the skip graph.

Yang Song, Li-wei He, "Optimal Rare Query Suggestion With Implicit User Feedback", ACM WWW 2010, April 26-30,2010

  However, according to the method of Non-Patent Document 1, since the edge of the skip graph is created and weighted to the URL that has not been clicked for the search result of the target query, the user who does not click the corresponding URL When the number increases, a large weight is given. In this case, when the user's unclicked URL is assumed to be a URL not related to the query, many unrelated queries are recommended, and the search efficiency of the user may be deteriorated.

  In addition, when creating a skip graph using a log file, there are many URLs that do not have a click relationship with the query in the log file, so there are many queries that are judged to be related queries. Many queries may be recommended.

  The present invention has been made to solve the above-described problems of the prior art, and it is a solution to provide a highly accurate recommendation query to the user and contribute to improvement of the search efficiency of the user. .

  Therefore, the present invention sequentially extracts the search log of each query linked by the clicked URL in the search log, and calculates the importance of the URL in the extracted search log, that is, the URL interposed between the queries. By calculating the feature quantity between queries in the skip graph based on the importance of this URL, the influence of the click status (for example, many URLs that have not been clicked) in the search result is reduced, and the degree of association between queries is calculated. Then, a highly accurate recommendation query is extracted. At this time, the feature amount between queries in the skip graph can also be calculated using the distance between queries. This makes it possible to reduce the feature amount of a query with a low degree of association.

  Specifically, one aspect of the query recommendation device according to the present invention extracts a search log of each query linked with the clicked URL from the search query storage unit, and determines the importance of the URL in the extracted search log. Query-related log extraction means that is calculated based on the click probability from each query, a click graph that links each query and the clicked URL based on the search log extracted by the extraction means, and each query and non-click Graph creation means for creating a skip graph obtained by connecting the URLs of URLs with edges, and click graph feature quantity extraction for calculating the feature quantity between queries in the click graph created by the graph creation means in accordance with the arrival probability via URL And the feature quantity between the queries in the skip graph created by the graph creation means is calculated as U calculated by the extraction means. A skip graph feature amount extraction means for calculating with the importance of L, and a feature amount synthesis means for calculating the feature amount between the final queries by combining the feature amounts calculated by the feature amount extraction means, A feature is that a recommended query is presented to a user in accordance with a feature quantity between queries synthesized by a feature quantity synthesis unit.

  In another aspect of the query recommendation device according to the present invention, when the search log of each query linked by the clicked URL is extracted from the search query storage unit, the position information of the query sequentially linked by the URL is extracted between the queries. Query-related log extraction means for setting the URL according to the number of URLs and calculating the importance of the URL based on the click probability from each query, and each query and the clicked URL based on the search log extracted by the extraction means A graph creation unit that creates a click graph linked by an edge, a skip graph in which each query and an unclicked URL are linked by an edge, and a feature amount between queries in the click graph created by the graph creation unit is a URL. Click graph feature amount extraction means that is calculated according to the via arrival probability, and skipping created by the graph creation means A skip graph feature amount extraction unit that calculates a feature amount between queries in F using a distance between queries based on position information set by the extraction unit and an importance level of the URL calculated by the extraction unit; A feature amount synthesizing unit that synthesizes the feature amounts calculated by the feature amount extraction units and calculates a feature amount between the final queries, and according to the feature amounts between the queries synthesized by the feature amount synthesis unit A recommendation query is presented to the user.

  One aspect of the query recommendation method according to the present invention is to extract a search log of each query linked with the clicked URL from the search query storage unit, and to determine the importance of the URL in the extracted search log as the click probability of each query. A query-related log extraction step that is calculated based on the search log, a click graph that links each query and the clicked URL based on the search log extracted in the extraction step, and each query and an unclicked URL at the edge A graph creation step for creating a connected skip graph, a click graph feature amount extraction step for calculating a feature amount between queries in the click graph created in the graph creation step according to an arrival probability via URL, and a graph creation step Extracting the feature quantity between queries in the skip graph created in step A skip graph feature amount extraction step for calculating the importance of the calculated URL, and a feature amount synthesis step for calculating the feature amount between the final queries by combining the feature amounts calculated in the feature amount extraction steps. And a recommended query is presented to the user according to the feature quantity between the queries synthesized in the feature quantity synthesis step.

  In another aspect of the query recommendation method according to the present invention, when the search log of each query linked by the clicked URL is extracted from the search query storage unit, the position information of the query sequentially linked by the URL is extracted between the queries. A query related log extracting step that sets the URL according to the number of URLs and calculates the importance of the URL based on the click probability from each query, and each query and the clicked URL based on the search log extracted in the extracting step. A graph creation step for creating a click graph linked by an edge, a skip graph by linking each query and an unclicked URL by an edge, and a feature amount between queries in the click graph created by the graph creation step Click graph feature amount extraction step calculated according to the via arrival probability, and graph creation step The skip graph feature is calculated by using the distance between queries based on the position information set in the extraction step and the importance of the URL calculated in the extraction step. An amount extraction step; and a feature amount synthesis step that calculates the feature amount between the queries by combining the feature amounts calculated in the feature amount extraction steps, and between the queries combined in the feature amount combination step A recommendation query is presented to the user in accordance with the feature amount.

  In addition, this invention is good also as an aspect of the program which makes a computer function as said apparatus. This program can be provided through a network or a recording medium.

  ADVANTAGE OF THE INVENTION According to this invention, a highly accurate recommendation query can be provided to a user and it can contribute to the improvement of a user's search efficiency.

Click graph (bipartite graph) example. Example of skip graph. The structural example of the query recommendation apparatus which concerns on embodiment of this invention. The flowchart of the same query recommendation method. 5 is a flowchart showing query position calculation in S01 in FIG. The flowchart which shows importance calculation of the URL. Explanatory drawing which shows the example of feature-value calculation between queries from the same click graph.

≪Example of device configuration≫
Based on FIG. 3, the structural example of the query recommendation apparatus which concerns on embodiment of this invention is demonstrated. This recommendation device 1 creates a click graph and a skip graph based on a search log of a search engine, and presents a recommended query to a user using the degree of association between queries obtained from the both graphs, that is, the feature quantity between queries. To do.

  Specifically, the recommendation device 1 is linked with a search engine that provides a search service, and includes hardware resources such as a CPU, a memory (RAM), a hard disk drive device, and a communication device. As a result of the cooperation between the hardware resource and the software resource (OS. Application, etc.), the recommendation device 1 has a search query storage unit 2, a query related log extraction unit 3, a click graph creation unit 4, and a click graph feature value extraction. 5, a skip graph creation unit 6, a skip graph feature amount extraction unit 7, and a feature amount synthesis unit 8 are implemented.

  The storage unit 2 stores a search log (log file) such as a search query group input to the search engine and a search result. Here, the search log is stored in a storage device such as a memory (RAM) or a hard disk drive device in conjunction with the search log of the search engine. Table 2 shows an example of data stored in the storage unit 2 and records information on items of “date”, “user ID”, “query name”, “click URL”, and “ranking rank” as in Table 1. ing.

  The extraction unit 3 sequentially extracts each query connected by the clicked URL with reference to the storage unit 2, that is, a search log of each query connected by the URL. At this time, the position information of the queries sequentially linked via the clicked URLs is set according to the number of URLs interposed between the queries. In addition, the importance of the URL in the extracted search log, that is, the URL interposed between the queries is calculated. Here, the URL click probability is calculated based on the query usage frequency (the number of URL clicks relative to the number of query inputs), and the URL entropy is calculated based on the click probability. This entropy is normalized with the maximum entropy value of each URL. Normalized entropy is used for URL importance.

  The graph creation unit 4 creates a click graph in which each query and the clicked URL are connected by an edge based on the search log extracted by the extraction unit 3, and the feature amount extraction unit 5 includes the graph creation unit 4 The feature amount between queries in the click graph created in step 1, that is, the explicit relevance between the queries is calculated based on the probability that one query will reach the other URL via the URL. At that time, the ratio of the number of URL clicks to the number of query inputs is used.

  The graph creation unit 6 creates a skip graph in which each query and an unclicked URL are connected by an edge based on the search log extracted by the extraction unit 3, and the feature amount extraction unit 7 creates a query in the skip graph. A feature amount between them, that is, an implicit relevance between queries, is calculated using a distance based on the importance of the URL calculated by the extraction unit and position information of each query.

  The synthesizing unit 8 synthesizes the feature quantities calculated by the feature quantity extraction units 5.7 to calculate a feature quantity between final queries. A recommendation query to be presented to the user is extracted according to the feature amount between the queries synthesized here. The extracted recommendation query is displayed on a user-owned terminal (PC, mobile phone, etc.) via the Internet.

≪Process step≫
Hereinafter, processing steps of the recommendation device 1 will be described with reference to FIG. Here, a case where the query q1 is a query to be processed will be described. The query q1 is input to an input unit (not shown) and processing is started.

S01: The extraction unit 3, a query q1 is input to the input unit (not shown), it accesses the storage unit 2, search log (click logs for each query from the query q ₁ tied at the clicked URL in order ) To extract a subset. At this time, the extraction unit 3 sets position information of each query in the search log, and calculates the importance of each URL. The subset, the position information of each query, and the importance of each URL are output to the graph creation unit 4.6.

A setting example of position information of each query will be described with reference to FIG. Here, position information of a query having a URL click relationship (so-called related query) is set in order from the query q ₁ . The minimum distance of the position information of the relevant query from the query q _1, that is set according to the minimum over the number URL from the query q _1. First, S12. The maximum value “N_max” of the number of processes in S13 is set. In this processing example, “N_max = 2” is set, and the processing is started with an initial value “N = 0” (S11). Note that the value of “N_max” can be adjusted as appropriate.

Next, “N = N + 1 = 1” is set, and the value “1” of N is set as position information in the search log of q ₁ (S12), and the first search log extraction for the query q1 is performed. That is, a URL is extracted from a search log having q ₁ as a query, and a search log including a query linked to the URL, that is, a query having a click relationship with the URL is extracted (S13).

After this extraction, it is confirmed whether “N = N_max” is satisfied (S14). As a result of the confirmation, if it is satisfied, the process proceeds to S15. Since “N = 1” at this stage, “N = N_max” does not hold. Therefore, returning to S12, “N = N + 1 = 2” is set, and the position information in the search log is set to the value “2” of N for a query (related query 1) other than q ₁ included in the extracted search log. Set (S12). A second search log extraction for a query other than q ₁ set here is performed.

That is, a search log including the query set in the position information “2” is extracted, and a query associated with the URL included in the search log is extracted (S13). Since “N = 2” at this stage, S14 is established, and the process proceeds to S15. In S15, “N = N + 1 = 3” is set, and the position information in the search log for the query (related query 2) extracted in S13 is set to the value “3” of N (excluding related query 1). End the process. As a result, it is possible to set the position information of each query sequentially linked via the URL from the query q ₁ according to the minimum number of URLs. In this setting example, the position information is set as “N + 1” (the number of URLs + 1).

A method for calculating the importance of the URL will be described with reference to FIG. First, each URL in the subset and a query having a click relationship are extracted at the start of processing (S21). Here, as an example, the query q _i (i = 1, 2,..., N) having the click relationship with the URL of the expression (1) is extracted.

Next, the URL click probability of the formula (1) is calculated from the usage frequency of the query q _i extracted in S21 (S22). That is, the click probability (P _ik ) for the URL of the formula (1) is calculated using the formula (2) based on the number of times the query q _i is inserted in the subset.

Here, for all the queries q _i (i = 1, 2,..., N) having a click relationship with the URL of Expression (1), the click probability (P _ik ) is calculated using Expression (2). To do. Then, based on the click probability (P _ik ) calculated in S22, the URL entropy E _k of equation (1) is calculated using equation (3) (S23).

Finally done for all URL included a process of S21~S23 in the subset, normalized by the formula (4) using the maximum value of the entropy E _k for each URL (S24). The normalized entropy E _k and the importance of each URL.

S02: The graph creating unit 4 creates a click graph based on the subset extracted in S01 among the output information from the extracting unit 3. A method of creating this click graph will be described. Here, if the subset is the search log of Table 2, from the query of q ₁ , u1. Since u4 URL of is clicked, as shown in the click graph of FIG. 1, u ₁ from q _1. Each linking the edge to u _4. Similarly, based on the search log in Table 2, q ₂ . A click graph is similarly created for q ₃ . That is, q ₂ to u ₃ . Connect edges to u ₅ , q ₃ to u ₁ . u ₅ to link the edges. A click graph is created in this way, and the created click graph is output to the feature amount extraction unit 5 together with the output information (including the click probability P _ik ) from the extraction unit 3.

  S03: The feature quantity extraction unit 5 calculates the edge weight in the click graph created in S02, that is, the feature quantity (relevance) between queries. The feature amount from the click graph is calculated using equations (5) to (7).

The feature quantity between all the queries in the click graph is calculated using Expressions (5) to (7). Here, “click_value _ij ” indicates a feature amount representing the relationship between the query q _i and the query q _j . Further, “k” indicates the URL of the formula (1), and “P _ik × P _kj ” starts from the query q _i and passes through the URL of the formula (1), and the query q _j from the URL of the formula (1). The probability of reaching is shown. FIG. 7 shows a specific calculation method for the feature quantity “click_value ₁₂ ” representing the relationship between the query q ₁ and the query q ₂ in the click graph. The calculated feature quantity between queries is output to the synthesis unit 8.

S04: The graph creating unit 6 creates a skip graph based on the subset extracted in S01 in the output information from the extracting unit 3. A method of creating this skip graph will be described. Here, the subset is the search log of Table 2 as in S02. At this time, focusing on q ₁ , u ₁ . u ₄ is clicked, but u ₂ . u ₃ . u ₅ has not been clicked. So, u ₂ that it has not been clicked from q _1. u ₃ . Edges are connected to u ₅ as shown in FIG.

Q ₂ . even for q _3, u ₁ that have not been clicked from q _2. u ₂ . Connect an edge to u ₄ and not clicked from q ₃ u ₂ . u ₃ . u ₄ to link the edges. The skip graph is generated in this way, and the generated skip graph is output to the feature amount extraction unit 7 together with the output information from the extraction unit 3.

S05: The feature quantity extraction unit 7 calculates the edge weight in the skip graph created in S04, that is, the feature quantity between queries (degree of association between queries). At this time, the distance between the queries based on the position information of each query (the position of q _i −the position of q _j ) is calculated from the output information from the graph creation unit 6.

Using the distance between the queries and the importance E _k of the URL in the output information, the feature quantity between the queries in the skip graph is calculated. For this calculation, equation (8) is used, and the calculated feature quantity between the queries is output to the synthesis unit 8.

  S06: The synthesizing unit 8 performs S03. In S05, the feature values of the click graph output from the feature value extraction unit 5.7 and the feature values of the skip graph are synthesized, and the feature values between the final queries are calculated. Formula (9) is used for this calculation.

  In Expression (9), “α” represents a real number from “0 to 1.0”, and can be set to an arbitrary value in advance. In the following, we set “α” to “0.7”, and the examples are “romance fortune-telling”, “psychological test”, “tarot fortune-telling”, “first-name judgment”, “birthday fortune-telling” as queries. explain. Here, it is assumed that feature quantities between queries shown in Table 3 are obtained from the click graph, and feature quantities between queries shown in Table 4 are obtained from the skip graph.

At this time, the feature value between queries for "love fortune telling"
"Love divination"=>"Psychologicaltest" = 0.7 x 0.0 + 0.3 x 0.4 = 0.12
"Love divination"=>"Tarotdivination" = 0.7 x 0.5 + 0.3 x 0.2 = 0.41
"Love fortune telling"=>"Last name judgment" = 0.7 x 0.0 + 0.3 x 0.3 = 0.09
“Love fortune telling” ⇒ “Birthday fortune telling” = 0.7 × 0.1 + 0.3 × 0.0 = 0.07
Is calculated.

  If the calculated feature amount exceeds the threshold value, it is extracted as a recommended query, while if it does not exceed the threshold value, it is not extracted as a recommended query. This threshold value can be appropriately adjusted according to the specification. For example, if the threshold is set to “0.1”, “psychological test” and “tarot fortune-telling” are extracted as recommendation queries. The extracted recommendation query is displayed on the user-owned terminal via the Internet and presented to the user.

According to the above recommendation apparatus 1, when calculating the feature quantity between the query in the skip graph, as shown in equation (8), for utilizing the importance E _k of URL linking between query, the search log The influence of the middle click status (such as many URLs that have not been clicked) is reduced. Therefore, as the number of unclicked URLs in the search log increases as in Non-Patent Document 1, the feature amount does not increase. This makes it possible to present a highly accurate recommended query to the user, contributing to improvement of the user's search efficiency. it can.

At this time, in Non-Patent Document 1, when creating a skip graph, edges are created for all URLs that do not have a click relationship with the query. according to equation recommendation apparatus 1 (8), because of the consideration of the distance (position of -q _j of q ₁₎ between the query to be bound by URL, it is possible to lower the degree of association between such queries Also in this respect, a highly accurate recommendation query can be presented to the user.

Note that the present invention is not limited to the above-described embodiment, and for example, only the URL importance level E _k can be used for calculating the feature quantity between queries in a skip graph. In this case, S12. Without performing setting of the positional information S15, it is sufficient to remove the "position of 1 / q ₁ position -q _j +1" from the formula (8). At this time, although the distance between the queries linked by the URL is not taken into consideration, it is possible to obtain an effect of reducing the influence of the click state in the search log (there are many URLs that have not been clicked).

≪Programs≫
The present invention can also be configured as a document search program that causes a computer to function as some or all of the units 2 to 8 of the recommendation device 1. According to this program, S01 to S06. S11-S15. It is possible to cause the computer to execute part or all of S21 to S24.

  The program can be provided through a network such as a website or e-mail. The program is stored in a recording medium such as a CD-ROM, DVD-ROM, CD-R, CD-RW, DVD-R, DVD-RW, MO, HDD, BD-ROM, BD-R, or BD-RE. It is also possible to record, save and distribute. This recording medium is read using a recording medium driving device, and the program code itself realizes the processing of the above embodiment, so that the recording medium also constitutes the present invention.

DESCRIPTION OF SYMBOLS 1 ... Query recommendation apparatus 2 ... Search query memory | storage part (search query memory | storage means)
3 ... Query related log extracting unit (query related log extracting means)
4. Click graph creation unit (graph creation means)
5 ... Click graph feature value extraction unit (click graph feature value extraction means)
6 ... Skip graph creation unit (graph creation means)
7 ... Skip graph feature value extraction unit (skip graph feature value extraction means)
8... Feature amount synthesis unit (feature amount synthesis means)

Claims (9)

A query recommendation device for presenting a recommendation query to a user based on a search log stored in a search query storage means,
Query-related log extracting means for extracting a search log of each query linked with the clicked URL from the search query storage means, and calculating the importance of the URL in the extracted search log based on the click probability from each query;
Creation of a graph that creates a click graph in which each query and clicked URL are connected by an edge based on the search log extracted by the extraction means, and a skip graph in which each query and an unclicked URL are connected by an edge Means,
Click graph feature quantity extraction means for calculating the feature quantity between queries in the click graph created by the graph creation means according to the arrival probability via URL;
A skip graph feature amount extraction unit that calculates a feature amount between queries in the skip graph created by the graph creation unit based on the importance of the URL calculated by the extraction unit;
A feature amount combining unit that calculates the feature amount between the final queries by combining the feature amounts calculated by the feature amount extraction units;
A query recommendation device that presents a recommended query to a user according to a result of comparing feature amounts between queries synthesized by a feature amount synthesis unit with a preset threshold value . A query recommendation device for presenting a recommendation query to a user based on a search log stored in a search query storage means,
When extracting the search log of each query linked by the clicked URL from the search query storage means, the position information of the query sequentially linked by the URL is set according to the number of URLs between the queries, and the importance of the URL Query-related log extraction means for calculating the log based on the click probability from each query,
Creation of a graph that creates a click graph in which each query and clicked URL are connected by an edge based on the search log extracted by the extraction means, and a skip graph in which each query and an unclicked URL are connected by an edge Means,
Click graph feature quantity extraction means for calculating the feature quantity between queries in the click graph created by the graph creation means according to the arrival probability via URL;
Skip for calculating the feature quantity between queries in the skip graph created by the graph creation means using the distance between queries based on the location information set by the extraction means and the importance of the URL calculated by the extraction means Graph feature extraction means;
A feature amount combining unit that calculates the feature amount between the final queries by combining the feature amounts calculated by the feature amount extraction units;
A query recommendation device that presents a recommended query to a user according to a result of comparing feature amounts between queries synthesized by a feature amount synthesis unit with a preset threshold value . The extraction means calculates a click probability of each URL based on the number of clicks of the URL with respect to the number of query inputs,
Calculate the entropy of each URL based on the click probability, normalize the entropy of each URL with the maximum value of each entropy,
The query recommendation device according to claim 1, wherein the normalized entry value is set as the importance of each URL. The click graph feature quantity extraction means calculates the click probability of each URL based on the number of clicks of the URL with respect to the number of queries input,
The query recommendation device according to any one of claims 1 to 3, wherein the arrival probability of reaching the other query via a URL from one query is calculated based on the click probability. A query recommendation method executed by an apparatus for presenting a recommendation query to a user based on a search log stored in a search query storage unit,
A query-related log extracting step of extracting a search log of each query linked with the clicked URL from the search query storage unit, and calculating the importance of the URL in the extracted search log based on a click probability of each query;
Creation of a graph that creates a click graph in which each query and clicked URL are connected by an edge based on the search log extracted in the extraction step, and a skip graph in which each query and an unclicked URL are connected by an edge Steps,
A click graph feature amount extraction step for calculating a feature amount between queries in the click graph created in the graph creation step according to the arrival probability via URL;
A skip graph feature amount extraction step of calculating a feature amount between queries in the skip graph created in the graph creation step with the importance of the URL calculated in the extraction step;
A feature amount combining step of combining the feature amounts calculated in each of the feature amount extraction steps to calculate a feature amount between the final queries,
A query recommendation method characterized by presenting a recommended query to a user according to a result of comparing feature amounts between queries combined in a feature amount combining step with a preset threshold value . A query recommendation method executed by an apparatus for presenting a recommendation query to a user based on a search log stored in a search query storage unit,
When extracting the search log of each query linked by the clicked URL from the search query storage means, the position information of the query sequentially linked by the URL is set according to the number of URLs between the queries, and the importance of the URL A query-related log extraction step for calculating the log based on the click probability from each query,
Creation of a graph that creates a click graph in which each query and clicked URL are connected by an edge based on the search log extracted in the extraction step, and a skip graph in which each query and an unclicked URL are connected by an edge Steps,
A click graph feature amount extraction step for calculating a feature amount between queries in the click graph created in the graph creation step according to the arrival probability via URL;
Skip to calculate the feature quantity between queries in the skip graph created in the graph creation step using the distance between queries based on the location information set in the extraction step and the importance of the URL calculated in the extraction step A graph feature extraction step;
A feature amount combining step of combining the feature amounts calculated in each of the feature amount extraction steps to calculate a feature amount between the final queries,
A query recommendation method characterized by presenting a recommended query to a user according to a result of comparing feature amounts between queries combined in a feature amount combining step with a preset threshold value . In the extraction step, a click probability of each URL is calculated based on the number of clicks of the URL with respect to the number of query inputs,
Calculate the entropy of each URL based on the click probability, normalize the entropy of each URL with the maximum value of each entropy,
7. The query recommendation method according to claim 5, wherein the normalized entry value is set as the importance of each URL. 8. In the click graph feature amount extraction step, the click probability of each URL is calculated based on the number of clicks of the URL with respect to the number of queries input,
The query recommendation method according to any one of claims 5 to 7, wherein the arrival probability of reaching the other query via a URL from one query is calculated based on the click probability.   A query recommendation program for causing a computer to function as the query recommendation device according to claim 1.

Images