JP2018097560A - Information processing device, information processing method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an information processing device capable of more suitably ranking a suggest query, an information processing method, and a program.SOLUTION: An information processing device includes a learning unit for learning weight of background of a suggest query on the basis of a set of suggest queries that are exhibited on the basis of a query input in an electronic store and in which positive examples and negative examples are distinguished, and a ranking unit for ranking the suggest query on the basis of the weight of the query learned by the learning unit.SELECTED DRAWING: Figure 1

Description

  The present invention relates to an information processing apparatus, an information processing method, and a program.

  Conventionally, a technique for ranking suggest queries in a shopping system (electronic store) has been disclosed (see Patent Document 1). In this technique, the display order of product information is determined based on the listing score. The listing score is a value indicating an evaluation determined for a product or a product page of the product. The listing score is determined such that the listing score increases as the number of accesses (viewing number) to the product page increases, or the listing score increases as the number of purchases of the product increases.

JP 2013-050905 A

  However, in the conventional technology, for example, when operation is performed on a portal site where a plurality of electronic stores are registered, a specific store intentionally performs an operation so that the number of accesses to its own product page increases. In some cases, suggesting queries could not be properly ranked.

  The present invention has been made in view of such circumstances, and an object of the present invention is to provide an information processing apparatus, an information processing method, and a program capable of appropriately ranking a suggestion query. To do.

  One aspect of the present invention is a set of suggestion queries that are presented based on a query input at an electronic store, and based on the set of suggestion queries that are distinguished from positive examples and negative examples, An information processing apparatus comprising: a learning unit that learns feature weights; and a ranking unit that ranks the suggest queries based on the feature weights learned by the learning unit.

  According to one aspect of the present invention, it is possible to rank the suggest queries more appropriately.

It is a figure which shows the structure and use environment of the ranking management apparatus 100 which is an example of information processing apparatus. It is a figure which shows an example of the interface screen provided by the electronic store server. It is a figure which shows an example of the content of the data for learning 154 typically. 6 is a flowchart illustrating an example of a flow of processing executed by a learning unit 120 and a ranking unit 130.

  Hereinafter, an information processing apparatus, an information processing method, and a program according to embodiments of the present invention will be described with reference to the drawings. The information processing apparatus according to the embodiment is realized by one or more processors. An information processing device is a device that ranks a suggestion query presented to an input query (hereinafter referred to as input query) in an electronic store that sells products or services (hereinafter referred to as products) via a network. is there. The information processing apparatus learns the feature weight of the suggestion query for each input query, for example, based on the set of suggestion queries in which the positive example and the negative example are distinguished, and, for example, the input query based on the learned feature weight Rank the suggestion query by each. Note that the information processing device and the device that operates the electronic store need not be separate devices, and may be realized by a single processor. Hereinafter, the functions of the information processing apparatus will be disclosed step by step.

  FIG. 1 is a diagram illustrating a configuration and usage environment of a ranking management apparatus 100 that is an example of an information processing apparatus. The ranking management apparatus 100 is connected to the network NW. The network NW includes, for example, a wireless base station, a Wi-Fi access point, a communication line, a provider, the Internet, and the like. One or more terminal devices 10 and the electronic store server 20 are connected to the network NW. Note that the terminal device 10 and the ranking management device 100 do not necessarily need to be able to communicate with each other, and the network NW may be divided into several sub-networks.

  The terminal device 10 is a terminal device used by a user who uses an electronic store. As the terminal device 10, for example, a mobile phone such as a smartphone, a personal computer, a tablet terminal, or the like is used. In the terminal device 10, a user agent (UA) such as a web browser or a news application is activated and accesses the electronic store server 20.

  The electronic store server 20 is a server device that operates an electronic store. The electronic store server 20 operates in a web server, an application server, or other forms. The electronic store server 20 provides the terminal device 10 with an interface screen for enabling the user of the terminal device 10 to shop on the network NW.

  FIG. 2 is a diagram illustrating an example of an interface screen provided by the electronic store server 20. In this screen, a query input window QW for inputting an input query is set. When an input query is input to the query input window QW, for example, a suggestion query list SQL that is list information of a suggestion query in which another word is added to the input query is displayed. The order of the suggest queries displayed in the suggest query list SQL is determined based on the ranking performed by the ranking management apparatus 100. When the search button SB is operated in a state where the suggestion query is selected, a transition is made to a screen that displays a search result using the selected suggestion query. In this way, shopping at the electronic store is started.

  The electronic store server 20 logs the suggestion query presented to the input query each time a search is performed by the user, and information on the suggestion query selected in the suggestion query list SQL each time the search is performed. The information 22 is stored in the storage unit. In addition to these pieces of information, the log information 22 may include information related to the purchase behavior of the user, such as whether or not the product has finally been purchased.

  Returning to FIG. 1, the ranking management apparatus 100 includes, for example, an information management unit 110, a learning unit 120, and a ranking unit 130. These functional units are realized, for example, when a processor such as a CPU (Central Processing Unit) executes a program (software). Some or all of these functional units may be realized by hardware such as LSI (Large Scale Integration), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), or software. And hardware may cooperate.

  The ranking management apparatus 100 includes a storage unit 150. The storage unit 150 is realized by, for example, an HDD (Hard Disk Drive), a flash memory, a RAM (Random Access Memory), a ROM (Read Only Memory), or the like. The storage unit 150 stores information such as ranking data 152, learning data 154, weight data 156, and the like.

  The configuration of the ranking management apparatus 100 shown in FIG. 1 may be distributed as appropriate. For example, the learning unit 120 and the ranking unit 130 may be realized by independent processors, and the storage unit 150 is a NAS (Network Attached Storage) accessible from the ranking management apparatus 100 via the network NW. Also good. Further, the electronic store server 20 and the ranking management apparatus 100 may be integrated and realized by a single processor.

  The information management unit 110 acquires the log information 22 from the electronic store server 20, processes the log information 22 as appropriate, and stores it in the storage unit 150 as learning data 154.

  The learning unit 120 learns the weight of the feature of the suggestion query, for example, for each input query based on the learning data 154. FIG. 3 is a diagram schematically illustrating an example of the content of the learning data 154. As shown in the drawing, the learning data 154 is information in which the suggested query and the suggested query selected therein are associated with each other when the search is performed for each input query. A flag indicating that it is a positive example is attached to the selected suggestion query, and a flag indicating that it is a negative example is attached to the suggestion query that is not selected.

  The feature of the suggestion query includes, for example, part or all of the following features. In the following, the binary, logarithmic value, and real value in parentheses indicate in what numerical format the content represented by the explanatory text is handled.

1. Edit distance (binary)
This feature is the minimum number of steps required to transform one character string (input query) into another character string (suggest query) by insertion / deletion / replacement of one character. For example, the edit distance is included in which range of values [0 to 2, 3 to 4, 5 to 6, 7 to 8, 9 to 10, 11 to 12, 13 to 15, 16 to]. Is represented by a vector in which the corresponding part is 1 and the non-applicable part is zero.

2. Display position (binary)
This feature is information indicating the number of suggested queries selected from among the suggested queries to which a fixed number is presented.

3. Suggest query character length (binary)
This feature is information indicating whether the number of characters in the suggestion query is a predetermined number (for example, 15) or more or less than the predetermined number.

4). Number of tokens in the suggestion query (real number)
This feature is the number of tokens obtained by morphological analysis of the suggestion query. A token is the smallest unit of a sequence of character strings that cannot be decomposed.

5. Number of searches aggregated using input query and selected suggestion query pairs as keys (binary)
For example, the number of searches is [0 to 2, 3 to 4, 5 to 6, 7 to 8, 9 to 10, 11 to 15, 16 to 20, 21 to 30, 31 to 50, 51 to 100. , 101-200, 201-], which is included in a range of values, is represented by a vector in which the corresponding part is 1 and the non-applicable part is zero.

6). Suggested CTR (Click Through Ratio) (real number)
This feature is represented by a value obtained by dividing the number of selections of the suggest query by the number of searches.

7). Number of input query searches on the assumption that a suggest query is selected (binary)
This feature is obtained, for example, by determining which value range of [0 to 10, 11 to 50, 51 to 100, 101 to 1000, 1001] the number of searches is, It is expressed as a vector with zero not being taken.

8). CTR (real number) obtained from the character string of the difference between the suggestion query and the input query
This feature is, for example, a CTR related to “case” when the input query is “smartphone” and the suggestion query is “smartphone case”.

9. Number of searches where the input query was searched without using a suggestion query (binary)
For example, the number of searches is [0, 1, 2, 3 to 4, 5 to 6, 7 to 8, 9 to 10, 11 to 15, 16 to 20, 21 to 30, 31 to 50, 51. ˜100, 101-200, 201]], which is included in a range of values, is represented by a vector in which the corresponding part is 1 and the non-applicable part is zero.

10. Number of search queries (including binary search) (including search by input query only and search by suggestion query)
For example, the number of searches is [0 to 2, 3 to 4, 5 to 6, 7 to 8, 9 to 10, 11 to 15, 16 to 20, 21 to 30, 31 to 50, 51. [100, 101-200, 201]], which is included in a range of values, is represented by a vector in which the corresponding part is 1 and the non-applicable part is zero.

11. Clicks on product links in search results when searching with a search query (binary)
This feature is counted up to one click per request, [0 to 2, 3 to 4, 5 to 6, 7 to 8, 9 to 10, 11 to 15, 16 to 20, 21 to 30, 31 to 50. , 51-100, 101-200, 201]], which is included in a range of values, is represented by a vector in which the corresponding part is 1 and the non-applicable part is zero.

  12 Total number of clicks on product links in search results when searching with search query (binary) This feature counts all clicks on the number of links on the product detail page. The maximum value is the same as the number of displayed product links. This number is [0 to 2, 3 to 4, 5 to 6, 7 to 8, 9 to 10, 11 to 15, 16 to 20, 21 to 30, 31 to 50, 51 to 100, 101 to 200, 201. To which value range is included, and is represented by a vector in which the corresponding part is 1 and the non-applicable part is zero.

13. Product details CTR when searching with a search query (real number)
This feature is represented by a value obtained by dividing the number of clicks of a product link (counting as 1 even if a plurality of links are clicked) by the number of requests.

14 Product details iCTR when searching with a search query (real number)
This feature is represented by a value obtained by dividing the number of clicks of the product link by the number of links of the product details.

15. Click Per Request (logarithm) of product details when searching with search query
This feature is represented by how many product detail links are clicked in one request. The maximum value is the number of clicks.

  16. Product hits (logarithm) when searching with a search query

17. Whether or not the search query corresponds to the upper predetermined number (for example, 5000) of queries with a large number of searches (binary)
This feature is represented by 1 when the above condition is met and zero when it is not met.

18. Whether or not the character string obtained by separating the search query with a space corresponds to the upper predetermined number (for example, 5000) (binary)
This feature is represented by 1 when the above condition is met and zero when it is not met.

19. Whether or not the noun extracted from the morphological analysis of the search query corresponds to the upper predetermined number (for example, 5000) (binary)
This feature is represented by 1 when the above condition is met and zero when it is not met.

  The learning unit 120 learns the feature weight of the above-described suggestion query. Since the learning data 154 is based on search log information, even the same suggestion query may be a positive example or a negative example. For this reason, the learning unit 120 uses, for example, an SVM (Support Vector Machine) to obtain the hyperplane so that the total distance from the hyperplane of the sample with the wrong determination is reduced.

  The ranking unit 130 ranks the suggest queries for each input query, for example, based on the feature weights learned by the learning unit 120. The ranking result is provided to the electronic store server 20. The electronic store server 20 determines a suggestion query to be presented for the input query based on the provided ranking result. As a result, the CTR of the suggestion query can be improved.

  FIG. 4 is a flowchart illustrating an example of the flow of processing executed by the learning unit 120 and the ranking unit 130. First, the learning unit 120 reads the ranking data 152 (S100). The ranking data 152 is data obtained by ranking the suggest queries for each input query, for example. The ranking data 152 as initial data is, for example, a predetermined number (for example, 200) of suggest queries arranged in the order of the frequency of the search query.

  Next, the learning unit 120 applies the SVM to the learning data 154, learns the weight of the feature of the suggestion query (S102), and stores it in the storage unit 150 as the weight data 156. The ranking unit 130 ranks the suggest queries included in the ranking data 152 based on the learned weights, and updates the ranking data (S104).

For example, the ranking unit 130 calculates the evaluation value Score represented by the expression (1) for each suggestion query, and ranks the suggestion queries in descending order of the evaluation value Score. In the formula, i is a feature index, w _i is a weight for the feature, and f _i is a feature value of the suggestion query (for example, 1 to 19 described above).
Score = Σw _i × f _i (1)

Specific examples are shown below. For example, the input query is “iPhone” (registered trademark), and [6. Suggested track record CTR] and [16. Assume that the number of product hits when a search query is used is adopted. Then, the learning unit 120 [6. The weight w ₆ for the suggest result CTR] is 0.5, [16. The weight w ₁₆ for the goods number of hits] at the time of the search query and the search to those learning and 0.3.

In this case, for the suggestion query [iPhone case], [6. Suggested track record CTR] is 0.1, [16. Assume that the number of product hits when searching with a search query is 3 (= log ₁₀ 1000). The evaluation value Score (iPhone case) for the suggestion query [iPhone case] on this occasion is calculated as follows.
Score (iPhone case) = 0.5 × 0.1 + 0.3 × 3 = 0.95

On the other hand, for the suggestion query [iPhone 7], [6. Suggested track record CTR] is 0.2, [16. Assume that the number of product hits when searching with a search query is 2 (= log ₁₀ 100). The evaluation value Score (iPhone 7) for the suggestion query [iPhone 7] on this occasion is calculated as follows.
Score (iPhone 7) = 0.5 × 0.2 + 0.3 × 2 = 0.7

  In the above case, Score (iPhone case) is higher than Score (iPhone 7). For this reason, the ranking unit 130 sets the ranking of the suggestion query [iPhone case] higher than the ranking of the suggestion query [iPhone 7].

  The applicant of the present application has confirmed that various index values are significantly improved by the method of the embodiment as compared with the ranking of the suggestion query simply arranged in the order of the frequency of the search query. The index values include average suggest CTR, average suggest MRR (mean reciprocal rank), and the like.

  According to the information processing apparatus, the information processing method, and the program according to the embodiments described above, it is a set of suggestion queries presented based on a query input in an electronic store, and a positive example and a negative example are distinguished. Based on the set of suggested queries, it is possible to rank the suggestion queries more appropriately by learning the feature weights of the suggestion queries and ranking the suggestion queries based on the learned feature weights. .

  As mentioned above, although the form for implementing this invention was demonstrated using embodiment, this invention is not limited to such embodiment at all, In the range which does not deviate from the summary of this invention, various deformation | transformation and substitution Can be added.

10 terminal device 20 electronic store server 22 log information 100 ranking management device (information processing device)
110 Information management unit 120 Learning unit 130 Ranking unit 150 Storage unit 152 Ranking data 154 Learning data 156 Weight data

Claims (7)

Learning to learn feature weights of a suggestion query based on a set of suggestion queries that are presented based on a query input at an electronic store and are distinguished from positive examples and negative examples And
A ranking unit that ranks the suggestion queries based on weights of features learned by the learning unit;
An information processing apparatus comprising: The learning unit learns the feature weight of the suggestion query for each query,
The ranking unit ranks the suggestion query for each query.
The information processing apparatus according to claim 1. The set of suggestion queries is distinguished from positive examples when selected by the user and negative examples when not selected.
The learning unit learns feature weights using a support vector machine, and determines a hyperplane so that the total distance from the hyperplane of a sample with a wrong determination is reduced.
The information processing apparatus according to claim 1 or 2. The feature includes information that does not depend on the number of searches.
The information processing apparatus according to any one of claims 1 to 3. The feature includes whether or not the noun extracted from the suggestion query is in the ranking of the number of searches.
The information processing apparatus according to any one of claims 1 to 4. Computer
A set of suggestion queries presented based on a query input at an electronic store, and learning the weight of the feature of the suggestion query based on a set of suggestion queries in which positive examples and negative examples are distinguished,
Ranking the suggestion queries based on the learned feature weights;
Information processing method. On the computer,
A set of suggestion queries presented based on a query input at an electronic store, and based on a set of suggestion queries in which positive examples and negative examples are distinguished, the feature weights of the suggestion queries are learned,
Ranking the suggested queries based on the learned feature weights;
program.

Images