JP7354329B2 - Information processing device, information processing method, and program - Google Patents

Description

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

BACKGROUND ART In fields such as electronic commerce, ranking objects (for example, products) are ranked based on input queries, and ranking objects with higher ranking results are presented to users. In this regard, the steps include receiving a query word string, retrieving a plurality of different product information entries associated with the query word string, and extracting from memory a first core product word from the query word string. , extracting from memory second core product words from the various product information entries; and, for each second core product word, placing each second core product word in a list of candidate product words associated with the key product word. a step of checking whether or not each second core product word exists; a step of reducing the weighting of the product information entry corresponding to each second core product word if each second core product word exists; and adjusting the product information entry. An invention of an online product search method is disclosed, which includes a step of sorting product information entries according to the given weights (see Patent Document 1).

Special table 2015-518220 publication

In conventional technology, ranking processing was performed that included factors not directly related to the query, such as the number of page views. In many cases, the ranking targets presented did not fully reflect the user's intentions.

The present invention has been made in consideration of such circumstances, and an object of the present invention is to provide an information processing device, an information processing method, and a program that can select ranking targets that are considered to be closer to the search intention. be one of the.

One aspect of the present invention is to derive a score of the ranking target by inputting an acquisition unit that acquires a query input by a user, and a feature amount of the ranking target that matches the query into a trained model; a ranking processing unit that performs a ranking process based on the score, and the trained model uses the feature quantity of the ranking target as learning data and uses a label based on the user's behavior history and the category of the ranking target as a teacher. This is an information processing device that is trained by machine learning using data.

According to one aspect of the present invention, it is possible to select a ranking target that is considered to be closer to the search intention.

1 is a diagram illustrating an example of the configuration and usage environment of a shopping server 100 using an information processing device according to an embodiment. It is a figure showing an example of search input screen IM1. It is a figure which shows an example of search result display screen IM2. It is a figure showing an example of product sales screen IM3. 3 is an image diagram showing an example of the contents of log data 152. FIG. 3 is a diagram illustrating an example of the flow of processing by the ranking processing unit 110. FIG. 3 is a diagram showing an example of the contents of product data 150. FIG. 3 is a diagram illustrating an example of the content of processing by the learning processing unit 130. FIG. 3 is a diagram for explaining a label determination method by the learning processing unit 130. FIG. FIG. 7 is a diagram illustrating an example implementation of an information processing device according to another embodiment.

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of an information processing apparatus, an information processing method, and a program according to the present invention will be described below with reference to the drawings. An information processing device is realized by one or more processors. The information processing device is a device that ranks ranking targets according to a query input by a user. Ranking targets include products (including services) traded in electronic commerce, articles provided via networks, etc. If categories that serve as objective indicators are assigned as meta information, It can be anything. The following description will be made assuming that products traded in electronic commerce are subject to ranking.

Electronic commerce includes shopping, auctions, flea markets, etc. conducted via networks. In the following explanation, shopping will be taken as an example. The information processing device applied to shopping may be a device included in a shopping server or the like that manages a shopping site. That is, the information processing device may be a virtual device. Further, the information processing device may be the shopping server itself. In the following, the information processing device will be described as forming part of a shopping server. A shopping server may provide a shopping site in response to a request from a browser, or may provide content similar to a shopping site in response to a request from an application program installed on a smartphone or the like. It's okay. In the following explanation, it is assumed that the shopping server is the former one.

[composition]
FIG. 1 is a diagram illustrating an example of the configuration and usage environment of a shopping server 100 using an information processing device according to an embodiment. Shopping server 100 communicates with terminal device 10 via network NW. The network NW includes, for example, the Internet, a WAN (Wide Area Network), a LAN (Local Area Network), a wireless base station, a provider terminal, a dedicated line, and the like.

The terminal device 10 is, for example, a personal computer, a mobile phone such as a smartphone, a tablet terminal, or the like. In these, a UA (User Agent) 20 such as a browser or an application program is started. The UA 20 displays a sales screen provided by the shopping server 100 and transmits a request to the shopping server 100 in response to an input operation performed by the user of the terminal device 10 .

The shopping server 100 includes, for example, a site providing section 102, a log collecting section 104, a ranking processing section 110, and a learning processing section 130. The ranking processing unit 110 includes, for example, a search execution unit 112, a feature amount acquisition unit 114, and a model execution unit 116. These components are realized by, for example, one or more hardware processors such as a CPU (Central Processing Unit) executing a program (software). Some or all of these components are hardware (circuit parts) such as LSI (Large Scale Integration), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), and GPU (Graphics Processing Unit). (including circuitry), or may be realized by collaboration between software and hardware. The program may be stored in advance in a storage device (a storage device with a non-transitory storage medium) such as an HDD (Hard Disk Drive) or flash memory, or may be stored in a removable storage device such as a DVD or CD-ROM. It is stored in a medium (non-transitory storage medium), and may be installed by loading the storage medium into a drive device.

In this embodiment, an information processing device that includes at least the site providing unit 102 functioning as a query acquisition unit and the ranking processing unit 110 is an example of an information processing device.

Furthermore, the shopping server 100 stores data such as a trained model 118, product data 150, and log data 152 in a storage device such as an HDD, flash memory, or RAM (Random Access Memory). This storage device may be attached to the shopping server 100, or may be a NAS (Network Attached Storage) that the shopping server 100 can access via the network NW.

The site providing unit 102 provides the terminal device 10 with various screens as a shopping site. FIG. 2 is a diagram showing an example of the search input screen IM1. The search input screen IM1 is provided with a query input field A1 for inputting a query for searching for products, and a search button B1 for executing a search using the input query. When a query is input into the query input field A1 and the search button B1 is operated, the screen changes to a search result display screen IM2. At this time, the site providing unit 102 acquires a query in the form of a form input (functions as a query acquisition unit), outputs the query to the ranking processing unit 110, and requests ranking processing. Note that when a category is selected instead of a query, the screen similarly transitions to the search result display screen IM2.

FIG. 3 is a diagram showing an example of the search result display screen IM2. The search result display screen IM2 includes a search result display column A2. The search result display field A2 displays images and descriptions of a predetermined number (or as many as possible) of products that can be viewed by scrolling in the ranking order determined by the ranking processing unit 110 (for example, the titles of the titles described below). (part or all) are displayed in order of position. Hereinafter, this predetermined number will be referred to as k, and the displayed products will be referred to as introduced products. The positions are defined as 1 for the top position, 2 for the second position, and so on. When an image or description of one product is operated (selected) on the search result display screen IM2, the screen changes to a product sales screen IM3.

FIG. 4 is a diagram showing an example of the product sales screen IM3. The product sales screen IM3 includes a product image display field A3-1, a title field A3-2, a detailed explanation field A3-3, and the like. In the title column A3-2 and detailed description column A3-3, the manufacturer of the selected product (selected product), the material of the product, its use, and other details are posted as text. The text displayed in the title column A3-2 is referred to as a "title", and the text displayed in the detailed explanation column A3-3 is referred to as a "detailed explanation". A purchaser purchases a product by operating the product sales screen IM3.

The functions of each part will be explained below. The log collection unit 104 collects the usage history of the shopping site for each user and registers it in the log data 152. FIG. 5 is an image diagram showing an example of the contents of the log data 152. As shown in the figure, the log data 152 includes information on various screens viewed by the user, queries entered, introduced products displayed, purchased products, etc. for the user ID, which is the user's identification information. Information is associated with each other. Note that the log data 152 does not have to be organized by user ID. Furthermore, in reality, introduced products and product information may be represented by a product ID, which will be explained below.

[Ranking process]
FIG. 6 is a diagram showing an example of the flow of processing by the ranking processing unit 110. First, the search execution unit 112 of the ranking processing unit 110 searches the product data 150 using the query passed from the site providing unit 102. FIG. 7 is a diagram showing an example of the contents of the product data 150. The product data 150 is, for example, a product ID that is product identification information, a category ID that is category identification information, a product image, a title, a detailed description, a price, and other data that are associated with each other. Among these, the product ID is information numbered by an exhibition reception unit (not shown), and the other information is input by the seller. The categories are set, for example, hierarchically, and in the illustrated example, "3" indicates "food and drink,""35" indicates "alcohol," and "356" indicates "red wine." A learning processing unit 130, which will be described later, performs processing by focusing on, for example, an arbitrary hierarchy of categories (category IDs) in learning processing. As an example, the learning processing unit 130 performs processing focusing on the lowest layer information (“356” in FIG. 5).

For example, the search execution unit 112 extracts records of product data 150 that include a query in the title or detailed description, and obtains the product ID. As a result, the search execution unit 112 obtains products that match the query. Note that the search execution unit 112 may obtain the product ID of the product matching the query using any method. In the example of FIG. 6, it is assumed that n product IDs have been acquired. n does not have to be a fixed value, and product IDs corresponding to the extraction conditions may be acquired.

The feature amount acquisition unit 114 acquires the feature amount used by the model execution unit 116 as input data for each product (for each product ID) acquired by the search execution unit 112. The feature amounts include, for example, the number of PVs (Page Views), degree of match with a query, number of clicks (number of selections), price, number of reviews, and the like. For example, the feature amount acquisition unit 114 performs necessary processing on the log data 152 to generate these data.

The feature amount acquisition unit 114 counts, for example, the number of times the product is displayed as being included in the search result display screen in the log data 152 as the number of PVs. Further, the feature acquisition unit 114 derives the degree of match with the query based on the degree of character-based correspondence, such as the number of matches between the word string in the query and the word string in the title or detailed description. Regarding other matters, the feature amount acquisition unit 114 refers to the log data 152 and product data 150, performs processing as necessary, and obtains feature amounts. Then, the feature amount acquisition unit 114 digitizes and combines various feature amounts, for example, to generate one product vector corresponding to the product.

The model execution unit 116 derives scores used in the ranking process by sequentially inputting each of the product vectors generated by the feature amount acquisition unit 114 to the trained model 118, for example. Then, the ranking processing unit 110 selects k product IDs in descending order of scores, for example, and outputs them to the site providing unit 102. The site providing unit 102 searches the product data 150 using the k product IDs obtained from the ranking processing unit 110, and displays the extracted product information on the search result display screen on the user's terminal device 10. let Through such processing, it is possible to select ranking targets that are considered to be closer to the search intention and provide information to the user.

[Learning process]
Hereinafter, a process for generating the trained model 118 that brings about the above effects will be described. The learning processing unit 130 performs machine learning using a plurality of data sets in which the feature values of products to be ranked are training data and the labels are training data to generate a trained model 118. The feature amount is the same as that used by the model execution unit 116 as input data, and the learning data may be the aforementioned product vector. Therefore, the learning processing unit 130 may acquire the processing results from the feature amount acquisition unit 114 and perform the processing. In this embodiment, the learning processing unit 130 is included in the shopping server 100 and the information processing device, but it may be configured as a separate information processing device (learning device).

FIG. 8 is a diagram illustrating an example of the content of processing by the learning processing unit 130. As illustrated, the learning processing unit 130 adjusts the parameters of the machine learning model so that the score output by inputting the product vector into the machine learning model becomes higher as the label becomes larger. The machine learning model is, for example, a model configured based on a neural network, but is not limited to this, and any model can be used.

FIG. 9 is a diagram for explaining a label determination method by the learning processing unit 130. For example, as shown in the upper part of Figure 9, the introduced products displayed on a certain search result display screen are (1) to (4), and the user selects product (2) and moves to the product sales screen. However, it was not purchased. It is assumed that product (1) is a product assigned the same category ID as product (2). In this case, the learning processing unit 130 assigns a label value of 2 to product (2), and also assigns a label value of 1 to product (1), for example. Further, as shown in the lower part of FIG. 9, the introduced products displayed on another search result display screen were similarly (1) to (4), and product (3) was selected and purchased. It is assumed that product (4) is a product assigned the same category ID as product (3). In this case, the learning processing unit 130 assigns a label value of 4 to product (3), and also assigns a label value of 1 to product (1), for example. In the latter case, an arbitrary label value greater than 1 and less than 4 may be assigned to the product (1). Note that these numbers such as 1, 2, and 4 are just examples; in the example in the upper part of Figure 9, the label value given to product (2) is greater than or equal to the label value given to product (1), and the product It is sufficient that a relationship is established in which the label value given to (1) is greater than or equal to the label value given to products (3) and (4), and in the example shown in the lower part of Figure 9, the label value given to product (3) is larger than the label value given to product (4), and the label value given to product (4) is larger than the label values given to products (1) and (2), then any You may use numbers. In addition, in the above, it was decided that "the same category was assigned" by focusing on the information at the lowest level of the hierarchical categories, but if the categories do not match at the lowest level but at a higher level, , a smaller label value (for example, 0.5) may be assigned than when the lowest category matches.

In this way, the learning processing unit 130 assigns a label value indicating a high evaluation to a product selected or purchased by the user (for which a predetermined action has been taken), and also assigns a label value indicating a high evaluation to the product selected or purchased by the user. A label value indicating the next highest evaluation is also assigned to the product. The label value showing the next highest evaluation is a label value that is lower than the selected or purchased product and higher than the product that is not selected or purchased and has a different category from the selected or purchased product. Therefore, the label value is for products that are selected or purchased and assigned the same category as the selected or purchased product, and for products that are not selected or purchased and assigned the same category as the selected or purchased product. It is rated higher than other ranking targets that are not ranked.

The learning processing unit 130 similarly processes the selection or purchase results on a plurality of search result display screens, assigns a label value to each product, and uses the label value as training data.

After determining the label of each product as described above, the learning processing unit 130 adjusts the parameters of the machine learning model so that the larger the label, the higher the score output by the machine learning model. For example, if the product vector of product (m) is expressed as Vm, the label as Lm, and the machine learning model is expressed as a function f(Vm), and the size relationship of the labels is L2>L1>L3>L4. , the learning processing unit 130 adjusts the parameters of the machine learning model in the direction of f(V2)>f(V1)>f(V3)>f(V4). The machine learning model at the time when all of the data sets prepared in advance are processed and adjustment is completed is used as the learned model 118.

By generating the trained model 118 as described above, it is possible to select a product that is considered to be closer to the search intention. Traditional methods set labels in such a way that only the products that were selected or purchased were evaluated. In this case, high labels will inevitably be set for popular or best-selling products, and it is natural that low-priced products will be selected and purchased more often, so labels for high-priced products will inevitably be set. There was a tendency for the value to decrease. Therefore, if a user entered a query with the intention of searching for a product that was not very famous, or if the user entered a query with the intention of looking for a high-priced product, products that were different from the search intent might be listed. .

In contrast, in the method of the embodiment, the label is set to give a certain degree of evaluation even for products in the same category as the selected or purchased product, so it is possible to select a product that is considered to be closer to the search intent. I can do it. The product selected or purchased has a high probability of meeting the search intent, and the fact that it is assigned the same category means that it is also likely to match the search intent. .

According to the embodiment described above, the score for each product is calculated by inputting the feature amount (product vector) of a product that matches a query to the trained model 118 that outputs a score when the feature amount (product vector) is input. The trained model 118 includes a ranking processing unit 110 that performs ranking processing based on the acquired scores and selects ranking targets with high ranking results. Since the product is generated using a label that gives a high rating to products that are assigned the same category as the product that was searched for, it is possible to select a product that is considered to be closer to the search intent.

<Other embodiments>
The information processing device may be configured by a processor separate from the shopping server. FIG. 10 is a diagram illustrating an example of implementation of an information processing device according to another embodiment. As shown in the figure, an information processing device (learning device) A including a learning processing section 130 and an information processing device (ranking processing device) B including a ranking processing section 110 are each configured separately from a shopping server C. These may be configured to communicate with each other via the network NW. In this example, information processing device A and information processing device B may be integrated into one device. Components in FIG. 10 denoted by the same reference numerals as in FIG. 1 have the same functions as those described with reference to FIG.

In the implementation example of FIG. 10, the learning processing unit 130 of the information processing device A acquires product data 150 and log data 152 from the shopping server C at a desired timing, and performs the same processing as described above. Then, the generated trained model is provided to the information processing device B.

Information processing device B includes a query acquisition unit 106. The query acquisition unit 105 acquires the ranking request and the query from the shopping server C, and passes the query to the ranking processing unit 110. The ranking processing unit 110 uses product data 150 and log data 152 that are obtained in advance from the shopping server C at a frequency of about once a day, and the learned model 118 that is obtained from the information processing device A. Perform the same processing as above. With this configuration, as well as in the above embodiment, it is possible to select a product that is considered to be closer to the search intention.

Although the mode for implementing the present invention has been described above using embodiments, the present invention is not limited to these embodiments in any way, and various modifications and substitutions can be made without departing from the gist of the present invention. can be added.

100 Shopping server (information processing device)
102 Site providing unit 104 Log collection unit 106 Query acquisition unit 110 Ranking processing unit 112 Search execution unit 114 Feature acquisition unit 116 Model execution unit 118 Learned model 130 Learning processing unit 150 Product data 152 Log data A, B Information processing device C shopping server

Claims (5)

an acquisition unit that acquires a query input by a user;
a ranking processing unit that derives a score of the ranking target by inputting a feature quantity of the ranking target that matches the query into a trained model, and performs a ranking process based on the score,
The trained model is learned by machine learning using the feature quantity of the ranking target as learning data and a label based on the user's behavior history and the ranking target category as training data,
The ranking target category is information that has been assigned in advance to all the ranking targets regardless of the query;
Information processing device. The label indicates that the same category is assigned to the ranking target that has been assigned the same category as the presented ranking target when the user performs a predetermined action in response to the ranking target being presented. It is rated higher than the ranking target that is not ranked.
The information processing device according to claim 1. further comprising a learning processing unit that generates the learned model by machine learning based on the learning data and teacher data;
The information processing device according to claim 1 or 2. The computer is
Get the query entered by the user,
An information processing method that derives a score of the ranking target by inputting a feature quantity of the ranking target that matches the query into a trained model, and performs a ranking process based on the score,
The trained model is learned by machine learning using the feature quantity of the ranking target as learning data and a label based on the user's behavior history and the ranking target category as training data,
The ranking target category is information that has been assigned in advance to all the ranking targets regardless of the query;
Information processing method. to the computer,
Get the query entered by the user,
A program that derives a score of the ranking target by inputting a feature quantity of the ranking target that matches the query into a trained model, and performs a ranking process based on the score, the program comprising:
The trained model is learned by machine learning using the feature quantity of the ranking target as learning data and a label based on the user's behavior history and the ranking target category as training data,
The ranking target category is information that has been assigned in advance to all the ranking targets regardless of the query;
program.

Images