JP6851944B2 - Search system and search method - Google Patents

Description

The present invention relates to a search system and a search method.

Conventionally, there is known a content recommendation system that aggregates vectorized situation data and content data, creates new aggregated data, and updates a content preference model based on the new aggregated data (see Patent Document 1). ).

Japanese Unexamined Patent Publication No. 2013-69266

However, with the conventional technology, it may not be possible to smoothly update the model related to the service provided in real time.

The present invention has been made in consideration of such circumstances, and one of the objects of the present invention is to provide a search system and a search method capable of smoothly updating a model.

In one aspect of the present invention, a first generation unit that generates first feature information by inputting the first information into the first model and stores it in the first storage unit, and a second information unit are stored in the first storage unit. The second generation unit that generates the second feature information by inputting to the model and stores it in the second storage unit, and the first feature information corresponding to the content of the inquiry by searching the first storage unit are obtained. The management unit that acquires and outputs the search request for which the first feature information is specified to the search unit, and the second feature information that is specified in response to the search request from the management unit. The first model has a relevance to the second model and includes a search unit that searches a storage unit and acquires a second feature information related to the designated first feature information. Each of the first generation unit and the second generation unit has a plurality of channels of different generations of the model used, and processes are executed in parallel in each channel, and the management unit has the plurality of channels. This is a search system in which a working channel is selected from the channels of the above, and the first model and the second model used in the non-working channel are updated to a new model at a predetermined timing. ..

According to one aspect of the present invention, it is possible to provide a search system and a search method capable of smoothly updating the model.

It is a figure which shows an example of the use environment of the recommendation management apparatus 30 which constitutes a part of a search system. It is a figure which shows an example of the content of the user information 22. It is a figure which shows an example of the content of the delivery article information 24. It is a figure which shows an example of the content of the action log 26. It is an overall block diagram of a search system. It is a figure which conceptually shows the content of the process of the article vector generation unit 42. It is a figure which conceptually shows the content of processing of a user vector generation unit 62. It is a figure which shows an example of the time schedule about the channel switching. It is a flowchart which shows an example of the flow of processing executed by the recommendation management apparatus 30.

Hereinafter, embodiments of the search system and search method of the present invention will be described with reference to the drawings.

[Summary]
The search system is realized by one or more processors. The search system is, for example, a system that searches for other feature information (second feature information) corresponding to a certain feature information (first feature information) in response to an inquiry from the outside.

The first feature information is, for example, a user vector generated by the first model based on the history of articles viewed by the user. The user vector is considered to indicate the object of interest of the user. Articles are electronic articles provided via networks such as the Internet and include textual information. The article may be provided as part of a web page played by a browser or as part of an app page displayed by an application program.

The second feature information is, for example, an article vector generated based on an article. The article vector is generated, for example, by extracting the words contained in the article by morphological analysis or the like, and then inputting the word vector distributed and expressed by a method such as word2vec into the second model.

Here, the user vector which is the first feature information is generated by using, for example, the article vector which is the second feature information. For example, the user vector is generated by inputting a set of article vectors generated for each of the articles viewed by the user into the first model. The user vector is also a vector represented by a distributed representation.

As described above, the user vector which is the first feature information and the article vector which is the second feature information, and the first model and the second model have a semantic relationship. This is because in the above example, when the definition of the article vector is changed, the model for generating the user vector from the article vector also needs to be changed.

When the article vector corresponding to the user vector is obtained, the search system returns the article corresponding to the obtained article vector to the inquiry source as a recommended article corresponding to the user. As a result, articles that are close to the content that the user is interested in will be provided to the user, and the attractiveness of the web page or application page can be maintained and improved.

In providing such a search service, the search system has a process and a mechanism for smoothly updating the model. In addition, characteristic processing is performed so that the search is not performed using the user data of the old version after the model update. These will be disclosed step by step in the following description.

[Constitution]
FIG. 1 is a diagram showing an example of a usage environment of the recommendation management device 30 that constitutes a part of the search system. In the illustrated example, one or more user terminals 10 are connected to the service server 20 via the network NW. The network NW includes, for example, the Internet, a WAN (Wide Area Network), a LAN (Local Area Network), a provider terminal, a wireless communication network, a wireless base station, a dedicated line, and the like. The recommendation management device 30 is also connected to the network NW and communicates with the service server 20. It is assumed that the components appearing in the present embodiment include a communication interface for connecting to a network NW or other network. The communication interface includes a network card such as a NIC (Network Interface Card), a wireless communication module, and the like.

[User terminal]
The user terminal 10 is a terminal device including an input device, a display device, a communication device, a storage device, and an arithmetic device, such as a mobile phone such as a smartphone, a tablet terminal, and various personal computers. At the user terminal 10, a UA (User Agent) such as a browser or an application program is activated, and a request according to the content input by the user is transmitted to the service server 20. In addition, the UA displays various images based on the information acquired from the service server.

[Service server]
The service server 20 is a web server that provides a web page to the user terminal 10 in response to a request from a browser, or an application server that provides an image or sound to a user terminal 10 in response to a request from an application program.

The service provided by the service server 20 is provided as a customized service for more individual users by, for example, entering a user ID and password to log in. If the service server 20 is a web server, the web page can be provided without logging in, but in that case, a general-purpose web page that has not been customized for the user is provided.

When the user logs in, the service server 20 transmits, for example, an inquiry specifying a user ID to the recommendation management device 30. The recommendation management device 30 returns the article ID of the article recommended to the user indicated by the user ID to the service server 20. The service server 20 posts an article indicated by an article ID received from the recommendation management device 30 on a predetermined screen (for example, a top page of a website or an application page) and transmits the article to the user terminal 10.

The service server 20 stores the user information 22, the delivered article information 24, and the action log 26 in a storage device (not shown) inside or outside the server.

FIG. 2 is a diagram showing an example of the contents of the user information 22. The user information 22 is, for example, information in which a password, age, gender, birthday, occupation, and other demographic information are associated with a user ID, which is user identification information.

FIG. 3 is a diagram showing an example of the contents of the delivered article information 24. The delivered article information 24 is, for example, information in which article data is associated with an article ID, which is article identification information. The article data is, for example, text data, image data, a combination thereof, or reference information such as a URL (Uniform Resource Locator).

FIG. 4 is a diagram showing an example of the contents of the action log 26. The action log 26 is, for example, information in which the browsing page ID, which is the identification information of the browsed page, the article ID of the article included in the browsed page, the browsing time, and the like are associated with each other for each user ID.

[Overall configuration of search system]
FIG. 5 is an overall configuration diagram of the search system 1. In addition to the recommendation management device 30, the search system 1 includes, for example, an article vector generation device 40 (an example of a second generation unit), an article vector storage unit 50 (an example of a second storage unit), and a user vector generation. A part or all of a device 60 (an example of a first generation unit), a user vector storage unit 70 (an example of a first storage unit), and an article vector search device 80 (an example of a search unit) are provided. .. That is, the search system 1 may or may not include the article vector storage unit 50 and the user vector storage unit 70. The recommendation management device 30 is an example of a management unit. It should be noted that these components may be integrated or decentralized as appropriate. For example, the recommendation management device 30, the article vector generation device 40, and the user vector generation device 60 may be realized by one integrated processor.

Prior to the configuration of the recommendation management device 30, the article vector generation device 40 will be described in order. Here, the article vector generation device 40, the article vector storage unit, the user vector generation device 60, and the user vector storage unit 70 have two channels, channel 1 and channel 2, each of which advances processing in parallel. There is. The number of channels is not limited to two, and three or more channels may be provided. In the following description, it is assumed that the numbers in parentheses represent the channels, and when the channels are not distinguished, the parentheses and the numbers in the parentheses are omitted.

The article vector generation device 40 includes an article vector generation unit 42 (1) and an article vector generation unit 42 (2). FIG. 6 is a diagram conceptually showing the contents of the processing of the article vector generation unit 42. The content (text) of the article is input to the article vector generation unit 42. The content of the article is obtained from, for example, the delivered article information 24.

The article vector generation unit 42 first performs morphological analysis or the like on the content of the article and divides it into words (morphemes). Next, the article vector generation unit 42 applies a method such as word2vec to each word to acquire a distributed expression word vector. One or both of the morphological analysis and word2vec may be realized by requesting the processing to an external device instead of the article vector generation unit 42 performing the processing by itself.

Then, the article vector generation unit 42 derives the article vector by inputting one or more word vectors obtained from the article into the second model. The second model is, for example, a model generated by machine learning. The second model is generated by finding the parameters of the activation function of the hidden layer so that the relationship between the known word vector and the article vector can be reproduced by a method such as deep learning.

Notwithstanding the above definition, the second model may include the identification of word2vec, that is, what kind of corpus is given to word2vec.

The article vector generation unit 42 stores the generated article vector in the article vector storage unit 50 together with the article ID. At this time, if it is channel 1, it is stored in the article vector storage unit 50 as channel 1 data, and if it is channel 2, it is stored in the article vector storage unit 50 as channel 2 data.

The article vector storage unit 50 is realized by, for example, a storage device such as an HDD (Hard Disk Drive), a flash memory, or a RAM (Random Access Memory). The article vector storage unit 50 separately stores the channel 1 data and the channel 2 data in the storage area. Each of the channel 1 data and the channel 2 data is data in which an article vector is associated with an article ID.

The user vector generation device 60 includes a user vector generation unit 62 (1) and a user vector generation unit 62 (2). FIG. 7 is a diagram conceptually showing the contents of the processing of the user vector generation unit 62. The user ID and the article ID of the article viewed by the user indicated by the user ID are input to the user vector generation unit 62. The article ID of the article viewed by the user is acquired from, for example, the action log 26.

First, the user vector generation unit 62 acquires an article vector corresponding to the article ID of the article viewed by the user from the article vector storage unit 50. At this time, if it is channel 1, the article vector is acquired from the channel 1 data of the article vector storage unit 50, and if it is channel 2, the article vector is acquired from the channel 2 data of the article vector storage unit 50.

Then, the user vector generation unit 62 derives the user vector by inputting the acquired one or more article vectors into the first model. The first model is, for example, a model generated by machine learning. The first model is generated by finding the parameters of the activation function of the hidden layer so that the relationship between the known article vector and the user vector can be reproduced by a method such as deep learning. The user vector generation unit 62 stores the generated user vector in the user vector storage unit 70 together with the user ID. At this time, if it is channel 1, it is stored in the user vector storage unit 70 as channel 1 data, and if it is channel 2, it is stored in the article vector storage unit 70 as channel 2 data. The user vector may be created by reflecting elements other than the article vector of the browsed article. The elements other than the article vector are, for example, a part or all of the query entered by the user, the browsing history of the web page having no article, and the attributes of the user (age, gender, occupation, etc.).

Hereinafter, the recommendation management device 30 will be described. The recommendation management device 30 includes, for example, a channel switching unit 32, a response unit 34, a user vector search unit 36, and a search request unit 38. These components are realized by, for example, a hardware processor such as a CPU (Central Processing Unit) executing a program (software). 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 channel switching unit 32 switches between the active system and the standby system for channel 1 and channel 2. Details will be described later.

The response unit 34 responds to the inquiry from the service server 20, outputs an operation instruction to the user vector search unit 36 and the search request unit 38, and recommends the search result (article vector) returned from the search request unit 38. Is returned to the service server 20. The inquiry from the service server 20 is made by designating, for example, a user ID. The response unit 34 outputs the user ID of the search target to the user vector search unit 36 to perform the search.

The user vector search unit 36 searches the user vector storage unit 70 using the input user ID, and acquires the user vector corresponding to the user ID. At this time, the user vector search unit 36 searches the data of the channel set in the active system by the channel switching unit 32. The user vector search unit 36 outputs the user vector obtained as a result of the search to the search request unit 38.

The search request unit 38 requests the article vector search device 80 to search the article vector storage unit 50 using the input user vector. At this time, the search request unit 38 requests the article vector search device 80 to search the data of the channel set in the active system by the channel switching unit 32.

The article vector search device 80 searches the article vector storage unit 50 upon request. The article vector search device 80 searches for article vectors having the smallest inter-vector distance from the user vector, and extracts a predetermined number of article vectors in the order of the smallest inter-vector distance, for example. At this time, the article vector search device 80 may perform a search using a method similar to k-Nearest neighbor search. The article vector search device 80 returns the search result to the search request unit 38. The search request unit 38 returns this to the response unit 34. This completes the provision of recommended articles for one user.

[Channel switching]
Hereinafter, channel switching in the search system 1 will be described. FIG. 8 is a diagram showing an example of a time schedule relating to channel switching. In channel 1 and channel 2, the models (first model and second model) are updated with a half cycle shift. Then, after the update is completed, the system is promptly switched to the working system. Model update means updating the training data to the latest one and retraining the model. Since this learning is not completed instantly, model update, that is, machine learning, is started at the timing obtained by back-calculating the time obtained by adding the margin time to the estimated required time from the switching timing in time for the switching timing to the working system. To.

As the article vector, a new one is generated every day as the article is submitted, and the article vector is added to the article vector storage unit 50. However, regarding the function of word2vec and the rule for deriving the article vector from the word vector obtained by word2vec, the same one is used unless the second model is updated. Therefore, the update of the second model includes, for example, the update of the corpus given to word2vec and the update of the rule for deriving the article vector from the word vector obtained by word2vec. Not limited to this, even if the corpus given to word2vec is fixed or changed at a cycle longer than the update cycle T of the second model, and the update of the second model does not include the change of the corpus given to word2vec. Good.

In addition, the user vector is updated by batch processing or the like each time the user browses the article. At this time, new determination information (for example, a time stamp) is added to the user vector. However, as for the rules for deriving the user vector from the article vector, the same rules are used unless the first model is updated. Therefore, the update of the first model means the update of the rule for deriving the user vector from the article vector.

The first model and the second model are updated for each update cycle T based on the data from the update start date to the date retroactive to the reference period X. At this time, the article vector created from the update start date to the date retroactive to the reference period X is recreated by the new second model. In addition, the user vector created from the update start date to the date retroactive to the reference period X is also recreated using the new first model and the recreated article vector. However, when the user vector is recreated with the model update, the time stamp is not changed. In FIG. 8, it is shown that the update cycle T and the reference period X coincide with each other, but the reference period X may be a longer period than the update cycle T or a shorter period.

The switching of the working system is executed, for example, in a cycle T / 2 which is half of the update cycle T. For example, T / 2 is a period of about two weeks. The switching date of the working system is set in advance on the calendar, and the model is updated in time for that date. Further, the recommendation management device 30 also grasps the switching date of the active system, and when, for example, midnight of that day arrives, the recommendation management device 30 switches the reference destination channel.

With such a mechanism, the search system 1 of the embodiment can switch between the active system and the standby system without delay, and can smoothly update the model.

[About time stamps]
Since the user vector is statistically useful information, it may be retained for a period longer than the update cycle T (for example, about one month). In this case, assuming that the state in which the user does not read any article continues for a period longer than the reference period X, the user vector created by the first model and the second model of the old version is the user vector storage unit. It may be stored in 70. In this case, since the presupposed article vector is different, the search by the article vector search device 80 may not be significant.

Therefore, in the search system 1 of the embodiment, each time the user vector generator 60 generates the user vector, the user vector is accompanied by determination information that can be confirmed to be the information created by the latest model. It is stored in the user vector storage unit 70. In the following description, the latest model refers to the latest first model and the latest second model for each channel. The determination information is a time stamp indicating a time in the vicinity of the time when the user vector is generated, such as a time when the user vector is generated and a time when the user vector is stored in the user vector storage unit 70.

Then, when the recommendation management device 30 acquires the user vector from the user vector storage unit 70 in response to the inquiry from the service server 20, the recommendation management device 30 also reads the determination information together, and the read user vector is read based on the determination information. Make sure that is the information created by the most recent model. The recommendation management device 30 compares the time stamp with the current time, and when the difference time exceeds the reference period X, it cannot always be confirmed that the information is created by the latest model, so the recommendation management device 30 specifies the user vector. The searched search request is not output to the article vector search device 80. In this case, the recommendation management device 30 returns, for example, a response indicating "no recommendation article" to the service server 20. As a result, the service server 20 distributes articles to users in order from articles having good index values such as the number of accesses and CTR (Click Through Ratio).

By such a mechanism, the search system 1 of the embodiment can prevent inappropriate recommended articles from being delivered to the user.

Further, by assigning the time stamp to the user vector, the recommendation management device 30 does not need to know whether the active channel is channel 1 or channel 2. Since the recommendation management device 30 knows the update cycle T, which channel is the active channel by comparing the time stamp given to the user vector in the user vector storage unit 70 with the current time. Can be determined.

Here, instead of the time stamp, the version information of the first model and the second model can be used as the determination information. However, in this case, since the user vector is not time-stamped, the recommendation management device 30 needs to recognize the information of the channel of the active system and the version information of the model, and specifies these information. It is necessary to output the data acquisition request to the user vector storage unit 70. Since this increases the processing load and the amount of communication, it is more preferable to use the time stamp as the determination information.

[Processing flow]
FIG. 9 is a flowchart showing an example of the flow of processing executed by the recommendation management device 30. The processing of this flowchart is started when the response unit 34 of the recommendation management device 30 receives an inquiry from the service server 20.

First, the user vector search unit 36 of the recommendation management device 30 acquires the user vector and the determination information from the user vector storage unit 70 (S100).

Next, the search request unit 38 of the recommendation management device 30 determines whether or not it can be confirmed that the user vector is the information created by the latest model based on the determination information (S102).

When it is confirmed that the user vector is the information created by the latest model, the search request unit 38 specifies the user vector and makes a search request to the article vector search device 80 (S104). If it is not confirmed that the user vector is the information created by the latest model, the processing of S104 is skipped.

Then, the response unit 34 returns the response of either the article vector acquired from the article vector search device 80 or "no recommended article" to the service server 20 (S106). This ends the processing of the flowchart.

According to the search system and the search method of the embodiment described above, the user vector is generated by inputting the history of the articles viewed by the user into the first model, and the user vector is stored in the user vector storage unit 70. The device 60, the article vector generation device 40 that generates an article vector by inputting an article distributed via a network into the second model and stores it in the article vector storage unit 50, and the user vector storage unit 70 are searched. Then, the user vector corresponding to the content of the inquiry is acquired, and the recommendation management device 30 that outputs the search request for which the user vector is specified to the article vector search device 80 and the recommendation management device 30 are designated according to the search request from the recommendation management device 30. It is provided with an article vector search device 80 that searches the article vector storage unit 50 using the specified user vector and acquires an article vector related to the specified user vector, and the first model is a second model. Relevant, each of the user vector generator 60 and the article vector generator 40 has a plurality of channels of different generations of the model used, and each channel executes processing in parallel, and the recommendation management device. 30 selects an active channel from a plurality of channels, and the first model and the second model used in the non-active channel are updated to a new model at a predetermined timing. Thereby, the search system and the search method of the embodiment can smoothly update the model.

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.

1 Search system 10 User terminal 20 Service server 30 Recommendation management device 32 Channel switching unit 34 Response unit 36 User vector search unit 38 Search request unit 40 Article vector generator 50 Article vector storage unit 60 User vector generator 70 User vector storage unit 80 Article vector search device

Claims (7)

A first generation unit that generates first feature information by inputting the first information into the first model and stores it in the first storage unit, and
A second generation unit that generates second feature information by inputting the second information into the second model and stores it in the second storage unit, and a second generation unit.
A management unit that searches the first storage unit to acquire the first feature information corresponding to the content of the inquiry and outputs a search request for which the first feature information is specified to the search unit.
In response to the search request from the management unit, the second storage unit is searched using the designated first feature information, and the second feature information related to the designated first feature information is acquired. With a search unit,
The first model is related to the second model and
Each of the first generation unit and the second generation unit has a plurality of channels of different generations of the model used, and processing is executed in parallel in each channel.
The management unit selects a working channel from the plurality of channels and selects an active channel.
The first model and the second model used in the non-working channel are updated to a new model at a predetermined timing.
Search system. Each of the first feature information and the second feature information is vector data.
The search system according to claim 1.
Search system. The first generation unit, together with the determination information that can confirm that the generated first feature information is the information created by the latest first model and the second model, is the first. Make it memorize in the storage part
Based on the determination information, the management unit may find that the first feature information acquired from the first storage unit is the information created by the latest first model and the second model. If confirmed, the search request for which the first feature information is specified is output to the search unit.
The search system according to claim 1 or 2. The determination information is a time stamp and is
The management unit compares the time stamp with the current time, and the difference between the time stamp and the current time is a reference to data used for updating the first model and the second model at the predetermined timing. If it is within the period, the search request for which the first feature information is specified is output to the search unit.
The search system according to claim 3. The first information is a history of articles viewed by the user.
The first feature information is information indicating the features of the user, and is
The first generation unit associates the first feature information with a user ID and stores it in the first storage unit.
The search system according to any one of claims 1 to 4. The second information is an article distributed via a network.
The second feature information is information indicating the feature of the article, and is
The second generation unit associates the second feature information with the article ID and stores it in the second storage unit.
The first model is related to the second model in that it generates the first feature information based on the second feature information corresponding to the article viewed by the user.
The search system according to claim 5. The first generation unit generates the first feature information by inputting the first information into the first model and stores it in the first storage unit.
The second generation unit generates the second feature information by inputting the second information into the second model and stores it in the second storage unit.
The management unit searches the first storage unit to acquire the first feature information corresponding to the content of the inquiry, and outputs a search request specifying the first feature information to the search unit.
In response to the search request from the management unit, the search unit searches the second storage unit using the designated first feature information, and the second feature related to the designated first feature information. It ’s a search method to get information.
The first model is related to the second model and
Each of the first generation unit and the second generation unit has a plurality of channels of different generations of the model used, and processing is executed in parallel in each channel.
The management unit selects a working channel from the plurality of channels and selects an active channel from the plurality of channels.
The first model and the second model used in the non-working channel are updated to a new model at a predetermined timing.
retrieval method.

Images