JP6933623B2 - Feature vector generator, feature vector generation method and feature vector generation program - Google Patents

Description

The present invention relates to a feature vector generator, a feature vector generation method, and a feature vector generation program.

Time series data analysis is used in a wide range of industries. In time series data analysis, future data is often estimated based on past data. Conventionally, a linear time series analysis method (for example, an autoregressive integrated moving average model) has been often used, but a machine learning method has also been proposed. Machine learning technology can effectively handle non-linear time-series data and time-series data with complex periods.

In recent years, a machine learning technique for clustering items and classifying time-series patterns by assigning feature vectors to each item of time-series data has been proposed. For example, in Non-Patent Document 1, Word2Vec (Doc2Vec) can be used to assign a feature vector to each item in the time-series data set of all users, and recommend items to users based on the feature vector. It is disclosed. Further, Non-Patent Document 2 discloses that a feature vector is assigned to each item in consideration of the interval between items of time series data.

Ozsoy, Makbule Gulcin. "From word embeddings to item recommendation." ArXiv preprint arXiv: 1601.01356, 2016 Hong, Shenda, et al. "Event2vec: Learning Representations of Events on Temporal Sequences." Asia-Pacific Web (APWeb) and Web-Age Information Management (WAIM) Joint Conference on Web and Big Data. Springer, Cham, 2017

In the conventional technique, the feature vector is assigned to each item in the time series data set corresponding to one domain, and the assignment of the feature vector to each item in a different domain is not considered. Therefore, when performing transfer learning between time-series data sets corresponding to each of the different domains, there is a problem that the relevance of these data sets cannot be taken into consideration and the transfer learning cannot be performed accurately.

Therefore, the present invention has been made in view of these points, and is a feature vector generator and feature vector generator capable of generating feature vectors by associating items of time series data corresponding to different domains with each other. It is an object of the present invention to provide a method and a feature vector generation program.

The feature vector generator according to the first aspect of the present invention includes item information indicating an item corresponding to an event, a time when the event occurs, and user identification information that identifies a user who has generated the event. A plurality of time-series data acquisition units for acquiring the first time-series data which is the time-series data of the first domain and the second time-series data which is the time-series data of the second domain. Correspondence relationship between the correspondence relationship specifying unit that specifies the correspondence relationship of the user identification information, the time included in the first time series data and the second time series data, and the user identification information specified by the correspondence relationship identification unit. Based on the above, an integrated unit that generates a plurality of partial time series data by integrating a part of the first time series data and a part of the second time series data, and a plurality of integrated units integrated by the integrated unit. Based on the partial time series data of the above, the feature vector generation unit for generating a feature vector indicating the feature of the item indicated by each of the plurality of item information included in each of the plurality of the partial time series data is provided.

The first time-series data includes at least one of item information corresponding to a predetermined event and item information corresponding to an event different from the predetermined event, and the integrated unit includes the first item. From the time-series data, the first partial time-series data including the occurrence time of the predetermined event included in the first time-series data and corresponding to the period before the occurrence time is extracted, and the first partial time-series data is extracted. 2. Extract the second partial time series data, which is the time series data corresponding to the period before the occurrence time, from the time series data, and integrate the first partial time series data with the second time series data. May generate the partial time series data.

When the first time-series data does not include the item information corresponding to the predetermined event, the integration unit obtains the time-series data corresponding to an arbitrary period from the first time-series data. In addition to extracting as one partial time series data, time series data corresponding to the arbitrary period is extracted as the second partial time series data from the second time series data, and the first partial time series data and the first part time series data are extracted. Partial time series data may be generated by integrating with two partial time series data.

The integrated unit may generate the partial time series data so that the number of the item information included in the partial time series data is a predetermined number.
The integration unit extracts the first partial time series data so that the number of item information included in the first partial time series data becomes the first number, and the item included in the second partial time series data. The second partial time series data may be extracted so that the number of information becomes the second number.

The integration unit specifies a period in which the number of item information included in the first partial time-series data is the first number in the first time-series data, and selects the second time-series data corresponding to the period. When the number of item information included in the second partial time series data is larger than the second number, the item information included in the second partial time series data is extracted as the second partial time series data. The item information may be reduced so that the number becomes the second number.

The integration unit specifies a period in which the number of item information included in the first partial time-series data is the first number in the first time-series data, and the second time-series data corresponding to the period is used. When the number of item information included is smaller than the second number, the period is extended so that the number of item information included in the second partial time series data becomes the second number. The second part time series data may be extracted.

The integrated unit may generate the partial time series data so that the period in which the event corresponding to the item information included in the partial time series data occurs is a predetermined predetermined period.
The integrated unit may generate the partial time series data so that the number of item information corresponding to the predetermined event included in the partial time series data is a predetermined number.

The generation unit may generate a feature vector of each of the plurality of items by analyzing the relationship between the items indicated by the plurality of item information included in the plurality of the partial time series data.

The integration unit identifies the second time-series data corresponding to the same user as the user corresponding to the first time-series data, based on the correspondence of the user identification information specified by the correspondence-specificing unit. , The partial time series data may be generated by integrating a part of the first time series data and a part of the second time series data.

The first time-series data includes at least one of a predetermined event and an event different from the predetermined event, and the generation unit includes a plurality of item information included in the first time-series data. Based on the feature vector of each of the plurality of items indicated by, the feature vector of the user corresponding to the first time series data is generated as the first feature vector, and a plurality of features included in the second time series data. Based on the feature vector of each of the plurality of items indicated by the item information, the feature vector of the user corresponding to the second time series data is generated as the second feature vector, and the feature vector generator is a plurality of the above. Based on the first feature vector and the result of whether or not the user corresponding to the first feature vector has generated a predetermined event, the user is sent to the predetermined event in response to the input of the user's feature vector. By generating a classifier that classifies the user who generated the above and the user who did not generate the predetermined event and inputting the second feature vector into the generated classifier, the second feature vector can be used. A prediction unit may be further provided to classify the corresponding users into a user who is predicted to generate the predetermined event and a user who is predicted not to generate the predetermined event.

In the feature vector generation method according to the second aspect of the present invention, the item information indicating the item corresponding to the event executed by the computer, the time when the event occurs, and the user who identifies the user who generated the event are identified. A plurality of steps for acquiring the first time-series data which is the time-series data of the first domain and the second time-series data which is the time-series data of the second domain, which is the time-series data including the identification information. Based on the step of specifying the correspondence relationship of the user identification information, the time included in the first time series data and the second time series data, and the correspondence relationship of the specified user identification information. A plurality of steps based on a step of generating a plurality of partial time series data by integrating a part of one time series data and a part of the second time series data, and a plurality of integrated partial time series data. It includes a step of generating a feature vector indicating the feature of the item indicated by each of the plurality of item information included in each of the partial time series data.

The feature vector generation program according to the third aspect of the present invention uses the computer to display item information indicating an item corresponding to an event, a time when the event occurs, and user identification information that identifies a user who has generated the event. A time-series data acquisition unit that acquires the first time-series data, which is the time-series data of the first domain, and the second time-series data, which is the time-series data of the second domain. Correspondence between the correspondence relationship identification unit for specifying the correspondence relationship of a plurality of the user identification information, the time included in the first time series data and the second time series data, and the user identification information specified by the correspondence relationship identification unit. Based on the relationship, the integration unit that generates a plurality of partial time series data by integrating a part of the first time series data and a part of the second time series data, and the integration unit are integrated. It functions as a feature vector generator that generates a feature vector indicating the feature of the item indicated by each of the plurality of item information included in each of the plurality of the partial time series data based on the plurality of the partial time series data. Let me.

According to the present invention, it is possible to generate a feature vector by associating items of time series data corresponding to different domains.

It is a figure explaining the outline of the feature vector generation apparatus which concerns on this embodiment. It is a figure which shows the structure of the feature vector generation apparatus which concerns on this embodiment. It is a figure which shows the example of the 1st time series data and the 2nd time series data which concerns on this embodiment. It is a figure which shows the generation example of the partial time series data which concerns on this embodiment. It is a figure which shows the example which arranged the feature vector generated based on the time series data shown in FIG. 3 in a feature space. It is a figure which shows the example which trained the classifier based on the 1st feature vector which concerns on this embodiment. It is a figure which shows the example which trained the classifier including the 1st feature vector generated based on the partial time series data including the 2nd partial time series data which concerns on this embodiment. It is a figure which shows the example which the 2nd feature vector was classified by the classifier which concerns on this embodiment. It is a flowchart which shows the flow of processing when the feature vector generation apparatus which concerns on this embodiment generates a feature vector of an item. It is a flowchart which shows the flow of processing when the feature vector generation apparatus which concerns on this embodiment predicts a user who generates a predetermined event.

[Overview of feature vector generator]
FIG. 1 is a diagram illustrating an outline of a feature vector generator according to the present embodiment. The feature vector generator is a computer that integrates time series data of different domains and generates a feature vector indicating the features of the items included in the time series data.

The feature vector generator acquires the time-series data of the first domain and the time-series data of the second domain ((1) in FIG. 1). In the present embodiment, the domain is an area for classifying time series data based on the type of time series data. In the present embodiment, the first domain is, for example, an area containing time-series data indicating the user's behavior regarding the purchase of an item on an EC (Electronic Commerce) site, and the second domain is, for example, a user on an arbitrary website. This is an area containing time-series data indicating the browsing behavior of.

In addition, the time-series data includes item information indicating an item corresponding to the event, a time when the event occurred, and user identification information for identifying the user who generated the event.

The feature vector generator identifies the time-series data of a common user among the time-series data of the first domain and the time-series data of the second domain. Then, the feature vector generator has a part of the time series data of the first domain of the common user and a part of the time series data of the second domain based on the time information included in the time series data of the common user. By integrating with and, a plurality of partial time series data are generated ((2) in FIG. 1).

The feature vector generator generates a feature vector indicating the feature of the item indicated by each of the plurality of item information included in each of the generated plurality of partial time series data ((3) in FIG. 1). By doing so, the feature vector generator can generate the feature vector by associating the items of the time series data corresponding to different domains. As a result, the feature vector generator can accurately perform transfer learning between time series data corresponding to different domains.
Hereinafter, the configuration of the feature vector generator will be described.

[Configuration example of feature vector generator 1]
FIG. 2 is a diagram showing a configuration of a feature vector generation device 1 according to the present embodiment. The feature vector generation device 1 includes a storage unit 11 and a control unit 12.

The storage unit 11 is, for example, a ROM (Read Only Memory), a RAM (Random Access Memory), or the like. The storage unit 11 stores various programs for operating the feature vector generation device 1. For example, the storage unit 11 causes the control unit 12 of the feature vector generation device 1 to function as a time series data acquisition unit 121, a correspondence relationship identification unit 122, an integration unit 123, a feature vector generation unit 124, and a prediction unit 125. Memorize the program.

The control unit 12 is, for example, a CPU (Central Processing Unit). The control unit 12 controls the function related to the feature vector generation device 1 by executing various programs stored in the storage unit 11. The control unit 12 functions as a time series data acquisition unit 121, a correspondence relationship identification unit 122, an integration unit 123, a feature vector generation unit 124, and a prediction unit 125 by executing a program stored in the storage unit 11.

[Generate item feature vector]
In the present embodiment, the time-series data acquisition unit 121, the correspondence relationship identification unit 122, the integration unit 123, and the feature vector generation unit 124 cooperate to generate the feature vector of the item included in the time-series data. Hereinafter, the functions of the time-series data acquisition unit 121, the correspondence relationship identification unit 122, the integration unit 123, and the feature vector generation unit 124 related to the generation of the feature vector of the item will be described.

The time-series data acquisition unit 121 acquires the first time-series data which is the time-series data of the first domain and the second time-series data which is the time-series data of the second domain. For example, the time-series data acquisition unit 121 acquires information indicating the browsing and purchasing history of items on the user's EC site as the first time-series data, and also obtains information indicating the browsing history of each of the websites of a plurality of users. Acquire as the second time series data. The time-series data acquisition unit 121 receives, for example, a plurality of first time-series data and a plurality of second time-series data from an information collection server (not shown) that collects the first time-series data and the second time-series data. And are acquired every predetermined time.

FIG. 3 is a diagram showing an example of the first time series data and the second time series data according to the present embodiment. FIG. 3A shows three first time series data D1A to D1C. FIG. 3B shows three second time series data D2A to D2C.

The first time-series data includes a plurality of event data in which item information indicating an item corresponding to an event, the time when the event occurred, and user identification information for identifying the user who generated the event are associated with each other. .. In the example shown in FIG. 3, v1 to v7, p1, p2, and w1 to w6 indicate event data, and the display form of the event data indicates the type of event.

Here, the item is, for example, a product or service. In the present embodiment, the code attached to the event data is used as the identification information for identifying the item. Items having different codes may be the same item or different items. For example, the item corresponding to the event data p1 and the item corresponding to the event data v1 may be the same or different.

The first time-series data includes at least one of item information corresponding to a predetermined event and item information corresponding to an event different from the predetermined event. Further, the second time series data also includes item information corresponding to an event different from the predetermined event.

For example, in the first time series data, a predetermined event is an event in which a user purchases an item on an EC site. Further, in the first time series data, an event different from the predetermined event is an event in which the user browses the EC site. In the example shown in FIG. 3A, the event corresponding to the event data v1 to v7 is an event to browse the EC site, and the event corresponding to the event data p1 and p2 is an event to purchase an item on the EC site. Is. Further, in the second domain, an event different from a predetermined event is an event for browsing a website. In the example shown in FIG. 3B, the event corresponding to the event data w1 to w6 is an event for browsing a website.

Further, the arrows shown in FIG. 3 correspond to the time when the event occurs. For example, in the first time series data D1A, it is shown that the items corresponding to the event data v1, v2, and v3 are sequentially browsed on the EC site, and then the items corresponding to the event data p1 are purchased. ..

The user identification information is information that can uniquely identify the user, and is, for example, an IP address assigned to the terminal used by the user. As the user identification information, the user ID used to identify the user on the EC site may be used, or the user ID used to identify the user on each website may be used.

A plurality of event data included in one time series data includes common user identification information. Therefore, in the example shown in FIG. 3, for convenience of explanation, the user corresponding to the user identification information is separately displayed. In the example shown in FIG. 3, user A, user B, user C, user a, user b, and user c are used as user identification information corresponding to the respective time series data.

The first time-series data is, for example, an access history on an EC site, and includes a plurality of event data in which the IP address of the terminal that accessed the EC site, the URL on the EC site, and the access time to the URL are associated with each other. ing. The URL on the EC site includes the URL of the purchase completion page displayed on the terminal when the item is purchased and the URL of the page explaining the item. The URL of the purchase completion page displayed on the terminal when the item is purchased corresponds to the item information corresponding to the item purchase event, and the URL of the page explaining the item corresponds to the item indicating the item browsing event. Corresponds to information. Also, the access time to the URL corresponds to the time when the event occurred.

The second time-series data is, for example, an access history on a website, and includes a plurality of event data in which the IP address of the terminal that accessed the website, the URL on the website, and the access time to the URL are associated with each other. ing. The URL on the website includes the URL of the page that describes the item. The URL of the page explaining the item corresponds to the item information corresponding to the item browsing event on the website, and the access time to the URL corresponds to the time when the event occurred.

The correspondence relationship specifying unit 122 specifies the correspondence relationship of a plurality of user identification information. For example, the correspondence relationship specifying unit 122 corresponds to a plurality of user identification information by identifying a match / mismatch between the user identification information included in the first time series data and the user identification information included in the second time series data. Identify the relationship. In the present embodiment, when the user identification information of the first time series data and the user identification information of the second time series data match, the correspondence identification unit 122 has the same user corresponding to these time series data. Determined to be a user.

When the user identification information is a user ID used to identify a user on an EC site and a user ID used to identify a user on each website, the storage unit 11 stores these user IDs. Correspondence information indicating the correspondence of is stored. Then, the correspondence relationship specifying unit 122 specifies the correspondence relationship of the plurality of user identification information by referring to the correspondence relationship information stored in the storage unit 11.

The integration unit 123 includes a part of the first time series data based on the time included in the first time series data and the second time series data and the correspondence relationship of the user identification information specified by the correspondence relationship identification unit 122. , A plurality of partial time series data are generated by integrating a part of the second time series data.

Specifically, the integration unit 123 includes the occurrence time of a predetermined event included in the first time series data from the first time series data, and is the time series data corresponding to the period before the occurrence time. Extract 1 partial time series data. For example, the integration unit 123 specifies a period from the first time series data to 30 minutes before the occurrence time of a predetermined event as a data extraction period, and sets the time series data corresponding to the data extraction period as the first part time series. Extract as data.

In the example shown in FIG. 3, the integration unit 123, from the first time series data D1A, event data p1 corresponding to the purchase event of the item as a predetermined event, and the item generated before the purchase event of the item occurs. The event data v1 to v3 corresponding to the browsing event of the above are extracted as the first partial time series data.

Subsequently, the integration unit 123 specifies the second time series data corresponding to the same user as the user corresponding to the first time series data, based on the correspondence relationship of the user identification information specified by the correspondence relationship specifying unit 122. .. Then, the integration unit 123 is a second time-series data corresponding to a period before the occurrence time of the predetermined event from the second time-series data corresponding to the same user as the user corresponding to the first time-series data. Extract partial time series data. For example, the integration unit 123 extracts time-series data corresponding to the same period as the data extraction period specified for the first partial time-series data from the second time-series data as the second partial time-series data.

In the example shown in FIG. 3, the user A, which is the user identification information corresponding to the first time series data D1A, and the user a, which is the user identification information corresponding to the second time series data D2A, correspond to each other. It is assumed that the users corresponding to the time series data are the same. In this case, the integration unit 123 extracts the event data w1 corresponding to the item browsing event and the event data w2 as the second partial time series data from the second time series data D2A. Here, since the event occurrence time corresponding to the event data w2 is later than the event occurrence time corresponding to the event data p1, the integration unit 123 does not include the event data w2 in the second part time series data. ..

Then, the integration unit 123 generates the partial time series data by integrating the first partial time series data and the second partial time series data. FIG. 4 is a diagram showing an example of generating partial time series data according to the present embodiment. In FIG. 4, the first partial time series data D1A-1 extracted from the first time series data D1A shown in FIG. 3 and the second partial time series data D2A-1 extracted from the second time series data D2A are integrated. It can be confirmed that the partial time series data d1 is generated.

Here, the integration unit 123 may generate the partial time series data so that the number of item information included in the partial time series data is a predetermined number. For example, the integration unit 123 extracts the first partial time series data so that the number of item information included in the first partial time series data is the first number, and the integration unit 123 extracts the first partial time series data and the items included in the second partial time series data. The second partial time series data may be extracted so that the number of information becomes the second number.

For example, the integration unit 123 specifies a data extraction period in which the number of item information included in the first partial time series data is the first number in the first time series data. Then, the integration unit 123 extracts the second time-series data corresponding to the specified data extraction period as the second partial time-series data. When the number of item information included in the extracted second partial time series data is larger than the second number, the integration unit 123 has a second number of item information included in the second partial time series data. Reduce event data to a number.

For example, when the integration unit 123 includes a plurality of event data indicating the same event and item, the integration unit 123 reduces the number of item information by deleting a part of the event data. By doing so, the feature vector generator 1 has a second number of item information even when the number of item information included in the second time series data is large during the data extraction period. Can be adjusted as follows. In addition, when the integration unit 123 includes a plurality of event data indicating the same event and item, the integration unit 123 does not delete the event data and collects the event data in which the event and the item do not overlap with other event data. The number of item information may be reduced by deleting the item.

Further, when the number of item information included in the second time series data corresponding to the specified data extraction period is smaller than the second number, the integration unit 123 may lengthen the data extraction period. For example, the integration unit 123 sets the end time of the data extraction period as it is, sets the start time to a time earlier than the start time, and sets the number of item information included in the data extraction period in the second time series data. The data extraction period is lengthened so that it becomes the second number. Then, the integration unit 123 extracts the second partial time series data so that the number of item information included in the second partial time series data is the second number.

By doing so, the feature vector generator 1 has a second number of item information even when the number of item information included in the second time series data is small during the data extraction period. Can be adjusted as follows.

Further, when the number of item information included in the second time series data corresponding to the specified data extraction period is smaller than the second number, the integration unit 123 includes the second time series data corresponding to the data extraction period. By including the item information included in the second part time series data in the second part time series data and duplicating the item information and including it in the second part time series data, the number of item information included in the second part time series data can be increased. It may be a second number.

In addition, the integration unit 123 specifies the data extraction period so that the number of item information included in the first partial time series data is the first number in the first time series data, and based on the data extraction period. Partial time series data was generated, but it is not limited to this. The integration unit 123 may generate the partial time series data so that the period in which the event corresponding to the event data included in the partial time series data occurs is a predetermined predetermined period. By doing so, the feature vector generator 1 sets the data extraction period based on the first number and the second number, so that the period becomes significantly longer and has no relation to the predetermined event. It is possible to prevent the item information corresponding to a non-event from being included in the partial time series data.

Further, the integration unit 123 may generate the partial time series data so that the number of event data corresponding to the purchase event of the item as a predetermined event included in the partial time series data is a predetermined number. .. For example, the integration unit 123 may generate the partial time series data so that the partial time series data includes only one event data corresponding to the purchase event. For example, if the purchase of an item is continuous for a short period of time, it is highly possible that the browsing event of the item before the purchase of the first item and the browsing event of the item before the purchase of the second item are not related. On the other hand, the feature vector generation device 1 partially collects only the event data corresponding to the item browsing event that is likely to be related to the item purchase event when the item purchase events are continuous in a short period of time. Can be included in.

Further, the integration unit 123 generates partial time series data so as to include event data corresponding to a predetermined event, but the present invention is not limited to this. When the first time-series data does not include the event data corresponding to a predetermined event, the integration unit 123 selects the time-series data corresponding to an arbitrary period from the first time-series data as the first part time-series. It may be extracted as data. In this case, the integration unit 123 extracts the time series data corresponding to the arbitrary period as the second partial time series data from the second time series data corresponding to the same user as the user corresponding to the first time series data. Then, the partial time series data may be generated by integrating the first partial time series data and the second partial time series data. By doing so, the feature vector generation device 1 can generate a feature vector based on the item browsing status when the item is not purchased.

Further, the integration unit 123 may generate partial time series data based on each of the first time series data and the second time series data corresponding to the user whose correspondence relationship is not specified. For example, the integration unit 123 may use the first partial time series data generated from the first time series data of the user whose correspondence with the user of the second time series data is not specified as it is as the partial time series data. good. Further, the integration unit 123 extracts and extracts the second partial time series data corresponding to an arbitrary period from the second time series data of the user whose correspondence relationship with the user of the first time series data is not specified. The second partial time series data may be used as the partial time series data.

The feature vector generation unit 124 is a feature vector indicating the characteristics of the items indicated by each of the plurality of item information included in each of the plurality of partial time series data based on the plurality of partial time series data integrated by the integration unit 123. To generate.

Specifically, the feature vector generation unit 124 extracts the item information included in the event data included in the plurality of partial time series data. The feature vector generation unit 124 generates a feature vector indicating the characteristics of each of the plurality of items by analyzing the relationship between the items indicated by the extracted plurality of item information.

For example, the feature vector generation unit 124 regards each of the plurality of items as one word, and generates a sentence in which the words are concatenated in each of the plurality of partial time series data. The feature vector generation unit 124 generates the feature vectors of each of the plurality of items for the generated sentences by using, for example, Word2Vec. It is assumed that the number of elements of the feature vector corresponds to, for example, the number of items.

FIG. 5 is a diagram showing an example in which a feature vector generated based on the time series data shown in FIG. 3 is arranged in a feature space. Note that FIG. 5 shows an example in which the feature space is compressed two-dimensionally and arranged in the feature space for convenience of explanation. In FIG. 5, a plurality of marks M1 and marks M2 are arranged. These marks indicate the position of the item in the feature space. The mark M1 indicates an item corresponding to the browsing event corresponding to the first time series data, and the mark M2 indicates an item corresponding to the browsing event corresponding to the second time series data.

Further, the marks M1 and the mark M2 are indicated with reference numerals v1 to v7 and w1 to w6. This code indicates an item and matches the code of the event data shown in FIG. For example, the mark M1 with v1 in FIG. 5 corresponds to the item indicated by the item information included in the event data v1 of the first time series data D1A shown in FIG. 3, and the mark M2 with w1 is attached. Corresponds to the item indicated by the item information included in the event data w1 of the second time series data D2A shown in FIG.

In the example shown in FIG. 3, the event data v1 and v3 indicating the event of browsing the item on the EC site and the event data w1 and w2 indicating the event of browsing the item on the website are the event p1 of purchasing the item on the EC site. Frequently appearing in front of. Corresponding to this, in FIG. 5, v1, v2, w1 and w2 are arranged at positions closer to each other in the feature space than other items, and it can be confirmed that they co-occur.

[Transfer learning and prediction of occurrence of predetermined events]
In the present embodiment, the feature vector generation unit 124 and the prediction unit 125 cooperate to perform transfer learning between time series data corresponding to different domains, and a user corresponding to the second time series data is predetermined. Predict whether or not to generate the event of. As a result, the feature vector generation device 1 predicts whether or not a learning device that performs transfer learning between time-series data corresponding to different domains and a user corresponding to the second time-series data will generate a predetermined event. Functions as a predictor. Hereinafter, the functions of the feature vector generation unit 124 and the prediction unit 125 related to transfer learning and prediction of the occurrence of a predetermined event will be described.

The feature vector generation unit 124 sets the user's feature vector corresponding to the first time series data as the first feature vector based on the feature vector of each of the plurality of items indicated by the plurality of item information included in the first time series data. Generate as.

For example, the feature vector generation unit 124, like the integration unit 123, includes the occurrence time of the purchase event of the item which is a predetermined event included in the first time series data from the first time series data, and the occurrence time thereof. The first partial time series data, which is the time series data corresponding to the previous period, is extracted. Here, the feature vector generation unit 124 extracts the first partial time series data so that the number of event data included in the first partial time series data is the first number.

Further, the feature vector generation unit 124 corresponds to a period in which the event data corresponding to the purchase event of the item, which is a predetermined event, is not included from the first time series data, and includes the first number of event data. Extract 1 partial time series data.

Then, the feature vector generation unit 124 calculates the average value (for example, the calculated average value or the weighted average value) of the feature vectors generated for each of the plurality of item information included in the first partial time series data. Generates the first feature vector.

The prediction unit 125 responds to the input of the user's feature vector based on the plurality of first feature vectors and the result of whether or not the user corresponding to the first feature vector has generated a predetermined event. A classifier is generated that classifies users into a user who has generated a predetermined event and a user who has not generated the predetermined event.

For example, the prediction unit 125 uses the partial time-series data including the item purchase event as regular example data in the first partial time-series data from which the first feature vector is generated, and does not include the item purchase event. Let the time series data be negative example data. Then, the prediction unit 125 performs machine learning based on these positive example data and negative example data, so that the user is referred to the user who generated the item purchase event in response to the input of the user's feature vector. , Generates a classifier that classifies items as users who did not generate a purchase event.

FIG. 6 is a diagram showing an example in which the classifier is trained based on the first feature vector according to the present embodiment. Note that FIG. 6 shows an example in which the first feature vector is compressed two-dimensionally and arranged in the feature space for convenience of explanation. The mark M3 shown in FIG. 6 indicates a first feature vector corresponding to the positive example data, and the mark M4 indicates a first feature vector corresponding to the negative example data. Further, the boundary line L indicates a boundary line when the first feature vector is classified into positive example data and negative example data by a classifier. The boundary line is shown for convenience of explanation, and the boundary line is not actually generated.

In generating the classifier, the prediction unit 125 generates a first feature vector based on the partial time series data including the second partial time series data generated by the integration unit 123, and generates the first feature vector. It may be included to generate a classifier. FIG. 7 is a diagram showing an example in which the classifier is trained including the first feature vector generated based on the partial time series data including the second partial time series data according to the present embodiment. In FIG. 7, a mark M3 indicating the first feature vector corresponding to the positive example data, a mark M4 indicating the first feature vector corresponding to the negative example data, and a mark M5 different from these marks are shown in FIG. , Mark M6 and so on are displayed.

The mark M5 shown in FIG. 7 is a first feature vector corresponding to the positive example data, and indicates a first feature vector generated based on the partial time series data including the second partial time series data. Further, the mark M6 is a first feature vector corresponding to the negative example data, and indicates a first feature vector generated based on the partial time series data including the second partial time series data. Further, the boundary line L2 indicates a boundary line when the first feature vector is classified into positive example data and negative example data by a classifier. In the example shown in FIG. 7, it can be confirmed that the position of the boundary line L2 is slightly different from that of the boundary line L because the positive example data and the negative example data are increased as compared with the example shown in FIG.

The feature vector generation unit 124 corresponds to the second time series data based on the feature vectors of the plurality of items indicated by the plurality of item information included in the second time series data acquired by the time series data acquisition unit 121. The user's feature vector is generated as the second feature vector. For example, the feature vector generation unit 124 generates the second partial time series data based on the event data included in the predetermined period from the latest time in the second time series data. Then, the feature vector generation unit 124 generates the second feature vector by calculating the average value of the generated feature vectors for each of the plurality of item information included in the second partial time series data. The time-series data acquisition unit 121 may acquire the second time-series data at the timing when the feature vector generation unit 124 generates the second feature vector.

By inputting the second feature vector into the generated classifier, the prediction unit 125 has to generate a user corresponding to the second feature vector, a user predicted to generate a predetermined event, and a predetermined event. It classifies it as a predicted user and outputs information indicating the classification result.

FIG. 8 is a diagram showing an example in which the second feature vector is classified by the classifier according to the present embodiment. The example shown in FIG. 8 shows an example in which the second feature vector is classified by the classifier corresponding to FIG. 6, and the same boundary line L as in FIG. 6 is displayed. The mark M7 shown in FIG. 8 indicates a second feature vector corresponding to the user predicted to generate a predetermined event. Further, the mark M8 indicates a second feature vector corresponding to a user who is predicted not to generate a predetermined event. By doing so, the feature vector generator 1 can accurately perform transfer learning between time series data corresponding to different domains.

[Process flow in feature vector generator 1]
Subsequently, an example of the processing flow in the feature vector generator 1 will be described. First, the flow of processing when the feature vector generation device 1 generates the feature vector of the item will be described. FIG. 9 is a flowchart showing a processing flow when the feature vector generation device 1 according to the present embodiment generates a feature vector of an item.

First, the time-series data acquisition unit 121 acquires a plurality of first time-series data and a plurality of second time-series data (S1).
Subsequently, the correspondence relationship specifying unit 122 specifies the correspondence relationship between the user identification information included in the plurality of first time series data and the user identification information included in the plurality of second time series data (S2).

Subsequently, the integration unit 123 together with a part of the first time series data based on the occurrence time of the event data included in the first time series data and the second time series data and the correspondence relationship specified in S2. Partial time series data is generated by integrating a part of the second time series data (S3). The integration unit 123 generates a plurality of partial time series data based on the plurality of first time series data and the plurality of second time series data.

Subsequently, the feature vector generation unit 124 generates a feature vector indicating the feature of the item indicated by each of the plurality of item information included in each of the plurality of partial time series data based on the plurality of partial time series data ( S4).

Next, a processing flow when the feature vector generation device 1 predicts a user who generates a predetermined event will be described. FIG. 10 is a flowchart showing a processing flow when the feature vector generation device 1 according to the present embodiment predicts a user who generates a predetermined event. At the start of this flowchart, the time series data acquisition unit 121 has acquired a plurality of first time series data and a plurality of second time series data, and the feature vector generation unit 124 generates feature vectors of a plurality of items. It is assumed that you are doing.

First, the feature vector generation unit 124 generates a plurality of first partial time series data based on the first time series data acquired by the time series data acquisition unit 121 (S11).
Subsequently, the feature vector generation unit 124 first, for each of the plurality of first partial time series data generated in S11, based on the feature vector of the item indicated by the item information included in the first partial time series data. A plurality of feature vectors are generated (S12).

Subsequently, the prediction unit 125 determines the result of whether or not the plurality of first feature vectors generated in S12 and the user corresponding to the first feature vector have generated a predetermined event (item purchase event). Based on this, a classifier that classifies the user into a user who generated a predetermined event and a user who did not generate the predetermined event is generated in response to the input of the feature vector of the user (S13).

Subsequently, the feature vector generation unit 124 generates the second partial time series data based on the second time series data acquired by the time series data acquisition unit 121 (S14).
Subsequently, the feature vector generation unit 124 uses the feature vector of the item indicated by the item information included in the second partial time series data for each of the plurality of second partial time series data generated in S14. A plurality of feature vectors are generated (S15).

Subsequently, the prediction unit 125 inputs the second feature vector generated in S15 to the classifier generated in S13 to generate a predetermined event for the user corresponding to the second feature vector. It is classified into a predicted user and a user who is predicted not to generate a predetermined event (S16).
Subsequently, the prediction unit 125 outputs the classification result in S16 (S17). For example, the prediction unit 125 generates a file including information indicating the classification result, and stores the file in the storage unit 11.

[Effect in this embodiment]
As described above, the feature vector generation device 1 according to the present embodiment includes the time included in the plurality of first time series data and the second time series data, and the user included in the first time series data and the second time series data. A plurality of partial time series data are generated by integrating a part of the first time series data and a part of the second time series data based on the correspondence of the identification information. Then, the feature vector generation device 1 generates a feature vector indicating the feature of the item indicated by each of the plurality of item information included in each of the plurality of partial time series data based on the integrated plurality of partial time series data. do. By doing so, the feature vector generation device 1 can generate the feature vector by associating the items of the time series data corresponding to different domains. As a result, the feature vector generator 1 can accurately perform transfer learning between time series data corresponding to different domains.

Although the present invention has been described above using the above-described embodiment, the technical scope of the present invention is not limited to the scope described in the above-described embodiment. It will be apparent to those skilled in the art that various changes or improvements can be made to the above embodiments.

For example, in the above-described embodiment, the event data is generated by the user, but the present invention is not limited to this, and the event data may be generated by the device. In this case, the user identification information included in the event data may be device identification information for identifying the device.

Further, in the above-described embodiment, the feature vector generator 1 generates partial time-series data by integrating a part of the first time-series data of the same user and a part of the second time-series data. , Not limited to this. For example, the storage unit 11 may store the attribute information indicating the user's attribute together with the user's correspondence information. Then, the feature vector generator 1 is a part of the first time series data and a part of the second time series data corresponding to a user who is different from the user of the first time series data and has similar attributes. Partial time series data may be generated by integrating with.

Further, in the above-described embodiment, the feature vector generator 1 integrates a part of the first time series data of the first domain and a part of the second time series data of the second domain to perform a partial time series. Data was generated, and a feature vector indicating the characteristics of the item was generated based on the partial time series data, but the present invention is not limited to this. The feature vector generator 1 generates partial time series data by integrating a part of time series data corresponding to each of three or more domains, and shows the features of the item based on the partial time series data. A feature vector may be generated.

Further, in particular, the specific embodiment of the distribution / integration of the apparatus is not limited to those shown above, and all or a part thereof may be arbitrarily added according to various additions or functional loads. It can be functionally or physically distributed / integrated in units.

1 ... Feature vector generator, 11 ... Storage unit, 12 ... Control unit, 121 ... Time series data acquisition unit, 122 ... Correspondence relationship identification unit, 123 ... Integration unit, 124・ ・ ・ Feature vector generation unit, 125 ・ ・ ・ Prediction unit

Claims (14)

Time-series data including item information indicating an item corresponding to the event, the time when the event occurred, and user identification information for identifying the user who generated the event, and is the time-series data of the first domain. A time-series data acquisition unit that acquires a certain first time-series data and a second time-series data that is time-series data of the second domain.
A correspondence relationship specifying unit that specifies a correspondence relationship between a plurality of the user identification information,
A part of the first time series data and the above based on the time included in the first time series data and the second time series data and the correspondence relationship of the user identification information specified by the correspondence relationship specifying unit. An integration unit that generates multiple partial time series data by integrating a part of the second time series data,
Based on the plurality of the partial time series data integrated by the integration unit, a feature that generates a feature vector indicating the characteristics of the item indicated by each of the plurality of item information included in each of the plurality of the partial time series data. Vector generator and
Feature vector generator with. The first time-series data includes at least one of item information corresponding to a predetermined event and item information corresponding to an event different from the predetermined event.
The integrated unit includes the occurrence time of the predetermined event included in the first time-series data from the first time-series data, and is the time-series data corresponding to the period before the occurrence time. The series data is extracted, and the second partial time series data, which is the time series data corresponding to the period before the occurrence time, is extracted from the second time series data, and the first partial time series data and the second part time series data are extracted. The partial time series data is generated by integrating with the time series data.
The feature vector generator according to claim 1. When the first time-series data does not include the item information corresponding to the predetermined event, the integration unit obtains the time-series data corresponding to an arbitrary period from the first time-series data. In addition to extracting as one partial time series data, time series data corresponding to the arbitrary period is extracted as the second partial time series data from the second time series data, and the first partial time series data and the first part time series data are extracted. Generate partial time series data by integrating with two partial time series data,
The feature vector generator according to claim 2. The integration unit generates the partial time series data so that the number of the item information included in the partial time series data is a predetermined number.
The feature vector generator according to claim 2 or 3. The integration unit extracts the first partial time series data so that the number of item information included in the first partial time series data becomes the first number, and the item included in the second partial time series data. The second partial time series data is extracted so that the number of information becomes the second number.
The feature vector generator according to claim 4. The integration unit specifies a period in which the number of item information included in the first partial time-series data is the first number in the first time-series data, and selects the second time-series data corresponding to the period. When the number of item information included in the second partial time series data is larger than the second number, the item information included in the second partial time series data is extracted as the second partial time series data. Reduce the item information so that the number is the second number,
The feature vector generator according to claim 5. The integration unit specifies a period in which the number of item information included in the first partial time-series data is the first number in the first time-series data, and the second time-series data corresponding to the period is used. When the number of item information included is smaller than the second number, the period is extended so that the number of item information included in the second partial time series data becomes the second number. Part 2 Extract time series data,
The feature vector generator according to claim 5. The integrated unit generates the partial time series data so that the period in which the event corresponding to the item information included in the partial time series data occurs is a predetermined predetermined period.
The feature vector generator according to claim 2 or 3. The integrated unit generates the partial time series data so that the number of item information corresponding to the predetermined event included in the partial time series data is a predetermined number.
The feature vector generator according to any one of claims 2 to 8. The generation unit generates a feature vector of each of the plurality of items by analyzing the relationship between the items indicated by the plurality of item information included in the plurality of the partial time series data.
The feature vector generator according to any one of claims 2 to 9. The integration unit identifies the second time-series data corresponding to the same user as the user corresponding to the first time-series data, based on the correspondence of the user identification information specified by the correspondence-specificing unit. , The partial time series data is generated by integrating a part of the first time series data and a part of the second time series data.
The feature vector generator according to any one of claims 1 to 10. The first time series data includes at least one of a predetermined event and an event different from the predetermined event.
The generation unit first obtains the feature vector of the user corresponding to the first time series data based on the feature vector of each of the plurality of items indicated by the plurality of item information included in the first time series data. The user's feature vector corresponding to the second time series data is generated as a feature vector and based on the feature vector of each of the plurality of items indicated by the plurality of item information included in the second time series data. Generated as a second feature vector
Based on the plurality of the first feature vectors and the result of whether or not the user corresponding to the first feature vector has generated a predetermined event, the user is subjected to the input of the user's feature vector. By generating a classifier that classifies a user who has generated a predetermined event and a user who has not generated the predetermined event, and inputting the second feature vector into the generated classifier, the second feature vector is input. It further includes a prediction unit that classifies the users corresponding to the feature vector into a user who is predicted to generate the predetermined event and a user who is predicted not to generate the predetermined event.
The feature vector generator according to any one of claims 1 to 11. Computer runs,
Time-series data including item information indicating an item corresponding to the event, the time when the event occurred, and user identification information for identifying the user who generated the event, and is the time-series data of the first domain. A step of acquiring a certain first time series data and a second time series data which is time series data of the second domain,
A step of identifying the correspondence between a plurality of the user identification information and
A part of the first time series data and the second time series data based on the correspondence between the time included in the first time series data and the second time series data and the specified user identification information. Steps to generate multiple partial time series data by integrating with a part of
Based on the plurality of integrated partial time series data, a step of generating a feature vector indicating the characteristics of the item indicated by each of the plurality of item information included in each of the plurality of the partial time series data, and a step of generating a feature vector.
A feature vector generation method comprising. Computer,
Time-series data including item information indicating an item corresponding to the event, the time when the event occurred, and user identification information for identifying the user who generated the event, and is the time-series data of the first domain. A time-series data acquisition unit that acquires a certain first time-series data and a second time-series data that is time-series data of the second domain.
Correspondence relationship identification unit that specifies the correspondence relationship of a plurality of the user identification information,
A part of the first time series data and the above based on the time included in the first time series data and the second time series data and the correspondence relationship of the user identification information specified by the correspondence relationship specifying unit. An integration unit that generates multiple partial time series data by integrating a part of the second time series data, and
Based on the plurality of the partial time series data integrated by the integration unit, a feature that generates a feature vector indicating the characteristics of the item indicated by each of the plurality of item information included in each of the plurality of the partial time series data. Vector generator,
Feature vector generator to function as.

Images