JP5469233B1 - RECOMMENDATION INFORMATION DISTRIBUTION DEVICE AND RECOMMENDATION INFORMATION DISTRIBUTION METHOD - Google Patents

Abstract

Kind Code: A1 The present invention provides as a recommendation information a combination of an area and an activity estimated to be of interest to a user including an area where the user is not present at present.
Means for acquiring actual area visit frequency data; means for acquiring actual activity frequency data; means for generating predicted regional visit frequency data; means for generating predicted activity frequency data; When there is a request for specifying a user identifier from means for generating activity matrix data, means for generating user-by-user activity × regional matrix data, means for generating final score matrix data for each user, and a service providing apparatus And a means for responding to the service providing apparatus with a combination of the area and activity that are expected to have the highest score.
[Selection] Figure 1

Description

  The present invention relates to a technique for distributing recommendation information via a network such as the Internet.

  Estimate the current location of the user who owns the terminal device (obtains a registered address from an IP address, obtains location information by GPS (Global Positioning System), etc.), and recommends (recommends) information about stores in the vicinity. ) Has already been performed (see, for example, Patent Document 1).

JP 2003-248776 A

  As mentioned above, recommending information related to the user's current location has already been done, but even in areas where the user is not currently located, commuting routes that the user frequently visits and places on holidays Technology for recommending useful information with high accuracy has not been established for areas belonging to such living areas.

  The present invention has been proposed in view of the above-described conventional problems, and its purpose is to include regions and activities (behaviors) that are estimated to be of interest to the user, including regions where the user is not currently located. Content) is provided as recommended information.

  In order to solve the above-described problems, in the present invention, means for obtaining actual area visit count data in which the actual visit count for each area of the user is associated with the user identifier and the regional identifier, and the actual results for each user activity A means for acquiring the actual behavior number data in which the number of behaviors is associated with the user identifier and the activity identifier, and a matrix of predicted values corresponding to the user × region matrix of the actual value corresponding to the actual region visit frequency data. Means for performing a convergence operation so that the value of the loss function indicating the difference between the actual value and the predicted value is minimized, and generating as predicted area visit number data, and the actual result corresponding to the actual action number data The value of the loss function that represents the difference between the actual value and the predicted value, expressed as the product of the two decomposed matrices, the matrix of the predicted value corresponding to the user x activity matrix of values Means for performing a convergence calculation so as to minimize and generating as predicted action frequency data, and for each user, for each area of the predicted area visit frequency data, the predicted behavior for each activity of the predicted action frequency data of the user Setting the number of times and removing the combination of the region and the activity that cannot be combined based on the region-activity correspondence data, generating means for each region × activity matrix data, and for each user, For each activity, a predicted number of visits for each region of the predicted region visit number data of the user is set for each activity, and combinations of activities and regions that cannot be combined based on the region-activity correspondence data are removed for each user. Activity x Regional procession Means for generating data, means for generating final score matrix data for each user by transposing and adding one of said region for each user × activity matrix data and activity for each user × region matrix data, and for each generated user When there is a request specifying the user identifier from the service providing device based on the final score matrix data, the service providing device responds to the service providing device with a combination of the region and the activity that is predicted to have the highest score by the user of the user identifier. Means.

  In the recommended information distribution apparatus and the recommended information distribution method of the present invention, it is possible to provide, as recommended information, a combination of an area and an activity estimated to be interesting to the user, including an area where the user is not present at the present time. it can.

It is a figure which shows the structural example of the system concerning one Embodiment of this invention. It is a figure which shows the example of a data structure of an area visit log. It is a figure which shows the data structure example of an action log. It is a figure which shows the example of a data structure of performance area visit frequency data. It is a figure which shows the data structure example of performance action frequency data. It is a figure which shows the example of a data structure of interest graph data. It is a figure which shows the data structural example of social graph data. It is a figure which shows the example of a data structure of regional affinity data. It is a figure which shows the example of a data structure of an area | region list. It is a figure which shows the example of a data structure of regional similarity data. It is a figure which shows the data structure example of activity affinity data. It is a figure which shows the data structure example of an activity list. It is a figure which shows the data structure example of activity similarity data. It is a figure which shows the example of a data structure of prediction area visit frequency data. It is a figure which shows the data structure example of prediction action frequency data. It is a figure which shows the example of a data structure of area for every user x activity matrix data. It is a figure which shows the example of a data structure of activity for every user x area matrix data. It is a figure which shows the example of a data structure of area-activity corresponding | compatible data. It is a figure which shows the example of a data structure of final score matrix data for every user. It is a flowchart (generation of prediction area visit frequency data) showing an example of processing of an embodiment. It is a flowchart (generation of prediction action frequency data) which shows an example of processing of an embodiment. It is explanatory drawing of the production | generation of the prediction area visit frequency data. It is explanatory drawing of the production | generation of prediction action frequency data. It is a figure which shows the example of a loss function. It is a flowchart (generation of final score matrix data for each user) showing a processing example of the embodiment. It is a figure (the 1) which shows the specific example of a process. It is FIG. (2) which shows the specific example of a process. It is FIG. (The 3) which shows the specific example of a process. It is FIG. (4) which shows the specific example of a process. It is FIG. (5) which shows the specific example of a process. It is a flowchart (at the time of recommendation information distribution) which shows the example of processing of an embodiment.

  Hereinafter, preferred embodiments of the present invention will be described.

<Configuration>
FIG. 1 is a diagram showing a configuration example of a system according to an embodiment of the present invention.

  In FIG. 1, the system includes a plurality of terminal devices 1 operated by a user, a plurality of service server devices 3 that provide an information providing service to a user who uses the terminal device 1, and a user who uses the terminal device 1. SNS (Social Networking Service) server device 4 that provides social network services to the service server, and distribution of recommendation information consisting of a combination of area and activity (behavior such as shopping and appreciation) when the service server device 3 provides the service A recommendation information distribution server device 5 is connected to a network 2 such as the Internet. Note that the recommendation information distribution server device 5 may have a path connected directly or indirectly (including offline data exchange) with the service server device 3 and the SNS server device 4.

  The terminal device 1, the service server device 3, the SNS server device 4, and the recommendation information distribution server device 5 have a hardware configuration of a general computer device.

  The recommended information distribution server device 5 is a functional unit mainly realized by software (computer program). The regional visit log acquisition unit 501, the behavior log acquisition unit 503, the actual region visit count data calculation unit 505, and the actual behavior count data calculation. Unit 507, interest graph data calculation unit 509, social graph data acquisition unit 511, region affinity data calculation unit 513, region similarity data calculation unit 516, activity affinity data calculation unit 518, activity similarity data calculation unit 521, and predicted region visit Number-of-times data calculation unit 523, predicted number-of-actions data calculation unit 525, region for each user × activity matrix data generation unit 527, activity for each user × region matrix data generation unit 529, filter unit 532, filter unit 534, and user A final score matrix data integration unit 536 for each user, a recommendation information request reception unit 538, a recommendation information determination unit 539, and a recommendation information distribution unit 540 are provided.

  The interest graph data calculation unit 509, the social graph data acquisition unit 511, the region affinity data calculation unit 513, the region similarity data calculation unit 516, the activity affinity data calculation unit 518, and the activity similarity data calculation unit 521 Although effective in increasing the accuracy of the number-of-times data 524 and the predicted number-of-actions data calculation unit 525, it can be omitted.

  The area visit log acquisition unit 501 acquires the area visit log (log acquired when accessing the service server apparatus 3 at various places) from the service server apparatus 3 online or offline. The local visit log 502 has a function of holding it in a storage area inside the recommendation information distribution server device 5.

  FIG. 2 is a diagram showing an example of the data structure of the area visit log 502, and has items such as “time stamp”, “user ID”, and “area ID”. “Time stamp” is information on the date and time when log data was acquired. “User ID” is information for identifying a user. The user ID is a user ID registered as a member (preliminary login is required to obtain the user ID) or an ID (also referred to as bcookie or webcookie) for identifying the terminal device 1 that has accessed. Anything that can be specified to some extent can be used. “Region ID” is information specifying the region visited by the user. When position information (latitude / longitude) based on GPS (Global Positioning System) information or the like is used in raw log data, it is converted to an area ID by associating the position information with the area ID.

  Returning to FIG. 1, the action log acquisition unit 503 acquires an action log that is a history of actions such as user browsing and purchase from the service server device 3 (obtained when browsing or purchasing by accessing the service server device 3. Log) is acquired online or offline, and is stored in a storage area inside the recommendation information distribution server device 5 as an action log 504.

  FIG. 3 is a diagram showing an example of the data structure of the action log 504, and includes items such as “time stamp”, “user ID”, and “activity ID”. “Time stamp” and “User ID” are the same as those in FIG. “Activity ID” is information for specifying the content of actions such as browsing and purchasing performed by the user. In addition, when the content ID browsed in the raw log data or the purchased product ID is used, the ID is converted into the activity ID by associating the ID with the activity ID.

  Returning to FIG. 1, the actual region visit count data calculation unit 505 calculates actual region visit count data in which the actual visit count is associated with each user ID and each region ID from the regional visit log 502, and the actual region visit count data 506. As a storage area in the recommendation information distribution server device 5. If the actual area visit count data 506 can be acquired from the service server device 3 online or offline, the actual area visit count data calculation unit 505 can be omitted.

  FIG. 4 is a diagram showing an example of the data structure of the actual area visit count data 506, and the actual visit count is recorded for the combination of the user ID and the area ID.

  Returning to FIG. 1, the actual action number data calculation unit 507 calculates actual action number data in which the actual action number is associated with each user ID and each activity ID from the action log 504, and recommend information distribution as the actual action number data 508. It has a function of holding it in a storage area inside the server device 5. Note that when the actual action number data 508 can be acquired from the service server device 3 online or offline, the actual action number data calculation unit 507 can be omitted.

  FIG. 5 is a diagram showing a data structure example of the actual action number data 508, and the actual action number is recorded for the combination of the user ID and the activity ID.

  Returning to FIG. 1, the interest graph data calculation unit 509 calculates interest graph data indicating the degree of association (visited region and / or commonness of behavior) between users from the regional visit log 502 and / or behavior log 504, The graph data 510 has a function of holding it in a storage area inside the recommendation information distribution server device 5. If the interest graph data 510 can be acquired from the service server device 3 or the SNS server device 4 online or offline, the interest graph data calculation unit 509 can be omitted.

  FIG. 6 is a diagram showing an example of the data structure of the interest graph data 510, in which the degree of association is recorded for combinations of user IDs and user IDs (including combinations of the same user IDs).

  Returning to FIG. 1, the social graph data acquisition unit 511 acquires the social graph data indicating the interchangeability between the users from the SNS server device 4 online or offline, and stores it as the social graph data 512 in the recommended information distribution server device 5. It has a function to hold in the area.

  FIG. 7 is a diagram illustrating an example of the data structure of the social graph data 512, and the degree of association is recorded for a combination of a user ID and a user ID (including a combination of the same user ID).

  Returning to FIG. 1, the region affinity data calculation unit 513 calculates region affinity data indicating the possibility that two regions are visited at the same time from the region visit log 502, and recommends information distribution server device 5 as region affinity data 514. It has a function of holding in an internal storage area. If the regional affinity data 514 can be acquired from the service server device 3 online or offline, the regional affinity data calculation unit 513 can be omitted.

  FIG. 8 is a diagram showing an example of the data structure of the region affinity data 514, in which the degree of association (co-occurrence degree, etc.) is recorded for the combination of the region ID and the region ID (including the combination of the same region ID).

  Returning to FIG. 1, the regional similarity data calculation unit 516 calculates regional similarity data indicating similarity between regions from the region list 515 held in the storage area inside the recommendation information distribution server device 5. The utility data 517 has a function of holding it in a storage area inside the recommendation information distribution server device 5. If the regional similarity data 517 can be acquired from the service server device 3 online or offline, the regional list 515 and the regional similarity data calculation unit 516 can be omitted.

  FIG. 9 is a diagram showing an example of the data structure of the area list 515, in which the area name, position information, etc. of the area ID are recorded for each area ID.

  FIG. 10 is a diagram showing an example of the data structure of the regional similarity data 517, and the degree of association (similarity of attributes such as region name and position) with respect to the combination of region ID and region ID (including the combination of the same region ID). Etc.) are recorded.

  Returning to FIG. 1, the activity affinity data calculation unit 518 calculates activity affinity data indicating the possibility that two activities will be acted on at the same time from the behavior log 504, and the recommendation information distribution server device 5 has the activity affinity data 519. It has a function to hold in the storage area. If the activity affinity data 519 can be acquired from the service server device 3 online or offline, the activity affinity data calculation unit 518 can be omitted.

  FIG. 11 is a diagram showing an example of the data structure of the activity affinity data 519, in which the relevance (co-occurrence, etc.) is recorded for the combination of activity ID and activity ID (including the same activity ID combination).

  Returning to FIG. 1, the activity similarity data calculation unit 521 calculates activity similarity data indicating similarity between activities from the activity list 520 held in the storage area inside the recommendation information distribution server device 5. The utility data 522 has a function of holding it in a storage area inside the recommendation information distribution server device 5. When the activity similarity data 522 can be acquired from the service server device 3 online or offline, the activity list 520 and the activity similarity data calculation unit 521 can be omitted.

  FIG. 12 is a diagram showing an example of the data structure of the activity list 520. The activity name, category, etc. of the activity ID are recorded for each activity ID.

  FIG. 13 is a diagram showing an example of the data structure of the activity similarity data 522, and the degree of association (activity similarity of attributes such as activity name and category) with respect to the combination of activity ID and activity ID (including the combination of the same activity ID) Etc.) are recorded.

  Returning to FIG. 1, the predicted region visit number data calculation unit 523 calculates predicted region visit number data from the actual region visit number data 506, the interest graph data 510, the social graph data 512, the region affinity data 514, and the region similarity data 517. Is calculated and stored in the storage area inside the recommendation information distribution server device 5 as the predicted area visit count data 524.

  That is, the actual area visit count data 506 includes only information about the area where each user actually visited and is data in a perforated state, so that the value of the hole portion is predicted from the entire data, and the user By further considering the information between and the information between regions, the process of filling the hole is performed. Details of the calculation will be described later.

  FIG. 14 is a diagram showing an example of the data structure of the predicted area visit count data 524, in which the predicted visit count is recorded for a combination of a user ID and an area ID.

  Returning to FIG. 1, the predicted action number data calculation unit 525 calculates predicted action number data from the actual action number data 508, the interest graph data 510, the social graph data 512, the activity affinity data 519, and the activity similarity data 522. In addition, the predicted action number data 526 has a function of holding it in a storage area inside the recommendation information distribution server device 5.

  In other words, since the actual action count data 508 includes only information about the activity that each user actually acted on and is data in a perforated state, the value of the hole portion is predicted from the entire data, and between users The process of filling the hole is performed by further considering the information of the above and the information between the activities. Details of the calculation will be described later.

  FIG. 15 is a diagram showing an example of the data structure of the predicted action number data 526, in which the predicted action number is recorded for the combination of the user ID and the activity ID.

  Returning to FIG. 1, the region for each user × activity matrix data generation unit 527 uses the predicted region visit count data 524 and the predicted behavior count data 526 for each user for the predicted behavior count data 526 for each region. By setting the number of predicted actions for each activity, a region for each user × activity matrix data is generated, and the region for each user × activity matrix data 528 is stored in a storage area inside the recommendation information distribution server device 5. .

  That is, the predicted action count for each activity of the user in the predicted action count data 526 is not directly linked to the area, but indicates the degree of interest in each activity of the user. Can be combined. However, since there are activities that do not exist in the area, they are excluded by the filter processing described later.

  FIG. 16 is a diagram showing an example of a data structure of “region for each user × activity matrix data 528”, and the number of predicted actions is recorded for each combination of region and activity for each user.

  Returning to FIG. 1, the user-by-user activity × region matrix data generation unit 529 calculates, based on the predicted action number data 526 and the predicted region visit number data 524, the predicted region visit number data of the user for each activity for each user. By setting the predicted number of visits for each area of 524, the function of generating per-user activity × regional matrix data and storing it in the storage area inside the recommended information distribution server device 5 as per-user activity × regional matrix data 530 is provided. Have.

  That is, the predicted number of visits for each area of the user in the predicted area visit number data 524 is not directly linked to the activity, but indicates the degree of interest in each area of the user. It can be linked to an activity. However, since there is an area where the activity does not exist, it is excluded by a filtering process described later.

  FIG. 17 is a diagram showing an example of the data structure of the activity for each user × the region matrix data 530, and the predicted number of visits is recorded for each user for the combination of activity and region.

  Returning to FIG. 1, the filter unit 532 is based on the region-activity correspondence data 531 that holds the correspondence relationship between the region and the activity (s) existing in the region. It has a function of removing matrix elements that cannot exist as a relationship and holding them in a storage area inside the recommendation information distribution server device 5 as user-specific regions × activity matrix data 533. The data structure of each user region × activity matrix data 533 is the same as that shown in FIG.

  FIG. 18 is a diagram showing a data structure example of the region-activity correspondence data 531, and one or a plurality of activity IDs are associated with the region ID.

  Returning to FIG. 1, the filter unit 534 determines the activity and the region from the per-user activity × region matrix data 530 based on the region-activity correspondence data 531 that holds the correspondence relationship between the region and the activity (s) present in the region. The matrix element that cannot exist as a relationship is removed and stored in the storage area inside the recommended information distribution server device 5 as per-user activity × regional matrix data 535. The data structure of the activity for each user × the region matrix data 535 is the same as that shown in FIG.

Returning to FIG. 1, the final score matrix data integration unit 536 for each user from the filtered region for each user × activity matrix data 533 and the filtered user activity × regional matrix data 535 for each user makes the final score matrix for each user. It has a function of calculating data and holding it as a final score matrix data 537 for each user in a storage area inside the recommendation information distribution server device 5. In other words, a region for each user of a filtered user × activity matrix data 533 is matrix A, a user's activity for each user × regional matrix data 535 is a matrix B, and the coefficients are α and β. The matrix S of the final score matrix data 537 for each user of
S = αA + βB ^T
Calculate as B ^T represents a transposed matrix of the matrix B. The reason for transposing is to match the region of the matrix component and the combination of activities. Although the matrix B is transposed, the matrix A may be transposed.

  The area for each user x activity matrix data 533 as the matrix A and the activity for each user x area matrix data 535 as the matrix B are based on different area visit logs 502 and action logs 504, and consider mutual information along the way. Since both are information obtained symmetrically and indicate the degree to which a specific user is interested in the combination of each region and each activity, one of them is transposed and the region of the matrix component and the activity By matching the combinations and adding them together at an appropriate ratio, it is possible to obtain a comprehensive score for each region and each activity combination.

  FIG. 19 is a diagram showing an example of the data structure of the final score matrix data 537 for each user, and scores are recorded for combinations of area IDs and activity IDs.

  Returning to FIG. 1, the recommendation information request receiving unit 538 has a function of receiving a recommendation information distribution request specifying a user ID from the service server device 3.

  When the recommendation information request reception unit 538 receives a recommendation information distribution request, the recommendation information determination unit 539 determines the region and activity with the highest score for the user with the specified user ID based on the final score matrix data 537 for each user. It has the function which determines and acquires the combination of as recommendation information.

  The recommendation information distribution unit 540 has a function of distributing the recommendation information acquired by the recommendation information determination unit 539 to the service server device 3 that is a request source.

<Operation: Generation of predicted regional visit count data and predicted behavior count data>
FIG. 20 is a flowchart showing an example of processing related to the generation of the predicted area visit count data 524 in the above embodiment, and FIG. 21 is a flowchart showing an example of processing related to the generation of the predicted behavior count data 526.

  In FIG. 20, when the process is started (step S101), the area visit log acquisition unit 501 acquires the area visit log from the service server device 3 (step S102) and holds it as the area visit log 502 (step S103).

  Next, the actual region visit count data calculation unit 505 aggregates the visit counts for each user and each region from the regional visit log 502 to generate actual region visit count data (step S104), and holds the actual region visit count data 506. (Step S105). When generating the actual area visit frequency data 506 of FIG. 4 from the area visit log 502 of FIG. 2, the area ID visited for each user ID is divided into the data of the area visit log 502, and the number is counted for each area ID. The actual area visit count data 506 is generated.

  Next, the area affinity data calculation unit 513 generates area affinity data from the area visit log 502 (step S106) and holds it as area affinity data 514 (step S107).

  On the other hand, the area similarity data calculation unit 516 generates area similarity data by obtaining the edit distance of the area name or the Euclidean distance of the position information between areas from the area list 515 held internally or acquired from the outside ( Step S108), and the region similarity data 517 is retained (Step S109). When calculating the region similarity data 517 of FIG. 10 from the region list 515 of FIG. 9, the degree of similarity between the region names and location information of the regions of the two region IDs (if the region name, the matching degree of the character string In the case of position information, the degree of proximity, etc.) is calculated as an edit distance or Euclidean distance.

  On the other hand, returning to FIG. 20, the interest graph data calculation unit 509 generates interest graph data indicating the degree of association (visited region and / or commonness of behavior) between users from the regional visit log 502 and / or the behavior log 504. (Step S110), and retained as interest graph data 510 (Step S111). When calculating the interest graph data 510 of FIG. 6 from the area visit log 502 of FIG. 2 and / or the action log 504 of FIG. 3, for example, a history of common area IDs or activity IDs of a predetermined number (for example, eight) or more. Assume that the degree of relevance is “1” for two users who have (default is “0”). Moreover, it is good also as a numerical value according to the number of common log | history instead of setting only a value "1" "0".

  On the other hand, returning to FIG. 20, the social graph data acquisition unit 511 acquires social graph data from the SNS server device 4 (step S112), and holds it as social graph data 512 (step S113).

  Next, the predicted area visit frequency data calculation unit 523 inputs the actual area visit frequency data 506, the interest graph data 510, the social graph data 512, the regional affinity data 514, and the regional similarity data 517, and the actual area visit frequency data 506. Is treated as user x area matrix data as it is.

  For the region affinity data 514 and the region similarity data 517, region x region matrix data (matrix data in which the degree of association between regions is associated with a biaxial region ID by selecting one of them or averaging both, etc. ) Is generated (step S114). For interest graph data 510 and social graph data 512, user × user matrix data (matrix data in which the degree of association between users is associated with a biaxial user ID) by selecting one of them or averaging both, for example. Is generated (step S115).

  Next, the predicted area visit frequency data calculation unit 523 decomposes the matrix of predicted values for the user × area matrix data, the area × area matrix data, and the user × user matrix data of the actual area visit frequency data, respectively. The convergence calculation is performed so that the loss function value (also including the regularization term) indicating the difference between the actual value and the predicted value is minimized, and the actual area visit count data (user x area) Matrix data) is generated (step S116), and is stored as predicted area visit count data 524 (step S117), and the process is terminated (step S118).

  In FIG. 21, when processing is started (step S201), the action log acquisition unit 503 acquires an action log from the service server device 3 (step S202) and holds it as an action log 504 (step S203).

  Next, the actual action number data calculation unit 507 aggregates the number of actions for each user and for each activity from the action log 504 to generate actual action number data (step S204), and holds the result action number data 508 (step S205). ). 3 is generated from the action log 504 of FIG. 3, the activity ID data 508 is divided into activity IDs for each user ID and the numbers are counted for each activity ID. Is generated.

  Next, the activity affinity data calculation unit 518 generates activity affinity data from the action log 504 (step S206) and holds it as activity affinity data 519 (step S207).

  On the other hand, the activity similarity data calculation unit 521 generates activity similarity data based on cosine similarity or the like from the activity list 520 held internally or acquired from the outside (step S208), and holds it as activity similarity data 522 (step S208). Step S209). When the activity similarity data 522 of FIG. 13 is calculated from the activity list 520 of FIG. 12, the similarity between the activity names and categories of the two activity IDs (if the activity name, the matching degree of the character string, the category If there is a hierarchical relationship, etc.), the cosine similarity is calculated.

  Returning to FIG. 21, the predicted action number data calculation unit 525 inputs the actual action number data 508, the interest graph data 510, the social graph data 512, the activity affinity data 519, and the activity similarity data 522, and the actual action number data 508. Is treated as user × activity matrix data as it is.

  For the activity affinity data 519 and the activity similarity data 522, activity × activity matrix data (matrix data in which the degree of association between activities is associated with two-axis activity IDs by selecting one of them or averaging both of them) ) Is generated (step S210). For the interest graph data 510 and the social graph data 512, the result of step S115 in FIG. 20 is used.

  Next, the predicted action frequency data calculation unit 525 is a matrix obtained by decomposing the matrix of predicted values for the user × activity matrix data, the activity × activity matrix data, and the user × user matrix data of the actual action frequency data. Convergence calculation is performed so that the value of the loss function (also including the regularization term) indicating the difference between the actual value and the predicted value is minimized, and the predicted action frequency data (user x activity matrix data) Is generated (step S211), retained as the predicted action frequency data 526 (step S212), and the process is terminated (step S213).

  FIG. 22 is an explanatory diagram of generation of the predicted area visit frequency data 524, and FIG. 23 is an explanatory diagram of generation of the predicted behavior frequency data 526.

In FIG. 22, the element of the user × region matrix data corresponding to the actual region visit count data 506 is y _ij , the region × region matrix data element corresponding to the region affinity data 514 and the region similarity data 517 is s _ii , Elements of user × user matrix data corresponding to the graph data 510 and the social graph data 512 are indicated by a _{jj ′} .

Also, the user × region matrix data element corresponding to the predicted region visit count data 524 is f _ij , the predicted value region × region matrix data element is h _ii , and the predicted value user × user matrix data element is g _jj. It is indicated by _'represent each a product of two matrices.

  From these, the value of the equation shown in the lower part of FIG. 22 is calculated, and the convergence operation (the value of the matrix element is changed little by little so that the value of the equation becomes the minimum value). When the change disappears, it is determined that the data has converged), and the predicted area visit number data 524 corresponding to the user × area matrix data is acquired.

  Note that λ is a constant determined in operation. Moreover, l (el) is a loss function, and a function as shown to Fig.24 (a)-(e) is employable. The regularization term is a value that depends on the norm (length) of the decomposed matrix.

In FIG. 23, the element of the user × activity matrix data corresponding to the actual action frequency data 508 is z _ij , the element of the activity × activity matrix data corresponding to the activity affinity data 519 and the activity similarity data 522 is r _ii , and the interest graph. Elements of user × user matrix data corresponding to the data 510 and the social graph data 512 are indicated by a _{jj ′} .

Also, the element of the user × activity matrix data corresponding to the predicted action number data 526 is e _ij , the element of the predicted value × activity matrix data is o _{ii ′} , and the element of the predicted value user × user matrix data is g _{jj ′.} Each is represented as the product of two matrices.

  From these, the value of the equation shown in the lower part of FIG. 23 is calculated, and the convergence calculation (the value of the matrix element is changed little by little so that the value of the equation becomes the minimum value). When the change disappears, it is determined that the data has converged), and the predicted action frequency data 526 corresponding to the user × activity matrix data is acquired.

<Operation: Generation of final score matrix data for each user>
FIG. 25 is a flowchart showing an example of processing related to generation of the final score matrix data 537 for each user.

  In FIG. 25, when the process is started (step S301), the region for each user × activity matrix data generation unit 527 generates the corresponding region for each user from the predicted region visit count data 524 and the predicted behavior count data 526. By setting the predicted number of actions for each activity in the predicted number-of-actions data 526 of the user, the area for each user × activity matrix data is generated and held as the area for each user × activity matrix data 528 (step S302).

  Next, the filter unit 532 exists as a relationship between the region and the activity from the region for each user × activity matrix data 528 based on the region-activity correspondence data 531 that holds the correspondence relationship between the region and the activity (s) present in the region. The matrix elements that cannot be processed are removed and stored as the region for each user × activity matrix data 533 (step S303).

  On the other hand, the per-user activity × area matrix data generation unit 529 determines, based on the predicted action count data 526 and the predicted area visit count data 524, for each activity, for each user's predicted area visit count data 524 for each activity. By setting the predicted number of visits to the area, user-by-user activity x area matrix data is generated and held as user-by-user activity x area matrix data 530 (step S304).

  Next, the filter unit 534 creates the relationship between the activity and the region from the per-user activity × region matrix data 530 based on the region-activity correspondence data 531 holding the correspondence relationship between the region and the activity (s) present in the region. Matrix elements that cannot exist are removed and held as per-user activity × regional matrix data 535 (step S305).

Thereafter, the final score matrix data integration unit 536 for each user calculates the final score matrix data for each user for each user from the filtered region for each user × activity matrix data 533 and the filtered user activity × regional matrix data 535. Then, it is stored in the storage area inside the recommendation information distribution server device 5 as the final score matrix data 537 for each user (step S306). That is, a region for each user of a filtered user × activity matrix data 533 is matrix A, a user's activity for each user × regional matrix data 535 is a matrix B, and coefficients are α and β. The matrix S of the final score matrix data 537 for each user is
S = αA + βB ^T
Calculate according to Then, the process ends (step S307).

<Operation: Specific example of generation of predicted area visit number data, predicted action number data, and final score matrix data for each user>
FIG. 26 shows an example where actual area visit count data 506 and actual action count data 508 are generated from the local visit log 502 and the action log 504, respectively. For easy understanding, regions and activities are indicated by region names and activity names instead of region IDs and activity IDs.

  In the actual area visit count data 506, for example, the user “user1” visits the area “Meguro” three times, the area “Ueno” has no data, the area “Ebisu” visits once, and the area “Roppongi” twice Indicates that you have visited. In the actual action number data 508, for example, the user “user1” acts twice for the activity “ramen”, acts three times for the activity “museum”, no data for the activity “Italian”, and acts five times for the activity “sports bar”. It shows that.

  In FIG. 27, predicted area visit number data 524 is generated from actual area visit number data 506, interest graph data 510, social graph data 512, area affinity data 514, and regional similarity data 517, and actual action number data 508, interest graph. An example in which predicted behavior number data 526 is generated from data 510, social graph data 512, activity affinity data 519, and activity similarity data 522 is shown.

  The elements without data indicated by “?” In the actual area visit number data 506 and the actual action number data 508 in FIG. 26 are filled in the predicted area visit number data 524 and the predicted action number data 526 in FIG. .

  In FIG. 28, per-user region × activity matrix data 528 is generated from the predicted region visit number data 524 and the predicted behavior number data 526, and per-user activity × region matrix data is calculated from the predicted behavior number data 526 and the predicted region visit number data 524. An example in which 530 is generated is shown.

  For example, for the user “user1”, the horizontal axis “Meguro Ueno Ebisu Roppongi” in the predicted area visit frequency data 524 is set on the vertical axis, and the user “user1” in the predicted action frequency data 526 is set as the horizontal axis. Is set in association with all the regions, the region for each user × activity matrix data 528 is generated.

  For the user “user1”, the horizontal axis of the activity number data 526 “Ramen Museum Italian Sports Bar” is set on the vertical axis, and the horizontal axis indicates the user “user1” in the predicted area visit frequency data 524. By setting the estimated number of visits “3 2.4 1 2” of “user1” in association with all activities, the per-user activity × regional matrix data 530 is generated.

  In FIG. 29, by filtering based on the region-activity correspondence data 531 from the region for each user × activity matrix data 528 and the activity for each user × region matrix data 530, the region for each user × activity matrix data 533 and each user are respectively filtered. An example in which activity × region matrix data 535 is generated is shown.

  In the illustrated example, in the region-activity correspondence data 531, the region “Meguro” has no activity “museum”, and the region “Ueno” does not have the activity “sports bar”. User-by-user area × activity matrix data 533 in which data on the combination of the region “Meguro” and the activity “museum” and the combination of the region “Ueno” and the activity “sports bar” in the region matrix data 530 is removed. And per-user activity × regional matrix data 535.

  FIG. 30 shows an example in which the final score matrix data 537 for each user is generated from the filtered area for each user × activity matrix data 533 and the filtered activity for each user × area matrix data 535. Here, α = 0.5 and β = 0.5 are set as the coefficients.

  From the final score matrix data 537 for each user in this example, it can be seen that the combination of the region “Meguro” and the activity “sports bar” has the score “4” and is the highest combination for the user “user1”.

<Operation: Distribution of recommendation information>
FIG. 31 is a flowchart showing an example of processing at the time of recommendation information distribution in the above embodiment.

  In FIG. 31, the recommendation information distribution receiving unit 538 of the recommendation information distribution server device 5 receives the recommendation information distribution request with the user ID from the service server device 3 and starts the process (step S401).

  The recommendation information determination unit 539 refers to the record of the corresponding user ID from the final score matrix data 537 for each user, and determines and acquires the combination of the region having the highest score for the user and the activity as recommendation information (step S402). . Note that not only the combination of the region having the highest score and the activity, but also a plurality of combinations having higher scores can be used as recommendation information.

  Next, the recommendation information distribution unit 540 responds the recommendation information acquired by the recommendation information determination unit 539 to the service server device 3 that requested the request (step S403), and ends the processing (step S404).

  Upon receiving the recommendation information, the service server device 3 embeds the recommendation information at a predetermined position on the Web screen for providing the service, and provides page data to the terminal device 1.

  Since the recommendation information embedded in the Web screen is content that the user of the terminal device 1 is likely to be interested in, the recommendation information is expected to be transferred to the action with high probability.

<Summary>
As described above, according to the present embodiment, it is possible to provide, as recommendation information, a combination of a region and an activity estimated to be interested for the user, including a region where the user is not present at the present time.

  The present invention has been described above by the preferred embodiments of the present invention. While the invention has been described with reference to specific embodiments, various modifications and changes may be made to the embodiments without departing from the broad spirit and scope of the invention as defined in the claims. Obviously you can. In other words, the present invention should not be construed as being limited by the details of the specific examples and the accompanying drawings.

DESCRIPTION OF SYMBOLS 1 Terminal apparatus 2 Network 3 Service server apparatus 4 SNS server apparatus 5 Recommendation information delivery server apparatus 501 Area visit log acquisition part 502 Area visit log 503 Action log acquisition part 504 Action log 505 Actual area visit frequency data calculation part 506 Actual area visit frequency Data 507 Actual behavior frequency data calculation unit 508 Actual behavior frequency data 509 Interest graph data calculation unit 510 Interest graph data 511 Social graph data acquisition unit 512 Social graph data 513 Regional affinity data calculation unit 514 Regional affinity data 515 Regional list 516 Regional similarity Data calculation unit 517 Regional similarity data 518 Activity affinity data calculation unit 519 Activity affinity data 20 Activity List 521 Activity Similarity Data Calculation Unit 522 Activity Similarity Data 523 Predicted Regional Visit Number Data Calculation Unit 524 Predicted Regional Visit Number Data 525 Predicted Behavior Number Data Calculation Unit 526 Predicted Behavior Number Data 527 Generation of Each Region × Activity Matrix Data Unit 528 Region per user × Activity matrix data 529 Activity per user × Region matrix data generation unit 530 Activity per user × Region matrix data 531 Region-activity correspondence data 532 Filter unit 533 Region per user × Activity matrix data 534 Filter unit 535 Activity per user x region matrix data 536 Final score matrix data integration unit per user 537 Final score matrix data per user 538 Rec Command information request receiving unit 539 recommendation information determining section 540 recommendation information distribution unit

Claims (6)

Means for obtaining actual area visit count data in which the actual visit count for each area of the user is associated with the user identifier and the area identifier;
Means for acquiring the actual action number data in which the actual action number for each activity of the user is associated with the user identifier and the activity identifier;
The value of the loss function indicating the difference between the actual value and the predicted value is represented by a product of the two values obtained by decomposing the matrix of the predicted value corresponding to the user of the actual value corresponding to the actual region visit count data × the region matrix. A means for performing a convergence calculation so as to generate as predicted area visit frequency data,
A matrix of predicted values corresponding to the user × activity matrix of actual values corresponding to the actual behavior frequency data is represented by a product of two decomposed matrices, and the value of the loss function indicating the difference between the actual values and the predicted values is minimized. A means for performing a convergence operation so as to generate the predicted action number data,
For each user, for each region of the predicted region visit number data, a predicted behavior number for each activity of the predicted behavior number data of the user is set, and regions and activities that cannot be combined based on the region-activity correspondence data A means for generating user region / activity matrix data by removing the combination of
For each user, setting the predicted number of visits for each region of the predicted region visit number data of the user for each activity of the predicted behavior number data, and an activity that cannot be combined based on the region-activity correspondence data; Means for generating per-user activity × regional matrix data by removing region combinations;
Means for generating final score matrix data for each user by transposing and adding one of the region for each user x activity matrix data and the activity for each user x region matrix data;
Based on the generated final score matrix data for each user, when there is a request for specifying a user identifier from the service providing device, the service providing a combination of an area and an activity that is predicted to have the highest score by the user of the user identifier A recommendation information distribution apparatus comprising: means for responding to the apparatus. In the recommendation information distribution apparatus according to claim 1,
Means for acquiring user × user matrix data in which the degree of association between users is associated with a two-axis user identifier;
Means for acquiring region × region matrix data in which the degree of association between regions is associated with two-axis region identifiers;
Means for acquiring activity × activity matrix data in which the degree of association between activities is associated with a two-axis activity identifier,
The means for generating the predicted region visit number data is based on the user × region matrix data corresponding to the result region visit number data, the user × user matrix data, and the region × region matrix data. A matrix of predicted values corresponding to the matrix of is represented by the product of two decomposed matrices, and a convergence operation is performed so that the value of the loss function indicating the difference between the actual value and the predicted value is minimized,
The means for generating the predicted action frequency data is based on the user × activity matrix data corresponding to the actual action frequency data, the user × user matrix data, and the activity × activity matrix data, and a matrix of these actual values. A recommendation information distribution device characterized in that a prediction value matrix corresponding to is represented by a product of two decomposed matrices, and a convergence operation is performed so that a loss function value indicating a difference between the actual value and the prediction value is minimized. . In the recommendation information distribution apparatus according to claim 2,
The user × user matrix data is either interest graph data indicating the commonality of content viewed by two users, social graph data indicating the interchangeability of two users, or data obtained by combining these. A recommendation information distribution apparatus characterized by the above. In the recommendation information distribution apparatus as described in any one of Claim 2 or 3,
The region × region matrix data is either region affinity data indicating the possibility of two regions being visited at the same time, region similarity data indicating the similarity of two regions, or data obtained by combining these. A recommended information distribution apparatus characterized by that. In the recommendation information distribution apparatus as described in any one of Claims 2 thru | or 4,
The activity × activity matrix data is either activity affinity data indicating the possibility of two activities being acted on at the same time, activity similarity data indicating the similarity of two activities, or a combination of these. A recommended information distribution apparatus characterized by that. It is a method executed by the recommended information distribution apparatus,
Obtaining the actual area visit count data in which the actual visit count for each area of the user is associated with the user identifier and the area identifier;
Obtaining actual behavior number data in which the actual behavior number for each activity of the user is associated with the user identifier and the activity identifier;
The value of the loss function indicating the difference between the actual value and the predicted value is represented by a product of the two values obtained by decomposing the matrix of the predicted value corresponding to the user of the actual value corresponding to the actual region visit count data × the region matrix. A process of performing a convergence calculation so as to generate as predicted area visit frequency data,
A matrix of predicted values corresponding to the user × activity matrix of actual values corresponding to the actual behavior frequency data is represented by a product of two decomposed matrices, and the value of the loss function indicating the difference between the actual values and the predicted values is minimized. A process of performing a convergence calculation so as to generate as predicted behavior frequency data,
For each user, for each region of the predicted region visit number data, a predicted behavior number for each activity of the predicted behavior number data of the user is set, and regions and activities that cannot be combined based on the region-activity correspondence data Generating a region for each user x activity matrix data by removing the combination of
For each user, setting the predicted number of visits for each region of the predicted region visit number data of the user for each activity of the predicted behavior number data, and an activity that cannot be combined based on the region-activity correspondence data; Generating per user activity x region matrix data by removing region combinations;
Generating per-user final score matrix data by transposing and adding one of the per-user region × activity matrix data and the per-user activity × regional matrix data;
Based on the generated final score matrix data for each user, when there is a request for specifying a user identifier from the service providing device, the service providing a combination of an area and an activity that is predicted to have the highest score by the user of the user identifier A recommendation information distribution method comprising: a step of responding to an apparatus.

Images