JP6100677B2 - Information recommendation device, information recommendation method, and information recommendation program - Google Patents

Description

  The present invention relates to an information recommendation device, an information recommendation method, and an information recommendation program for recommending contents and items (hereinafter referred to as “items”) of a plurality of services, genres, and domains (hereinafter referred to as “domains”) in information recommendation.

  Graph mining method by constructing a graph with metadata representing the item and item attributes such as users and web pages as nodes in order to make it possible to recommend items for multiple domains such as different service categories In particular, there is a method for realizing highly accurate information recommendation in a situation where the item usage history of the user does not exist or in a situation where the recommendation accuracy is likely to decrease with a small amount (see, for example, Patent Document 1). Also, there is a method of using a user usage history common to a plurality of service domains, such as a portal site usage history, in a similar manner (see, for example, Non-Patent Document 1).

JP 2012-150561 A

Junta Tsutsuda, Makoto Nakatsuji, Toshiro Uchiyama, Hiroyuki Toda, Kei Uchiyama, “Recommendation of Information in a Cross-Domain Environment Using Access Logs”, 154th DBS / 107th IFAT Joint Research Presentation, August 2012

  In the above document, the method used for graph mining is a calculation method called Random Walk with Restarts (also called “Personalized PageRank”, hereinafter “RWR”). This is one of the methods that are attracting attention as a method for calculating the degree of association between nodes. Unlike the shortest distance between nodes, which has been often used in graph theory, RWR can calculate relevance based on the structural features of the entire graph.

  In RWR, a walker that transitions each node on the graph is assumed, and finally, an existence probability representing which node the walker is likely to exist in is calculated and used as a relevance score. As an initial state, it is assumed that a walker exists in a node corresponding to a user or item as a reference for measuring the degree of association. Then, the walker moves between nodes with a probability proportional to the edge weight connecting the nodes, and after that, the walker moves to each node in a steady state. Calculate the probability of existence (existence probability). As a result of this calculation, the obtained existence probability is used as the degree of association with the user or item corresponding to the reference node.

A specific calculation formula is that a 1 × N column vector p representing the arrival probability (relevance) from the origin node a to each node, an N × N transition matrix A, and only the origin node representing the recommended user is 1 Assuming that 1 × N column vectors q and α are constants satisfying 0 <α <1, the following expression 1 is obtained. The number of repeated calculations here is, for example, 30. However, this number is not limited to this, and may be a sufficiently large value.

  In Patent Literature 1 and Non-Patent Literature 1, when calculating the probability of transition from a starting node to each node, what domain (for example, different service category, etc.) each node belongs to is not considered. However, in the actual usage scene, when transitioning from one domain A to another domain B, the selection of the transition destination domain is affected by the transition source domain A, that is, there is a context effect, It may not be possible. For example, when browsing a web page, it is assumed that the user is interested in a certain topic, such as checking something, and it is easy to transition to a related topic, This is because it seems difficult to be interested in topics that seem unrelated. However, in Patent Document 1, in a history of a certain user for a certain period of time, if the page of the “health” domain is viewed m times and the page of the “car” domain is viewed the same number of times as m times, the user “diet” When visiting a domain page, the probability that a “healthy” domain page is recommended is equal to the probability that a “car” domain page is recommended.

  If the context should be considered as described above, it is more effective to present the “health” domain page than the “car” domain page when browsing the “diet” domain page. Expected to be However, in the conventional technology such as Patent Document 1, only the number of times in the history is used in the process of constructing the transition matrix used for RWR calculation, and the combination (direction) of the transition source / transition destination domain is not considered.

  The present invention has been made in view of the above circumstances, and an object of the present invention is to provide an information recommendation device, an information recommendation method, and information for performing more effective information recommendation in consideration of a combination of transition source / transition destination domains. To provide a recommendation program.

  In order to solve the above-mentioned problem, the present invention according to claim 1 is an information recommendation device, wherein an access history database storing access history data of access to an item by a user, and a user using the access history database A transition matrix creating unit that creates a transition matrix indicating the access relationship between items and the access history database, and calculating the number of transitions between domains based on the number of transitions, and a portion for each domain combination in the transition matrix A domain combination weight calculating unit that calculates a domain combination weight for multiplying the matrix; and a partial matrix weight calculating unit that performs weighting on the partial matrix using the domain combination weight.

  According to a second aspect of the present invention, in the information recommendation device, the domain combination weight calculation unit corrects the domain combination weight using a ratio between the number of times of use of all items and the number of times of use of items in the transition destination domain.

  According to a third aspect of the present invention, in the information recommendation device, the domain combination weight calculation unit may multiply a transition within the same domain by a predetermined coefficient to underestimate the transition between different domains. Next, the domain combination weight is corrected.

  The invention according to claim 4 is the information recommendation device, wherein the domain combination weight calculation unit specifies a transition source item using an HTTP referrer included in access history data, and specifies a domain of the transition source item. Thus, the number of transitions between domains is estimated.

  The present invention according to claim 5 is the information recommendation device, wherein the domain combination weight calculating unit uses a plurality of items having a common link source from an access history using information on the presence / absence of a link between items, When there is also an access history to the corresponding link source, it is estimated that there is a transition from the link source to a plurality of items, and the number of transitions between domains is estimated.

  The present invention according to claim 6 is an information recommendation method performed by an information recommendation device, wherein the information recommendation device includes an access history database storing access history data when a user accesses an item, and the access history database is stored in the information recommendation device. A transition matrix creating step for creating a transition matrix indicating an access relationship between a user and an item, and calculating the number of transitions between domains using the access history database, and a domain in the transition matrix based on the number of transitions A domain combination weight calculating step for calculating a domain combination weight for multiplying a partial matrix for each combination, and a submatrix weight calculating step for weighting the partial matrix using the domain combination weight are performed.

  According to a seventh aspect of the present invention, in the information recommendation method, the domain combination weight calculating step corrects the domain combination weight using a ratio between the number of times of use of all items and the number of times of use of items in the transition destination domain.

  In the information recommendation method according to the eighth aspect of the present invention, in the information recommendation method, the domain combination weight calculating step may multiply a transition within the same domain by a predetermined coefficient to underestimate the transition between different domains. Next, the domain combination weight is corrected.

  The present invention according to claim 9 is the information recommendation method, wherein the domain combination weight calculating step specifies a transition source item using an HTTP referrer included in access history data, and specifies a domain of the transition source item. Thus, the number of transitions between domains is estimated.

  In the information recommendation method according to claim 10, in the domain combination weight calculation step, there is a plurality of items having a common link source from an access history using information on the presence / absence of a link between items, When there is also an access history to the corresponding link source, it is estimated that there is a transition from the link source to a plurality of items, and the number of transitions between domains is estimated.

The present invention according to claim 11 is an information recommendation program for causing a computer to function as the information recommendation device.

  According to the present invention, it is possible to provide an information recommendation device, an information recommendation method, and an information recommendation program that perform more effective information recommendation in consideration of a combination of transition source / transition destination domains.

It is a block diagram of the information recommendation apparatus in one embodiment of this invention. It is an example of the table comprised by the access history database. It is a flowchart of the whole process of an information recommendation apparatus. It is a flowchart of a transition matrix creation process. It is an image figure of UI matrix before weighting calculation. It is an image figure of UI matrix after weighting calculation. It is an image figure of UI matrix after weighting calculation. It is an image of recommended user ID data. It is an image of the output of a recommendation calculation part. It is a flowchart of a process of a domain combination weight calculation part. It is an image figure of the transition between domains. It is an image figure of the appearance number of the transition between domains. It is an image figure which shows the ratio which occupies in the origin domain of the transition between domains.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  FIG. 1 shows a configuration of an information recommendation device according to an embodiment of the present invention. The information recommendation apparatus shown in FIG. 1 includes an access history database 200, a calculation request unit 100, a transition matrix creation unit 101, a domain combination weight calculation unit 102, a transition matrix storage unit 103, and a submatrix weight calculation unit 104. And a recommendation condition setting unit 105 and a recommendation calculation unit 106.

  The access history database 200 is connected to the access log collection system 10 which is an external system, and stores access history (access log) data periodically supplied from the access log collection system 10.

  The calculation request unit 100 sends an instruction to the transition matrix creation unit 101 to create a transition matrix. The transition matrix creation unit 101 creates a transition matrix indicating the access relationship between the user and the item using the access history database 200 and stores it in the transition matrix storage unit 103. The domain combination weight calculation unit 102 uses the access history database 200 to calculate the number of transitions between domains, and uses the number of transitions to determine a domain combination weight that is multiplied by a submatrix for each domain combination in the transition matrix. .

  The submatrix weight calculation unit 104 weights the submatrix using domain combination weights. The recommended condition setting unit 105 receives the recommended user ID data 201 supplied from the external access log collection system 10. The recommendation calculator 106 performs a recommendation calculation using the transition matrix weighted by the submatrix weight calculator 104 and outputs a recommendation result. The output of the recommendation calculation unit 106 is a list of recommended items for the recommendation target user, and the list is sent to the recommendation result presentation system 20 in the subsequent stage and finally displayed on a terminal (not shown) such as a user. .

  Further, the information recommendation device may be connected to the item structure database 30. The item structure database 30 stores information as to which item each item has a link from.

  In the present embodiment, it will be described below that an access history (access history database 200) by which a user accesses an item on a website is used as a usage history used for recommendation calculation. It is not limited to.

  FIG. 2 is an image diagram of a table of the access history database 200 in the present embodiment. The access history database 200 shown in the figure is a database in which access history data (record) including a user ID, access time, item ID, domain ID of the domain to which the item belongs, and item metadata is stored. . As shown in FIG. 2, the access history database 200 is not limited to a case where the access history database 200 is realized by only one table, and may be divided into an appropriate size and configured from a plurality of tables or files. The tables or files may be on different computers.

  As the information recommendation device described above, for example, a general-purpose computer system including a CPU, a memory, and an external storage device such as an HDD can be used. In this computer system, each function of the information recommendation device is realized by the CPU executing the program of the information recommendation device loaded on the memory. The program for the information recommendation apparatus can be stored in a computer-readable recording medium such as a hard disk, a flexible disk, a CD-ROM, an MO, a DVD-ROM, or can be distributed via a network. Note that the information recommendation apparatus may be an information recommendation system including a plurality of computers.

  Next, the processing of this embodiment will be described.

  FIG. 3 is a flowchart illustrating an example of overall processing of the information recommendation apparatus.

  Step 310) The calculation requesting unit 100 sends an instruction to the transition matrix creating unit 101 to create a transition matrix. The calculation requesting unit 100 may be a module that runs a script on time, for example.

  Step 320) Upon receiving the command, the transition matrix creation unit 101 acquires access history data including a user ID, an item ID, an item domain ID, and the like from the access history database 200. An image of data stored in the access history database 200 is as shown in FIG. In the access history database 200, which user accessed what item in time series is stored together with additional information such as the domain of the item and metadata.

  The transition matrix creation unit 101 creates a transition matrix A ′ necessary for recommendation calculation described later, using data stored in the access history database 200. The method described in Non-Patent Document 1 is used as a method for creating the transition matrix A ′. However, in this embodiment, since the weighting process considering the domain combination is performed for each submatrix, out of the processes described in Non-Patent Document 1, the normalization for each submatrix is performed. The normalization is performed after the weight calculation of the submatrix is completed. Among the processes described in Non-Patent Document 1, the outline of the process performed by the transition matrix creating unit 101 of this embodiment is shown in the flowchart of FIG.

  FIG. 4 is a flowchart showing processing performed by the transition matrix creation unit 101.

  Step 410) The transition matrix creation unit 101 creates respective partial matrices for any two combinations of user, item, and metadata of the data acquired from the access history database 200. For example, a user-item submatrix, a user-metadata submatrix, an item-metadata submatrix, etc. Here, since the weight of the domain combination of the transition source / transition destination of this embodiment is mainly used as a user-item submatrix (hereinafter referred to as a UI matrix), the UI matrix is taken as an example hereinafter. I will discuss it.

In the UI matrix, when access history data (record) to an item i _k of a certain user u _j exists in the access history database 200, the corresponding element UI (j, k) in the matrix is set to 1, and 0 does not exist. Is created. However, the value is not 1 and may be another fixed value. Further, in the case where a plurality of access history data from the user u _j to item i _k may be 1 or a fixed value value obtained by adding the number of the access history data (frequency). This process is repeated for all combinations of users and items.

  Sub-matrices other than the UI matrix are created in the same way. However, in the case of a paired matrix such as a UI matrix and an IU matrix, the IU matrix may be created by simply exchanging the rows and columns of the UI matrix.

  Step 420) The transition matrix creation unit 101 performs normalization processing for each domain so that the sum of column values becomes 1 for each domain. Examples of domains include a group of subscription services such as users of service A and users of service B in the case of users, and categories of content such as “diet” and “car” described above in the case of items. Depending on the service, it is assumed that the number of components may vary greatly. Also, the number of domains may be different, for example, there is only one user, but there are v items.

  For example, in the case of a UI matrix, the processing represented by the following equation 2 is performed. Here, x and y respectively represent domain IDs of domains to which the user and item belong, and j and k respectively represent arbitrary row numbers and column numbers in the matrix. That is, UI (x, y, j, k) represents an element in the j-th row and the k-th column of the UI matrix in which the domain ID of the user (before normalization) is x and the domain ID of the item is y. An item set with a domain ID y is represented as I ′ (y).

Normalization is performed in the same manner for matrices containing UI matrices, items, and metadata other than x and y.

  The transition matrix creation unit 101 stores the matrix created in step 420 in the transition matrix storage unit 103. Hereinafter, the process of FIG. 3 will be described again.

Step 330) The domain combination weight calculator 102 calculates the domain combination weight w using the access history for each user ID obtained from the access history database 200. The domain combination weight w is a weight to be applied to the corresponding submatrix determined by the combination of the transition source domain and the transition destination domain. For example, when the user u _j browses an item in the “Diet” domain and then transitions to the “Health” domain, the transition source domain is “Diet” and the transition destination domain is “Health”. If there are v types of item domains, the transition from “diet” to “health” and the transition from “health” to “diet” is considered, and transitions within the same domain such as “diet” to “diet” are also considered. The combination of the transition source domain and the transition destination domain is v ² , that is, the weight w is also calculated as v ² . The process of calculating the weight w will be described later with reference to FIG.

  Step 340) The partial matrix weight calculation unit 104 reads the transition matrix created in step 320 from the transition matrix storage unit 103. Next, in the transition matrix, for the part to which the domain combination weight w is applied (here, only the UI matrix is used), the matrix is duplicated for the number of types of domain (v types) of the item. This is to apply different transition matrices to which different domain combination weights w obtained in step 330 are applied depending on which domain item is being viewed by the user to be recommended at the time of recommendation (transition source domain). .

Then, as shown in Expression 3, the weight w calculated in step 330 is multiplied by the corresponding matrix. However, the user is a single domain (for example, all users are registered users of service A), and the items are divided into v types of domains. In this case, D (a, b) in Equation 3 is a matrix for a user who browses an item in domain A (ID is a), and a UI matrix when transitioning to an item in domain B (ID is b) It is. w (a, b) is the domain combination weight when transitioning from domain A to B, and ui (a, b, j, k) is the jth row and kth column of the UI matrix when transitioning from domain A to B Represents an element. The number of users is u, and the number of items is m.

Here, FIG. 5 to FIG. 7 are used to illustrate an example in which the weighted matrix value changes when the transition source domain is different in the weighting calculation in step 340. FIG. 5 shows a UI matrix before weighting calculation. U _i written as a row number / column number represents the i-th user, and i _i ^a represents the i-th item of the domain A (ID is a). The number of users is n, the number of items in domain A is m, and the number of items in domain B is g.

FIG. 6 and FIG. 7 show UI matrices after weighting calculations, but show different domain combinations. In FIG. 6, the domain combination weight w (a, a) when the transition source is domain A and the transition destination is domain A is 3, and the weight w (a, b) when the transition destination is domain B is 1. It shall be / 2. Note that u _i ^a represents a case where the i-th user is browsing an item in domain A (domain ID is a). On the other hand, when the transition source is domain Z and the transition destination is domain a, the domain combination weight w (z, a) is 3/2, and when the transition destination is domain B, the weight w (z, b) is 1. At this time, it becomes as shown in FIG.

  In this way, the plurality of transition matrices that are created are selectively used depending on the difference in the domain that the recommended user browses. That is, the recommendation result (p in Formula 1) is different depending on the domain of the item being browsed by the user.

  Next, the weighted matrix is combined to create a UI matrix A ′ (a, UI) of the transition matrix A ′ when the transition source domain is A (ID is a) as shown in Equation 4. . D (a, b) is a weighted UI matrix when the domain combination is domain A to B. This is the same as the matrix D in Equation 3. In this way, the UI matrix including all items is combined.

Further, as shown in Expression 4, the matrix is combined with other matrices including items and metadata stored in the transition matrix storage unit 103. In Equation 4, U represents a user, I represents an item, and M represents metadata. For example, IM is an item-metadata matrix. In this way, a transition matrix A ′ including all elements is created.

Next, as in Non-Patent Document 1, the entire transition matrix A ′ shown in Expression 5 is normalized to create a final transition matrix A. However, although A (a) is a normalized transition matrix, it represents a transition matrix when the user is browsing the domain A (ID is a). A ′ (a, j, k) represents the element in the j-th row and the k-th column of the transition matrix A ′ whose transition source domain is A.

  Step 350) Step 340 is repeated by the number of types of transition source domains to create transition matrices A (a) to A (z) for all transition source domains.

  Step 360) The recommendation condition setting unit 105 acquires a recommendation condition necessary for the subsequent recommendation calculation from the external system, and sets the corresponding recommendation condition. First, the recommendation condition setting unit 105 reads out a recommended user ID from the recommended user ID data 201 acquired from the external system.

  FIG. 8 shows an image of the recommended user ID data 201. As shown in the figure, the recommended user ID data stores the ID of the user to be recommended. The recommended user can be arbitrarily set based on service requirements such as the user agreeing to receive the recommended service, contract details, conditions, and the like.

  Step 370) The recommendation calculation unit 106 sets the recommendation condition acquired in Step 360 for the transition matrices A (a) to A (z) created in Step 340, and performs the recommendation calculation. In this recommendation calculation method, Equation 1 is calculated in the same manner as in the method of Non-Patent Document 1 except that the transition source domain of the user is assumed.

  Finally, as shown in FIG. 9, the combination of the user transition source domain (the domain of the item being viewed at the time of recommendation), the recommended item ID, the domain ID, and the evaluation value for each recommended user is an external system. Is sent to the recommendation result presentation system 20.

  The recommendation result presentation system 20 identifies the domain of the item being browsed by the recommended user, searches the recommendation calculation result of the corresponding user in which the transition source domain that matches the domain is selected, and selects the recommended item Have

  Next, the process of calculating the domain combination weight w in step 330 will be described in more detail. FIG. 10 is a flowchart of step 330.

  Step 610) The domain combination weight calculation unit 102 acquires data including a combination of a user ID, an access time, and an item domain ID from the access history database 200.

  Step 620) The domain combination weight calculation unit 102 sorts the data acquired in Step 610 in ascending order using the user ID as the primary key and the access time as the secondary key. Then, it is counted from which domain to which domain in the same user. However, since the clear transition source is not known from the access history data illustrated here, it is assumed here that the domains of two data (records) whose access times are adjacent are the transition source and the transition destination, and the transition between the domains. It is considered.

For example, when the sorted data of the user u ₁ is arranged in the order of “music domain ID”, “movie domain ID”, “movie domain ID”, “music to movie” and “movie to movie” It is considered that there was a transition. In addition, a threshold may be provided for the difference in access time between adjacent data, and only when the difference in access time is within the threshold may be subject to this processing as “transition between domains”, or in the middle May be omitted, and the domain of the item with the oldest access time may transition to the domain of the item with the newest access time.

Further, the domain combination weight calculation unit 102 may be connected to the item structure database 30 that is an external system in which the link structure of items is stored, and the link structure may be used for estimating the transition source. Here, for example, item i _j has links to items i _k , i _g , and i _x , item i _k has links to items i _m and i _p , and item i _g has Indicates information that there is a link to items i _p and i _r . If there is a plurality of items having a common link source (item i _j , item i _{g in the} above example) in the access history within a certain period of time, and there is an access history to the common link source, a common link An item with a source may be regarded as a transition from the link source.

Specifically, the user u _j is, within a certain period of time, the item to the old order of access time i _j, item i _k, and have a history of access to the item i _g. At this time, when the link structure of each item is acquired in the item structure database 30 and the common link source is analyzed, the item i _k and the item i _g have a link (common link source) from the item i _j , and since the user u _j is also access history for the item i _j, the transition may be estimated to be the item i _g items i _k and items i _j from the item i _j.

  Alternatively, when access history data that clearly indicates the transition source can be acquired, the information may be used. For example, when HTTP referrer information can be used, it is analyzed whether the transition source address corresponds to any domain, and if it is applicable, the domain name is acquired, and the current domain is transitioned from the domain. May be considered.

  Then, as shown in FIG. 11, for example, an inter-domain transition such as “movie to music” is extracted.

Step 630) The weight of the domain combination is obtained by correcting the ratio of the combination of certain domains in each starting domain. This will be specifically described below. If there are v types of service domains that the access history database 200 targets, how many transitions can be considered for each of the ^two combinations (including transitions in the same domain, including transition directions)? , Aggregate across all users. The image is shown in FIG. In this table, for each domain as a starting point such as music and cooking, the number of times each transition has occurred is tabulated over all users.

  Furthermore, the ratio that each transition occupies in the origin domain is calculated. The image is shown in FIG. The ratio obtained here is represented as R (a, b), where a is the domain ID of the transition source and b is the domain ID of the transition destination. In this embodiment, the following correction is applied to R (a, b) to obtain the domain combination weight w (a, b), but R (a, b) is converted to the domain combination weight w without correction. It may be (a, b). Further, both the following first correction and second correction may be performed, or only one of the corrections may be performed.

<First correction: correction based on the number of items in the destination domain>
Here, the total usage count of items belonging to each domain differs depending on the domain, so depending on the domain, there are things that are easy to transition because there are not only context but simply the number of times the item is used Let's go. In order to reduce this influence, the domain combination weight is corrected using the ratio between the number of times of use of all items and the number of times of use of items in the transition destination domain. That is, as shown in Expression 6, R (a, b) is corrected by the ratio of the number of times of use of the item in the transition destination domain b to the total number of times of item use. Here, the number of times the item is used can be calculated by the sum of elements in the range of the corresponding domain of the item in the UI matrix. It may be the sum of only elements whose value is not 0 (has an access history). In Equation 6, m _b represents the sum of elements whose domain ID of the UI matrix item is b, and M represents the sum of elements of all domains of the UI matrix. As a result, R (a, b) is highly evaluated as the proportion of all items in domain b is small.

<Second correction: suppression of intra-domain transition>
Furthermore, the trend of R (a, b) is the same as the domain of the transition source as represented by the situation where a certain site is read across multiple pages if the data has a property such as a web page browsing history. There should be many things that transition to the domain (hereinafter, the self domain). On the other hand, as described in Patent Document 1 and Non-Patent Document 1, these technologies are for enabling recommendation across a plurality of domains and for promoting migration on a user's Web site or the like. Therefore, if the proportion of each transition occupies the weight corresponding to the domain combination as it is, the return to the own domain is promoted and it is contrary to the original recommendation intention, so it is not suitable. Therefore, in this embodiment, in order to suppress weighting when returning to the own domain, a transition within the same domain is multiplied by a predetermined coefficient (here, self-loop suppression weight 1-γ), The domain combination weight is corrected so as to underestimate the transition between different domains.

Specifically, as shown in Equation 7, the weight w (a, b) when transitioning from domain A to domain B is the transition source domain and the transition destination domain for any γ where 0 <γ <1 If the two are different, the relative weight is adjusted by multiplying the weight R (a, b) obtained in the previous step by the self-loop suppression weight 1-γ if they are the same. For example, when γ = 0.8 or the like, the weight of transition in the same domain can be made smaller than in the case of transition between different domains.

  In the present embodiment, the domain combination weight w is newly calculated every time the calculation request unit 100 starts processing, but the domain combination weight w is obtained once in a span longer than the recommendation calculation. The data may be stored in a storage unit (not shown) and used.

  In the present embodiment described above, the number of transitions between domains is calculated using the access history database 200, and the domain combination weight w for multiplying the submatrix for each domain combination in the transition matrix is calculated based on the number of transitions. The partial matrix is weighted using domain combination weights.

  Thereby, in this embodiment, the weight by the combination of a transition origin / transition destination domain can be reflected in a transition matrix, and the more effective recommendation result suitable for the context of a user's information browsing action can be obtained. . That is, in the transition matrix, the combination of the transition source / transition destination domain is considered by weighting the partial matrix defined by the combination of the transition source / transition destination domain according to the corresponding combination. More effective information recommendation can be realized. For example, a transition matrix that takes into account the combination (direction) of the transition source / destination domain, such as multiplying the “diet” domain to the “health” domain by a greater weight than the “diet” domain to the “car” domain. Can be built.

  Therefore, the present embodiment provides a recommendation calculation mechanism including a technique for calculating a weight for a submatrix defined by a combination of a transition source / transition destination domain and a technique for creating a transition matrix using the weight. be able to.

  In addition, this invention is not limited to the said embodiment, A various change and application are possible within a claim.

10: Access log collection system
20: Recommendation result presentation system
30: Item structure database
100: Calculation request department
101: Transition matrix generator
102: Domain combination weight calculator
103: Transition matrix storage
104: Submatrix weighting calculator
105: Recommendation condition setting section
106: Recommendation calculator
200: Access history database

Claims (11)

An information recommendation device,
An access history database that stores access history data of items accessed by users;
Using the access history database, a transition matrix creating unit that creates a transition matrix indicating an access relationship between a user and an item
A domain combination weight calculating unit that calculates the number of transitions between domains using the access history database, and calculates a domain combination weight for multiplying a partial matrix for each domain combination in the transition matrix based on the number of transitions;
An information recommendation device comprising: a partial matrix weighting calculation unit that weights the partial matrix using the domain combination weight. The information recommendation device according to claim 1,
The domain combination weight calculation unit corrects the domain combination weight by using a ratio between the number of times of use of all items and the number of times of use of items in the transition destination domain. The information recommendation device according to claim 1 or 2,
The domain combination weight calculation unit corrects the domain combination weight so that the transition within the same domain is multiplied by a predetermined coefficient and is underestimated than the transition between different domains. apparatus. The information recommendation device according to any one of claims 1 to 3,
The domain combination weight calculation unit specifies a transition source item using an HTTP referrer included in access history data, and estimates the number of transitions between domains by specifying a domain of the transition source item. Information recommendation device. The information recommendation device according to any one of claims 1 to 3,
The domain combination weight calculation unit uses the information on the presence / absence of a link between items, and when there are a plurality of items having a common link source from the access history and there is also an access history to the corresponding link source, the link An information recommendation device that estimates that there are transitions from the original to multiple items and estimates the number of transitions between domains. An information recommendation method performed by an information recommendation device,
The information recommendation device includes an access history database that stores access history data in which a user accesses an item,
Using the access history database, a transition matrix creating step for creating a transition matrix indicating an access relationship between a user and an item
A domain combination weight calculating step of calculating the number of transitions between domains using the access history database and calculating a domain combination weight for multiplying a partial matrix for each domain combination in the transition matrix based on the number of transitions;
A submatrix weighting calculation step of weighting the submatrix using the domain combination weights. An information recommendation method according to claim 6,
The domain combination weight calculating step corrects the domain combination weight using a ratio of the number of times of use of all items and the number of times of use of items in the transition destination domain. The information recommendation method according to claim 6 or 7,
The domain combination weight calculation step corrects the domain combination weight so that a transition within the same domain is multiplied by a predetermined coefficient and is underestimated than a transition between different domains. Method. The information recommendation method according to any one of claims 6 to 8,
The domain combination weight calculating step specifies a transition source item using an HTTP referrer included in access history data, and estimates the number of transitions between domains by specifying a domain of the transition source item. How to recommend information. The information recommendation method according to any one of claims 6 to 8,
The domain combination weight calculating step uses information on the presence / absence of a link between items, and when there are a plurality of items having a common link source from the access history and there is also an access history to the corresponding link source, the link An information recommendation method characterized by estimating the number of transitions from the original to multiple items and estimating the number of transitions between domains. An information recommendation program for causing a computer to function as the information recommendation device according to any one of claims 1 to 5.

Images