JP2017182724A - Item recommendation program, item recommendation method, and item recommendation apparatus - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce calculation load on item recommendation.SOLUTION: An item recommendation program causes a computer to execute generation processing, clustering processing, and extraction processing. The generation processing determines, for each user, a vector which indicates preference of a user with respect to each of characteristics of an item by multiplying history information of the item selected by a user by an item characteristic matrix reduced with at least one piece of group information, out of the item characteristic matrix having characteristics of each of items and group information organizing characteristics, and adds user profile to the vector, to generate a user preference characteristic vector. The clustering processing clusters users, on the basis of the user preference characteristic vectors of the users. The extraction processing aggregates history information of users belonging to each of clusters, to extract an item selected with high frequency as a recommendation item.SELECTED DRAWING: Figure 1

Description

  Embodiments described herein relate generally to an item recommendation program, an item recommendation method, and an item recommendation device.

  2. Description of the Related Art Conventionally, there is an item recommendation device that searches and recommends an item suitable for a user from a large number of items such as a program. Searches for recommended items include a user base, an item base, and a hybrid base combining a user base and an item base.

  In the user-based item search, users with similar interests are extracted from the user's behavior history, and the items accessed by the extracted users are recommended. The item-based item search recommends items that the user has accessed in the past or items similar to the selected item. In the hybrid-based item search, an item is recommended based on a similarity obtained by combining a user profile, an item content, and a user action history. In the user base, it may be difficult to recommend items to new users with a small behavior history. In addition, in the item base, the user information similar to the user is not reflected in the item recommendation. On the other hand, in the hybrid base, it is possible to realize item recommendation for a new user with a small action history and item recommendation reflecting information of a user similar to himself / herself.

JP 2002-369090 A JP 2000-13708 A

  However, in the item recommendation by the hybrid base, the amount of data of the multidimensional matrix (higher order tensor, multidimensional array) due to the profile of each user, the contents of the item, and the elements of the user's behavior history becomes enormous, and the calculation for item recommendation There is a problem that the load is large. Thus, when the calculation load for item recommendation is large, the calculation time becomes long, and it becomes difficult to finish the processing within the practical time.

  In one aspect, an object is to provide an item recommendation program, an item recommendation method, and an item recommendation device that can reduce the calculation load for item recommendation.

  In the first plan, the item recommendation program causes the computer to execute a process to create, a process to cluster, and a process to extract. The processing to be created includes, for each user, an item characteristic matrix contracted by using at least one group information among item characteristic matrices having each characteristic of the item for each item and group information for collecting the characteristics. Is multiplied by the history information of the selected item to obtain a vector indicating the user's preference for each characteristic of the item, and the user's profile is added to the vector to create a user preference characteristic vector. In the clustering process, the users are clustered based on the user preference characteristic vector created for each user. In the process of extracting, for each cluster, the history information of users belonging to the cluster is aggregated and an item with a high selection frequency is extracted as an item to be recommended.

  According to one embodiment of the present invention, the calculation load for item recommendation can be reduced.

FIG. 1 is a block diagram illustrating a functional configuration of an item recommendation device according to the embodiment. FIG. 2 is a flowchart illustrating an operation example of the item recommendation device according to the embodiment. FIG. 3 is an explanatory diagram illustrating creation of a user preference characteristic vector. FIG. 4 is an explanatory diagram for explaining the user behavior analysis table. FIG. 5A is a graph illustrating the relationship between the number of clusters and SD (K). FIG. 5B is a graph illustrating the relationship between the number of clusters and SSW (K). FIG. 6 is a flowchart illustrating an example of the clustering process. FIG. 7 is an explanatory diagram for explaining the creation of the basic recommendation list table. FIG. 8 is an explanatory diagram for explaining the creation of the user recommendation list table. FIG. 9 is a block diagram illustrating an example of a hardware configuration of the item recommendation device according to the embodiment.

  Hereinafter, an item recommendation program, an item recommendation method, and an item recommendation device according to embodiments will be described with reference to the drawings. In the embodiment, configurations having the same functions are denoted by the same reference numerals, and redundant description is omitted. Note that the item recommendation program, the item recommendation method, and the item recommendation device described in the following embodiments are merely examples, and do not limit the embodiments. In addition, the following embodiments may be appropriately combined within a consistent range.

  FIG. 1 is a block diagram illustrating a functional configuration of an item recommendation device according to the embodiment. As shown in FIG. 1, the item recommendation device 1 includes an analysis unit 10, a clustering unit 20, and an output unit 30, and searches for items suitable for the user from among a large number of items (programs and the like) in the item DB 11. It is a device that outputs a user recommendation list table 17.

  In this embodiment, a video program (content) provided by video on demand (VOD) is exemplified as an item recommended to the user. The type of item recommended to the user is not limited to a video program, and may be a product handled on a mail order site, for example.

  The analysis unit 10 includes an item keyword table 12 for each video program in which the contents (characteristics) of the item (video program) are extracted from the item DB 11, and a viewing history for each user indicating a history of programs selected (viewed) by the user in the past. The user profile 14 for each user loaded from the data 13 and the user profile DB 18 is input.

  The contents (characteristics) for each video program in the item keyword table 12 include the genre of the video program, the performers, the keywords included in the description, the appearance area, and the like. The characteristics of the video program are not limited to the above-mentioned types, and any characteristic may be used as long as it shows the characteristics of the video program.

  The user profile in the user profile 14 includes the user's gender, age, location, and the like. Note that the user profile is not limited to the above-mentioned type, and may be any as long as it shows the user's characteristics (features).

  The analysis unit 10 analyzes the similarity of the user's preference using the user profile 14, the similarity of the content of the video program using the item keyword table 12, and the similarity of the viewing history using the viewing history data 13 on a hybrid basis. To do. The analysis unit 10 outputs the analysis result as the user behavior analysis table 15.

  The clustering unit 20 creates a cluster of similar users based on the user behavior analysis table 15. Then, for each cluster, the clustering unit 20 aggregates the viewing histories of users belonging to the cluster and extracts video programs with a high selection frequency as recommended video programs. The clustering unit 20 arranges the video programs extracted for each cluster in descending order of selection frequency and outputs the basic recommendation list table 16.

  The output unit 30 outputs the user recommendation list table 17 of the video program extracted from the cluster to which the user belongs based on the basic recommendation list table 16. Specifically, the output unit 30 refers to the viewing history data 13 of the user, and excludes video programs that the user has already viewed (selected) from the list of video programs extracted from the cluster to which the user belongs. Then, the output unit 30 outputs a list of video programs that the user has not viewed (unselected) as the user recommendation list table 17. Thereby, the item recommendation device 1 can recommend an unviewed video program.

In the item recommendation on a hybrid basis, the calculation load increases as the number of elements in the multidimensional matrix increases. For example, assume a case in which a program is recommended to a user from about 50,000 programs in a video-on-demand service with 10 million users. In this case, the number of elements of the two-dimensional matrix (array) holding the viewing history for each user is (item, user) → 5 · 10 ^ 4 × 1 · 10 ^ 7 = 5 · 10 ^ 11 ("^" Sign represents power, and so on). The number of elements of a multidimensional matrix (higher tensor, multidimensional array) including information such as sex (two levels of men and women) and age (stratified by every five years and 20 levels) as the user profile is as follows: As shown in the figure, it becomes 10 tera orders.
(Item, user, gender) → 5 · 10 ^ 11 × 2 = 1 · 10 ^ 12
(Item, user, gender, age) → 1 · 10 ^ 12 × 20 = 2 · 10 ^ 13

  Furthermore, as the contents of the items are added, the data amount of the multidimensional matrix exceeds the peta order. When calculating the degree of similarity between users using such a multidimensional matrix, tensor singular value decomposition obtained by extending the singular value decomposition of the matrix to a tensor is used. Since the calculation amount of this singular value decomposition is proportional to the square of the number of elements, when the data amount is peta-order (10 ^ 12), the calculation amount required for singular value decomposition is (10 ^ 12) ^ 4 = 10. ^ 48, which increases the calculation load.

  Accordingly, the analysis unit 10 of the item recommendation device 1 multiplies the item keyword table 12 (item characteristic matrix) having each characteristic for each video program (item) by the user's viewing history data 13 to correspond to each characteristic of the video program. A vector indicating the user's preference is obtained, and the user profile 14 is assigned to create a user preference characteristic vector. That is, the analysis unit 10 reduces the multidimensional matrix (tensor) in the item keyword table 12, the viewing history data 13, and the user profile 14.

  The analysis unit 10 obtains a user preference characteristic vector for each user and outputs it as the user behavior analysis table 15. The clustering unit 20 refers to the user behavior analysis table 15 and clusters users based on user preference characteristic vectors having profiles and preferences for each user. As described above, the item recommendation device 1 reduces the amount of data required for calculation by reducing the multidimensional matrix (tensor) in the item keyword table 12, the viewing history data 13, and the user profile 14, thereby reducing the calculation load. Can be small.

Here, the detail of operation | movement of the item recommendation apparatus 1 is demonstrated. Note that symbols used in the following description are defined as follows.
I: number of items (video programs) J: number of users K: number of groups of item characteristics / number of clusters of users L [k]: number of items of groups of k-th item characteristics M: number of items of user profile N: in item Number of items for all characteristics

  FIG. 2 is a flowchart illustrating an operation example of the item recommendation device 1 according to the embodiment. As shown in FIG. 2, when the process is started, the analysis unit 10 receives input of the item keyword table 12, the viewing history data 13, and the user profile 14. Next, the analysis unit 10 creates a viewing time ratio vector having a viewing time ratio for each video program as an element for all users based on the viewing history data 13 (S1).

Specifically, the analysis unit 10 creates a viewing time ratio matrix R for all users based on the viewing history data 13 of each user. The viewing time ratio matrix R identifies each user by a column index and each item (video program) by a row index, and describes them as follows.
Viewing time ratio matrix: R [i, j] = vtime [i, j] / ptime [i]
User index: j = 1, 2,..., J
Item index: i = 1, 2,..., I

  Here, vtime [i, j] is the time when the user (j) viewed the item (i). Also, ptime [i] is the playback time of item (i). The viewing history data 13 used for calculating the viewing time ratio matrix R [i, j] uses a history that goes back from the present for a certain period (for example, one year). It is assumed that the period for using this history can be arbitrarily set such as one month, three months, and half a year. In addition, the analysis unit 10 recalculates the viewing time ratio matrix R [i, j] every certain period (for example, three months). This recalculation period can also be set arbitrarily such as one month, two weeks.

  In this embodiment, the viewing time of the video program is calculated based on the item history information (video program viewing history). However, if the item is a product, the user (j) ) May be obtained as an element value.

Next, the analysis unit 10 creates an item characteristic matrix X [i, g [k, l]] that represents the contents (characteristics) of all items as a matrix based on the item keyword table 12 (S2). In the item characteristic matrix X [i, g [k, l]], “1” is included if the item (i) includes the item content item G [k, l], and “0” otherwise. It has a value and is defined as follows.
Item characteristic matrix: X [i, g [k, l]] = 0 or 1
Item characteristic item matrix: G [k, l]
Item characteristic serial number index: g [k, l]
Item characteristic group index: k = 1, 2,..., K
In-group index: l = 1, 2,..., L [k]

  L [k] has a different value for each group. A group is a genre (k = 1), performer (k = 2), director (k = 3), explanatory keyword (k = 4), appearance area (k = 5), release year (k = 6) )... Is a characteristic used to classify items (video programs). In the present embodiment, for the sake of explanation, a large group including a very large number of flags, such as performers and directors, is simply expressed, but each group has a degree of user characteristics when clustered. A subdivided group may be prepared.

  If the item content item is a genre, G [1,] = [SF, sports, historical drama, document, music,...], If the performer is G [2,] = [XXXXXX, XXXXXX, ..], If it is a genre (k), it is a specific item that characterizes the item, such as G [k,] = [AAAA, BBBB,. The keywords are nouns, adjectives, verbs, etc. included in the description of the item.

About the keyword of an item, it collects by performing following (1)-(4).
(1) Collecting proper nouns, common nouns, verbs, and adjectives by text mining from explanations (EPG (electronic program guide) for TV programs, content introduction text for on-demand video programs) that describe item contents To do.
(2) A new keyword that cannot be collected in (1) is collected by searching a review sentence on SNS (Social Networking Service) related to the item and performing text mining.
(3) The registered keyword is added because it is determined that the item provider should add the item.
(4) A keyword obtained by text mining from an opinion commented by the viewer on the item is added.

  The item characteristic serial number index is an index representing the item number when an item content item is viewed through all groups. For example, if there are 100 genres, 1000 performers, and 100 directors, g “1,100” = 100, g [2,1] = 100 + 1 = 101, g [2,2] = 100 + 2 = 102 , G [3,1] = 100 + 1000 + 1 = 1101.

  The analysis unit 10 may limit the number of content items when creating the item characteristic matrix X [i, g [k, l]] from the item keyword table 12 by the above-described method. Specifically, the analysis unit 10 can reduce the number of items by creating a union of keywords representing characteristics for all items and using them as common content keywords.

  Since the number of elements of the common content keyword may be enormous, the analysis unit 10 provides a selection rule in advance to prevent the number of elements from exploding. For example, rules such as discarding old reviews, discarding keywords with low appearance frequency when creating a union, and discarding keywords related to items with a small number of views may be applied. In addition, the number of elements of characteristic items may be limited because clustering of user preference characteristic vectors is completed within a practical calculation time.

  Next, the analysis unit 10 multiplies the item characteristic matrix X [i, g [k, l]] created in S2 and the viewing time ratio matrix: R [i, j], to calculate the preference characteristic matrix V0 [j, g [k, l]] (a vector in which each row of the matrix indicates the preference of each user) is created (S3).

  Specifically, the analysis unit 10 creates the preference characteristic matrix V0 [j, g [k, l]] by the following equation (1).

Next, the analysis unit 10 adds a user characteristic (user profile 14) for each user and extends the user preference characteristic vector (S4). Specifically, the analysis unit 10 creates a preference characteristic matrix V [j, n] obtained by adding a user profile characteristic matrix based on the user profile 14 to the preference characteristic matrix V0 [j, g [k, l]]. . The user profile characteristic matrix is a matrix for identifying each user by a column index and each profile by a row index, and is as follows.
User profile characteristic matrix: P [j, m]
User index: j = 1, 2,..., J
Profile index: m = 1, 2,..., M

  The user profile indicated by each element of the user profile characteristic matrix may be a non-numeric value such as gender or prefecture name, but is digitized so that distance calculation in clustering is possible. For example, in the case of males and females, the numerical values are such that male = 1 and female = 0. In the case of the name of the prefecture, the numbers from Hokkaido to Okinawa are digitized by assigning serial numbers such as Hokkaido = 1, Aomori = 2, etc. In addition, elements such as age and annual income that have a large numerical number are appropriately divided into hierarchies and assigned numbers. For example, in the case of age, it is digitized every 10 years, and in the case of annual income, it is digitized every 1 million yen. It should be noted that these rules for digitization can be arbitrarily set.

For the preference characteristic matrix V [j, n], for example, the user profile element (user profile characteristic matrix P [j, m] before the item content element (preference characteristic matrix V0 [j, g [k, l]]) is used. ]). That is, the preference characteristic matrix V [j, n] is as follows.
(P [j, 1], P [j, 2], ..., P [j, M],
V0 [j, [1,1]], V0 [j, [1,2]], ..., V0 [j, [1, L1]], V0 [j, [2,1]], V0 [j, [2,2]], ..., V0 [j, [2, L2]], ...,
V0 [j, [K, 1]], V0 [j, [K, 2]], ..., V0 [j, [K, LK])

  Here, it is assumed that the number of characteristics of the user profile is M, the number of item groups of the item content characteristics is K, and the number of items of each group is L1, L2,. A matrix having the preference characteristic vector for each user as a row element is a preference characteristic matrix V [j, n] (j = 1,..., J, n = 1,..., N = M + L1 +... + LK). The analysis unit 10 outputs the created preference characteristic matrix V [j, n] as the user behavior analysis table 15.

  FIG. 3 is an explanatory diagram illustrating creation of a user preference characteristic vector. The data table T10 shown in FIG. 3 is obtained by multiplying the item characteristic matrix of each item (Item1, Item2 ... ItemI) by the viewing time ratio vector for one user. As illustrated in FIG. 3, the analysis unit 10 multiplies the item characteristic matrix by the user viewing time ratio vector and aggregates the columns for each column to obtain a vector (bottom row) indicating the user preference.

  FIG. 4 is an explanatory diagram for explaining the user behavior analysis table 15. As shown in FIG. 4, the user behavior analysis table 15 is obtained by adding a user profile (characteristic 1 to characteristic M) to a vector (the lowest level in FIG. 3) indicating the user's preference obtained for each user. Note that the user profile is actually digitized.

  Note that the user preference characteristic vector to which the user characteristic is added in S4 has different numerical values in the elements of the user profile and the user's preference with respect to the items, and is not uniform. When clustering is performed based on such user preference characteristic vectors, clusters may be determined with characteristics having large numerical values. Therefore, the analysis unit 10 normalizes the user preference characteristic vector in S4 and outputs it as the user behavior analysis table 15.

  Specifically, the analysis unit 10 assumes that the standardized user preference characteristic matrix is NV [j, n] (j = 1,..., J, n = 1,..., N = M + L1 +... + LK). It normalizes using the following formula | equation (2).

  In addition, when creating a list of recommended items, it may be artificially adjusted which characteristic group is important when clustering users. In order to perform such adjustment, a weight coefficient vector may be introduced for each group of characteristics.

Specifically, the following weight coefficient vector is introduced for the user profile or item content. For example, if wc [k = 4] = 0, clustering is performed without considering the fourth characteristic group.
User profile weight coefficient: wu [m], m = 1,..., M, (0 ≦ wu [m] ≦ 1)
Item content group weight coefficient: wc [k], k = 1,..., K, (0 ≦ wc [k] ≦ 1)

  Next, the clustering unit 20 refers to the user behavior analysis table 15 created by the analysis unit 10 through the processing up to S4, and performs clustering processing to divide users into user groups (clusters) (S5).

Specifically, the clustering unit 20 performs clustering of row vectors constituting the user preference characteristic matrix NV [j, n] in the user behavior analysis table 15 in user clustering. In order to perform clustering, first, a user preference characteristic matrix WV [j, n] multiplied by the above-described weighting factor is calculated. Note that the element value of each row of the weighted user preference characteristic matrix WV [j, n] is calculated as follows.
User profile part:
WV [j, n] = wu [n] × NV [j, n], n = 1,.

Item content part:
Group 1: WV [j, n] = wc [1] × NV [j, n], n = M + 1,..., M + L1
Group 2: WV [j, n] = wc [2] × NV [j, n], n = M + L1 + 1,..., M + L1 + L2
:
Group K: WV [j, n] = wc [K] × NV [j, n], n = N−LK,.

That is, the row vector element is as follows.
(Wu [1] NV [j, 1], wu [2] NV [j, 2], ..., wu [M] NV [j, M],
wc [1] NV [j, M + 1], wc [1] NV [j, M + 2], ..., wc [1] NV [j, M + L1],
wc [2] NV [j, M + L1 + 1], wc [2] NV [j, M + L1 + 2], ..., wc [2] NV [j, M + L1 + L2],
:
wc [K] NV [j, N-LK], wc [K] NV [j, N-LK + 1], ..., wc [K] NV [j, N])

  The clustering unit 20 performs clustering using the weighted user preference characteristic matrix WV [j, n]. There are many models of clustering methods. In the present embodiment, since the number of users (j) is 1 million or more and may reach 10 million orders in some cases, the hierarchical clustering method in which the calculation amount is O (j ^ 2) or more is not realistic. . Therefore, a non-hierarchical division optimization clustering method that can perform clustering with a calculation amount equal to or less than O (number of clusters × number of users (j)) is adopted.

In the clustering calculation, since the distance of a multidimensional vector that is an index of similarity is used, a distance calculation method is first defined. When a row vector of the j-th row composed of row elements of the weighted user preference characteristic matrix WV [j, n] is written as wv _j , it can be expressed as the following equation (3).

When the distance between the row vector wv _j and the row vector wv _i is written as d _ji (wv _j , wv _i ), the squared Euclidean distance and the Manhattan distance are defined as the following Expression (4). Note that the user may determine whether to adopt the squared Euclidean distance or the Manhattan distance, based on whether the clustering result is good or bad.

  For example, the k-means method is adopted as the division optimization clustering method. The k-means method minimizes the objective function of the following equation (5).

However, the meaning of wv _j εc _k means taking the sum of users belonging to the kth cluster. Further, it is assumed that the number of clusters is K, and _ck is the k-th cluster center coordinate vector.

The actual cluster division processing is as follows (1) to (4).
(1) K central coordinates _ck are randomly generated.
(2) If it is not at the time of the first calculation, the center coordinates are calculated by the following equation (6).

(3) The user vector wv _j is assigned to the cluster k that becomes min _k d _jk (wv _j , c _k ).
(4) If there is no change from the previously calculated cluster member, the calculation is terminated. If there is a change, the process returns to (2).

  Since the k-means method randomly determines the initial center point, it does not guarantee that a global minimum value is obtained. Therefore, in practice, clustering is performed using several sets of initial center points, and cluster division with the smallest objective function is adopted.

When the calculation time becomes a problem in practice, about 10% of users are randomly sampled, the k-means method is performed, and the center point _ck is fixed, and then all users are clustered. In this case, in the clustering for all the users, the cluster allocation process is performed only once, so that the calculation time can be shortened.

  In the description so far, it is assumed that the number of clusters (K) is given in advance by the creator who creates the list of items. However, in practice, it may be difficult to give the number of clusters in advance. However, the number of clusters may be set if it can be set based on knowledge in a field such as video on demand. For example, since there are hundreds of combinations of major classifications and middle classifications for genres such as movies and video programs, such a value is set. On the other hand, in a field where the number of clusters cannot be determined based on knowledge, some methods are used to estimate the optimal number of clusters.

  The method for estimating the number of clusters is described in “Tsunenori Ishioka (2006): Proposal for improving the x-means method—sequential iteration of k-means method and re-merging of clusters—,“ Computer Statistics ”, 18 (1) There are many models such as “p.3-13”. Here, an integrated classification likelihood (ICL) standard method and an elbow-point method are used, but there is no limitation on the use of other methods. For example, with respect to the ICL standard method, “Biernacki, C., et.al (1998): Assessing a Mixture for clustering the Integrated Classification Derivatives, Rapport, 21”. The elbow-point method is described in “Salvador, S. and Chan, P. (2004): Determining the Number of Clusters in Segmental on the Ill. Of the Institute for Education.” Intelligence, pp. 576-584 ”is known.

In the ICL standard method, the number of clusters with the maximum ICL value is adopted as the optimum number of clusters. The calculation of the ICL value uses the following approximate expression (7). Here, K is the number of clusters, J is the number of users, J _K is the number of users belonging to the cluster K, and N is the dimension of the user preference characteristic vector.

  The elbow-point method is called a knee-point method or an L-shaped method. When the sum SSW (K) of intra-cluster distances is a function of the number of clusters, the transition point from a rapidly decreasing region to a region that hardly decreases. This is a method of obtaining the optimum number of clusters by the number of clusters. Since SSW (K) is not a continuous function of the number of clusters, a secondary differential coefficient is obtained by difference approximation, and the number of clusters whose value is the maximum negative is determined as the optimum number of clusters. That is, K that minimizes the following equation (8) is obtained. Here, since only the relative size of SD is seen, the divisor 2 which is a common factor of the difference approximation formula is dropped.

  FIG. 5A is a graph illustrating the relationship between the number of clusters and SD (K). In the example of FIG. 5A, the minimum k = 6 is set as the optimum number of clusters. However, the relationship between the number of clusters and SSW (K) may be confirmed, and the user may determine whether the optimum number of clusters is correct.

  FIG. 5B is a graph illustrating the relationship between the number of clusters and SSW (K). As shown in FIG. 5B, a graph showing the relationship between the number of clusters and SSW (K) may be displayed on the monitor 103 (see FIG. 9) or the like, and the user may confirm it visually. If it is determined from the graph indicating the relationship between the number of clusters and SSW (K) that the data is not correct, the user inputs the optimum change point from the input device 102 (see FIG. 9).

  Further, the clustering unit 20 may use a dimensional reduction method in clustering in order to suppress the occurrence of a so-called “curse of dimension” that occurs in clustering of high-dimensional data.

For example, the clustering unit 20 uses any one of the following three methods i to iii as the dimension reduction method.
i. The first method is a method of collecting user characteristics that are included in the user preference vector for each predetermined group that focuses on the type of characteristics, etc., and that is reduced by removing groups that are not valid using preset clustering. This is a method for obtaining a vector.
ii. In the second method, each characteristic included in the user preference vector is clustered after being reduced for each predetermined group focusing on the type of characteristic, and the user preference characteristics before reduction are clustered in each cluster. This is a method of clustering users based on vectors (details will be described later). In this method, clustering is performed using all group characteristics. However, since a small difference in item content characteristics cannot be reflected, two-stage clustering is executed.
iii. The third method is a vector sumX [g [k, l]] obtained by summing each column in the item characteristic matrix X [i, g [k, l]] (each element value is the number of appearances of each characteristic item). This is a method of excluding items having a relatively small number of occurrences in each group.

  Here, the second method (ii) described above will be described in detail. In the second method (ii), the clustering unit 20 collectively reduces each characteristic included in the user preference characteristic vector for each predetermined group focusing on the characteristic type and the like. Next, the clustering unit 20 clusters users based on the reduced user preference characteristic vector (first clustering). Next, the clustering unit 20 clusters users based on the user preference characteristic vector before contraction in each cluster after the first clustering (second clustering). As described above, by performing clustering for each multi-level hierarchy, the item recommendation device 1 can suppress the occurrence of a so-called “dimensional curse” that occurs in clustering of high-dimensional data.

FIG. 6 is a flowchart illustrating an example of the clustering process. As shown in FIG. 6, when the process is started, the clustering unit 20 collects the elements of the user preference characteristic vector wv _j for each characteristic group (S10).

Here, the calculation method of the vector element put together for every characteristic group about the user preference characteristic vector wv _j is demonstrated. This calculation method, taking the sum of the element value of each characteristic group of the user preference feature vector wv _j, is to convert the Dimension Reduction vector rv _j as shown in formula (9).

  Note that (NV [j, 1], P [j, 2],..., NV [j, M]) in the user profile characteristic portion is one group 1 characteristic, and therefore, it is not necessary to take the sum of elements. Further, since the number of characteristic items differs for each group, each group element value is divided by (L1, L2,..., LK) for standardization.

However, if this dimension reduction vector rv _j is simply used as it is for clustering, a problem may occur. For example, since there is one director per item in the director group, the vector elements collected for each characteristic group are not effective for clustering. In the case of a performer group, clustering may be performed depending on whether there are many performers. The same applies when the item is not a video program but a product. For example, assume that raw materials of processed food are grouped and each raw material is a characteristic item. Here, taking the sum within a group means how many ingredients a certain processed food uses, and does not reflect which ingredients the purchaser likes. In this way, information about user preferences may be lost due to dimension reduction.

  Therefore, in this embodiment, clustering is performed by introducing user evaluation weights for each group. This evaluation weight reflects which group's characteristics the user tends to select items (video programs). Specifically, the clustering unit 20 performs weight setting for each characteristic group based on the user setting from the input device 102 (see FIG. 9) (S11). By this weight setting, even if the dimension is reduced, it can be classified into users having similar selection criteria.

  The clustering unit 20 introduces user evaluation weights wg [k], k = 1,..., K, (0 ≦ wg [k] ≦ 1) for each group based on the user settings. Further, the clustering unit 20 sets the mixing ratio to δ (0 ≦ δ ≦ 1), creates a vector shown in the following equation (10), and performs clustering (first stage) (S12). The mixing ratio δ is set by the user in S11.

Next, the clustering unit 20 performs clustering on the users in each cluster obtained by the first-stage clustering using all the characteristic items (user preference characteristic vector wv _j ) before reduction (second stage). (S13).

  Returning to FIG. 2, for each cluster, the clustering unit 20 aggregates the viewing histories of users belonging to the cluster and extracts video programs with a high selection frequency as recommended video programs. The clustering unit 20 creates the basic recommendation list table 16 by arranging the video programs extracted for each cluster in descending order of selection frequency (S6).

  FIG. 7 is an explanatory diagram for explaining the creation of the basic recommendation list table 16. As illustrated in FIG. 7, the clustering unit 20 generates a data table T11 indicating the cluster to which the user belongs for each user by clustering. Next, based on the data table T11, the clustering unit 20 aggregates viewing histories of users belonging to the cluster for each cluster and extracts video programs with a high selection frequency. For example, in the illustrated example, for the cluster “1”, the viewing history of “user5, user7” is aggregated to extract the video program “11, 25, 98, 4, 62” having a high selection frequency, and the selection frequency of The recommendation order is given in descending order. In this way, the clustering unit 20 obtains a list of recommendation ranks for all clusters and outputs it as the basic recommendation list table 16.

  Next, the output unit 30 outputs the user recommendation list table 17 for each user based on the basic recommendation list table 16 and the viewing history data 13 (S7). FIG. 8 is an explanatory diagram for explaining the creation of the user recommendation list table 17.

  As illustrated in FIG. 8, the output unit 30 refers to the user's viewing history data 13, and the video program that the user has already viewed (selected) from the basic recommendation list table 16 extracted from the cluster to which the user belongs. exclude. Then, the output unit 30 outputs a list of video programs that the user has not viewed (unselected) as the user recommendation list table 17.

  For example, in the example of FIG. 8, the user “1” belongs to the cluster “4”, and the program “6, 12, 25, 62, 99” has been viewed. For the cluster “4”, “6, 88, 62, 13, 78” is tabulated as a recommendation list. Therefore, for the user of “1”, “88, 13, 78” which has not been viewed in the totaled recommendation list is output. Thereby, the item recommendation device 1 can recommend an unviewed video program.

  The output unit 30 may calculate and output an evaluation value for an item recommended for each user in the user recommendation list table 17. Specifically, the output unit 30 calculates a precision, a recall, and an F-measure as evaluation values of recommended items. Each definition is as the following formula | equation (11).

  For example, the calculation of the evaluation value is performed for all users and users for each cluster using the top N items of the user recommendation list table 17 of the users. The value of “N” for the top N may be arbitrarily set by the user, and for example, “10” is set as an implicit value.

  The viewing history data 13 used for calculating the evaluation value is different before and after the creation of the user recommendation list table 17. Before the user recommendation list table 17 is created, the viewing history data 13 may be divided into appropriate evaluation periods for evaluation. However, the division ratio is arbitrarily set by the user. Since the new viewing history data 13 can be acquired after the user recommendation list table 17 is created, the evaluation is performed when the viewing history data 13 for a certain period can be accumulated. However, the period required for this accumulation is arbitrarily set by the user.

  By checking the output evaluation value, the user can update various settings from the input device 102 (see FIG. 9) and try to improve the process. For example, when the user determines that the evaluation value is small and the recommendation effect is bad, the user performs adjustment of the group characteristic weight or setting for adding a new characteristic item from the input device 102 (see FIG. 9). You can try to improve.

  Various processing functions performed by the item recommendation device 1 may be executed entirely or arbitrarily on a CPU (or a microcomputer such as an MPU or MCU (Micro Controller Unit)). In addition, various processing functions may be executed in whole or in any part on a program that is analyzed and executed by a CPU (or a microcomputer such as an MPU or MCU) or hardware based on wired logic. Needless to say, it is good. Further, various processing functions performed in the item recommendation device 1 may be executed by a plurality of computers in cooperation with each other by cloud computing.

  By the way, the various processes described in the above embodiments can be realized by executing a program prepared in advance by a computer. Therefore, in the following, an example of a computer (hardware) that executes a program having the same function as the above embodiment will be described. FIG. 9 is a block diagram illustrating an example of a hardware configuration of the item recommendation device 1 according to the embodiment.

  As illustrated in FIG. 9, the item recommendation device 1 includes a CPU 101 that executes various arithmetic processes, an input device 102 that receives data input, a monitor 103, and a speaker 104. In addition, the item recommendation device 1 includes a medium reading device 105 that reads a program and the like from a storage medium, an interface device 106 for connection to various devices, and a communication device 107 for communication connection with an external device by wire or wireless. Have. The item recommendation device 1 also includes a RAM 108 that temporarily stores various types of information and a hard disk device 109. Each unit (101 to 109) in the item recommendation device 1 is connected to the bus 110.

  The hard disk device 109 stores a program 111 for executing various processes in the analysis unit 10, the clustering unit 20, and the output unit 30 described in the above embodiment. The hard disk device 109 stores various data 112 (the item keyword table 12, the viewing history data 13 and the like) referred to by the program 111. For example, the input device 102 receives input of operation information from an operator of the item recommendation device 1. The monitor 103 displays various screens operated by the operator, for example. The interface device 106 is connected to, for example, a printing device. The communication device 107 is connected to a communication network such as a LAN (Local Area Network), and exchanges various types of information with an external device via the communication network.

  The CPU 101 reads out the program 111 stored in the hard disk device 109, develops it in the RAM 108, and executes it to perform various processes. Note that the program 111 may not be stored in the hard disk device 109. For example, the program 111 stored in a storage medium readable by the item recommendation device 1 may be read and executed. The storage medium readable by the item recommendation device 1 corresponds to, for example, a portable recording medium such as a CD-ROM, a DVD disk, a USB (Universal Serial Bus) memory, a semiconductor memory such as a flash memory, a hard disk drive, and the like. Alternatively, the program 111 may be stored in a device connected to a public line, the Internet, a LAN, or the like, and the item recommendation device 1 may read and execute the program 111 therefrom.

1 ... Item recommendation device 10 ... Analysis unit 11 ... Item DB
12 ... Item keyword table 13 ... Viewing history data 14, 18 ... User profile 15 ... User behavior analysis table 16 ... Basic recommendation list table 17 ... User recommendation list table 20 ... Clustering unit 30 ... Output unit 101 ... CPU
102 ... Input device 103 ... Monitor 104 ... Speaker 105 ... Media reader 106 ... Interface device 107 ... Communication device 108 ... RAM
109 ... Hard disk device 110 ... Bus 111 ... Program 112 ... Various data T10, T11 ... Data table

Claims (8)

For each user, the item selected by the user in the item characteristic matrix contracted using at least one group information out of the item characteristic matrix having each characteristic in the item for each item and group information for grouping the characteristics. To obtain a vector indicating the user's preference for each characteristic of the item by multiplying the history information, to create a user preference characteristic vector by adding the user profile to the vector,
Clustering users based on user preference characteristic vectors created for each user,
An item recommendation program that causes a computer to execute a process of collecting, for each cluster, history information of users belonging to the cluster and extracting items with high selection frequency as recommended items. The clustering process includes a first clustering that reduces the characteristics included in the user preference characteristic vector collectively for each predetermined group, and clusters users based on the reduced user preference characteristic vector; 2. The item recommendation program according to claim 1, wherein in each cluster after the first clustering, second clustering is performed for clustering users based on the user preference characteristic vector before contraction. The clustering process is performed by reducing each characteristic included in the user preference characteristic vector by excluding a predetermined group, and clustering users based on the reduced user preference characteristic vector; and 2. The item recommendation program according to claim 1, wherein in each cluster after the first clustering, second clustering is performed for clustering users based on the user preference characteristic vector before contraction. The item recommendation program according to claim 2 or 3, wherein in the clustering process, a weight is set for each group to be reduced in the first clustering. 5. The process according to claim 1, wherein the creating process normalizes each element indicating a user preference and a profile of the user included in the user preference characteristic vector. 5. Item recommendation program given in one paragraph. The computer according to any one of claims 1 to 5, wherein the computer further executes a process of outputting an unselected item by the user based on the history information from recommended items extracted from a cluster to which the user belongs. Item recommendation program given in any 1 paragraph. For each user, the item selected by the user in the item characteristic matrix contracted using at least one group information out of the item characteristic matrix having each characteristic in the item for each item and group information for grouping the characteristics. To obtain a vector indicating the user's preference for each characteristic of the item by multiplying the history information, to create a user preference characteristic vector by adding the user profile to the vector,
Clustering users based on user preference characteristic vectors created for each user,
An item recommendation method characterized in that, for each cluster, a computer executes a process of aggregating history information of users belonging to the cluster and extracting an item with a high selection frequency as an item to be recommended. For each user, the item selected by the user in the item characteristic matrix contracted using at least one group information out of the item characteristic matrix having each characteristic in the item for each item and group information for grouping the characteristics. A vector indicating user preferences for each characteristic of the item by multiplying the history information, and an analysis unit for creating a user preference characteristic vector by adding the user profile to the vector;
A clustering unit for clustering users based on the user preference characteristic vector created for each user, and for each cluster, collecting history information of users belonging to the cluster and extracting items with a high selection frequency as recommended items; An item recommendation device comprising:

Images