JP4307220B2 - Content recommendation target user selection apparatus and method, program, and content recommendation system - Google Patents

Description

  The present invention relates to a technique for automatically selecting a user (recommended user) to be recommended when recommending a certain content to a user, and in particular, efficiently recommending a recommended user of a new content that has not been used. The present invention relates to a technique suitable for selection.

  Conventionally, as a technique for automatically selecting a recommended user to be recommended when recommending a certain content to a user on the Internet or the like, for example, the technique described in Non-Patent Document 1 (generally content- based filtering).

  This technology recommends content when it is similar to the preference information of a certain user. In this technology, based on the results of questionnaires to the recommended user and the past use of the content, Preference information is generated, and it is determined whether or not the recommendation target content is recommended to the recommended user based on the degree of similarity between the description text or the description keyword of the recommendation target content and the preference information of the recommended user.

  Further, for example, in the technique described in Non-Patent Document 2 (generally called collaborative filtering (collaborative filtering method)), the usage tendency of a plurality of users is similar to that of the recommended user. Other users who are present are selected as collaborative users, and the collaborative users have a usage record and the recommended users have no usage record.

  In Patent Document 1 (“Recommendation Engine, Recommendation Method, Recommendation Program”), a target user's collaborative user is selected, and content that the selected collaborative user uses but the target user does not use is recommended to the target user. In this case, a technique for displaying a usage ranking when recommendation is not possible, that is, a recommendation (recommendation) technique using the technique of Non-Patent Document 2 is described.

  Patent Document 2 (“Information Recommendation Server Device”) selects a collaborative user whose profile is similar to that of the recommended user, and determines a recommendation based on the similarity between the profile of the selected collaborative user and the recommended content. If the user has a plurality of profiles and uses a profile that is not similar to the profile of the collaborative user, and the profile is similar to the recommendation target content, the recommendation target content is recommended to the recommended user. The technology to do is described.

  The feature of the technique described in Patent Document 2 is that content and user profiles can be described by vectors, so even content that has not been used by a collaborative user is recommended to a recommended user. Is possible. In this technique, the techniques of both Non-Patent Document 1 and Non-Patent Document 2 are used.

  However, in the case of the technique described in Non-Patent Document 1 (“content based filtering”), there is a first problem that only content similar to a questionnaire result or usage record answered by the recommended user tends to be recommended. .

  In the case of the technique described in Non-Patent Document 2 (“collaborative filtering”), other users who have similar usage trends to the recommended user are selected as cooperative users using the usage histories of a plurality of users. Since this is a technique for recommending content that has been used by a collaborative user and not used by a recommended user, there is a second problem in that content that has not been used by a collaborative user cannot be recommended. Furthermore, there is a third problem in that nobody has a track record of use, and new content cannot be recommended (for example, when new content is recommended to users who are suitable all at once).

  In addition, when selecting a recommendation target user of a certain recommended content from among many users, the processing amount tends to increase in proportion to the number of users in both the non-patent document 1 and the non-patent document 2. There is a fourth problem.

  As a result, the technique described in Patent Document 1 has the second, third, and fourth problems, and the technique described in Patent Document 2 has the fourth problem.

JP 2002-334257 A JP 2002-215665 A Marko Balabonavic and Yoav Shoham. "A Content-Based, Collaborative Recommendation", Communication of ACM, Vol.40, No.3, PP.66-72 (1997) P. Resnick, N. lacovou, M.M. Suchak, P. Bergstrom, and J.M. Riedl. "GroupLens: Open Architecture for Collaborative Filtering of Netnews", Conference on Computer Supported Cooperative Work, PP. 175-186 (1994)

  The problems to be solved are that, in the conventional technology, (1) only content similar to the questionnaire results and usage results answered by the recommended users tends to be recommended, and (2) the cooperative users also have usage results. No content can be recommended, (3) Nobody has a track record of use, new content cannot be recommended, and (4) When selecting recommended users of recommended content from many users, the amount of processing is large. This is a point that tends to increase in proportion to the number of users.

  In order to achieve the above object, in the present invention, content recommendation target users are selected using the three functions of score generation means (score generation block 2), learning means (learning block 3) and selection means (selection block 4). To do.

  That is, as shown in FIG. 1, first, in the score generation block 2, the operation information input means 2a is used to input operation information 10 (information such as user operation history and questionnaire results), and the content usage log storage unit 2b. The log information is sequentially stored as log information, and the format conversion means 2c performs weighting and normalization of the content usage history, data format conversion, and the like for each log keyword. On the other hand, score data between the user and the keyword in which the score is recorded is generated and stored in the storage device by the score storing unit 2d between the user and the keyword.

  Next, in the learning block 3, the keyword relation group extracting means 3a reads the user / keyword score data stored in the user / keyword score storage unit 2d, and groups and extracts the keywords from which the relation is obtained, The result is stored in the storage device by the keyword relation group extraction result storage unit 3b. Furthermore, based on the keyword-related group extraction result stored in the keyword-related group extraction result storage unit 3b and the user-keyword score data stored in the user-keyword score storage unit 2d by the group-specific user score calculation means 3c, The score of each user is calculated by performing an appropriate calculation for each keyword-related group, and the result is stored in the storage device by the group-specific user-specific score storage unit 3d.

  In the selection block 4, a set of keywords, which are metadata of recommended content, is input by the content keyword information input unit 4a, and the keyword-related group extraction result storage unit 3b and the group-specific group are calculated by the user-specific score calculation unit 4b. The user-specific score is stored by referring to the user-specific score storage unit 3d and performing appropriate calculations and processing. Thereafter, the recommendation target user information output means 4c selects one or a plurality of users having scores that match the selection condition (4d) as the recommendation target user 12 (recommended user).

  According to the present invention, a group (keyword relationship) having a relationship between keywords based on “user / keyword score data” generated based on content usage history (information such as user operation history and questionnaire results). Group) is extracted, and recommended users of content are selected based on the score of each user in the “keyword related group”. In other words, content recommendation is performed using a “keyword-related group” that reflects the trends of other users, so the first problem (only content similar to the questionnaire results and usage results answered by the recommended user) Is likely to be recommended) is resolved.

  Moreover, the recommended user of the content is selected using a keyword-related group reflecting the trends of other users, and the cooperative user is not selected. Therefore, the second problem that occurs to make recommendations based on the history of the selected collaborative user (content that does not have a collaborative user's usage record cannot be recommended) and the third problem (nobody has a use record, new It cannot be recommended for content).

  In the present invention, in the learning block (3), the scores of all users are calculated and held in advance for each keyword relation group, and it is not necessary to evaluate each user. That is, the calculation performed when selecting a target user (recommended user) who recommends content is a similarity calculation process with a predetermined number of keyword-related groups. Therefore, as in the technique described in Non-Patent Document 1 (“content based filtering”), the calculation of similarity with all users is performed, or the technique described in Non-Patent Document 2 (“collaborative filtering”). When the number of users is larger than the predetermined number of keyword-related groups as compared to the case where similar users are selected for all users, the amount of calculation processing in the present invention is smaller. Furthermore, in the present invention, the amount of processing does not increase in proportion to the number of users. As a result, the fourth problem (when a recommended target user of a recommended content is selected from many users, the processing amount tends to increase in proportion to the number of users) is solved.

  The best mode for carrying out the present invention will be described below with reference to the drawings.

  FIG. 1 is a block diagram showing a configuration example of a content recommendation target user selection device according to the present invention, and FIG. 2 shows a configuration example of a content recommendation system incorporating a content recommendation target user selection device according to the present invention. FIG. 3 is a block diagram showing the detailed configuration of the content recommendation system in FIG. 2, FIG. 4 is an explanatory diagram showing a specific example of operation information input by the operation information input means of the score generation block in FIG. FIG. 5 is an explanatory diagram showing a specific example of the content usage log stored in the content usage log storage unit of the score generation block in FIG. 1, and FIG. 6 is a user output by the format conversion means of the score generation block in FIG. An explanatory diagram showing a specific example of the score between keywords, FIG. 7 is a flowchart showing an example of processing operation of the learning block in FIG. FIG. 8 is an explanatory view showing a specific example of the keyword relation group generation result extracted by the keyword relation group extraction means of the learning block in FIG. 1 and stored in the keyword relation group extraction result storage unit, and FIG. FIG. 10 is an explanatory diagram showing a specific example of the group-specific user score generation result calculated by the group-specific user score calculation means of the learning block and stored in the group-specific user score storage unit, FIG. FIG. 11 is an explanatory diagram showing a specific example of recommendation target content information input to the content distribution control device of the content recommendation system in FIG. 3, and FIG. 12 is a content recommendation target user according to the present invention. Explanatory drawing which shows the specific example of the content keyword information input into a selection apparatus, 13 is a flow chart showing another processing operation example of the select block in Figure 1.

  First, the content recommendation system will be described with reference to FIGS. As shown in FIG. 2, the content recommendation system of this example includes a content recommendation target user selection device 1, a content distribution control device 20, a content database (denoted as “content DB” in the figure) 21, a network 22, and terminals 23a to 23a. 23c.

  Each of the content recommendation target user selection device 1, the content distribution control device 20, and the terminals 23a to 23c has a computer configuration including a CPU (Central Processing Unit), a main memory, a display device, an input device, an external storage device, and the like. After installing a program or data recorded in a storage medium such as a CD-ROM via a drive device or the like into the external storage device, the program is read from the external storage device into the main memory and processed by the CPU, so that each processing unit Perform the function.

  In the content recommendation system of this example, the content distribution control device 20 transmits the content stored in the content database 21 to each of the terminals 23a to 23c. The content recommendation target user selection device 1 is a content distribution control device. 20 selects a target terminal (or recommendation target user) to transmit content information.

  In this example, the content recommendation target user selection device 1, the content distribution control device 20, and terminals 23 a to 23 c used by the recommended user are connected via a network 22. In this example, the content database 21 is connected to the content distribution control device 20.

  In this example, the user ID for identifying the user and the ID for identifying the terminals 23a to 23c are uniquely associated. If the recommended user is found, the terminals 23a to 23c used by the user are identified. Is uniquely determined, and the recommendation target content can be recommended to the terminals 23a to 23c.

  If the user ID for identifying the user and the ID for identifying the terminals 23a to 23c are not uniquely associated with each other, a user ID and an ID for identifying the terminals 23a to 23c used by the user are separately provided. It is sufficient to prepare a table in which is associated with each other and refer to it.

  3, details of the content recommendation system shown in FIG. 2 are shown and the operation thereof will be described. In the content recommendation system, the content distribution control device 20 acquires the recommendation target content information 24 from the content database 21, and the content as a set of keywords that are metadata of the recommendation target content information 24 for the content recommendation target user selection device 1. The keyword information 11 is transmitted, and the recommendation target user information 12 for identifying the recommended user suitable as the user who transmits the recommendation target content information 24 is acquired from the content recommendation target user selection device 1 correspondingly.

  Then, the content distribution control device 20 recommends, for example, the terminals 23a and 23c used by the recommendation target user (recommended user) specified by the recommendation target user information 12 acquired from the content recommendation target user selection device 1. The target content information 24 is transmitted.

  The content recommendation target user selection device 1 has processing function blocks of a score generation block 2, a learning block 3, and a selection block 4. In the score generation block 2, the operation information 10 (user operation history and content) is received from the terminal 2d. Information such as usage history and questionnaire results) is input as needed, and user / keyword score data used by the learning block 3 is generated and stored in the user / keyword score storage unit 2d.

  In the learning block 3, the user-keyword score data generated and stored in the score generation block 2 is read, and the keyword relation group in which related keywords are grouped, and the score for each user is recorded for each group. User-specific score data is generated as needed at appropriate times, and stored in the keyword-related group extraction result storage unit 3b and the group-specific user score storage unit 3d, respectively.

  In the selection block 4, the content keyword information 11 transmitted from the content distribution control device 20 is input by the content information input means 4a, and the keyword-related group extraction result storage unit 3b and the group-specific score storage unit in the learning block 3 With reference to the contents stored in 3d, the recommendation target user (recommended user) is selected by the appropriate calculation and processing by the recommendation target user information output means 4c, and the selection result is used as the recommendation target user information 12, and the content distribution control apparatus 20 Output for.

  Details of the content recommendation target user selection device 1 will be described below with reference to FIG. As shown in FIG. 1, the content recommendation target user selection device 1 has processing function blocks of a score generation block 2, a learning block 3, and a selection block 4. The score generation block 2 includes operation information input means 2a, It has a content usage log storage unit 2b, a format conversion unit 2c, and a user / keyword score storage unit 2d. The learning block 3 includes a keyword relationship group extraction unit 3a, a keyword relationship group extraction result storage unit 3b, and a user by group. There is a separate score calculation means 3c, a group-specific user score storage section 3d, and the selection block 4 is a content keyword information input means 4a, a user-specific score calculation means 4b, a recommendation target user information output means 4c, a selection condition storage. Part (described as “selection condition” in the figure) 4d.

  Details of each block of the content recommendation target user selection device 1 will be described. In the score generation block 2, the operation information input means 2a reads the operation information 10 shown in the specific example in FIG. 4, and the content use log storage unit 2b records the content use log information having the specific example in FIG. To do.

  Further, the format conversion means 2c refers to the content usage log information illustrated in FIG. 5 recorded in the content usage log storage unit 2b, performs weighting and normalization for content usage history, data format conversion, etc. Processing such as scoring is performed to generate user / keyword score data in which a score is recorded for each keyword that is content metadata for each user. In the user / keyword score storage unit 2d, FIG. The content is stored in the storage device with a specific example.

  For example, if there is a keyword whose operation content is “deleted” in the content usage log illustrated in FIG. 5, the format conversion unit 2 c calculates the score of the corresponding keyword of the corresponding user in the “user-keyword score” in FIG. 6. “−10”. Similarly, when there is a keyword whose operation content is “save icon” in the content usage log, the score of the corresponding keyword of the corresponding user of “user-keyword score” is “+10”.

  The format conversion means 2c can customize the score assignment method, conversion method, and the like according to the type and purpose of the operation information 10.

  Further, in the content usage log of FIG. 5, operation information 10 acquired in different usage scenes may be mixed. For example, information on “intermediate viewing has been interrupted” in the viewing history of a TV program, information on “icon deleted” in the operation history of a content icon on the terminal, and “10-minute icon in book usage history data” Information may be mixed.

  In this case, by defining a score for each usage scene in the format conversion means 2c in advance, it is possible to generate a “user-keyword score” even for different usage scenes. For example, “icon save = + 10 points”, “icon deletion = −10 points”, “icon disappearance = −5 points”, “a minute icon ownership = + log 10 (a) points”, “content activation = + 20 points”, “ It is defined that “intermediate content interruption = −5 points” and “content complete use = + 5 points”.

  Next, in the learning block 3, the keyword-related group extracting means 3a reads the user-keyword score data (see FIG. 6) stored in the user-keyword score storage unit 2d, and the relevant keywords should be determined appropriately. The keywords extracted by performing the operation are grouped, the group ID (identifier) is given to the group, and the result is the keyword relationship group extraction result storage unit 3b with the contents illustrated in FIG. Store in.

  Furthermore, based on the keyword-related group extraction result stored in the keyword-related group extraction result storage unit 3b and the user-keyword score data stored in the user-keyword score storage unit 2d by the group-specific user score calculation means 3c, The score of each user is calculated by performing an appropriate calculation for each keyword-related group, and the result is stored in the group-by-user score storage unit 3d with the contents illustrated in FIG.

  In the selection block 4, the content keyword illustrated in FIG. 12 corresponding to the recommendation target content information 24 illustrated in FIG. 11 is received from the external device such as the content distribution control device 20 in FIGS. The set 11 is input, and the user-specific score calculation means 4b reads the information stored in each of the keyword-related group extraction result storage unit 3b and the group-specific user score storage unit 3d, and performs an appropriate calculation using each information. Processing is performed to calculate a score for each user.

  Thereafter, the recommendation target user information output means 4c selects one or a plurality of users (recommended users) whose scores calculated by the user-specific score calculation means 4b match the selection condition 4d, and the recommended target user information 12 is Output to an external device such as the content distribution control device 20 in FIGS.

  In the operation information illustrated in FIG. 4, the user with the user ID “user-c” uses an activation icon that activates the content “Sick Kitten” displayed on the terminal (23 d) used by the user. In this example, the “save” operation is performed at 15:55. Here, it is assumed that the keywords “K1” and “K34” are preliminarily assigned to the content “I will raise a kitten”.

  Such operation information (10) can be acquired from an operation on the screen of the terminal (23d). For example, icons for activating various contents are displayed on the screen of the terminal (23d), and the user can use the contents by pressing the icons as needed, delete the icons, If it can be saved, the operation for the icon can be acquired as the operation information (10).

  As the operation information (10), various data such as a viewing history, an operation history, an input history, an acquisition history, a usage history, and an action history can be used.

  In this example, in the operation information (10), a keyword that is metadata for explaining the content is assigned in advance, and the keyword is used, but a keyword for explaining the content is assigned. If a description of the content is given instead, a separate process is required to extract a keyword that is metadata of the content by performing morphological analysis on the description.

  The content usage log illustrated in FIG. 5 is an accumulation of excerpts of items necessary for generating score data between users and keywords that can be used by the learning block 3 from the operation information (10) input as needed. Each item includes “time”, “user ID”, “operation target keyword”, and “operation content”.

  The score data between the user and the keyword illustrated in FIG. 6 is stored in a format referred to in the learning block 3 by performing scoring processing from the content use log shown in FIG. Yes, the score data between the user and the keyword in this example is a two-dimensional array variable of the user and the keyword, and a score (numerical value) is recorded for each user and each keyword.

  Hereinafter, an example of processing operation in the learning block 3 based on the content usage log shown in FIG. 6 will be described with reference to FIG. In the example of FIG. 7, clustering processing is used for the operation for generating the keyword relationship group in the keyword relationship group extraction unit 3 a.

  As described above, when the clustering process is used for the operation for generating the keyword relation group, the keyword relation group extraction unit 3a first reads the user / keyword score data stored in the user / keyword score storage unit 2d ( In step S701, a clustering process is performed on the user-keyword score data to generate a cluster of keyword sets (step S702). Then, the generated cluster of the keyword set is transferred to and stored in the keyword relationship group extraction result storage unit 3b as a keyword relationship group (step S703).

  For example, as shown in FIG. 8, the generated keyword-related groups are “group A = {K1, K2, K8},“ group B = {K3, K9, K10, K12, K15} ”, and so on. Become.

  In this clustering process, clustering conditions such as the number of generated clusters and the presence / absence of data normalization can be described in a setting file or can be defined by hard coding. In addition to the generated keyword relation group, a keyword relation group registered in advance can be used.

  Next, in the learning block 3, the group-specific user score calculation means 3c calculates group-specific user score data that is a user-specific score for each keyword-related group. That is, the group-by-user score calculation means 3c reads the user / keyword score data from the user / keyword score storage unit 2d one record at a time (step S704).

  For example, when the keyword is “key”, the user is “u”, and the user-keyword score data is “SCORE (key, u)”, when one record of the data shown in FIG. 6 is read, “SCORE (K1, user) -a) = 2 "," SCORE (K1, user-b) =-5 "," SCORE (K1, user-c) = 2 "," SCORE (K1, user-d) = 1 "," SCORE ( K1, user-e) = 2 "and" SCORE (K1, user-f) = 0 ".

  Further, the group-specific user score calculation means 3c uses any one of the keyword-related groups generated in step S702 and stored in the keyword-related group extraction result storage unit 3b in step S703. And “gr”, which is the group ID of the keyword related group including the key, is acquired (step S705), and “by user by group” of the user u of the keyword related group ID “gr” The score SCORE (key, u) read in step S704 is added to the score “SC (gr, u) (“ SC (gr, u) = SC (gr, u) + SCORE (key, u) ”) (step S705).

  For example, the keyword relation group generated in step S702 and stored in the keyword relation group extraction result storage unit 3b in step S703 is the keyword relation group illustrated in FIG. 8, and in step S704, between the user and the keyword illustrated in FIG. When the first record in the score data is read, since the keyword K1 read in step S704 is included in the keyword relation group A in FIG. 8, the user of the group A item in the group-specific user score data shown in FIG. The score of each user of the keyword K1 is added to each score.

  In this example, the calculation of the group-specific user score SC (gr, u) is “SC (group A, user-a) = SC (group A, user-a) +2”, “SC (group A, user). -b) = SC (group A, user-b) -5 "," SC (group A, user-c) = SC (group A, user-c) +2 "," SC (group A, user-d) " = SC (group A, user-d) +1 "," SC (group A, user-e) = SC (group A, user-e) + 2 "," SC (group A, user-f) = SC (group A, user-f) +0 ".

  In the group-by-user score calculation means 3c, when all the records have been read from the user / keyword score storage unit 2d, the process of the next step S707 is performed. If not, the process from step S704 is performed again. Run repeatedly. In addition, as a branching condition at this time, not only the case where all records have been read from the user / keyword score storage unit 2d, but also a case where the preset condition is met.

  In step S707, the group-specific user score calculation means 3c passes the group-specific user score SC (gr, u) generated in step S705 to the group-specific user score storage unit 3d for storage.

  The keyword relation group data shown in FIG. 8 is data describing the keyword relation group extracted by the keyword relation group extraction means 3a. In the keyword-related group data of this example, for each keyword-related group, the group ID and the keyword included in each group are described.

  In addition to the keyword relationship group extracted by the keyword relationship group extraction unit 3a, a keyword relationship group defined in advance can be stored as the keyword relationship group data. For example, as illustrated in FIG. 8, in addition to the keyword related group extracted by the keyword related group extracting unit 3a, a keyword related group called “all keywords” in which all keywords are grouped can be registered in advance. .

  The group-specific user-specific score data shown in FIG. 9 is data in which the score of each user is stored for each keyword-related group generated by the keyword-related group extracting unit 3a. Here, the score data generated by the keyword-related group extracting unit 3a is used. In each of the keyword related groups “group A”, “group B”, “group C”, “group D” and “all keywords” including all keywords, the users “user-a”, “user-b”, The scores of “user-c”, “user-d”, “user-e”, and “user-f” are stored in association with each other.

  As described in the example of FIG. 8, “all keywords” is not a keyword relationship group extracted by the keyword relationship group extraction unit 3a, but a keyword relationship group defined separately. The group-specific user-specific score is empty when the learning block 3 is executed, and is generated by adding the score in the group-specific user-specific score calculation means 3a.

  Next, an example of processing operation in the selection block 4 based on the keyword-related group data shown in FIG. 8 generated in the learning block 3 and the group-specific user-specific score data shown in FIG. 9 will be described based on FIG.

  First, in the content keyword information input means 4a, content keyword information as a set of keywords that is metadata of content to be recommended is input (step S801). Next, the keyword-related group calculation unit 4b reads keyword-related group data from the keyword-related group extraction result storage unit 3b (step S802).

  Further, the score calculation unit 4b for each user calculates the similarity between the content keyword information input in step S801 and each group of the keyword-related group data read in step S802 using a predetermined calculation formula (step S803). . The similarity is calculated for each group gr of the keyword related group.

  As an example of the similarity calculation formula, “similarity α (gr) = (number of keywords included in both content keyword information and keyword relationship group gr) / ÷ (number of keywords included in keyword relationship group gr)” There is. In addition, since the content keyword information and the keyword relation group are each a set of keywords, the content keyword information and the keyword relation group may be considered as a keyword vector, and the similarity may be obtained by an inner product of the keyword vectors or a cosine calculation.

  When the above calculation formula is used, for example, when the content keyword information shown in FIG. 12 is input in step S801 and the keyword-related group data shown in FIG. 8 is read in step S802, the number of keywords belonging to group A is “3. (= K1, K2, K8) ”and the number of keywords included in both the content keyword information and group A in FIG. 12 is“ 0 ”, so“ similarity α (group A) = 0 ÷ 3 = 0 ”. Become.

  Similarly, the number of keywords belonging to group B is “5 (= K3, K9, K10, K12, K15)”, and the number of keywords included in both the content keyword information and group B is “2 (= K3, K15)”. Therefore, “similarity α (group B) = 2 ÷ 5 = 0.4”, the number of keywords belonging to group C is “2”, and the number of keywords included in both the content keyword information and group C is “1”. Therefore, “similarity α (group C) = 1 ÷ 2 = 0.5”, the number of keywords belonging to group D is “5”, and the number of keywords included in both the content keyword information and group D is “1”. , “Similarity α (group D) = 1 ÷ 5 = 0.2”, the number of keywords belonging to all keyword-related groups is “15”, content keyword information and all keywords Number of keywords included in both the engagement groups since "4", and "similarity alpha (total keyword) = 4 ÷ 15 = 0.267".

  After calculating the similarity in this way, the score calculation unit 4b for each user selects the group ID of the keyword-related group with the highest similarity (step S804). In the case of the above-described example, the group ID of the keyword-related group having the highest similarity is group C (similarity = 0.5).

  Then, the user-specific score calculation means 4b refers to the group-specific user score “SC (group gr, user u)” stored in the group-specific user-specific score storage unit 3d, and determines the keyword-related group selected in step S804. The score S (u) for each user is acquired.

For example, as in the above example, when the group C is selected in step S804, the read user-specific group score C from a group by subscriber score data in Figure 9, so that the user-specific scores S (u to obtain ) “S (user-a) = SC (group C, user-a) = − 6”, “S (user-b) = SC (group C, user-b) = + 7”, “S (user -c) = SC (group C, user-c) =-8 "," S (user-d) = SC (group C, user-d) = + 9 "," S (user-e) = SC (group C, user-e) = − 13 ”and“ S (user-f) = SC (group C, user-f) = − 10 ”.

  The recommendation target user information output unit 4c has a score that satisfies the selection condition set in advance in the selection condition storage unit 4d from the user-specific score S (u) calculated by the user-specific score calculation unit 4b in this way. A user is selected (step S806), and an identifier (user ID) of the selected user is output (step S807).

  For example, the user-specific scores acquired in step S805 are “S (user-a) = SC (group C, user-a) = − 6” and “S (user-b) = SC (group C, user- b) = + 7 ”,“ S (user-c) = SC (group C, user-c) = − 8 ”,“ S (user-d) = SC (group C, user-d) = + 9 ”,“ In the case of S (user-e) = SC (group C, user-e) = − 13 ”and“ S (user-f) = SC (group C, user-f) = − 10 ”, the selection target user When the setting condition is “user with score of 0 or more”, “user-b” and “user-d” that satisfy the selection condition are selected and output.

  The selection conditions set in the selection condition storage unit 4d may be hard-coded in advance. It may also be described in a configuration file. Furthermore, the selection condition may be described by a mathematical expression using values such as the number of users and the number of keywords included in the group as variables.

  The recommendation target content information shown in FIG. 11 describes information related to the content recommended to the user. In this example, the content name (“GoGo Tora Story Episode 3”) and the URL (link) for accessing the content are shown. The destination URL), the URL of the icon of the content (icon URL), and one or more keywords (K3, K4, K13, K15) expressing the contents of the content are registered. In addition, as a keyword of such recommendation target content information, it is possible to use a morphological analysis of the description of the content.

  The content keyword information shown in FIG. 12 is information obtained by extracting only the keyword from the recommendation target content information shown in FIG. 11, and includes the keywords “K3, K4, K13, K15”.

  The processing operation example of the selection block 4 in FIG. 1 is not limited to the processing operation shown in FIG. 10, and for example, the processing operation shown in FIG. In the processing operation example shown in FIG. 13, only the processing part of steps S804 and S805 in the processing flow of FIG. 10 is changed to step S1304 in FIG. 13, and steps S1301 to S1303 in FIG. The processes in steps S1305 and S1306 correspond to the processes in steps S801 to S803 and steps S806 and S807 in FIG. 10, respectively. Only the processing part of step S1304 that is different will be described below.

  Here, it is assumed that the keyword-related group is “gr”, the similarity between the content keyword information and the keyword-related group gr is “α (gr)”, and the group-specific user score is “SC (gr, u)”. 13, the user-specific score “S (u)” is calculated as “S (u) = Σgr {SC (gr, u) × α (gr)}. This method is used. This makes it possible to select users reflecting the characteristics of a plurality of keyword-related groups.

  As described above with reference to FIGS. 1 to 13, in this example, when a user (target user) to be recommended for content distributed via a network is selected by computer processing, first, score generation is performed. In block 2, based on the operation content of the user's content and the keyword (operation information and content usage log information shown in FIGS. 4 and 5) of the content, a score corresponding to the operation content for each user and for each keyword Is accumulated to generate a table (score information between users and keywords shown in FIG. 6).

  Next, in the learning block 3, the keywords are grouped based on the user / keyword score information shown in FIG. 6 (keyword related group information in FIG. 8), and the keyword related group information in FIG. 8 and the user in FIG. Based on the inter-keyword score information, the score of each keyword under the group is accumulated for each group and for each user to generate a table (score information for each user in FIG. 9).

  Then, in the selection block 4, the similarity between the keyword group that is the metadata of the content to be recommended and the keyword group of each group in the keyword-related group information in FIG. 8 is calculated. For example, the group of the most similar keyword group is determined. A user having a score that matches a predetermined condition such as, for example, the score being equal to or higher than a certain value among the scores of each user identified and associated with this group by the group-specific user score information shown in FIG. Selected users as target users.

  As described above, according to the configuration and processing of this example, user recommendation (recommendation) is performed using a collaborative method using a keyword relation group generated based on the usage history of all users. Even content that has not been used can be recommended to the recommended user. In addition, since the content is expressed as a keyword vector, it is possible to recommend new content that has not been used by all users. Furthermore, even when selecting a user suitable for recommendation from a large number of users, there is no need to select a collaborative user and only the process of selecting a group to which the content belongs is performed. The user selection process can be performed at high speed without increasing in relation to the number.

  In addition, this invention is not limited to the example demonstrated using FIGS. 1-13, In the range which does not deviate from the summary, various changes are possible. For example, in this example, a clustering processing technique is used as a technique for grouping related keywords in the keyword relation group extracting unit 3a. However, in addition to this clustering technique, grouping is performed by an association rule extraction technique for data mining. It is also possible to use a method for generating.

It is a block diagram which shows the structural example of the content recommendation object user selection apparatus concerning this invention. It is a block diagram which shows the structural example of the content recommendation system incorporating the content recommendation object user selection apparatus concerning this invention. It is a block diagram which shows the detailed structure of the content recommendation system in FIG. It is explanatory drawing which shows the specific example of the operation information input by the operation information input means of the score generation block in FIG. It is explanatory drawing which shows the specific example of the content utilization log stored in the content utilization log storage part of the score generation block in FIG. It is explanatory drawing which shows the specific example of the score between user and keywords output by the format conversion means of the score generation block in FIG. It is a flowchart which shows the processing operation example of the learning block in FIG. It is explanatory drawing which shows the specific example of the keyword relation group production | generation result extracted by the keyword relation group extraction means of the learning block in FIG. 1, and stored in the keyword relation group extraction result storage part. It is explanatory drawing which shows the specific example of the score production | generation result classified by user classified by group computed by the score calculation means classified by user classified by group of the learning block in FIG. 1 and stored in the score storage part classified by user classified by group. It is a flowchart which shows the processing operation example of the selection block in FIG. It is explanatory drawing which shows the specific example of the recommendation object content information input into the content delivery control apparatus of the content recommendation system in FIG. It is explanatory drawing which shows the specific example of the content keyword information input into the content recommendation object user selection apparatus concerning this invention. It is a flowchart which shows the other process operation example of the selection block in FIG.

Explanation of symbols

  1: content recommendation target user selection device, 2: score generation block, 2a: operation information input means, 2b: content use log storage section, 2c: format conversion means, 2d: score storage section between user and keyword, 3: learning block 3a: Keyword-related group extraction means, 3b: Keyword-related group extraction result storage section, 3c: User-specific score calculation means by group, 3d: User-specific score storage section by group, 4: Selection block, 4a: Content keyword information input Means 4b: user-specific score calculation means 4c: recommendation target user information output means 4d: selection condition storage unit 10: operation information 11: content keyword information 12: recommendation target user information 20: content distribution control device , 21: content database (“content DB”), 22: network Work, 23a~23d: terminal, 24: recommendation target content.

Claims (9)

A content recommendation target user selection device for selecting users who are content recommendation targets,
Each time a user uses a content, a keyword that is metadata of the content, a user operation content and user identification information for the content, and a content usage log in which each operation content is associated with each user and each keyword First means for generating and recording information in a storage device;
A score corresponding to each operation content in the content usage log information is obtained for all content usage log information recorded in the storage device, and the obtained score is accumulated for each user and each keyword, and the accumulated score is Second means for generating user-keyword score information associated with each user and each keyword and recording it in the storage device;
A third means for calculating the respective associations of the keywords to divide them into groups, generating keyword-related group information in which each keyword is associated with each group, and registering the information in a storage device;
Referring to the keyword-related group information and the user-keyword score information, for each group, the cumulative score of each user associated with each keyword in the same group is further accumulated, and the accumulated score A fourth means for generating score information for each group and for each group and for each user and storing them in the storage device;
One or more keywords that are metadata of content to be recommended are acquired, one or more groups including the acquired keywords are specified with reference to the keyword-related group information, and each user associated with the specified group A fifth means for obtaining a score for each user by referring to the group-specific score information for each user and, if there are a plurality of identified groups, for each user,
A content recommendation target user selection device, comprising: a sixth means for selecting a user whose calculated score satisfies a predetermined score condition as a user who is a recommendation target of the recommended content. A content recommendation target user selection device for selecting users who are content recommendation targets,
Each time a user uses a content, a keyword that is metadata of the content, a user operation content and user identification information for the content, and a content usage log in which each operation content is associated with each user and each keyword First means for generating and recording information in a storage device;
A score corresponding to each operation content in the content usage log information is obtained for all content usage log information recorded in the storage device, and the obtained score is accumulated for each user and each keyword, and the accumulated score is Second means for generating user-keyword score information associated with each user and each keyword and recording it in the storage device;
A third means for calculating the respective associations of the keywords to divide them into groups, generating keyword-related group information in which each keyword is associated with each group, and registering the information in a storage device;
Referring to the keyword-related group information and the user-keyword score information, for each group, the cumulative score of each user associated with each keyword in the same group is further accumulated, and the accumulated score A fourth means for generating score information for each group and for each group and for each user and storing them in the storage device;
A fifth means for acquiring a keyword set which is metadata of content to be recommended, and specifying a keyword set group most similar to the acquired keyword set with reference to the keyword-related group information;
A sixth user who refers to the group-specific user score information and selects a user whose cumulative score associated with the specified group satisfies a predetermined score condition as a user who is a recommendation target of the recommended content. And a content recommendation target user selection device. A content recommendation target user selection device for selecting users who are content recommendation targets,
Each time a user uses a content, a keyword that is metadata of the content, a user operation content and user identification information for the content, and a content usage log in which each operation content is associated with each user and each keyword First means for generating and recording information in a storage device;
A score corresponding to each operation content in the content usage log information is obtained for all content usage log information recorded in the storage device, and the obtained score is accumulated for each user and each keyword, and the accumulated score is Second means for generating user-keyword score information associated with each user and each keyword and recording it in the storage device;
A third means for calculating the respective associations of the keywords to divide them into groups, generating keyword-related group information in which each keyword is associated with each group, and registering the information in a storage device;
Referring to the keyword-related group information and the user-keyword score information, for each group, the cumulative score of each user associated with each keyword in the same group is further accumulated, and the accumulated score A fourth means for generating score information for each group and for each group and for each user and storing them in the storage device;
Obtain a keyword set that is metadata of content to be recommended, calculate the similarity between the acquired keyword set and the keyword set included in each group read from the keyword-related group information for each group,
Using the calculated similarity for each group, a weighting operation is performed on each score associated with each user for each group in the group-specific user score information, and the sum of the weighted scores for each user A fifth means for obtaining
A content recommendation target user selection device comprising: a sixth means for selecting a user whose total score satisfies a predetermined score condition as a user who is a recommendation target of the recommended content. The content recommendation target user selection device according to any one of claims 1 to 3, wherein the third means reads the user / keyword score information and uses the user / keyword score information. perform one of the keywords of the clustering operations or data mining association rules extraction, the content recommendation target user selection apparatus characterized by performing a grouping for each keyword. The content recommendation target user selection device according to any one of claims 1 to 4,
The fourth means is
Read each user included in the user-keyword score information and each group included in the keyword-related group information, and generate a table in which each is two-dimensionally arranged,
Read the score of each user for each keyword sequentially with reference to the user-keyword score information, and each time the read keyword group is identified with reference to the keyword-related group information,
A content recommendation target user selection device, wherein the score information for each group user is generated by adding the score of each read user in association with the specified group and each read user in the table. 6. The content recommendation target user selection device according to claim 1, wherein the first means associates a keyword, which is metadata of content used by the user, with the content in advance. obtained from granted keyword, even properly, the content recommendation target user selection apparatus characterized by obtaining a description that is previously assigned to the content by morphological analysis.   The program for functioning a computer as each means in the content recommendation object user selection apparatus in any one of Claims 1-6. A content recommendation system for recommending content to a user via a computer network,
The content recommendation target user selection device according to any one of claims 1 to 6, comprising:
A content recommendation system, wherein a user who recommends the content is selected by the content recommendation target user selection device, and the content is recommended for the selected user. A content recommendation target user selection method for an apparatus that performs selection of a user to be recommended for content distributed via a network by programmed computer processing,
As a process execution procedure of the programmed computer,
The content recommendation object user selection method characterized by including the process procedure which each means in the content recommendation object user selection apparatus in any one of Claims 1-6 performs .

Images