JP2008009729A - Content distribution method and device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a content distribution method and device which allows content distribution on which user's taste is sufficiently reflected and even user's potential taste is reflected. <P>SOLUTION: The content distribution device includes: a content database 101 with which keywords and content vectors are registered per content; a history acquisition part 102a for acquiring a content viewing history of the user; a keyword acquisition part 103a for acquiring keywords of contents which the user has viewed; a clustering part 103b for classifying contents which the user has viewed, to taste clusters on the basis of the acquired keywords; a taste cluster vector generation part 103d for generating a taste cluster vector of each taste cluster; a similarity calculation part 102b for calculating similarities between a content vector of each new content and respective taste cluster vectors; and a priority setting part 102c for setting priorities to respective new contents on the basis of the similarities. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a content distribution method and apparatus, and in particular, dynamically generates and updates a preference cluster based on a user's content viewing history, and performs priority distribution that reflects the user's preference based on the preference cluster. The present invention relates to a content distribution method and apparatus that enable the content distribution.

  Services that customize information according to the interests of users are being studied. In Patent Literature 1, distribution history information including identification information of news articles distributed to users and user profile information created based on user registration information received from each user prior to service provision are disclosed. A technology that enables news distribution customized for each user is disclosed.

In Patent Document 2, every time a viewer browses a web page, the words appearing on the page are extracted and classified into genres, and the forgetting concept obtained from the browsing time of each web page is determined as the frequency of appearance of words and genres. Apply it to obtain personal preference information, and based on this personal preference information, display a word that matches the viewer's preference on the Web page by color, or perform a keyword search with multiple words that match the viewer's preference A technique for executing in the background is disclosed.
JP 2005-242758 A JP 2002-073677 A

  The technique disclosed in Patent Document 1 is troublesome because the user himself / herself has to input in advance the news genre and keyword that he / she wants to browse. In addition, since keyword selection greatly affects search results in keyword search, if the keyword selection is inappropriate, desired search results may not be obtained.

  In the technology disclosed in Patent Document 2, preference information is generated based on words appearing on a web page. Since preference is an abstract concept in the first place, such abstract concepts are classified by words. If the preference information is generated based on the classification result, the preference of the user cannot be sufficiently reflected in the preference information, and the potential preference of the user cannot be reflected.

  An object of the present invention is to provide a content distribution method and apparatus that solves the above-described problems of the prior art and enables content distribution that sufficiently reflects the user's preference and also reflects the user's potential preference. There is.

  In order to achieve the above-described object, the present invention dynamically generates and updates a preference cluster based on a user's content viewing history, and sets a priority for content distributed to the user based on the preference cluster. The content distribution method includes the following procedure.

  (1) For each content, a procedure for generating a keyword and a content vector representative of the content and registering the content in the content database; a procedure for extracting a keyword of content viewed by the user from the content database; and A procedure for classifying each content into at least one of a plurality of preference clusters, and a procedure for generating a preference cluster vector representing the characteristics of each preference cluster based on the content classified into each preference cluster; For each new content to be distributed, a procedure for calculating the similarity between the content vector and each preference cluster vector, a procedure for setting a priority for each new content based on the similarity, Procedures to provide to users according to priority. The features.

  (2) The method further includes a step of integrating a plurality of preference clusters having similar classified contents into one preference cluster.

  (3) The method further includes a step of setting a weight value for each preference cluster. In the step of setting the priority for each new content, the weight value of the preference cluster is reflected in the similarity between the new content and the preference cluster. And a priority is set.

According to the present invention, the following effects are achieved.
(1) For each preference cluster, a preference cluster vector representing the feature is generated, and based on the comparison result between this preference cluster vector and the content vector of the new content, a new similarity with any preference cluster is high Since the content is preferentially provided as content that matches the user's preference, it is possible to preferentially distribute the content that sufficiently reflects the user's preference and also reflects the user's potential preference.
(2) Since preference clusters similar to each other are integrated into one preference cluster, it is possible to reduce the storage area and the processing time.
(3) Each preference cluster is weighted so that the similarity to the new content is calculated, so it is possible to set priorities that take importance into account

  DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, the best embodiment of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing a configuration of a main part of a content distribution apparatus to which the present invention is applied. Here, news that adaptively distributes online news as content to a user's portable terminal 2 via the Internet. The distribution device 1 will be described as an example.

  In the news distribution apparatus 1, in the news database 101, distributed news articles and newly acquired undelivered news articles are managed in a database.

  FIG. 2 is a diagram showing an example of information managed in the news database 101, in which news article text, article vector files, keywords, and the like are managed with article IDs unique to each news article. News article texts are classified into titles, headlines and detailed articles.

  FIG. 3 is a diagram showing an example of the article vector file. For each main term, the TF (Term Frequency: frequency of terms) and TF / IDF (Inverse Document Frequency: ratio of documents in which terms appear) ) Is registered. TF / IDF is the product of TF and IDF, and represents the importance of the term in the document.

  Here, the appearance frequency tfij of the term tj in the article Di is represented by the following expression (1), the appearance frequency idfj of the article Di is represented by the following expression (2) using the number of articles dfj and the total number N of articles in which the term tj appears, Furthermore, if the weight Wtj of the term tj of the article Di is expressed by the following equation (3) and all the different terms in the entire article are n, the article vector of the article Di can be expressed by the equation (4).

  The online news server 102 is a Web server that returns a news article in response to a request from the mobile terminal 2 that can be connected to the Internet 3, and includes a history acquisition unit 102a, a similarity calculation unit 102b, a priority setting unit 102c, a priority A distribution unit 102d and a news distribution unit 102e are provided.

  The history acquisition unit 102 a acquires a user's viewing history and provides it to the preference extraction management unit 103. In this embodiment, when the user selects an article title and views up to the headline, the viewing history is left for the first time.

  FIG. 4 is a diagram illustrating an example of the viewing history, in which a plurality of pieces of information including at least an article ID unique to a news article viewed by a user and time information at which the news article is accessed are associated with each other. Is remembered. In the present embodiment, as the time information, the difference between UTC (universal coordinate time) UTC on January 1, 1970 and the current item is registered in milliseconds.

  The similarity calculation unit 102b, as will be described in detail later, is a preference cluster representing each feature of a plurality of preference clusters generated based on a content vector of new content to be newly distributed and a user's viewing history. The vector is compared and the similarity between the two is calculated. The priority setting unit 102c sets a priority for each new content based on the similarity. The priority delivery unit 102d preferentially selects new content with high priority from the news database 101 and delivers it to the user.

  The preference extraction management unit 103 includes a keyword acquisition unit 103a, a clustering unit 103b, a preference cluster integration unit 103c, a preference cluster vector generation unit 103d, and a preference cluster weighting unit 103e, and refers to the news database 101 based on the history information. Then, preference information of each user is extracted according to an algorithm described later.

  The keyword acquisition unit 103 a acquires the news keywords viewed by the user from the news database 101. The clustering unit 103b classifies the content viewed by the user into at least one preference cluster based on the acquired keyword. Here, the preference cluster is a group of articles with related contents and each classified as a preference.

  The preference cluster integration unit 103c integrates a plurality of preference clusters having similar contents. The preference cluster vector generation unit 103d generates a preference cluster vector for each preference cluster based on the feature of articles classified into each preference cluster. The preference cluster weighting unit 103e weights each preference cluster.

  FIG. 5 is a flowchart showing a procedure for extracting and updating preference information in the preference extraction management unit 103. In this embodiment, a plurality of preference clusters are generated as preference information for each user, and a user ID is assigned as a preference cluster group. Managed by each. Furthermore, this preference cluster is continuously updated based on the user's subsequent viewing history.

  When a history of news articles viewed by the user is notified from the history acquisition unit 102a and this is detected in step S1, the news database 101 is searched in step S2 using the article ID registered in the viewing history as a search key. And the title of the article is extracted. In step S3, keywords are extracted from the article title.

  FIG. 6 is a flowchart showing in detail the procedure of the keyword extraction process executed by the keyword acquisition unit 103a.

  In step S101, the article title is divided by a space. In step S102, the article title is morphologically analyzed. In the present embodiment, the “chasm (registered trademark)” (Matsumoto Yuji, Kitauchi Kei, Yamashita Tatsuo, Hirano Yoshitaka, Matsuda Hiroshi, Takaoka Kazuma, Asahara Masayuki, 2.3.3 Instructions for use, “Nara Institute of Science and Technology, Graduate School of Information Science, Natural Language Processing, 2003) is used.

  In step S103, prefixes, nouns (excluding non-independent and pronouns), and unknown words are extracted. In step S104, consecutive prefixes, nouns, and unknown words are connected and extracted as keywords. In step S106, nouns other than the last (except for the number, suffix, and change connection) and unknown words are extracted as keywords from consecutive words.

  Returning to FIG. 5, in steps S4 to S8, as will be described in detail below, each of the keywords is classified (clustered) into at least one of a plurality of preference clusters by the clustering unit 103b.

  In step S4, one of the keywords extracted in step S3 is selected as the current keyword of interest. In step S5, it is determined whether a preference cluster keyword matching the current attention keyword or a preference cluster keyword including the current attention keyword has already been registered. If the preference cluster keyword has already been registered, the process proceeds to step S7, and the article ID of the keyword of interest is additionally registered as an element in the preference cluster.

  On the other hand, if such a preference cluster keyword is not registered, the process proceeds to step S6, where a new preference cluster having the current attention keyword as a preference cluster keyword is generated, and an article ID is newly added to this preference cluster as an element. be registered. Thus, in this embodiment, since preference clusters are generated for all keywords included in articles viewed by the user, minor user preferences can be extracted without omission.

  FIG. 7 is a diagram schematically showing an example of a preference cluster, in which a plurality of preference clusters having each article ID as an element are generated using each keyword as a preference cluster keyword, and these are set as preference cluster groups for each user. Managed. Note that the IDs of articles that include a plurality of preference cluster keywords in the title are registered in duplicate in each preference cluster.

  Returning to FIG. 5, in step S7, it is determined whether or not the processing has been completed for all the keywords of the viewed article. If not completed, the process returns to step S4, and the above-described procedures are repeated while switching the keywords. . When the processing is completed for all keywords, the process proceeds to step S9, and preference cluster integration processing is executed.

  FIG. 8 is a flowchart showing a procedure of preference cluster integration processing. Preference clusters having a large number of article IDs are integrated into one preference cluster by the preference cluster combining unit 103c.

  In step S201, one of the preference clusters is selected as the current attention cluster. In step S202, one of the preference clusters other than the target cluster is selected as a comparison control cluster. In step S203, it is determined whether the number of article IDs common to the target cluster and the comparison control cluster exceeds a predetermined reference rate (50% in this embodiment) in any cluster.

  If it exceeds the reference value, the process proceeds to step S205, and the current attention cluster and the comparison control cluster are integrated into one preference cluster. In step S206, the preference cluster keyword is updated. In the present embodiment, if the keyword of one cluster to be integrated includes all the keywords of the other cluster, the keyword of one cluster is adopted as the preference cluster keyword after integration. The keyword becomes a new preference cluster keyword.

  In step S207, it is determined whether or not there are other comparison target clusters. The above-described procedures are repeated until the comparison with all preference clusters is completed, and then the process proceeds to step S208. In step S208, it is determined whether or not integration has been completed for all combinations of preference clusters. If not completed, the process returns to step S201, and the above-described processes are repeated while switching the cluster of interest.

  Returning to FIG. 5, in step S10, the preference cluster vector generation unit 103d generates a preference cluster vector representing the characteristics of each preference cluster by summing the article vectors of all articles included in each preference cluster. The In step S11, each preference cluster is weighted by the preference cluster weighting unit 103e. Here, the weighting method for each preference cluster will be described using three typical methods as examples.

  The first weighting method is weighting according to the number of articles included in the preference cluster. The more the number of articles included in the preference cluster, the higher the weight of the preference cluster.

  The second weighting method is weighting based on the concept of forgetting a preference cluster. The smaller the sum of the time differences between the access time of each article included in the preference cluster and the current time, the higher the weight of the preference cluster.

  More specifically, the weight Ri of an article (article vector Di) included in a preference cluster is defined by the following equation (5) by normalizing the current time τ and the time Ti when the article was viewed by the time T. The Note that λ (0 <λ <1) is a forgetting constant indicating how much forgetting at time T, and the weighting factor R0 = 1.

  The weight Rc of preference cluster C using this weight is given by the following equation (6) where m is the number of articles included in the cluster.

  The third weighting method is weighting according to the viewing rate of the preference cluster. The higher the viewing rate of each article in the preference cluster, the higher the weight of the preference cluster.

  The preference cluster (preference cluster vector) and the weight value R of each preference cluster generated as described above are managed as preference information for each user by the preference extraction management unit 103 for each user ID.

  Next, a method for selecting a news article using the preference information obtained as described above will be described with reference to the flowchart of FIG.

  When the user requests news distribution from his / her mobile terminal 2 and this is detected in step S301, the user ID is notified from the online news server 102 to the preference extraction management unit 103 in step S302. In the preference extraction management unit 103, preference information associated with the notified user ID is provided to the online news server 102. In step S303, in the online news server 102, one of the news articles new to the current user is selected on the news database 101 as the current notice article, and the article vector is extracted.

  In step S304, one preference cluster vector is selected as the current attention cluster from the preference information corresponding to the user ID. In step S305, in the similarity calculation unit 102b, the similarity sim (Qc, Di) between the preference cluster vector Qc of the cluster of interest and the article vector Di is set to the weight value Rc assigned to the preference cluster, and the preference cluster. The product of the vector Qc and the cosine distance between the article vector Di of the article of interest is obtained by the following equation (7).

  In step S306, it is determined whether or not the similarity calculation with all preference clusters has been completed for the current article of interest. If not completed, the process returns to step S304, and the above-described processes are repeated while switching the cluster of interest. When the similarity calculation between the current attention article and all the preference clusters is completed, the process proceeds to step S307, and the priority setting unit 102c registers the maximum similarity as the priority of the current attention article.

  In step S308, it is determined whether there are other new news articles, and the process returns to step S303 until the similarity calculation with all new news articles is completed, and the above-described processing is repeated while switching the article of interest.

  As described above, when the priorities of all new news articles are obtained, the news articles are sorted in descending order of priority in step S309. In step S310, the sorted news articles are distributed to the user from the priority distribution unit 102d. On the user's mobile terminal 2, articles with high priority are displayed on the initial screen, and articles with low priority can be viewed by scrolling the screen.

  In the above-described embodiment, the case where the content is online news has been described as an example. However, the present invention is not limited to this, and the present invention can be similarly applied even when the content is a mail magazine.

  Further, in the above-described embodiment, it has been described that all news articles are distributed regardless of the priority level, and only the display order and display position are changed according to the priority. However, the present invention is limited to this. Instead, only articles with a higher priority may be selectively distributed, and other articles may not be distributed.

It is a block diagram of a content distribution apparatus to which the present invention is applied. It is the figure which showed an example of the information managed by the news database. It is the figure which showed an example of the article vector file. It is the figure which showed an example of viewing history. It is the flowchart which showed the extraction and update procedure of preference information. It is the flowchart which showed the procedure of the keyword extraction process in detail. It is the figure which expressed typically an example of a taste cluster. It is the flowchart which showed the procedure of preference cluster integration processing. It is the flowchart which showed the procedure which selects a news article using preference information.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1 ... News distribution apparatus, 2 ... Portable terminal, 3 ... Internet, 101 ... News database, 102 ... Online news server, 103 ... Preference extraction management part

Claims (12)

In a content distribution method for dynamically generating and updating a preference cluster based on a user's content viewing history, and setting a priority for content distributed to the user based on the preference cluster,
For each content, a procedure for generating a keyword and content vector representing the content and registering it in the content database;
A procedure for extracting keywords of content viewed by the user from the content database;
A procedure for classifying each content into at least one of a plurality of preference clusters based on the extracted keywords;
A procedure for generating a preference cluster vector representing the characteristics of each preference cluster based on the content classified into each preference cluster;
For each new content to be distributed, a procedure for calculating the similarity between the content vector and each preference cluster vector;
A procedure for setting a priority for each new content based on the similarity;
And providing each new content to the user according to the priority. The procedure of classifying each content into preference clusters is as follows:
A procedure for classifying the content into the registered preference cluster when the preference cluster associated with the keyword of the content is already registered;
And a procedure for creating a new preference cluster associated with the keyword and classifying the content into the preference cluster when the preference cluster associated with the content keyword is unregistered. The content delivery method according to 1.   The content distribution method according to claim 1, further comprising a step of integrating a plurality of preference clusters having similar classified contents into one preference cluster. Further comprising the step of setting a weight value for each preference cluster;
4. In the procedure for setting a priority for each new content, the priority is set by reflecting the weight value of the preference cluster in the similarity between the new content and the preference cluster. The content distribution method according to any one of the above.   The content distribution method according to claim 4, wherein the weight value is a function of the number of contents included in the preference cluster, and the weight of the preference cluster having a larger number of contents included in the preference cluster is increased.   5. The content according to claim 4, wherein the weight value is a function of a time difference between an access time of each content included in the preference cluster and a current time, and the preference cluster having a smaller time difference has a higher weight. Delivery method.   The content distribution method according to claim 4, wherein the weight value is a function of a browsing rate of each content included in the preference cluster, and a preference cluster having a higher browsing rate of each content has a higher weight. .   The content distribution method according to claim 1, wherein the content is an online news article. In a content distribution apparatus that dynamically generates and updates a preference cluster based on a user's content viewing history, and sets a priority for content distributed to the user based on the preference cluster,
For each content, a content database in which keywords and content vectors representing the content are registered,
History acquisition means for acquiring content viewing history for each user;
A keyword acquisition means for referring to a content database based on the viewing history and acquiring a keyword of content viewed by the user;
Clustering means for classifying content viewed by the user into at least one preference cluster based on the acquired keyword;
A preference cluster vector generating means for generating a preference cluster vector representing the characteristics of each preference cluster based on the content classified into each preference cluster;
Similarity calculation means for calculating the similarity between the content vector and each preference cluster vector for each new content to be distributed;
Priority setting means for setting a priority for each new content based on the similarity;
A content distribution device comprising: content distribution means for distributing each new content according to the priority. The clustering means includes
Means for classifying content into the registered preference cluster when the preference cluster associated with the keyword of the content is already registered;
And a means for creating a new preference cluster associated with the keyword when the preference cluster associated with the keyword of the content is not registered, and classifying the content into the preference cluster. 10. The content distribution device according to 9.   The content distribution apparatus according to claim 9, further comprising an integration unit that integrates a plurality of preference clusters whose classified contents are similar to each other into one preference cluster. A weighting means for setting a weight value for each preference cluster;
12. The content distribution apparatus according to claim 9, wherein the similarity calculation unit reflects the weight value of the preference cluster in the similarity between the new content and the preference cluster.

Images