JP2010181966A - Device and method for evaluating recommendation information - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a device for performing evaluation such that recommendation information such as a content potentially desired by a user can be distributed without exception. <P>SOLUTION: A characteristic vector generation part 209 collects a content of a Web page or the like in prescribed timing, extracts a keyword indicating a characteristic of the content, and generates a characteristic vector. The generated characteristic vector is stored in a content characteristic vector management table 207b, and is updated. The timing of the collection is optional timing or timing wherein the Web page is updated (this is caused by reception or the like of an update notification signal). The content is evaluated by use of such the updated content characteristic vector. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a recommendation information evaluation apparatus and a recommendation information evaluation method for evaluating content.

  In order to distribute recommended information that matches the user's characteristics, a Content-basedFiltering method is considered as an information filtering method for extracting the recommended information. In this method, when a user takes an action such as browsing a Web page, the browsed target document is analyzed, and words (keywords) included in the document are extracted. These words are distinguished for each user and recorded as interest / preference information called a user profile. One concept of interest is expressed in the user profile, and it is considered to provide information to the user using this concept. For example, a system is generally known that makes it possible to recommend an appropriate advertisement by matching a keyword assigned to an advertisement product or the like with a user profile (expressed by a keyword).

  Moreover, the technique described in patent document 1 is mentioned as a technique relevant to this. This Patent Document 1 describes a technique for determining and distributing a distribution destination to which recommendation information is distributed and recommendation information to be distributed in accordance with a predetermined recommendation extraction rule.

JP 2007-148878 A

  However, in the background art described above, recommendation information is extracted by performing a matching process, or in the technique described in Patent Document 1, recommendation information is determined based on a predetermined recommendation extraction rule. Since the feature space vector is not updated depending on the recommendation target, there is a possibility that the recommendation information may be leaked.

  That is, in the technique of Patent Document 1, if a hobby / preference is registered in the recommendation extraction rule, recommendation information suitable for the hobby / preference is determined and distributed, and the purchase history and the like are also included. Based on this, when purchasing a product, information related to the purchased product will be distributed, but this distributed information is based on direct information such as hobbies / preferences, purchase history, or access history. Since the feature space vector is not updated according to the recommendation target, the content that the user wants may be leaked.

  In particular, when recommendation information is distributed and used based on the access history, it is considered that the leakage is particularly large. That is, when a user browses a site or the like, the browsed site itself is not represented by a single keyword, but is usually represented by a plurality of keywords. For example, when a user searches and browses information related to the movie a, the browsing site introduces the movie a, introduces the original, or introduces the entertainment movie theater. Or introducing performers, and various information (vector information). In this case, when the user is searching for an original book, it is desirable for the user to perform a filtering process specialized for the original book and distribute recommendation information only for that purpose.

  However, with the technology described in Patent Document 1, the entire site being browsed is viewed and information close to the entire site is distributed as recommendation information. Could not deliver.

  Therefore, the present invention provides a recommendation information evaluation apparatus and a recommendation information evaluation method that can perform evaluation so that recommendation information such as content that a user potentially desires can be delivered without omission. With the goal.

  In order to solve the above-described problem, a recommendation information evaluation apparatus according to the present invention includes a feature vector generation unit that generates a feature vector constituting a feature space of a selection target item, and a feature vector generated by the feature vector generation unit. Feature vector storage means for storing, reception means for receiving selection of one selection target item from a plurality of selection target items by a user operation, selection target item selected and received by the reception means, and selected A history storage unit that distinguishes and stores the selection target item that has not been stored, and a satisfaction feature vector of the selected selection target item and an unsatisfied feature vector of the selection target item that has not been selected, which are stored by the history storage unit. The user characteristic vector with the orthogonal vector orthogonal to the separation plane for separation within the user characteristic vector And an evaluation unit that evaluates an unknown selection target item based on the user characteristic vector generated by the user characteristic vector generation unit, and the feature vector generation unit adds the generated feature vector to the generated feature vector. Based on the feature vector, the feature vector stored in the feature vector storage means is updated.

  In the recommendation information evaluation method of the present invention, a feature vector generation step for generating a feature vector constituting a feature space of a selection target item, and a feature vector generated in the feature vector generation step are stored in a feature vector storage unit. A feature vector storage step, a reception step for receiving selection of one selection target item from a plurality of selection target items by a user operation, and a selection target item selected and not selected in the reception step A history storage step of storing the target item in the history storage means, a satisfaction feature vector of the selected selection target item and an unsatisfied feature vector of the selection target item not selected stored in the history storage means, Based on orthogonal vector orthogonal to the separation plane for separation within A user characteristic vector generation step for generating a user characteristic vector; and an evaluation step for evaluating recommendation information based on the user characteristic vector generated by the user characteristic vector generation step, wherein the feature vector generation step comprises: Based on the generated feature vector, the feature vector stored in the feature vector storage means is updated.

  According to the present invention, the feature vector constituting the feature space of the selection target item is generated, and the generated feature vector is stored, so that the already stored feature vector is updated based on the generated feature vector. . On the other hand, when receiving selection of one selection target item from a plurality of selection target items by a user operation, the selected selection target item and the selection target item that has not been selected are distinguished and stored. Then, an orthogonal vector orthogonal to the separation plane for separating the satisfaction feature vector of the selected selection target item and the unsatisfied feature vector of the selection target item not selected in the feature space is defined as a user characteristic vector, and this user characteristic Based on the vector, an unknown selection target item can be evaluated.

  As a result, it is possible to evaluate unknown selection target items, for example, recommendation information such as contents or operation menu items, according to a vector indicating the user's characteristics, and the user's hobbies / preferences, direct access history, etc. The recommendation information to be recommended based on the information is not evaluated, and the recommendation information can be provided to the user without omission. In addition, recommendation information such as contents that the user potentially desires can be distributed without omission. In other words, the feature vector is updated according to the type of content such as the selection target item, for example, TV program or CD introduction information, so that recommendation information specialized to the type of content can be provided to the user.

  Also, the recommendation information evaluation apparatus of the present invention separates the satisfaction feature vector of the selected selection target item stored in the history storage unit and the unsatisfied feature vector of the selection target item that has not been selected in the feature space. It is preferable that the apparatus further includes a separation plane generation unit that generates a separation plane for generating the user characteristic vector according to the separation plane generated by the separation plane generation unit.

  According to the present invention, the separation plane for separating the satisfactory feature vector of the selected selection target item and the unsatisfactory feature vector of the unselected selection target item in the feature space is generated, and based on this, the user characteristics By generating the vector, the user characteristic vector can be generated with high accuracy.

  In the recommendation information evaluation apparatus according to the present invention, it is preferable that the feature vector generation unit obtains information necessary for updating from document information that is a selection item target.

  According to the present invention, a feature vector can be generated by acquiring information necessary for updating from document information that is a selection item target. This document information is information describing a document that can be analyzed such as a Web page, word processor data (data created by a word processor), spreadsheet data (data created by a spreadsheet application), and the like.

  In the recommendation information evaluation apparatus according to the present invention, the feature vector generation unit further includes a feature extraction unit that extracts a characteristic word for each part of speech for each attribute of the selection target item, and is extracted by the feature extraction unit. It is preferable to generate a feature vector based on the word.

  According to the present invention, it is possible to generate an appropriate feature vector by extracting a characteristic word for each attribute of the selection target item for each part of speech, and generating a feature vector based on the extracted word. it can.

  Moreover, in the recommendation information evaluation apparatus of this invention, it is preferable that the said feature extraction means extracts a characteristic word for every attribute of a selection object item based on TF / IDF method.

  According to the present invention, an appropriate feature vector can be generated by extracting a characteristic word for each attribute of a selection target item based on the TF / IDF method.

  In the recommendation information evaluation apparatus according to the present invention, it is preferable that the feature extraction unit extracts a characteristic word for each attribute of the selection target item based on the PMM method.

  According to this invention, an appropriate feature vector can be generated by extracting a characteristic word for each attribute of a selection target item based on the PMM method.

  In the recommendation information evaluation apparatus according to the present invention, it is preferable that the feature vector generation unit generates a feature vector at an arbitrary timing.

  According to the present invention, an appropriate feature vector can be generated by generating a feature vector at an arbitrary timing. Therefore, it is possible to acquire the content based on the selection item, for example, a Web page, following the latest information on the selection target item, for example, the introduction information of the TV program or CD, and generate the feature vector.

  According to the present invention, it is possible to evaluate recommendation information such as an unknown selection target item, for example, content or an operation menu item, according to a vector indicating the user's characteristics, and the user's hobbies / preferences, access history, etc. The recommendation information to be recommended is not evaluated based on the direct information, and the recommendation information can be provided to the user without omission.

1 is a system configuration diagram illustrating a system configuration of a recommended information distribution system including a mobile phone 100 and an information distribution server 200 according to the present embodiment. 2 is a block diagram showing a functional configuration of an information distribution server 200. FIG. 2 is a hardware configuration diagram of an information distribution server 200. FIG. FIG. 10 is an explanatory diagram illustrating an example of a menu screen displayed on the mobile phone 100. It is a conceptual diagram which shows the processing concept when producing | generating a separation plane. It is a conceptual diagram which shows the concept when calculating a user characteristic vector based on a separation plane, and also evaluating based on a user characteristic vector. It is explanatory drawing which shows the memory content of the content management table 205a. It is explanatory drawing which shows the memory content of the content feature vector management table 207b. 2 is a block diagram showing a functional configuration of a mobile phone 100. FIG. FIG. 11 is a sequence diagram showing processing when content is distributed in response to a request from mobile phone 100. It is a flowchart which shows the calculation process of a user characteristic vector. It is a sequence diagram when distributing recommendation information. It is a flowchart which shows the content evaluation process. It is explanatory drawing which shows the menu screen in the log | history event a and the log | history event b. It is a figure which shows the positional relationship in the feature space of the feature vector of the content selected in the menu screen. It is a figure showing the feature vector of content in history space a and b in relative space. It is the figure which showed the concept when evaluating based on the feature vector of each content in absolute space. It is explanatory drawing when performing a process using relative space when a separation plane cannot be generated. It is a conceptual diagram which shows the process which calculates | requires the gravity center of a feature vector in consideration of browsing frequency. It is a conceptual diagram which shows the process in which a user characteristic vector is calculated | required by the NN method in the time series accessed. It is explanatory drawing which shows the parameter of the feature vector in the feature space regarding operation. It is a block diagram which shows the function of the mobile telephone 100a which can display recommendation operation information. It is a flowchart which shows the process which produces | generates a feature vector.

  Embodiments of the present invention will be described with reference to the accompanying drawings. Where possible, the same parts are denoted by the same reference numerals, and redundant description is omitted.

  FIG. 1 is a system configuration diagram illustrating a system configuration of a recommended information distribution system including a mobile phone 100 and an information distribution server 200 according to the present embodiment. In response to a request from the mobile phone 100, the information distribution server 200 distributes content that serves as recommendation information stored in advance to the mobile phone 100. The recommended information to be distributed has been edited based on the access history of the user of the mobile phone 100 and is information that seems necessary for the user. The information distribution server 200 acquires content from other content providers and distributes the acquired content in response to an access request from the mobile phone 100, and the access history is collected at this time. Further, the mobile phone 100 may be configured to transmit and acquire the access history.

  The configuration of such an information distribution server 200 will be described with reference to FIG. FIG. 2 is a block diagram showing a functional configuration of the information distribution server 200. This information distribution server 200 includes a content request receiving unit 201 (reception unit), a recommendation request receiving unit 202, a history class separation unit 203 (separation plane generation unit), a user characteristic vector calculation unit 204 (user characteristic vector generation unit), a distribution An information storage unit 205, a content evaluation unit 206 (evaluation unit), a history storage unit 207 (feature vector storage unit, history storage unit), a content transmission unit 208, and a feature vector generation unit 209 (feature vector generation unit) are configured. Has been.

  FIG. 3 is a hardware configuration diagram of the information distribution server 200. As shown in FIG. 3, the information distribution server 200 shown in FIG. 2 physically includes a CPU 11, a RAM 12 and a ROM 13 that are main storage devices, an input device 14 such as a keyboard and a mouse that are input devices, a display, and the like. The computer system includes an output device 15, a communication module 16 that is a data transmission / reception device such as a network card, and an auxiliary storage device 17 such as a hard disk. Each function described in FIG. 2 has the input device 14, the output device 15, and the communication module 16 under the control of the CPU 11 by reading predetermined computer software on the hardware such as the CPU 11 and the RAM 12 shown in FIG. 3. This is realized by reading and writing data in the RAM 12 and the auxiliary storage device 17. Hereinafter, each functional block will be described based on the functional blocks shown in FIG.

  The content request receiving unit 201 is a part that acquires content from a content provider on the network in response to a content request from the mobile phone 100, and receives an access history by receiving the content request and its URL from the mobile phone 100. Can be collected. The collected access history (such as URL) is output to the history storage unit 207. The content request receiving unit 201 includes, as an access history, content (URL) that is actually accessed in a certain menu screen (event) and content that is displayed on the menu screen but not actually accessed. Collect separately.

  Further, the content request receiving unit 201 may receive not only the access request but also access history information collected by the mobile phone 100 and output this to the history storage unit 207. Similarly, in this case, the mobile phone 100 is configured to distinguish the content selected on a certain menu screen from the content not selected and transmit it to the information distribution server 200 as access history information.

  The recommendation request receiving unit 202 is a part that receives a distribution request for recommendation information from the mobile phone 100 together with the user ID of the user of the mobile phone 100. When receiving the distribution request, the recommendation request receiving unit 202 notifies the content evaluation unit 206 to that effect, and evaluates the content stored in the distribution information storage unit 205 based on the user characteristic vector calculated in advance. To instruct As a modification, the fact that there is a recommendation request is output to the user characteristic vector calculation unit 204, the current user characteristic vector is calculated, the result is notified to the content evaluation unit 206, and the recommendation information is transmitted to the content transmission unit 208 You may make it deliver.

  The history class separation unit 203 stores content browsed by a certain user based on the feature vector of the content included in the browsing history information stored in the history storage unit 207 (browsing history table 207c: history storage unit). This is a part for generating a separation plane for separating the satisfaction class and the dissatisfaction class in the feature space constituted by the feature vectors stored in the feature vector management table 207b. Here, the satisfaction class refers to a class including only the actually selected content among a plurality of contents displayed in a selectable manner on a menu screen for accessing certain content. A class that contains only content that was not selected on the menu screen. The history class separation unit 203 uses the SVM method or the NN method to generate a separation plane that can separate feature vectors of each content into a satisfaction class and a dissatisfaction class, thereby separating these classes. it can. Here, the SVM (Support Vector Machine) method is a method in which a plane that maximizes the minimum distance between nodes of each class (between feature vectors) is a separation plane, and the NN method is the center of gravity of each class. In other words, a plane that is perpendicular to the straight line connecting the centroid of the satisfaction class and the centroid of the dissatisfaction class calculated based on the feature vector of each content is used as the separation plane. In the NN method, it is not always necessary to obtain a separation plane, and a straight line connecting the centroid of the unsatisfactory class to the centroid of the satisfactory class may be used as the user characteristic vector.

  Here, a specific method for generating the separation plane will be described. FIG. 4 is an explanatory diagram illustrating an example of a menu screen displayed on the mobile phone 100. As shown in FIG. 4, Node1 to Node5 are displayed. This Node indicates an item to be selected, and is, for example, an item distinguished for each category on a menu screen on the web, or a title of document information such as a web page. For example, Node 1 is content related to automobiles, and Node 2 is content related to finance. The user can browse the corresponding content by selecting any one of Node1 to Node5.

  FIG. 5 is a conceptual diagram showing a processing concept when generating a separation plane. FIG. 5 shows the positional relationship between the feature vectors of each content in the feature space for the content selected by the user and the content not selected on the menu screen in FIG. As shown in FIG. 5, the parts indicated by circles are satisfactory contents (Node3, Node4), and the parts indicated by triangles are unsatisfactory contents (Node1, Node2, Node5), and are rectangular. The parts indicated by are unknown unevaluated contents (Node8, Node7). A separation plane is generated so as to separate the satisfactory content and the unsatisfied content. Here, the unevaluated content is not taken into consideration when the separation process is performed. In the figure, the feature space is represented by a two-dimensional plane for convenience. However, there are vector axes corresponding to the parameters constituting the vector, and the feature space is usually composed of vectors of 20 dimensions or more. Yes.

  The user characteristic vector calculation unit 204 calculates an orthogonal vector orthogonal to the separation plane generated by the history class separation unit 203 in the feature vector space configured by the feature vector generated by the feature vector generation unit 209, This is the part that uses the user characteristic vector. A specific example is shown in FIG. FIG. 6 is a conceptual diagram illustrating a concept when a user characteristic vector is calculated based on the separation plane and evaluation is performed based on the user characteristic vector. As shown in FIG. 6, a user characteristic vector w that is an orthogonal vector orthogonal to the separation plane is calculated by the user characteristic vector calculation unit 204. The user characteristic vector calculation unit 204 is updated every time browsing history is collected, and updates the user characteristic vector stored in the user characteristic vector management table 207a.

  The distribution information storage unit 205 is a part that stores content to be distributed (recommendation information) and management information thereof, and stores a content management table 205a. FIG. 7 is an explanatory diagram showing the stored contents of the content management table 205a. As shown in FIG. 7, the content management table 205a stores a content ID, a category, a title, and a recommended text in association with each other. The content ID is identification information for uniquely identifying the content, and the category indicates the information type. For example, information related to CD, information related to DVD, information related to books, information related to TV programs, etc. It is information for indicating. The title is heading information to be displayed as recommendation information. When the recommendation information is distributed, this title portion is displayed. The recommendation sentence is a main part of the recommendation information and is useful information for the user. Information stored in the content management table 205a may be input in advance by an operator, or may be information that is automatically retrieved and extracted from a content provider and narrowed down by a predetermined filter.

  When the recommendation request is received by the recommendation request receiving unit 202, the content evaluation unit 206 applies the user characteristic vector calculated by the user characteristic vector calculation unit 204 to the content stored in the distribution information storage unit 205. This is a part for generating recommendation information in which the titles of the contents are arranged in the descending order of evaluation. Specifically, the content evaluation unit 206 calculates an evaluation value by calculating the inner product of the user characteristic vector and the feature vector of each content, and recommends information in which the content titles are arranged in descending order of the evaluation value. Generate.

  Conceptually, it is as follows. For example, in FIG. 6, Node 7 and Node 8 are treated as unknown content, assuming that the information is stored in the content management table 205a. Here, the point perpendicular to the user characteristic vector w from Node 7 and Node 8 is the evaluation point for the user characteristic vector w. The recommendation information is generated so that the content (title portion) is displayed in the order in which the position of the intersection of the perpendicular and the user characteristic vector is higher. Note that display or non-display may be performed with reference to the separation plane. For example, since the intersection with the perpendicular from Node 8 is positioned higher than the separation plane as a reference, it is determined that the evaluation is high for the user and is selected as recommendation information. On the other hand, since Node 7 is below the standard, the evaluation is low, and it may be recommended information that is not selected as recommended information.

  The history storage unit 207 stores each piece of history information, and stores a user characteristic vector management table 207a, a content feature vector management table 207b, and a browsing history table 207c.

  The user characteristic vector management table 207a is a part that stores the user characteristic vector calculated by the user characteristic vector calculation unit 204 in association with the user ID for specifying the user.

  The content feature vector management table 207b is a part that stores the feature vectors of content stored in the content management table 205a in association with information (for example, content ID, URL, etc.) for specifying the content. The content feature vector is acquired by the content request receiving unit 201 together with the content when the user accesses it, and is set by the operator when registering the recommendation information. For example, the content feature vector management table 207b stores information shown in FIG. As shown in FIG. 8, a plurality of parameters constituting the feature vector are stored in association with the content ID. In FIG. 8, although eight parameters are described, the feature vector is usually composed of more parameters. In addition to the defined parameters, keywords extracted according to morphological analysis for each content may be added as parameters. For each parameter constituting the feature vector, a numerical value is described between 0 and 1, indicating which feature vector has the feature.

  The browsing history table 207c is a part for storing the destination or collected access history information received by the content request receiving unit 201. For example, the content ID (or the URL of the content) and its content (viewed, viewed This is a part for storing the feature vectors of each of them that have not been assigned. This feature vector of content is acquired together with the content acquired by the content request receiving unit 201.

  The content transmitting unit 208 is a part that transmits content acquired in response to a content request received by the content request receiving unit 201 or a recommendation request received by the recommendation request receiving unit 202. For example, the content transmission unit 208 transmits the content acquired in response to the content request received by the content request reception unit 201, and the content stored in the distribution information storage unit 205 according to the recommendation request reception unit 202 The recommendation information generated in accordance with the evaluation result by the content evaluation unit 206 is transmitted.

  A feature vector generation unit 209 generates content feature vectors by collecting content information from document information such as Web pages, performing keyword analysis on the collected content information, and calculating the feature vectors. It is. In addition to the Web page, the document information includes information describing a document that can be analyzed such as word processing data (data created by a word processor), spreadsheet data (data created by a spreadsheet application), PDF data, and the like. Is included. Detailed processing is as follows. Then, the feature vector generation unit 209 updates the feature vector stored in the content feature vector management table 207b or the browsing history table 207c as necessary using the generated feature vector.

  Here, a specific method for generating a feature vector will be described. FIG. 23 is a flowchart showing a feature vector calculation process. As shown in FIG. 23, first, content information (for example, a web page) is collected from a web page by the content information collection unit 209a (S501).

  The collected content information is classified into parts of speech (nouns, proper nouns, verbs, particles, etc.) by the feature keyword extraction unit 209b by a morphological analysis algorithm (S502). Then, a characteristic keyword according to the attribute of the content information is extracted from the content information by the feature extraction algorithm (S503). The feature extraction algorithm of the feature keyword extraction unit 209b can extract a keyword that is characteristic for each attribute of content information for each part of speech using the TF / IDF method or the PMM method.

  Here, the TF / IDF method is a method in which words that do not frequently appear in other documents among words that appear frequently in a document are used as a characteristic word in the document. PMM (Parametric Mixture Model) ) Method is a method of simultaneously extracting multiple topics from a document by constructing a probabilistic model of documents having multiple topics. In the feature extraction algorithm, it is not always necessary to extract using the TF / IDF method or the PMM method.

  A feature vector is generated by the feature vector calculation unit 209c based on the extracted keyword (S504). The generated feature vector is stored in the content feature vector management table 207b.

  The feature vector generation unit 209 updates the feature vector stored in the feature vector management table 207b or the browsing history table 207c as necessary whenever the information on the Web page is updated or at an arbitrary timing.

  The information distribution server 200 configured as described above can distribute the recommendation information evaluated based on the user characteristic vector in response to a request from the mobile phone 100. Therefore, it is not necessary to register information such as tastes and hobbies, and it is possible to distribute recommended information without omission, without extracting only content directly related to accessed content.

  Next, the mobile phone 100 will be described. FIG. 9 is a block diagram showing a functional configuration of the mobile phone 100. As shown in FIG. 9, the mobile phone 100 includes a browser 101, a history transmission unit 102, a recommendation request transmission unit 103, and a distribution information display unit 104. The cellular phone 100 is configured by hardware such as a CPU, RAM, and ROM, and the CPU is configured to operate according to a stored program. Specifically, it has the same configuration as that of the information distribution server 200 and is realized by the hardware configuration shown in FIG. Each configuration will be described below.

  The browser 101 is an application part for connecting to the Internet, and accesses a site arranged on the Internet and allows a user to view content held on the site. The browser 101 is configured to hold an access history.

  The history transmission unit 102 is a part that transmits an access history (information indicating selected content and unselected content on the same menu screen) of content accessed using the browser 101. The history transmitting unit 102 is not an essential configuration when it has a function of collecting history information every time it is accessed on the information distribution server 200 side.

  The recommendation request transmission unit 103 is a part that transmits a request for recommendation information to the information distribution server 200 together with a user ID when a user operates an operation unit (not shown).

  The distribution information display unit 104 is a part that displays the menu screen and content (Web page) acquired by the browser 101.

  The mobile phone 100 configured as described above enables access to the Internet using the browser 101 and holds the access history. The history transmission unit 102 accesses the request from the information distribution server 200 or at a predetermined timing. History can be sent.

  Next, operations of the mobile phone 100 and the information distribution server 200 will be described. FIG. 10 is a sequence diagram illustrating processing performed by the mobile phone 100 and the information distribution server 200 when content is distributed in response to a request from the mobile phone 100.

  As shown in FIG. 10, in response to a request from the browser 101 in the mobile phone 100, a menu screen is distributed from the information distribution server 200 by the content transmission unit 208 (S101). The user of the mobile phone 100 selects any one content (corresponding to the above-mentioned Node) from the menu screen, and when the browser 101 in the mobile phone 100 accepts this (S102), a URL for the content connection request and destination Is transmitted to the information distribution server 200 (S103).

  In the information distribution server 200, the URL is received by the content request receiving unit 201 and stored as an access history in the browsing history table 207c. Based on the access history, the user characteristic vector is calculated and stored (S104). Then, content based on the destination requested in S103 is acquired by the content transmission unit 208 from a content provider (not shown) and distributed (S105). Note that the order of the user characteristic vector calculation process and the content distribution process may be reversed.

  Here, the process will be described in detail for the user characteristic vector calculation in S104. FIG. 11 is a flowchart showing the calculation process of the user characteristic vector. As shown in FIG. 11, first, access history information indicated by an access request or access history received by the content request receiving unit 201 is stored in the browsing history table 207c as a newly added access history (S201). .

  Then, based on the access history information stored in the browsing history table 207c, the history class separation unit 203 extracts feature vectors of contents belonging to each satisfaction class or dissatisfaction class (S202). Then, the history class separation unit 203 generates a separation plane for separating the satisfaction class and the dissatisfaction class based on these feature vectors (S202). Next, an orthogonal vector orthogonal to the generated separation plane is calculated by the user characteristic vector calculation unit 204 to obtain a user characteristic vector (S204). The user characteristic vector obtained here is held in the user characteristic vector management table 207a and is used when evaluating the content stored in the distribution information storage unit 205 at the time of distribution of recommendation information.

  Next, content evaluation processing when distributing recommendation information in the present embodiment will be described. FIG. 12 is a sequence diagram when distributing recommendation information. As shown in FIG. 12, a recommendation request including a user ID is transmitted by the recommendation request transmitting unit 103 in the mobile phone 100, and the recommendation request is received by the recommendation request receiving unit 202 in the information distribution server 200 (S301). Then, the information distribution server 200 evaluates the content stored in the distribution information storage unit 205 (S302). Here, the content evaluation unit 206 evaluates according to the user characteristic vector, and generates recommendation information in which the titles of the content are arranged in the order of high evaluation. And the recommendation information produced | generated in this way is delivered by the content transmission part 208 (S303).

  Here, further detailed processing for content evaluation in S302 will be described. FIG. 13 is a flowchart showing content evaluation processing. As shown in FIG. 13, the content evaluation unit 206 acquires a corresponding user characteristic vector from the user characteristic vector management table 207a based on the user ID transmitted together with the recommendation request (S401). Next, the content evaluation unit 206 acquires the feature vector of each content from the content feature vector management table 207b (S402).

  Then, the content evaluation unit 206 calculates the inner product of the user characteristic vector and the content feature vector (S403). The titles corresponding to the content IDs stored in the content management table 205a are extracted in descending order of the inner product calculated by the content evaluation unit 206, and recommendation information is generated by arranging the titles in descending order of evaluation. Then, the recommended information generated is transmitted to the mobile phone 100 by the content transmitting unit 208 (S404).

  Thus, according to the user characteristic vector, recommendation information can be provided, and recommendation information can be provided to the user without omission.

<Generation of separation plane in relative space>
The above generation method of the separation plane is performed in accordance with the feature vector of the content in the absolute space. However, as a method for improving the accuracy, a method using a relative space for which a relative positional relationship is obtained can be considered. Hereinafter, a method for generating the separation plane using the relative space will be described. Note that the processing configuration at this time is the same as that described above, and only the processing content in the history class separation unit 203 is different.

  FIG. 14 shows a configuration example of the menu screen in the history event a and the history event b, and FIG. 14A is a diagram showing the menu screen in the history event a. This menu screen is composed of Node1 to Node3. Yes. FIG. 14B is a diagram showing a menu screen in the history event b, and this menu screen is composed of Node 6 to Node 9. FIG. 14A shows that Node 3 has been selected, and FIG. 14B shows that Node 9 has been selected.

  FIG. 15 is a diagram showing the positional relationship in the feature space of the feature vector of the content selected on each menu screen shown in FIG. FIG. 15A shows the feature space (absolute space) of the content selected in the history event a, and FIG. 15B shows the feature space (absolute space) of the content selected in the history event b. . As described above, the selected content is defined as satisfactory content, and the content not selected is defined as unsatisfactory content, and can be classified into a satisfaction class and a dissatisfaction class, respectively.

  FIG. 16 is a diagram showing the feature vectors of the contents in the same feature space for the respective history events a and b. When the feature vector of the unsatisfactory content is taken as the origin, the feature vector of the satisfactory content is shown. It is a figure showing the relative space which shows a position. For example, for relative comparison between Node 2 and Node 3 and relative comparison between Node 1 and Node 3, a feature space representing the position of Node 3 when Node 1 and Node 2 are set as the origin is used as a relative space. More specifically, the relative space can be formed by subtracting the node 1 feature vector from the node 3 feature vector. Similarly, for other Nodes, a relative space can be formed by performing a process of subtracting the feature vector of the target content.

  In this relative space, as shown in FIG. 16, the separation plane for separating the satisfaction class and the unsatisfied class based on the feature vector of the satisfactory content and the feature vector of the unsatisfactory content as shown in FIG. 203. Then, the user characteristic vector calculation unit 204 calculates a user characteristic vector w that is an orthogonal vector orthogonal to the separation plane.

  As described above, even when the historical events are different, the user characteristic vector can be generated using the relative space, and the user characteristic vector with higher discriminability can be generated.

  FIG. 17 is a diagram illustrating a concept when the evaluation is performed based on the feature vector of each content in the absolute space. For example, Node 6 to Node 8 which are unknown contents stored in the distribution information storage unit 205 are as follows. Evaluation is performed based on the generated user characteristic vector. Here, as described above, the content evaluation unit 206 can calculate the evaluation value by calculating the inner product of the user characteristic vector and the feature vector of each content. Conceptually, in the same way as described above, the priority order of the contents is determined in the order of higher rank in the vector direction at the point projected onto the user characteristic vector, that is, the point where the perpendicular is dropped.

  In this way, a separation plane can be generated using the relative space, and thereby a user characteristic vector can be generated. 15 to 17, a separation plane can be generated even if handled as an absolute space, but a feature space composed of feature vectors that cannot generate a separation plane is also conceivable. In that case, the separation plane can be generated by using the relative space described above. The method will be described below.

  FIG. 18A is a diagram illustrating feature vectors of content in an absolute space. In FIG. 18A, it is assumed that Node1 and Node4 are contents existing in the same event (menu screen), and Node2 and Node3 are contents existing in the same event. In addition, the feature vector (satisfaction class) of the content in which the circular one is selected, and the feature vector (unsatisfied class) of the content in which the triangular one is not selected.

  As is clear from FIG. 18A, in the absolute space, a separation plane cannot be generated for the content of feature vectors having such a positional relationship. This is because the separation plane is a plane for separating the satisfaction class and the dissatisfaction class, and in the example of FIG. Therefore, it is necessary to convert to a relative space representing a relative positional relationship so that a separation plane can be generated.

  FIG. 18B is a diagram showing the feature vector of the content in the relative space. In the example of FIG. 18B, Node 2 is moved in parallel so that Node 1 and Node 2 are at the same position. Node3 corresponding to is translated. By converting to relative space in this way, it is possible to generate a separation plane v that can separate Node 1 (Node 2) that is unsatisfactory content from Node 3 and Node 4 that are satisfactory content.

  As described above, even in the case where the separation plane cannot be generated in the absolute space, the separation plane can be generated by converting to the relative space, and thus the user characteristic vector can be calculated.

<Method of generating separation plane considering browsing frequency>
Next, processing of the history class separation unit 203 when performing weighting processing using content browsing frequency and generating a separation plane using the weighted feature vector will be described. Here, generation of a separation plane using the NN method will be described.

  FIG. 19 is a conceptual diagram illustrating a process for obtaining the center of gravity of a feature vector in consideration of the browsing frequency, and is a conceptual diagram illustrating a feature space representing the positions of the feature vectors of the contents c1 to c3 corresponding to the Nodes A to C. The browsing frequency of the content c1 is 3 times, the browsing frequency of the content c2 is 1 time, and the browsing frequency of the content c3 is 10 times. The browsing history table 207c preferably stores the browsing history of the browsed content and the non-viewing history of the content that has not been browsed, and it is necessary to store each frequency so that it can be held or calculated. It is.

なお、ｃｉは閲覧されたコンテンツの特徴ベクトル、ｎｉは閲覧頻度である。 In this case, the center of gravity of the contents c1 to c3 is calculated by the following equation (1).

Here, ci is the feature vector of the browsed content, and ni is the browsing frequency.

  By performing this calculation, the center of gravity is determined. Further, for the content that is not shown but is not browsed, the center of gravity is obtained using the feature vector of the content that has not been browsed using Equation (1). Then, a straight line connecting the centroid of the browsed content to the centroid of the browsed content can be obtained as the user characteristic vector. This straight line has the same direction as the straight line orthogonal to the separation plane, and the separation plane may be calculated to obtain an orthogonal vector orthogonal to the separation plane. However, in the NN method, the separation is not necessarily performed. A plane is not required.

  Thereby, the user characteristic vector can be obtained using the browsing frequency of the content, and the user characteristic vector close to the user's preference can be obtained as compared with the case where the browsing frequency is not used. Therefore, it is possible to evaluate the content according to the user's preference.

<Method for generating separation plane in consideration of selection order>
Next, processing of the history class separation unit 203 when generating a separation plane in consideration of the order of the browsed contents will be described. In general, it can be considered that the most recently accessed content is more important than the previously accessed content or closer to the user's preference. As this time difference becomes larger, the user's preference changes, and it becomes more prominent that content accessed more recently is important. Here, processing when generating a separation plane in consideration of the time difference (time-series component) will be described.

  FIG. 20 is an explanatory diagram showing processing when the user class vector is generated by the NN method in the accessed time-series order when the history class separation unit 203 generates the separation plane, and FIG. FIG. 20 is an explanatory diagram showing a menu screen at that time, FIG. 20B is an explanatory diagram showing a selection history of content selected by the user, and FIG. 20C is a feature of each content in the feature space. FIG. 20D is a conceptual diagram for explaining the generation of a separation plane in consideration of the selection order. The selection history is stored in the browsing history table 207c.

  As shown in FIG. 20A, a menu screen is configured from Node A to Node F. Here, as shown in FIG. 20B, contents are selected in the order of Node A, Node B, and Node C. Therefore, NodeA, NodeB, and NodeC are contents classified into the satisfaction class, and NodeD, NodeE, and NodeF are contents classified into the dissatisfaction class.

  Next, with reference to FIG. 20C and FIG. 20D, the concept indicating the process when generating the separation plane will be described. As shown in FIG. 20C, Node A to Node F are positioned on the feature space according to their feature vectors. Then, the calculation is performed so that the influence of the feature vector of the most recently selected content is increased. As shown in FIG. 20D, in order to obtain the center of gravity in the satisfaction class, first, the midpoint w1 between Node A and Node B is calculated. Then, a midpoint w2 between the midpoint w1 and NodeC is calculated.

  On the other hand, in the dissatisfied class, the center of gravity w3 is calculated based on NodeD, NodeE, and NodeF. Since there is no concept of selection order in the dissatisfied class, the centroid w3 based on the three nodes is calculated as in the normal processing. Then, a separation plane v in which a line connecting the middle point w2 and the center of gravity w3 is a perpendicular is generated. In the present embodiment, an orthogonal vector orthogonal to the separation plane v is used as a user characteristic vector, which is the same as a straight line connected from the center of gravity w3 toward the middle point w2. For the convenience of explanation in the embodiment, the separation plane is obtained. However, when the NN method is used, obtaining the separation plane is not essential.

  In this way, the midpoint between the feature vector and the feature vector is calculated, the midpoint between the midpoint and the feature vector is calculated, and the midpoint is calculated sequentially in time series (the NN method is applied). Thus, the weight becomes smaller as the previous access history, and the center of gravity can be calculated in consideration of the time series order.

<Apply to operation menu>
By the way, in each of the above-described methods, the menu screen transmitted from the information distribution server 200 which is a website has been described as an example of the menu screen. However, the present invention is not limited to this, and for example, for the operation menu of the mobile phone 100 Then, the recommendation operation information may be displayed as the recommendation information.

  For example, a feature vector is assigned to each operation item in the operation menu, a user characteristic vector related to the user operation is generated according to the user operation history, and when the user issues a request for recommendation operation information, the user characteristic Recommendation operation information may be generated based on a vector. The operation history information is transmitted from the mobile phone 100 to the information distribution server 200 at a predetermined timing and stored.

  FIG. 21 is an explanatory diagram illustrating feature vector parameters in a feature space related to an operation. As shown in FIG. 21, “Chaku-moji”, “Videophone”, “One Seg”, “Screen Memo”, and “Text Memo” are defined as the terminal functions of the mobile phone 100. Also, “telephone”, “Internet”, “text”, “memo”, and “television” are defined as parameters constituting the feature vector. Thus, feature vectors are defined for the functions of the mobile phone 100. For example, in “Chaku-moji”, 1 is assigned to the parameters of “telephone” and “text”, indicating that the function is related to telephone and text. All the terminal function information is stored in the distribution information storage unit 205, and when the user performs an operation to display the operation menu, the fact is notified, and the content evaluation unit 206 evaluates the operation information and selects an appropriate one. The recommendation operation information is generated, and the content transmission unit 208 can transmit the recommendation operation information and display it on the mobile phone 100.

  As described above, by applying the evaluation method of the present embodiment to the operation menu and providing the recommended operation information to the user, it is possible to notify the user of an operation similar or related to the operation frequently used by the user.

  It is also conceivable to perform processing within the mobile phone without connecting to the information distribution server 200. FIG. 22 is a block diagram showing functions of the mobile phone 100a capable of displaying recommendation operation information. This hardware configuration is the same as in FIG.

  The mobile phone 100a includes an operation unit 101a, a recommendation operation display unit 104a, a history class separation unit 105, a user characteristic vector calculation unit 106, an operation information evaluation unit 107, an operation information storage unit 108, and a history storage unit 109. Has been.

  Each configuration will be described below. The operation unit 101a is a part for executing each function mounted on the mobile phone 100a, and includes a numeric keypad, a touch panel, and the like.

  The recommendation operation display unit 104a is a part for displaying recommendation operation information, and is a part for displaying recommendation operation information indicating an operation unknown to the user determined according to the user operation history as will be described later.

  The history class separation unit 105, the user characteristic vector calculation unit 106, the operation information evaluation unit 107, the operation information storage unit 108, and the history storage unit 109 are each a history class separation unit of the information distribution server 200 shown in FIG. 203, a user characteristic vector calculation unit 204, a distribution information storage unit 205, a content evaluation unit 206, and a history storage unit 207. In the history class separation unit 105, the functions and operations operated in the same operation menu are performed. A separation plane that separates the unsatisfied functions into satisfactory functions and unsatisfactory functions is generated, and the user characteristic vector calculation unit 106 sets an orthogonal vector orthogonal to the generated separation plane as a user characteristic vector. The operation information evaluation unit 107 performs evaluation based on each function feature vector in accordance with the calculated user characteristic vector, and edits the menu screen so that a highly evaluated function is displayed on the operation menu. When displaying the operation menu, the user can confirm the highly evaluated function on the operation menu.

  Next, functions and effects of the mobile phone 100 and the information distribution server 200 of this embodiment will be described. In the information distribution server 200, at a predetermined timing, the feature vector generation unit 209 collects content such as a Web page, extracts a keyword indicating the feature of the content, and generates a feature vector. The generated feature vector is stored and updated in the content feature vector management table 207b.

  The collection timing is an arbitrary timing or a timing at which the Web page is updated (for example, by receiving an update notification signal). In this way, collection processing and feature vector generation processing are performed at an arbitrary timing, so that content based on the selection item, for example, Web, follows the latest information on the selection target item, for example, TV program or CD introduction information. A page can be acquired and its feature vector can be generated.

  On the other hand, in the mobile phone 100, when the user operates the browser, a menu screen is displayed, and a plurality of selection target items representing contents such as Node1 to Node5 are displayed. When any one of these contents is selected by the user operation, the mobile phone 100 transmits the selected content and the unselected content to the information distribution server 200, and the information distribution server 200 (content request receiving unit) 201) accepts this. The content request receiving unit 201 acquires the selected content from the content provider, and the content transmitting unit 208 transmits this to the mobile phone 100.

  Also, the browsing history table 207c stores the feature vector of the selected content and the feature vector of the content that has not been selected. The history class separation unit 203 generates a separation plane for separating the feature vector of the selected content stored in the browsing history table 207c and the feature vector of the unselected content in the feature space. Then, the user characteristic vector calculation unit 204 generates a user characteristic vector based on the orthogonal vector orthogonal to the generated separation plane, and stores this in the user characteristic vector management table 207a. Thereafter, when the recommendation request receiving unit 202 receives the recommendation request, the content evaluation unit 206 evaluates the content based on the stored user characteristic vector.

  Thereby, it is possible to evaluate the content according to the vector indicating the user's characteristics, without evaluating the content to be recommended based on the user's hobbies / preferences, direct information such as the access history, Content without omission can be provided to the user. In addition, recommendation information such as contents that the user potentially desires can be distributed without omission. In other words, the feature vector is updated according to the type of content such as the selection target item, for example, TV program or CD introduction information, so that recommendation information specialized to the type of content can be provided to the user.

  Further, in the information distribution server 200 according to the present embodiment, the history class separation unit 203 calculates the relative positional relationship between the feature vectors of all the contents not selected and the feature vectors of the selected contents, It is preferable to generate a separation plane based on the relationship. That is, it is preferable to calculate the positional relationship so that another feature vector is superimposed on the feature vector of the content that has not been selected or selected, to form a relative space, and to generate a separation plane in the relative space. Thereby, since the separation plane can be generated in the relative space representing the relative positional relationship and the user characteristic vector can be generated based on the separation plane, it is possible to evaluate the content with higher accuracy.

  In the information distribution server 200 of the present embodiment, the browsing history table 207c stores the browsing frequency of content and the non-browsing frequency that is not browsed. In addition, you may make it calculate a browsing frequency and a non-browsing frequency as needed based on a browsing history, without memorize | storing such a browsing frequency and a non-browsing frequency. Then, the feature vector of the content with high browsing frequency is weighted according to the frequency, the feature vector of the content with high non-viewing frequency is weighted according to the frequency, and based on the weighted feature vector, A separation plane can be generated for separating the satisfaction class and the dissatisfaction class. As a result, the user characteristic vector can be calculated according to the frequency, the user characteristic vector with higher accuracy can be calculated, and the content can be evaluated more accurately.

  In the information distribution server 200 of this embodiment, the content management table 205a stores the time when each content is selected on the menu screen together with the feature vector of the content, and is weighted according to the degree of newness of the selection time. The separation plane can be generated based on the feature vector of the content. For example, in the NN method, the centroid is obtained in order from the feature vector of the content that is older in time, and as a result, a weighting process according to the degree of newness of the selection time is performed. As a result, it is possible to perform weighting in consideration of time-series components, and it is possible to evaluate content with higher accuracy.

  In the information distribution server 200 of the present embodiment, the content transmission unit 208 transmits recommended information generated according to the evaluation result by the content evaluation unit 206, so that the user can acquire useful information.

  Further, in the information distribution server 200 of the present embodiment, the feature vector generation unit 209 acquires information necessary for updating from a Web page that is a selection item target. Therefore, an appropriate feature vector can be generated.

  In the information distribution server 200 of the present embodiment, the feature vector generation unit 209 includes a feature keyword extraction unit 209b that extracts a characteristic word for each part of speech for each attribute of the selection target item, and the feature keyword extraction unit 209b Based on the extracted keyword, the feature vector calculation unit 209c can calculate a feature vector.

  According to the present invention, a characteristic keyword is extracted for each part of speech for each attribute of the selection target item, and an appropriate feature vector can be generated by generating a feature vector based on the extracted keyword. it can.

  The feature keyword extraction unit 209b may extract a characteristic keyword for each attribute of the selection target item based on the TF / IDF method, or for each attribute of the selection target item based on the PMM method. A characteristic keyword may be extracted.

DESCRIPTION OF SYMBOLS 100 ... Mobile phone, 100a ... Mobile phone, 101 ... Browser, 101a ... Operation part, 102 ... History transmission part, 103 ... Recommendation request transmission part, 104 ... Delivery information display part, 104a ... Recommendation operation display part, 105 ... History class Separation unit 106 ... User characteristic vector calculation unit 107 ... Operation information evaluation unit 108 ... Operation information storage unit 109 ... History storage unit 200 ... Information distribution server 201 ... Content request reception unit 202 ... Recommendation request reception unit , 203 ... history class separation unit, 204 ... user characteristic vector calculation unit, 205 ... distribution information storage unit, 205a ... content management table, 206 ... content evaluation unit, 207 ... history storage unit, 207a ... user characteristic vector management table, 207b ... Content feature vector management table, 207c ... Browsing history Buru, 208 ... content transmission unit, 209 ... feature vector generation unit.

Claims (9)

Feature vector generating means for generating a feature vector constituting the feature space of the selection target item;
Feature vector storage means for storing the feature vector generated by the feature vector generation means;
Accepting means for accepting selection of one selection target item from a plurality of selection target items by a user operation;
A history storage means for distinguishing and storing selected selection target items and unselected selection target items received by the reception means;
The orthogonal vector orthogonal to the separation plane for separating the satisfactory feature vector of the selected selection target item and the unsatisfactory feature vector of the unselected selection target item stored in the history storage means in the feature space. User characteristic vector generating means as a characteristic vector;
Evaluation means for evaluating an unknown selection target item based on the user characteristic vector generated by the user characteristic vector generation means,
The feature vector generation unit updates the feature vector stored in the feature vector storage unit based on the generated feature vector. Separation plane generating means for generating a separation plane for separating in the feature space the satisfaction feature vector of the selected selection target item stored in the history storage means and the unsatisfactory feature vector of the selection target item that has not been selected. Further comprising
The recommendation information evaluation apparatus according to claim 1, wherein the user characteristic vector generation unit generates a user characteristic vector according to the separation plane generated by the separation plane generation unit.   The recommendation information evaluation apparatus according to claim 1, wherein the feature vector generation unit acquires information necessary for updating from document information that is a selection item target.   The recommendation information evaluation apparatus according to claim 3, wherein the document information is a Web page. The feature vector generation means further includes feature extraction means for extracting a characteristic word for each part of speech for each attribute of the selection target item,
The recommendation information evaluation apparatus according to claim 1, wherein a feature vector is generated based on the word extracted by the feature extraction unit.   The recommendation information evaluation apparatus according to claim 5, wherein the feature extraction unit extracts a characteristic word for each attribute of the selection target item based on a TF / IDF method.   The recommendation information evaluation apparatus according to claim 5, wherein the feature extraction unit extracts a characteristic word for each attribute of the selection target item based on a PMM method.   The recommendation information evaluation apparatus according to claim 1, wherein the feature vector generation unit generates a feature vector at an arbitrary timing. A feature vector generation step for generating a feature vector constituting the feature space of the selection target item;
A feature vector storage step of storing the feature vector generated in the feature vector generation step in a feature vector storage means;
A reception step for accepting selection of one selection target item from a plurality of selection target items by a user operation;
A history storage step of storing in the history storage means the selected selection target item and the non-selected selection target item received in the reception step;
An orthogonal vector orthogonal to a separation plane for separating the satisfaction feature vector of the selected selection target item and the unsatisfactory feature vector of the selection target item not selected, stored in the history storage unit, in the feature space. Generating a user characteristic vector based on the user characteristic vector generating step;
An evaluation step for evaluating recommendation information based on the user characteristic vector generated by the user characteristic vector generation step, and
In the feature vector generation step, the feature vector stored in the feature vector storage unit is updated based on the generated feature vector.

Images