JPH11161670A - Method, device, and system for information filtering - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide the high-precision information filtering device which can automatically reflect a user's interest. SOLUTION: A profile management part 17 generates a user profile from a correlation matrix obtained by dimension deletion from a feature vector set showing features of an input document. A filtering process part 18 calculates the projection of feature vectors by documents and corresponding user profiles and filters the document on the basis of the calculated values. A correlation matrix is generated again from a feature vector set as filtering results and the profile management part 17 updates the corresponding user profiles. Further, the corresponding user profiles are put together to generate a common profile and when a user profile is updated, the corresponding common profile is updated.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for filtering digitized information having a theme of high interest to a user from a large amount of digitized information.

[0002]

2. Description of the Related Art In recent years, with the spread of large-scale and high-speed networks typified by the Internet, an environment has been provided in which end users can easily obtain various types of electronic information in various forms. However, while the promotion of computerization of information plays a part in the information society, the enormous amount of information has the adverse effect of exceeding the amount that can be managed by humans. Development of the system is desired. In addition, with the distribution of digitized information,
There is a need to select only necessary information from a large amount of electronic information. The selection operation in this case is too burdensome to be performed manually, and thus is automated by a computer device, for example, information that automatically selects a large amount of electronic information that flows in according to a theme of interest to the user. Discussions have been made on filtering methods.

In general, in the information filtering method, in order to quantify the degree of interest of a user and perform computer processing, a reference vector (user profile vector,
A user profile, or simply a profile) is used. The user profile is obtained, for example, in advance, for each word, the appearance frequency of a word included in a set of a plurality of text data included in digitized information of interest to the user, and a dimension corresponding to the obtained word type, for example, the word If there are 10 types, the vector is converted into a 10-dimensional vector, and this is normalized.

A subspace method (also called a subspace classification method) is known as one form of a pattern recognition method using vectors. This method is a statistical method in which a category to be classified is determined through projection onto a subspace formed from the distribution of feature vector components. The eigenvector calculation of the vector component to be converted in this case includes, for example,
Karhunen Rebe, a quantization algorithm
KL analysis by en-Loeve) conversion is employed. Representative methods in this subspace method include CLAFIC (C
LAss-Featuring Information Compression method, Averaged Learning Sub-space Metho
d, ALSM). CLAFIC method and ALSM
ALSM is an adaptive learning method that has similar classification criteria and also considers opposing categories. The subspace method is described in detail, for example, in "Pattern Recognition and Subspace Method" (Erkki Oya, Sangyo Tosho).

When actually performing information filtering,
A threshold is set in advance so that the user is determined to be interested, and by referring to the user profile based on the threshold, the target digitized information group is ranked in descending order of similarity. Is done. For example, a predetermined number of digitized information is selected from the top of the rank,
Presented to the user.

[0006]

By the way, in general, a filtering error is a "leakage" that drops necessary information.
It is well known that there is a trade-off between these and "noise" that takes in unnecessary information. However, the conventional filtering is a one-sided filtering that focuses only on the degree of interest of the user, that is, “necessary information”, and considers only a decrease in “leakage” of information. for that reason,
Since the removal of "noise" was not directly considered, there was a limit in improving the filtering accuracy.

In addition, the conventional filtering has a limitation that it is not possible to flexibly cope with a case where there is a plurality of interests of the user or a temporal change of the interest. Specifically, tasks and hobbies are long-term interests of the user, and incidents are temporary interests, but with conventional filtering, the degree of interest must be determined by uniform keyword input, etc. This makes automatic filtering in response to changing user interests impossible.

Accordingly, an object of the present invention is to enable automatic creation of a user profile, which is information of a user's interest, by a learning function that follows a changing user's interest without the need for the user to set and evaluate the profile. An object of the present invention is to provide an information filtering method capable of maintaining the accuracy of filtering at a certain value or more. Another object of the present invention is to provide an information filtering device suitable for implementing the above information filtering method.

[0009]

In order to solve the above-mentioned problems, the present invention extracts a learning vector in which redundant dimensions are reduced from digitized information to which identification information for identifying whether or not the user is interested is added. Then, by applying a predetermined subspace classification criterion to this learning vector, "interesting" or "not interested"
The process of creating a user profile corresponding to any of the categories, and when new digitized information to be sorted is input, a target vector representing the feature of the new digitized information is extracted, and this target vector and The feature difference from the created user profile is obtained by projecting the feature profile into a subspace corresponding to the user profile, and based on this feature difference, the new digitized information is either "interesting" or "not interested". Providing an information filtering method including a step of selecting into categories and a step of creating an update profile in the same format as the user profile from the digitized information after the selection and updating the user profile using the update profile. .

[0010] Preferably, the information filtering method further includes a step of integrating a plurality of mutually related user profiles to create a shared profile having a shared relationship with each user profile. In this case, the selecting step selects the new digitized information into one of the categories of “interested” or “not interested” based on a characteristic difference from the shared profile or the user profile.

According to another aspect of the present invention, there is provided an information filtering apparatus for extracting a vector from which redundant dimensions have been deleted from the characteristics of digitized information, and identifying whether or not the user is interested. A predetermined subspace classification criterion is applied to the learning vector extracted by the vector processing means from the digitized information to which the identification information is added to correspond to either the "interesting" or "not interested" category. A profile creating unit for creating a user profile to be performed, and a feature difference between the target vector extracted by the vector processing unit from the new digitized information to be sorted and the user profile to a subspace corresponding to the user profile. It is obtained by projection, and based on this feature difference, the new digitized information is "interested" or "not interested".
And a new learning vector is extracted from the result of the selection by the selection unit, and is guided to the profile creation unit.

[0012] More preferably, the apparatus further comprises a user management means for managing the user profile for each user, and when the user management means recognizes that this is the first selection by a new user, an initial profile for the new user is provided. The configuration data is interactively fetched, and the profile creation means is configured to create the user profile for the new user.

The profile creation means creates a correlation matrix from the extracted learning vectors based on a subspace classification criterion or based on adaptive learning conditions of a predetermined average learning subspace method. It is constituted so that. The former is effective when a new user profile is created, and the latter is effective when updating a user profile. Further, the vector processing means is configured to perform KL analysis by orthonormal transform to reduce the redundant dimension.

Another information filtering apparatus according to the present invention comprises:
When creating and saving a shared profile that forms a shared relationship with the user profile before integration by integrating a plurality of the user profiles that are related to each other, when at least one of the user profiles related to the shared profile is updated, A shared profile processing unit for reflecting the update in the shared profile;
The selection means performs the selection by selectively using the user profile or the shared profile. Note that the shared profile processing means includes:
For example, for each of a plurality of user profiles that are integration candidates, a distance value focusing on the difference between the presence or absence of interest is calculated,
It is determined whether or not to integrate from the sum of the distance values.

According to another aspect of the present invention, there is provided an information filtering system, wherein the information filtering device is connected to a communication line, and digitized information circulated through the communication line is taken into the information filtering device. Things. In this case, it is preferable that the information filtering device is configured to perform filtering of the digitized information captured through an agent unit.

[0016]

Embodiments of the present invention will be described below in detail with reference to the drawings. (First Embodiment) FIG. 1 is a functional block diagram of an information filtering apparatus to which the present invention is applied. In the figure, the solid line represents the flow of processing, and the broken line represents the flow of data. The information filtering apparatus 1 includes, for example, a document database (not shown) and a profile database (hereinafter, profile DB) 20 built inside or outside a stand-alone computer device, and the computer device reads and executes a predetermined program. The data input unit 11, the user management unit 12, the feature vector extraction unit 13, the dimension processing unit 14, the processing selection unit 15,
It comprises a correlation matrix creation unit 16, a profile management unit 17, a filtering processing unit 18, and a result output unit 19. Although not shown, a display device equipped with a setting interface for interactively capturing initial profile setting data and grouping reference setting data, which will be described later, an input device for capturing a document, and a filtering result are output. An output device is also provided.

The above program is usually stored in an internal storage device or an external storage device of the computer device, and is read and executed as needed. However, a recording medium separable from the computer device, for example, a CD-ROM. RO
It is recorded on a portable recording medium stored in an M or FD or a program server connected to a private network, read at the time of use, installed in the internal storage device or the external storage device, and provided for execution at any time. It may be something.

The profile DB 20 stores user interest information on digitized information (hereinafter, singular or plural documents unless otherwise specified), ie, a user who represents a category of “interested” or “not interested”. A profile and a shared profile having a shared relationship with a plurality of user profiles are stored for each user. The user profile and the like are created based on the above-mentioned subspace method, and are updated each time a user accesses and performs selection (filtering).

The data input unit 11 takes in the document when it is accessed by the user, that is, when the user reads the document through the input device such as a scanner, and inputs the document to the user management unit 12. is there. The user management unit 12 manages a user profile and the like for each user, and determines whether or not the user is a pre-registered user in response to access from the user, and performs subsequent processing according to the determination result. Perform selectively.

In the case of access by a new user, after assigning initial profile setting data to the document input from the data input unit 11, this is input to the feature vector extracting unit 13. Further, the processing selection unit 15 is notified that the access is from a new user. The initial profile setting data is identification information of “interested” or “not interested” of the new user interactively acquired through the setting interface. On the other hand, in the case of access by a registered user, the input document is input to the feature vector extraction unit 13 and the identification information of the user is notified to the process selection unit 15. The user management unit 12 also sets a grouping criterion for creating a shared profile, which will be described later, used in the profile management unit 17. This grouping criterion is a criterion indicating whether or not the profiles are related to each other, for example, which users have common interests and which are set in advance through system parameters. Individual users can also explicitly set through the setting interface.

The feature vector extracting unit 13 extracts a vector (a set of feature vectors or one vector representing the set) representing the features of the input document. Specifically, the type of a keyword (hereinafter, word) appearing in the document is set as the number of dimensions, and one or a plurality of vectors obtained by weighting the appearance frequency of each word are obtained by arithmetic processing.
In this case, the weighting of the word is performed using a well-known “TF / IDF”.
Method. The extracted vector is
It is input to the dimension processing unit 14.

The dimension processing unit 14 includes the feature vector extraction unit 1
The vector extracted in Step 3 is subjected to the above-described KL analysis, that is, principal component analysis by orthonormal transform, to reduce redundant dimension vectors having relatively low weights (also referred to as dimension compression). . The vector whose dimension has been reduced is input to the processing selection unit 15.

The processing selection section 15 selects the subsequent processing for the vector whose dimension has been reduced by the dimension processing section 14. More specifically, a profile creation process is selected when a new user accesses, a filtering (selection) process is performed when there is a remaining input document in the data input unit 11, and a profile update process is performed after filtering is completed. To do.

When the process selected by the process selecting unit 15 is profile creation or profile update, the correlation matrix creation unit 16 creates a correlation matrix corresponding to each dimension-reduced vector.

The profile management unit 17 mainly includes a profile DB 20 for profiles and shared profiles.
To save, read, and update. That is, the correlation matrix created by the correlation matrix creating unit 16 is stored as a user profile indicating the interest of the user in the profile DB 22 for each user, and this profile is stored in the profile DB 20 at the time of filtering.
Is read as appropriate. In the profile DB 20, a plurality of user profiles corresponding to the grouping criterion set by the user management unit 12 are integrated (sum of a plurality of corresponding correlation matrices) to create a shared profile of the group identification information. If the new or updated user profile is related to the shared profile, the corresponding shared profile is updated with the information of the user profile. By providing a shared profile, it is possible to provide information owned by users and to share information of interest between these users. In other words, even if there is interest information that is not included in the specific user profile, if the interest information exists in the shared profile, it can be complementarily referred to.

When a filtering process is selected by the process selecting unit 15, the filtering processing unit 18 uses the user profile of the user stored in the profile DB 22 or the shared profile to which the user belongs to to save the document (actually, the dimension is reduced). Vector). The filtering result is output to the result output unit 1.
9 to an output device or a document database. The result output unit 19 also has a function of feeding back the data to the data input unit 11 in order to reflect the filtering result on the existing user profile.

Next, the concept of profile and filtering according to the present invention will be described. FIG. 2 is a processing procedure diagram of the profile creation. The profile in this case is created based on the document input by the new user, and the document is provided with the identification information of “interested” or “not interested” as described above. Note that a new user is a user who is already managed by the user management unit 12, but includes a case in which there is no access from the user and filtering is performed for the first time.

When a document is input from a new user (step S101), the information filtering apparatus 1 sets the type of words appearing in this document as the number of dimensions and weights the frequency of appearance of each word (characteristic). Vector set)
Is extracted (step S102). The above-described dimension reduction is performed on the extracted vector (step S10).
3), a correlation matrix is created from the vector thus obtained (step S104). The correlation matrix is created based on a predetermined subspace classification criterion, that is, a criterion based on a subspace method used when performing pattern recognition of a vector space model. For this subspace classification criterion, for example, the description in “Pattern recognition and subspace method” (Erkki Oya, Sangyo Tosho) can be referred to. This correlation matrix is stored as a user profile (step S105). As described above, by using the correlation matrix as the user profile, the interest information indicating the interest of the user is an expression focusing on the co-occurrence relationship between words appearing in the document.

FIG. 3 is a flowchart of the filtering process. When a document is input (step S201), the information filtering apparatus 1 performs the same processing as that in steps S102 to S103 on the document to extract a dimension-reduced vector (steps S202 to S203).
Then, the above-described KL analysis is performed on the extracted vector and the user profile or the shared profile read from the profile DB 20 using the user identification information as a search key to calculate eigenvalues and eigenvectors, respectively. The document is sorted by calculating and extracting the projection of (step S204).

FIG. 4 is a processing procedure diagram of the profile update. The document input in this case is a filtering result obtained by the processing procedure of FIG. This filtering result is also provided with identification information of “interesting” or “not interested”. When a document as a filtering result is input (step S301), the information filtering apparatus 1 performs the above-described step S10 on the document.
The same processing as in steps 2 to 103 is performed to extract the dimension-reduced vector (steps S302 to S303). Thereafter, a correlation matrix is re-created from the extracted vectors based on the adaptive learning conditions of ALSM (step S30).
4). Further, the correlation matrix corresponding to the user profile already stored and the corresponding subspace are updated with the recreated correlation matrix (step S305). In this way, the user profile is automatically updated for each feedback.

Next, the sharing profile will be described in more detail. The creation, filtering, and updating of the shared profile can also be performed basically according to the processing procedures of FIGS. Here, description will be made focusing on creation of a shared profile in consideration of both “interesting” and “not interested” given to a document. Note that the category corresponding to the identification information “interest” is “positive interest”, and the category corresponding to “non-interest” is “negative interest”. When integrating a plurality of related user profiles, it is important to consider the degree of similarity between the user profiles, that is, the magnitude of the distance value, in order to effectively create a shared profile.

For example, when extracting the distance value of the user profile A from the document profile D with respect to the user profile B, a common file is created with high accuracy by focusing on “positive interest” and “negative interest”. can do. This point will be described. When the document K is included in the document set D, first, regarding the document K in the profile B, the difference between the projection value of “positive interest” and the projection value of “negative interest” is calculated (first projection difference value). ). Next, for the document K in the profile A, a difference between the projection value of “positive interest” and the projection value of “negative interest” is calculated (second projection difference value). Further, a difference between the second projection difference value of the profile A and the first projection difference value of the profile B is calculated.
A value obtained by squaring the calculated value is defined as a difference distance. This difference distance is expressed as a Euclidean distance and is a difference between the presence or absence of interest in the user profiles A and B for the document K.

Next, similarly, the difference distance is set to the document set D
Is calculated for all the documents included in, and the sum of these difference distances is calculated. The value calculated by this sum is the distance value of the user profile A to the user profile B in the document set D, and is a measure of the difference between the user profiles when adding documents having similar degrees of interest in the order of integration, that is, when integrating the documents. It is what becomes. When the distance value is large, the degree of similarity between the target user profiles is small, while when the distance value is small, the degree of similarity between the user profiles is large. When a shared file is created, a configuration in which the similarity is within a predetermined range (for example, one having a predetermined threshold or less) is integrated.

Next, an information filtering method using the information filtering apparatus 1 will be described with reference to FIGS. First, an example of creating an initial profile will be described. A document is input through the data input unit 11,
After confirming that the user is a new user (step S40)
1, S402), the user management unit 12 adds the data for initial profile setting to the document, and sends the document to the feature vector extraction unit 13. The feature vector extracting unit 13 extracts a vector from this document (step S403). further,
The dimension processing unit 14 reduces redundant dimensions from the extracted vector (step S404). The correlation matrix creation unit 16
A correlation matrix is created from the reduced vector (step S406) and sent to the profile management unit 17. The profile management unit 17 uses this correlation matrix as a user profile (initial profile)
22 (step S407).

Referring to FIG. 6, when the creation of the initial profile is completed, it is checked whether a shared profile to which the user belongs is set. If the shared profile has been set (step S408: Yes),
It is checked whether or not the shared profile already exists. If not (step S409: No), the shared profile processing unit 21 newly creates a corresponding shared profile (step S410). If the shared profile already exists (step S409: Yes),
The existing shared profile is updated with the information of the initial profile (step S411). If there is a next document, the process returns to step S401 (step S412: Ye
s) Repeat the above process. If only the initial profile is to be created, the process ends (step S412: N
o).

Returning to FIG. 5, when filtering is performed using this initial profile, an updated and saved user profile, or a shared profile (hereinafter simply referred to as a profile), a document to be sorted is selected. Perform the processing of steps S402 to S403,
The vector thus obtained is filtered by the filtering processing unit 1
Send to 8. The filtering processing unit 18 calculates the projection of this vector and the corresponding profile (Step S).
413), document selection is performed based on the calculation result (step S414). The selection result is presented to the user through the result output unit 19. The process returns to step S401 for updating the profile (step S415).

In the profile update, the processes of steps S402 to S403 are performed on the document as the selection result, and the vector obtained by this process is sent to the correlation matrix creating unit 16. The correlation matrix creation unit 16 re-creates a correlation matrix based on this vector (step S416). The profile management unit 17 updates the corresponding profile with the re-created correlation matrix (step S417). The processing related to the shared profile (see FIG. 6) is the same as the case of creating the initial profile.

In this embodiment, the correlation matrix is a user profile. This focuses on the co-occurrence of words in the input document, and reduces the processing load when synthesizing each user profile as a shared profile. It is intended for. However, in addition to this method, for example, it is also possible to perform KL analysis on the correlation matrix and create the resulting vector space as a user profile. The amount of information in this case is almost the same as the case where a correlation matrix is used for a user profile, but processing for unifying the number of dimensions is required when combining each user profile as a shared profile.

(Second Embodiment) The present invention functions as a system for automatically filtering a large amount of electronic information circulated through a public network such as the Internet as a communication line, for example, functions as the information filtering device. It is also possible to implement the present invention in the form of an information filtering system in which an information filtering server that performs the operation and a client having a function of acquiring information from a public network are deployed.

In this case, the information filtering server:
For example, it can be positioned as an information providing support server that selects and provides optimum information to a client from computerized information from various information providing servers including a plurality of large-scale databases on the Internet environment, that is, an information navigation server. it can.

As an example of the configuration in this case, the same database as the above-mentioned profile DB 20 is constructed in the internal or external storage device of the computer device, and the communication is performed between the client and the various information providing servers via the public network. A control unit is provided. Further, functional blocks similar to those of the information filtering apparatus 1 described above, that is, the data input unit 11, the user management unit 12, the feature vector extraction unit 1
3, a dimension processing unit 14, a processing selection unit 15, a correlation matrix creation unit 16, a profile management unit 17, a filtering processing unit 18, and a result output unit 19.

This information filtering server differs from the information filtering device 1 in that it includes a known communication control unit for performing communication control. By inputting the digitized information group distributed through the communication control unit to the data input unit 11 and receiving an information acquisition request from a client, information filtering using a network becomes possible. In this case, the information acquisition request can be input using, for example, a browser in a WWW environment as an interface. Also, the computerized information group from the various information providing servers is configured to always pass through the information filtering server, and to provide the client with the selection result for the computerized information group via the communication control unit. Further, the information filtering server, for example, by integrating with the agent technology of the server in the Internet environment, it is possible to construct a system for automatically filtering a large amount of computerized information distributed.

Next, the result of actually performing an evaluation experiment in the information filtering device 1 of the present embodiment will be described. In this experiment, as a document for evaluation,
Using newspaper articles written in Japanese for one year from January to October 1996, the first half of the year was used as a document set of initial profile setting data (hereinafter referred to as a training set), and the second half was used for evaluation experiments. Document set (hereinafter referred to as evaluation set).

These articles belong to any of the five genres shown in FIG. 7, and the number of articles in the figure represents the number of data for each genre. Each article is composed of a title and a text, and is given a genre name to which the article belongs. In this experiment, words were specified, and those containing the words in the training set were divided into "interested" articles and others were classified as "uninterested" articles for convenience to generate user profiles. Here, these words are referred to as "topics" as items of interest, and an experimental result of filtering an evaluation set with a profile based on the corresponding topic is shown.

FIG. 8 shows titles of 25 articles out of 1225 articles obtained by filtering the word “import” into topics. According to the result, it can be seen that articles related to the word “trade” having a similar meaning to the topic “import” and articles related to the word “illness” which is not necessarily closely related in vocabulary are relatively conspicuous. This is an "import"
It is interpreted that "infection" was also considered to be a matter of interest because many of the articles containing the word "infection" among articles containing the word "infection" at the same time. For this reason, according to the information filtering device 1 of the present embodiment, even if the user does not clearly recognize the content and number of items of interest, the target article can be automatically obtained through the connection of words in the learned article set. It can be seen that it is extracted.

FIGS. 9 and 10 show the distribution of articles including the words “professional baseball” and “earthquake” in the articles filtered by the word “Hanshin” as a topic in the “positive interest” subspace. It is represented by values up to the third component. In this case, the number of articles including the words “professional baseball” and “earthquake” is “1064” and “123” respectively in 46 articles filtered by “Hanshin” as a topic.
It is. Each point in the figure corresponds to each article, and represents a mapping result to each component.

In the distribution results of FIG. 9, the first component and the second component have a spread, and in particular, the third component has an uneven distribution. On the other hand, in the distribution result of FIG. 10, all the third components have values close to zero, and have a contrasting distribution that greatly spreads in the plane of the first and second components. Therefore, correspondence between the word connection in the article and the axis of the subspace is taken, and it can be seen that this enables filtering of a plurality of interests as described above.

As described above, in the information filtering apparatus 1 according to the present embodiment, the co-occurrence relation of words in a document is introduced, and filtering is performed based on a classification criterion in consideration of both interest and non-interest. Compared with such one-sided information filtering, it is possible to obtain a selection result with higher accuracy.

Further, since the created correlation matrix is used as a user profile, it is possible to obtain a filtering according to their relations by a single user profile, without being limited to the number and contents of the user's interests. And the extent of interest of the user can be known from the number of dimensions in the subspace.

Also, since adaptive learning for the correlation matrix can flexibly follow changes in the user's interest without changing the expression of the user profile, automatic filtering corresponding to temporal changes in interest can be performed. Becomes possible.

Further, since a common profile can be generated from the sum of the correlation matrices in each user profile, and the items of interest can be shared among the users, the information is omitted because it is based on the interest information specified and narrowed down to the individual user. Complementary filtering can be performed on information that is interest information close to the user or information that cannot be extracted from the interest information of the user.

Further, from the information of each individual user profile, the system operation manager such as a service provider can grasp the trend of the interest of the entire user, and, for example, public notice or direct mail according to the target user. It can be used as a research material for direct marketing and marketing research for new product development.

Further, it is easy to construct a system that operates independently of a plurality of existing information providing service systems and the like and to incorporate the system into an existing system.

[0054]

As is clear from the above description, according to the present invention, there is an effect that a user profile, which is information of interest of a user, is automatically created. Further, at the time of information filtering, both "interesting" and "non-interesting" are considered as the user's interest, so that the accuracy of the filtering can be maintained at a certain value or more.

When the present invention is applied in a network environment, by this information filtering, it is possible to reliably and timely obtain necessary information based on the user's interest from a large amount of continuously flowing computerized information. Because it can be obtained, effective use of information is promoted. From this,
It is possible to provide a system in which access efficiency and practicality are significantly improved.

[Brief description of the drawings]

FIG. 1 is a functional block diagram of an information filtering device according to an embodiment of the present invention.

FIG. 2 is a processing procedure diagram of profile creation.

FIG. 3 is a processing procedure diagram of filtering.

FIG. 4 is a processing procedure diagram at the time of updating a profile.

FIG. 5 is an explanatory diagram of a procedure of a filtering method by the information filtering device of the present invention.

FIG. 6 is an explanatory diagram of a procedure of a filtering method by the information filtering device of the present invention.

FIG. 7 is a table showing evaluation experiment data of the information filtering device of the present invention.

FIG. 8 is a list of titles of an evaluation set filtered by the topic “import”.

FIG. 9 is a distribution result of articles including the word “professional baseball” filtered on the topic “Hanshin”.

FIG. 10 is a distribution result of articles including the word “earthquake” filtered on the topic “Hanshin”.

[Explanation of symbols]

 Reference Signs List 1 information filtering device 11 data input unit 12 user management unit 13 feature vector extraction unit 14 dimension processing unit 15 processing selection unit 16 correlation matrix creation unit 17 profile management unit 18 filtering processing unit 19 result output unit 20 profile DB (database)

Claims (11)

[Claims] 1. A learning vector in which redundant dimensions are reduced is extracted from digitized information to which identification information for identifying whether or not a user is interested is provided, and a predetermined subspace classification criterion is applied to the learning vector. "Interested" or "not interested"
A process of creating a user profile corresponding to any of the categories, and when new digitized information to be sorted is input, a target vector representing the feature of the new digitized information is extracted,
The characteristic difference between the target vector and the created user profile is obtained by projecting the characteristic vector into a subspace corresponding to the user profile. Based on the characteristic difference, the new digitized information is determined as “interested” or “not interested”. Information filtering, comprising the steps of: selecting from any of the categories: creating an update profile in the same format as the user profile from the digitized information after the selection; and updating the user profile using the update profile. Method. 2. The method according to claim 1, further comprising the step of: integrating a plurality of mutually related user profiles to create a shared profile having a shared relationship with each of the user profiles, wherein the selecting step includes the step of combining the shared profile or the user profile. 2. The information filtering method according to claim 1, wherein the newly digitized information is classified into one of "interesting" and "not interested" categories based on the feature difference. 3. Vector processing means for extracting a vector from which redundant dimensions have been deleted from the characteristics of the digitized information, and said vector processing from the digitized information to which identification information for identifying whether or not the user is interested is added. A profile creating means for creating a user profile corresponding to either the "interesting" or "not interested" category by applying a predetermined subspace classification criterion to the learning vector extracted by the means; A characteristic difference between the target vector extracted by the vector processing means from the new digitization information and the user profile is obtained by projecting the feature vector into a subspace corresponding to the user profile, and based on the characteristic difference, the new digitization information is obtained. And selecting means for selecting the category of “interested” or “not interested”. An information filtering device configured to extract a new learning vector from the information and guide the extracted learning vector to the profile creation unit. 4. The apparatus according to claim 1, further comprising a user management unit for managing the user profile for each user, wherein when the user management unit recognizes that the selection is an initial selection by a new user, the user management unit sets an initial profile for the new user. 4. The information filtering apparatus according to claim 3, wherein the apparatus is configured to interactively fetch data to cause the profile creation means to create the user profile for the new user. 5. The information filtering apparatus according to claim 3, wherein the profile creation means is configured to create a correlation matrix from the extracted learning vectors based on a subspace classification criterion. 6. The profile generating means is configured to generate a correlation matrix from the extracted learning vectors based on adaptive learning conditions of a predetermined average learning subspace method. The information filtering device according to claim 3, wherein 7. The vector processing device according to claim 3, wherein said vector processing means is configured to perform KL analysis by orthonormal transform to reduce said redundant dimension. Information filtering device. 8. A plurality of user profiles related to each other are integrated to create and store a shared profile having a shared relationship with a user profile before integration, and at least one of the user profiles related to the shared profile is updated. The system further comprises: a shared profile processing unit that reflects the update to the shared profile when the update is performed, wherein the selection unit performs the filtering by selectively using the user profile or the shared profile. 8. The information filtering device according to any one of items 3 to 7. 9. The shared profile processing means calculates a distance value focusing on a difference in presence or absence of interest for each of a plurality of user profiles as integration candidates, and determines whether or not to integrate from the sum of the distance values. 9. The information filtering apparatus according to claim 8, wherein the information filtering apparatus is configured as follows. 10. An information filtering apparatus according to claim 3, wherein the information filtering apparatus is connected to a communication line, and digitized information circulated through the communication line is taken into the information filtering apparatus. Information filtering system. 11. The information filtering system according to claim 9, wherein the information filtering device is configured to perform filtering on the digitized information captured through an agent means.