JP2002269143A - System, method and program for filtering information - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an information filtering system that efficiently realizes accurate information filtering on the basis of both contents attributes and different users' evaluated values to the contents. SOLUTION: This information filtering system is provided with a storing part 30 provided with data of evaluated values showing the evaluation of a user to each of contents and attribute data being data representing the contents of each of the contents by preliminarily recording the data of evaluated values and the attribute data, an attribute compressing and converting part 22 for calculating compressed attribute data obtained by removing parts that do not affect evaluation from the attribute data of each of the contents on the basis of the attribute data and evaluated value of each of contents, an evaluation characteristic calculating part 23 for calculating user evaluation characteristic data showing the characteristic of evaluation shown by each user and contents evaluation characteristic data on the basis of information of both the compressed attribute data and the evaluated value, and an evaluation predicting part 24 for predicting and calculating evaluation given to each of the contents by each user on the basis of the user evaluation characteristic data and the contents evaluation characteristic data.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to information filtering for selecting / recommending contents, merchandise, and the like that match the interests and preferences of a user. The present invention relates to an information filtering system, a filtering method, and an information filtering program that perform filtering based on both.

[0002]

2. Description of the Related Art Conventionally, in an information filtering system which is a system for selecting and recommending contents, products, and the like that match the interests and tastes of a user, a database or WW
A large amount of information and documents provided by an information system such as W (World Wide Web), a large number of services and products, and the like (hereinafter, these are referred to as contents) are to be searched.

[0003] A conventional information filtering system selects contents required by a user by estimating an evaluation value that the user will give to a content not yet evaluated by the user. In general, a method for predicting an evaluation given to a content by a user includes CBF (Constant Bass)
ed Filtering) method and SIF (Social Information Fil)
tering) method.

[0004] The CBF method uses data such as a keyword and a word frequency included in content (hereinafter, this data is referred to as an attribute in this specification).
This is a method of filtering the contents by expressing the contents of each content and matching the interests and preferences of each user with those expressed by weights of keywords.

[0005] In the SIF system, an evaluation value indicating the degree of interest or preference for each content by each user (hereinafter, this data is referred to as evaluation in the present specification) is stored in advance. , Comparing the evaluations given by the user (requester of information filtering) in the past with the evaluations of users other than the requester, searching for and finding users who have similar preferences to the requester, and who have similar preferences In this method, content that has been highly evaluated by a user is selected and recommended to a requester, thereby filtering the content.

Conventionally, there are two types of devices for improving the accuracy of such information filtering.

[0007] First, there is a method of performing a combination of filtering methods of the CBF method and the SIF method. That is, each content is searched and extracted by both the CBF method and the SIF method.

[0008] Second, similar data is referred to for filtering, and data that is not useful for filtering is deleted (hereinafter, this processing is referred to as compression conversion). For example, if compression conversion is used in CBF, information filtering is performed by identifying and deleting useless data from data used for filtering of attributes of each user or content, and performing filtering using only important data. Accuracy can be improved. Also, due to the similarity of compression conversion, words having similar meanings such as "computer" and "computer" can be combined into one, thereby reducing filtering omissions and increasing the accuracy of information filtering. be able to.

[0009] An example of an information filtering system incorporating this compression conversion process is described in "IJCIA 99 Workshop, Machine Learning for Information Filtering, 86-91, August 1999". Page (IJCAI99 Workshop Machine L
earning for Information Filtering, pp.86-91, Augus
t, 1999) ".

Here, an information filtering system described in this document will be described in detail with reference to the drawings. FIG. 7 is a block diagram showing the configuration of this conventional information filtering system 100b.

Referring to FIG. 1, this conventional information filtering system 100b includes an input / output unit 10, a data processing unit 40, and a storage unit 50.

The input / output unit 10 includes a request input unit 11, a content presentation unit 12, and an evaluation input unit 13.

The data processing unit 40 includes an attribute extraction unit 21, a profile learning unit 41, a profile conversion unit 42, an attribute conversion unit 43, an evaluation prediction unit 44, and a content selection unit 25.

The attribute extracting unit 21 extracts a characteristic attribute from the content data 31 stored in the storage unit 50, and stores this as attribute data 32 in the storage unit 50.

The profile learning unit 41 uses the data of the evaluation value 33 based on the evaluation given to the content by the user and the attribute of the content evaluated by the requester among the attributes of the content in the attribute data 32. Then, a profile representing the relationship between the requester's evaluation and the attribute is learned, and the result is stored in the storage unit 50 as profile data 51.

The profile conversion section 42 compresses and converts the user's profile data 51 and stores the result in the storage section 50 as conversion profile data 54. Further, the profile conversion unit 42 stores, in the storage unit 50, information specifying a method of compressing and converting a profile, that is, which attributes are to be combined and which attributes are to be deleted, as conversion method data 52.

The attribute conversion unit 43 performs compression conversion of the attribute of each content indicated as the attribute data 32 by using the conversion method specified by the conversion method data 52.
The compressed attribute data is stored in the storage unit 50 as the conversion attribute data 53.

When an information filtering request of a requester is received from the request input unit 11 of the input / output unit 10, the evaluation prediction unit 44 converts conversion profile data 54, which is profile data obtained by compression conversion of the requester, and By comparing the converted attribute data 53 which is compressed attribute data of the content, the evaluation value of each content with respect to the requester is predicted and the value is calculated.

The content selection section 25 includes an evaluation prediction section 44
From the registered content recorded as the content data 31, a content having a high predicted evaluation value is selected, and is requested through the content presentation unit 12 of the input / output unit 10. To provide.

The storage unit 50 stores content data 31, attribute data 32, evaluation values 33, profile data 51,
It includes conversion method data 52, conversion attribute data 53, and conversion profile data 54.

The content data 31 is data of the registered content 60 to be recommended.

The evaluation value 33 is an evaluation value given to the content by the user, and includes an evaluation by the requester and a user other than the requester. When a new evaluation is input from the evaluation input unit 13, the value of the evaluation value 33 is added or updated.

The attribute data 32 is attribute data of the content calculated by the attribute extracting unit 21.

The profile data 51 is user profile data generated by the profile learning section 41.

The conversion method data 52 is data designating a compression conversion method of the profile, and is generated by the profile conversion unit 42.

The conversion attribute data 53 is attribute data obtained by compression-converting the content, and is data that has been compression-converted by the attribute conversion unit 43.

The conversion profile data 54 is compression-converted profile data, and is data that has been compression-converted by the profile conversion unit 42.

Next, the operation of the prior art shown in FIG. 7 will be described with reference to the flowcharts of FIGS.

The operation is roughly divided into a preparatory operation to be executed in advance and a filtering operation to be executed in response to a filtering request from a user.

First, the preparation operation will be described. FIG.
5 is a flowchart for explaining a preparation operation of a conventional information filtering system.

Referring to FIG. 8, in the preparation operation, first, the attribute extracting unit 2 extracts the contents in the contents data 31 from the contents.
1 extracts the characteristic attribute, and stores it in the storage unit 50 as the attribute data 32 (step 801).

As the "attribute" here, a keyword or a word frequency characterizing the content of the content can be used. For example, "Some content contains" baseball "
And the keyword "soccer". ], An attribute indicating the characteristic of the content is extracted.

Next, based on the attribute data 32 of the content evaluated by the requester and the evaluation value 33 given to each content by the user, the profile learning unit 41 compares the evaluation of the requester with the attribute. A profile representing the relationship between the two is learned, and the result is stored as profile data 51 in the storage unit 5.
0 is stored (step 802).

The learning result of this profile is, for example,
In the past, the requester rated content containing the keyword "baseball" high, while conversely content containing the keyword "soccer" rated low, indicating that the relationship between the requester's rating and the attribute was low. In other words, it is data expressing the interest of the user by the weight of a word (keyword) included in the content.

Next, the profile conversion section 42 compresses and converts the attribute of the user profile data 51 and stores it in the storage section 50 as the conversion profile data 54 (step 803). Further, the profile conversion unit 42 stores, in the storage unit 50, information that specifies which attribute is to be combined and which attribute is to be deleted, and specifies a profile compression conversion method as conversion method data 52.

The contents of the compression conversion process are as follows. For example, when keywords related to baseball such as "baseball", "baseball", and "pitcher" and their weights are stored in a certain user's profile, these are combined into one. In summary, the degree of interest of the user on the topic of [baseball] is expressed. Also, for example,
Delete keywords that are general and do not help identify the topic, such as "person" or "thing". In this prior art, the profile conversion unit 42 executes these compression conversion processes using a mathematical method called “singular value decomposition”.

Next, the attribute conversion unit 43 compresses and converts the attribute data 32 of each content in accordance with the conversion method specified in the conversion method data 52, and stores the result in the storage unit 50 as conversion attribute data 53 ( Step 80
4).

Subsequent to the preparatory operation, the actual filtering operation will be described. FIG. 9 is a flowchart for explaining a filtering operation of the conventional information filtering system.

First, when a requester's information filtering request is received from the request input unit 11 (step 901),
The evaluation prediction unit 44 compares the conversion profile data 54, which is the profile of the requester that has been compressed and converted, with the conversion attribute data 53, which is the compressed attribute of each content, and calculates the evaluation value of each content for the requester. Predict and output the predicted evaluation value (step 902).

Next, the content selection section 25 receives the predicted evaluation value for the requester of each content from the evaluation prediction section 44, and selects the content having a high predicted evaluation value from the content data 31 (step 903).

Finally, the content having a high predicted evaluation value selected by the content selection section 25 is provided to the requester through the content presentation section 12 (step 90).
4).

For example, when the requester is interested in the content dealing with the topic of [baseball] and not interested in the content dealing with the topic of [soccer], the evaluation prediction unit 4
4 receives the compressed and converted profile indicating the information to that effect. Then, the evaluation prediction unit 44 matches this profile with the conversion attribute data of each content, and outputs this as a predicted evaluation value. At this time, the predicted evaluation value of the content dealing with the topic of [baseball] becomes higher, and conversely, the predicted evaluation value of the content dealing with the topic of [soccer] becomes lower. In this conventional technique, a prediction evaluation value is calculated using a correlation coefficient between a profile of a requester that has undergone compression conversion and an attribute of the content that has undergone compression conversion.

Then, upon receiving the prediction result of the evaluation prediction section, the content selection section 25 receives the content data 3
1, the content that has a high predicted evaluation value and has not been evaluated by the requester in the past is selected. Therefore, the content dealing with the topic of “baseball” is selected and output to the content presentation unit 12. Conversely, content dealing with the topic of [soccer] is not selected.

Next, the requester evaluates whether the content presented on the content presentation unit 12 matches his / her interests and preferences, and inputs the evaluation from the evaluation input unit 13. The evaluation input from the evaluation input unit 13 is stored in the storage unit 50 as the evaluation value 33.

[0045]

Next, problems of the above-described conventional techniques, that is, the CBF system alone, the SIF system alone, and the system combining both systems will be described.

In the case of the CBF system alone, the relationship between the attribute of the content and the evaluation is learned, and the content is filtered. In other words, a content whose attribute is similar to a content that has been highly evaluated by the user has a high predicted evaluation value, and is recommended to the user.

However, the similarity between the attribute and the evaluation is
Does not exactly match. In other words, the information of the attribute includes both a part related to the evaluation and a part not related to the evaluation. In this CBF method, the data of the part not related to the evaluation is also used for the evaluation prediction. Would. For this reason, in the conventional CBF method, there is a problem that filtering accuracy is reduced due to the influence of a part of the information of the attribute which is not related to the evaluation.

Further, even in the case of a system combining the SIF system and the CBF system, there is a problem of the CBF system.
That is, even in the combined method, the information filtering by the CBF method is performed, so that the information included in the attribute that is not related to the evaluation affects the evaluation prediction.

In the case of the SIF system alone, information on content evaluated by another user and not yet evaluated by the requester is filtered. For this reason, the conventional SI
The F method has a problem in that appropriate filtering cannot be performed unless a certain amount of evaluation is collected in advance.

This is because the SIF method uses the evaluation of each user to perform filtering. As described above, in the SIF system, when the number of evaluators is small, the accuracy of filtering becomes poor, and there is a large problem that new information that no one has yet evaluated cannot be recommended. The problem with the SIF method is that
It can be reduced by combining with the CBF method, but in that case, the problem of the method combining the SIF method and the CBF method becomes a problem.

A first object of the present invention is to solve the above-mentioned drawbacks of the prior art, and realize efficient and accurate information filtering based on both the attribute of the content and the evaluation value of the content of another user. To provide an information filtering system, a filtering method thereof, and an information filtering program.

[0052]

In order to achieve the above object, an information filtering system according to the present invention predicts a user's evaluation of individual contents, and detects contents suitable for the user from the contents. In the information filtering system, data of an evaluation value indicating an evaluation of each of the contents of each of the users, which has been shown so far by each of the users, and related data which is information related to the evaluation are recorded in advance. A storage unit, based on the relevant data and the evaluation value of each of the contents, a compression conversion unit that calculates compressed data that is data obtained by removing a portion that does not affect evaluation from each of the relevant data, A user evaluation indicating characteristics of the evaluation indicated by each user based on information of both the compressed data and the evaluation value Evaluation feature calculating means for calculating collection data and content evaluation feature data; and predicting and calculating an evaluation given to each content by each user based on the user evaluation feature data and the content evaluation feature data. It is characterized by comprising an evaluation prediction means for performing the evaluation.

The information filtering system according to a second aspect of the present invention is characterized in that the related data includes attribute data expressing the contents of each of the contents.

The information filtering system according to a third aspect of the present invention is characterized in that the related data includes user data indicating information on each of the users.

In the information filtering system according to a fourth aspect of the present invention, the compression conversion means compresses and converts the relevant data based on the relevant data and the evaluation value, thereby converting each of the relevant data into an evaluation. The method is characterized in that the compressed data, which is data obtained by removing a portion having no influence, is calculated.

According to a fifth aspect of the present invention, in the information filtering system according to the fifth aspect, the compression conversion unit performs a compression conversion method for compressing the evaluation value based on the evaluation value so as to remove a portion that is not effective for evaluation prediction from the evaluation value. An evaluation compression conversion unit that calculates evaluation conversion method data, which is data to be specified, and a compression conversion method that calculates the compressed data from the relevant data based on the relevant data and the evaluation conversion method data. A second compression conversion unit that calculates compression conversion method data that is data to be specified; and a compression calculation unit that calculates the compressed data from the related data based on a calculation method indicated in the compression conversion method data. It is characterized by.

According to a sixth aspect of the present invention, in the information filtering system, the compression conversion means compresses and converts the related data and the evaluation value by singular value decomposition to calculate the compressed data.

According to a seventh aspect of the present invention, in the information filtering system, the evaluation feature calculating means converts data indicated by using information of both the compressed data and the evaluation value into data.
The compression conversion is performed to calculate user evaluation feature data indicating the evaluation feature indicated by each user and content evaluation feature data.

According to an eighth aspect of the present invention, in the information filtering system according to the present invention, the evaluation feature calculating means converts data indicated by using information of both the compressed data and the evaluation value,
The compression conversion is performed by singular value decomposition, and user evaluation characteristic data indicating the evaluation characteristics indicated by each user and content evaluation characteristic data are calculated.

According to a ninth aspect of the present invention, the information filtering system according to the present invention inputs and records an evaluation value, which has been shown by each user so far, indicating each user's evaluation of each of the contents. It is characterized by comprising evaluation input means and attribute extracting means for extracting attribute data which is data representing the content of each of the contents.

[0061] An information filtering system according to a tenth aspect of the present invention is characterized in that the content published on the Internet is filtered.

According to an information filtering method of the present invention, there is provided an information filtering method for predicting a user's evaluation of individual contents and detecting contents suitable for the user from the contents. Data of an evaluation value indicating an evaluation of each of the contents of each of the users shown so far by the user, and related data that is information related to the evaluation, referring to pre-recorded data, A compression conversion step of calculating compressed data that is data obtained by removing a portion that does not affect evaluation from each of the related data based on the related data and the evaluation value of each of the contents;
Based on both information of the compressed data and the evaluation value, user evaluation feature data indicating an evaluation feature indicated by each user, and an evaluation feature calculation step of calculating content evaluation feature data, An evaluation prediction step of predicting and calculating an evaluation given by each user to each content based on the user evaluation characteristic data and the content evaluation characteristic data is provided.

A twelfth aspect of the information filtering method according to the present invention is characterized in that the related data includes attribute data representing the contents of each of the contents.

According to a thirteenth aspect of the information filtering method of the present invention, the related data includes user data indicating information on each of the users.

In the information filtering method according to the present invention, in the compression conversion step, the relevant data may be subjected to compression conversion based on the relevant data and the evaluation value, so that each of the relevant data is converted into an evaluation. The method is characterized in that the compressed data, which is data obtained by removing a portion having no influence, is calculated.

In the information filtering method according to the present invention, the compression conversion step may include a compression conversion method for compressing the evaluation data based on the evaluation value so as to remove a portion that is not valid for evaluation prediction from the evaluation data. An evaluation compression conversion step of calculating evaluation conversion method data, which is data to be specified, and a compression conversion method of calculating the compressed data from the relevant data based on the relevant data and the evaluation conversion method data. A second compression conversion step of calculating compression conversion method data, which is data to be specified; and a compression calculation step of calculating the compressed data from the related data based on a calculation method indicated in the compression conversion method data. It is characterized by.

The information filtering method according to the present invention is characterized in that the compression conversion step includes compression conversion of the related data and the evaluation value by singular value decomposition to calculate the compressed data.

In the information filtering method according to the present invention, in the evaluation feature calculating step, each of the compressed data and the evaluation value is compressed and converted into data. It is characterized in that user evaluation feature data indicating the evaluation feature indicated by the user and content evaluation feature data are calculated.

In the information filtering method according to the present invention, the evaluation feature calculating step may include a step of compressing and transforming the data indicated by using the information of both the compressed data and the evaluation value by singular value decomposition. The method is characterized in that user evaluation characteristic data indicating evaluation characteristics indicated by each user and content evaluation characteristic data are calculated.

The information filtering method of the present invention according to claim 19 has been described so far by each of said users.
An evaluation input step of inputting and recording an evaluation value indicating an evaluation of each of the contents of each of the users, and an attribute extracting step of extracting attribute data which is data expressing the contents of each of the contents. Features.

A twentieth aspect of the information filtering method according to the present invention is characterized in that the content published on the Internet is filtered.

According to a twenty-first aspect of the present invention, an information filtering program predicts a user's evaluation of individual contents by controlling a computer, and detects information suitable for the user from the contents. In the filtering program, data of an evaluation value indicating an evaluation for each of the contents of each of the users that has been shown so far by each of the users,
Based on the process of referring to pre-recorded data of the related data that is the information related to the evaluation, and based on the related data and the evaluation value of each of the contents, do not affect the evaluation from each of the relevant data. A compression conversion process for calculating compressed data that is data obtained by removing a portion, and a user evaluation feature indicating an evaluation feature indicated by each user based on information of both the compressed data and the evaluation value. An evaluation feature calculation process of calculating data and content evaluation feature data; and estimating and calculating an evaluation given by each user to each content based on the user evaluation feature data and the content evaluation feature data. It is characterized in that an evaluation prediction process is executed.

In the information filtering system of the present invention, both information of the attribute data of the content and the evaluation value indicated by each user so far are referred to, and only the part effective for the evaluation prediction in both information is compressed. The content is extracted by the conversion, and the evaluation of each user predicted for each content is calculated by using those extracted by the compression conversion.
As described above, since the evaluation prediction of each user is calculated by using the data obtained by deleting the part that is not effective in the evaluation prediction by the compression conversion, the adverse effect due to the data of the part that is not effective in the evaluation prediction is eliminated. Are solved. In addition, since the evaluation prediction is calculated based on both the attribute data and the evaluation value, even if the content has not yet been evaluated by each user, it can be appropriately determined by referring to the characteristics of each content, The problem of the conventional SIF method is solved. Further, in the information filtering system of the present invention, the omission of filtering can be suppressed to a small extent due to the property of combining similar compression conversions.

[0074]

Embodiments of the present invention will be described below in detail with reference to the drawings.

According to the present invention, in information filtering for estimating a user's evaluation of each content and detecting a content suitable for the user from each content, each information previously shown by each user is used. Based on the evaluation value data indicating the user's evaluation of each content and the relevant data, which is information related to the evaluation, calculate compressed data that is data obtained by removing a portion that does not affect the evaluation from each relevant data. Then, based on the information of both the compressed data and the evaluation value, an evaluation given by each user to each content is predicted and calculated.

Here, as the related data which is information related to the evaluation, for example, attribute data which is data expressing the content of each content, and user data which is information on each user (content preference etc.) Can be used. As a method for calculating compressed data, a method such as singular value decomposition can be used as shown in the following embodiments. In this way, by compressing attribute data and user data, etc., and generating compressed attribute data and compressed users by removing portions that do not affect the evaluation, and performing evaluation using the compressed attribute data and user, evaluation prediction can be performed. It is possible to eliminate an adverse effect due to the data of an invalid portion. In addition, since the evaluation prediction is calculated based on both the evaluation value by the user and the attribute data, which is other related data, even if the content has not been evaluated by each user yet, the characteristics of each content can be determined. An appropriate determination can be made with reference to this.

FIG. 1 is a block diagram showing the configuration of the information filtering system according to the first embodiment of the present invention. In the present embodiment, attribute data, which is data expressing the content of each content, is used as the related data. Also, in FIG. 1, the same reference numerals are given to the same configurations as those of the above-described conventional technology shown in FIG. In such a configuration similar to that of the related art, the description thereof will be omitted unless particularly necessary.

Referring to FIG. 1, the information filtering system according to the present embodiment includes an input / output unit 10, a data processing unit 20, and a storage unit 30.

The input / output unit 10 includes a request input unit 11, a content presentation unit 12, and an evaluation input unit 13.

The data processing unit 20 includes an attribute extraction unit 21, an attribute compression conversion unit 22, an evaluation feature calculation unit 23, and an evaluation prediction unit 2.
4. A content selection unit 25 is provided.

The attribute extracting unit 21 extracts the characteristic attribute from the content data 31 of the registered content, and stores it in the storage unit 30 as attribute data 32.

The attribute compression conversion section 22 compresses and converts the attribute data based on the attribute data 32 and the evaluation value 33 of the content in order to delete a portion that is not valid for the evaluation prediction from the attribute data. Is stored in the storage unit 30.

The evaluation feature calculation unit 23 calculates the compression attribute data 3
Based on 4 and the evaluation value 33, a user characteristic effective for evaluation prediction is calculated and stored in the storage unit 30 as user evaluation characteristic data 35. In addition, the evaluation feature calculation unit 23 calculates a content feature effective for the evaluation prediction, and stores the content feature in the storage unit 30 as the content evaluation feature data 36.

The evaluation prediction unit 24, based on the user evaluation characteristic data 35 and the content evaluation characteristic data 36, receives the filtering request of the requester from the request input unit 11 of the input / output unit 10. The evaluation value given by the requester to the content is predicted and the value is calculated.

The content selection section 25 includes an evaluation prediction section 44
From the registered content recorded as the content data 31, a content having a high predicted evaluation value is selected and provided to the requester through the content presentation unit 12 of the input / output unit 10.

The storage unit 30 stores content data 31, attribute data 32, evaluation value 33, compressed attribute data 34, user evaluation characteristic data 35, and content evaluation characteristic data 36.
I remember.

The contents data 31 is data of the registered contents 60 to be recommended.

The attribute data 32 is content attribute data calculated by the attribute extracting unit 21.

The evaluation value 33 is an evaluation value given to the content by the user, and includes an evaluation by the requester and a user other than the requester. The evaluation input unit 1 of the input / output unit 10
When a new evaluation is input from No. 3, the evaluation value 33
Is added / updated.

The compressed attribute data 34 is stored in the attribute compression
2 is attribute data in which a portion that is not effective for evaluation due to compression conversion is deleted.

The user evaluation characteristic data 35 is data indicating characteristics of user evaluation calculated by the evaluation characteristic calculation unit 23.

The content evaluation feature data 36 is data indicating the content evaluation feature calculated by the evaluation feature calculation unit 23.

Next, the operation of the information filtering system of this embodiment will be described with reference to the flowcharts of FIGS.

The operation is roughly divided into a preparatory operation to be executed in advance and a filtering operation to be executed in response to a user's filtering request.

First, the preparation operation will be described. FIG.
5 is a flowchart for explaining a preparation operation of the information filtering system according to the present embodiment.

In the preparatory operation, first, the attribute extracting unit 21 extracts the characteristic attribute from the content data 31 of the registered content, and stores it as attribute data 32 in the storage unit 30 (step 201).

Next, the attribute compression conversion unit 22 compresses and converts the attribute data based on the attribute data 32 of the content and the evaluation value 33 in order to delete a portion that is not valid for evaluation prediction from the attribute data. It is stored in the storage unit 30 as the attribute data 34 (Step 202).

Next, the evaluation feature calculation unit 23 calculates a user feature effective for evaluation prediction based on the compressed attribute data 34 and the evaluation value 33, and stores it in the storage unit 30 as user evaluation feature data 35. It is stored (step 203). In addition, the evaluation feature calculation unit 23 calculates a content feature effective for the evaluation prediction, and stores the content feature in the storage unit 30 as the content evaluation feature data 36.

Following the preparation operation, an actual filtering operation will be described. FIG. 3 is a flowchart for describing a filtering operation of the information filtering system according to the present embodiment.

First, when an information filtering request of a requester is received from the request input unit 11 of the input / output unit 10 (step 301), the evaluation prediction unit 24 converts the user evaluation characteristic data 35 and the content evaluation characteristic data 36. Based on this, we estimate the value given by the requester for each content,
The predicted evaluation value is calculated (Step 302).

Next, the content selection unit 25 receives the predicted evaluation value from the evaluation prediction unit 44, and
A content having a high predicted evaluation value is selected from 1 (step 303).

Finally, the content selected by the content selection section 25 is provided to the requester through the content presentation section 12 of the input / output section 10 (step 304).

Then, the requester enters the content presentation unit 12
Is evaluated to see if it matches the user's interests and preferences, and inputs the evaluation from the evaluation input unit 13. The evaluation input from the evaluation input unit 13 is stored in the storage unit 30 as an evaluation value 33.

Next, in the information filtering system according to the present embodiment, the attribute compression conversion unit 22, which is a component for executing main processing, will be described in more detail with reference to the drawings. FIG. 4 is a block diagram illustrating an example of a configuration of the attribute compression conversion unit 22 according to the present embodiment.

Referring to the example of FIG.
2 is an evaluation compression converter 26, a second attribute compression converter 27,
A compression attribute calculator 28 is provided, and internally stores an evaluation conversion method data 37 and an attribute conversion method data 38.

The evaluation compression conversion unit 26 calculates data specifying a calculation method for deleting a portion that is not valid for evaluation prediction from the evaluation data based on the evaluation value 33, and stores the data as evaluation conversion method data 37. .

The second attribute compression conversion unit 27 generates data specifying a calculation method for deleting a portion that is not valid for evaluation prediction from the attribute data based on the attribute data 32 of the content and the evaluation conversion method data 37. It is calculated and stored as attribute conversion method data 38.

The compression attribute calculation unit 28 applies the conversion method indicated by the attribute conversion method data 38 to the attribute data 32 of the content, deletes a portion that is not valid for the evaluation prediction from the attribute data, and stores it as compressed attribute data 34. Stored in the unit 30.

The evaluation conversion method data 37 is data that specifies a method of deleting a portion that is not valid for evaluation prediction from the evaluation data calculated by the evaluation compression conversion unit 26.

The attribute conversion method data 38 is data that is calculated by the second attribute compression conversion unit 27 and specifies a method of deleting a part that is not valid for evaluation prediction from the attribute data.

Next, the operation of the attribute compression converter 22 will be described. FIG. 5 shows the attribute compression converter 22 according to the present embodiment.
5 is a flowchart for describing an example of the operation of FIG.

First, the evaluation compression conversion unit 26 calculates the evaluation value 33
Based on, calculate data that specifies a calculation method for removing portions that are not valid for evaluation prediction from evaluation data,
It is stored as the evaluation conversion method data 37 (step 50).
1).

Next, the second attribute compression conversion unit 27 specifies a calculation method for deleting, from the attribute data, a part that is not valid for evaluation prediction, based on the attribute data 32 of the content and the evaluation conversion method data 37. Is calculated and stored as attribute conversion method data 38 (step 50).
2).

Next, the compression attribute calculation unit 28 applies the conversion method indicated by the attribute conversion method data 38 to the attribute data 32 of the content, deletes the portion that is not valid for the evaluation prediction from the attribute data, and deletes the compressed attribute data. The storage unit 30 as 34
(Step 503).

Next, the processing in the attribute compression conversion section 22 will be described in more detail.

As an example of the compression conversion processing method according to the present invention, the attribute compression conversion unit 22 and the evaluation feature calculation unit 23 of the present embodiment use the singular value decomposition (SVD).
The compression conversion is performed using a mathematical method called "ecomposition".

First, the singular value decomposition will be described.

The information of the attribute data 32 and the evaluation value 33 in the present embodiment can be mathematically considered as a data matrix. For example, the attribute data 32 assigns each attribute to each row, assigns each content to each column, and assigns each element of the matrix to the attribute value of the attribute corresponding to the row of the content corresponding to the column. It can be considered as a data matrix of attribute values. Similarly, the evaluation value 33 is considered to be a data matrix by assigning each user to each row, assigning each content to each column, and assigning each element of the matrix to an evaluation value by the corresponding user of the corresponding content. be able to.

By applying singular value decomposition to such a data matrix, the following processing can be easily realized.・ Compression of the data matrix into essential parts

The singular value decomposition can be mathematically defined as follows.

Now, the matrix A is represented by rank × m columns × rank
Let (A) = r (the rank is r) a data matrix. Matrix A
Can be interpreted as expressing n types of data in an m-dimensional space or expressing m types of data in an n-dimensional space.

The singular value decomposition is a linear algebra technique for decomposing the matrix A as follows. From any matrix A,
Three matrices U, R, and V satisfying the following conditions are calculated. Here, “┬” at the right shoulder of the matrix is a symbol indicating a transposed matrix. _{A m × n = U m ×} r · R r × r · V ┬ r × n (1) where the matrix U, V is orthonormal matrix, the matrix R is a diagonal matrix. U ^┬ · U = I _r (I _r is an r-th unit matrix) V ^┬ · V = I _r R = diag (d ₁ , d ₂ ,..., _Dr ), d ₁ ≧ d ₂
≧ ・ ・ ・ ≧ d _r > 0

D ₁ , d ₂ ,..., _Dr are diagonal matrices R
And the amount of information on each axis of the space after compression (described later).

By using the singular value decomposition, the data matrix A can be compressed as follows.

First, focusing on A and U in equation (1), considering the process of calculating matrix U from matrix A in singular value decomposition A _{m × n} → U _{m × r} , the number of columns in row example A is What is compressed from n to r can be called a matrix U.

The vectors of each column in the original data matrix A have high similarity and low similarity. However, by the conversion for obtaining U from A by the singular value decomposition, similar column vectors in the matrix A are put together in the same column, and in U after compression, each column vector becomes orthogonal. In other words, m data is represented in the n-dimensional space in A, but can be represented in the r-dimensional space in U.

Similarly, focusing on A and V, considering a process of calculating a matrix V from a matrix A in singular value decomposition A _{m × n} → V ^┬ _{r × n} , the number of rows of the matrix A is changed from m to r Can be called a matrix V.

Even when it is considered that a row is compressed, it is compressed based on the similarity of each row vector of the matrix A. in this case,
While n pieces of data are represented in the m-dimensional space in A, they can be represented in the r-dimensional space in V.

Further, both U and V are r-dimensional spaces, but in singular value decomposition, the two spaces are not simply the same in the number of dimensions, but U and V are the same space.

Further, each diagonal element d ₁ , d ₂ ,.
·, D _r indicates the amount of information possessed by the axes of the r-dimensional space after compression. Therefore, if only a specified number (for example, k) of axes having a large diagonal element value (that is, a large amount of information) among the r axes is used, the following equation (2) is obtained. ), A matrix A ′ can be obtained by compressing the matrix A to k dimensions. _{A 'm × n = U'} m × k · R 'k × k · V' ┬ k × n (2) Here, however, the matrix U ', V', R 'is the original matrix each U, V , R using only the first to k-th columns. In this compression, the matrix A is converted into the matrix A 'by removing non-essential parts without losing essential parts of the information held by the matrix A.

By using this singular value decomposition, missing values in the data matrix can be predicted as follows.

When there is a missing value in the data matrix A, the missing value can be predicted using the equation (2). The matrix A ′ in the equation (2) is obtained by recombining A using only the main axis having a large amount of information among the axes of the compression space by the singular value decomposition. Therefore, by reading the value of the element corresponding to the missing value from A ′, the missing value of A can be predicted. The predicted value at this time is a value that is most consistent with the essential part of A represented in the k-dimensional space.

When data is added or updated in the data matrix after performing the singular value decomposition, the addition or update of the data can be reflected by the following procedure.

When new data is added and the value of the cell of the matrix A is updated or a column or a row is added, it is originally necessary to perform singular value decomposition again to generate a new compression space. However, if the position of new data in the compression space of A before data addition can be easily obtained, the singular value decomposition does not need to be performed again when the amount of added / updated data is not large.

In the equation (1), A = U · R · V^┬From, V and R
A · (V^┬)^-1・ R^-1= U. Further, since U and V are orthonormal matrices,
The column is equal to the transposed matrix, so the previous equation is: AVR^-1= U. New data is added as shown in this formula
The element of A was updated or a new row was added
If the amount of change is small,
VR for matrix A^-1  Is calculated by multiplying by
Efficiently find the matrix U (a matrix that sufficiently approximates)
Can be.

Also, in the case where the data has already been compressed to k dimensions, V ′ · R ′ ⁻¹ (3) can be similarly multiplied from equation (2) from the right.

Similarly, when V is updated, R ⁻¹ · U
^{Since -1} · A = R ⁻¹ · U ^┬ · A = V ^{て も} , even if the cell of A is updated or a new column is added by adding data, if the amount of change is small, the latest The updated matrix V (a matrix that sufficiently approximates the updated matrix V) can be obtained efficiently by a simple operation by multiplying the matrix A by R ⁻¹ · U ^} from the left.

[0138] When compressed to k-dimensional Again, it is multiplied by R ^'-1 · U' ^┬ (4) from the left from equation (2).

Next, an example of compression processing by singular value decomposition of the information filtering system of this embodiment will be described in detail.

Here, the storage unit 30 previously stores the appearance rate of each word for each content as the attribute data 32 of the content, and stores the evaluation value 33 recommended by each user and given to the content. Is stored.

The attribute compression converter 22 first evaluates the evaluation data matrix (hereinafter referred to as E and E) in which the content is assigned to each row and the user is assigned to each column based on the evaluation value in the evaluation value 33 by the evaluation compression converter 26. Call). Here, for example, when the number of contents is m and the number of users who have evaluated the contents is n, the evaluation data matrix E is an m × n matrix.

Next, the evaluation compression conversion section 26 performs singular value decomposition of the evaluation data matrix E. E = D ₁・ R ₁・ U ₁ ^┬ (5)

Next, the number of dimensions after compression is determined (hereinafter, the number of dimensions after compression is referred to as k ₁ ), and E is set to (K ₁ ) U ₁ ′.
Calculated ^┬ matrix for converting ^{_{R 1 '-1 · D 1'}} ┬ (6), the storage section 30 as evaluation conversion method data 37.

It is considered that each user's evaluation, that is, each column of the evaluation data matrix E expresses each user's interest. The calculation using this singular value decomposition (calculation multiplying the equation (6) to E), considering that conversion to U ₁ ^'┬ from E, E
So that the number of rows have been compressed to k ₁ that was content number m. In order to determine the interest of each user, it is not necessary to refer to all m pieces of content, and the same applies to the case where only _one type of value representing essential information is referred to. Is realized.
Now, k ₁ rows of U ₁ ^'┬ is considered to represent the features of the interests of each user. And equation (5) is
Since the evaluation data matrix E is compression-converted into U ₁ ^} , it can be considered as an expression for calculating the characteristic of the user tendency. Since each line of E represents a content, what each line of U ₁ ^┬ ′ represents is hereinafter referred to as characteristic content.

Next, the second attribute compression converter 27 first
From the word appearance rate data of each content stored as the attribute data 32, an attribute data matrix (hereinafter, this matrix is referred to as F) in which the content is assigned to each row and the word (keyword) is assigned to each column is generated.

If the number of contents is m and the number of words is p in all, the attribute data matrix F is an m × p matrix.

[0147] Next, by applying the stored equation (6) as evaluation conversion method data 37 to the attribute data matrix F, the number of rows of the matrix F to generate a matrix F ₁ compressed from m to k _1.

[0148] Next, singular value decomposition matrix _{F 1.} F ₁ = D ₂ · R ₂ · T ₂ ^┬ (7)

Next, the number of dimensions after compression is determined (hereinafter, the number of dimensions after compression is referred to as k ₂ ), and F _{1 is set} to (k ₂ order)
A matrix T ₂ · R ₂ ' ^-1 (8) for conversion into D ₂ ' is calculated and stored in the storage unit 30 as attribute conversion method data 38.

[0150] Attribute data matrix F represents the word occurrence rate of each content, the matrix F ₁ of the matrix F compressed by Equation (6) represents the word occurrence rate characteristics content.

Considering the conversion from the matrix F ₁ to the matrix D ₂ ′ using this singular value decomposition, the number of columns of the matrix F ₁ was reduced to k ₂ instead of the number p of word types. become. This characteristic content to determine the interests of each user, in order to determine the word appearance rate, all words p type is not required, indicates that you can determine as well by using no k ₂ pieces of values ing. From this, it is considered that the k ₂ columns of the matrix D ₂ ′ represent the features of each feature content. Then, equation (8) expresses matrix F ₁ as matrix D
Since it is converted to ₂ ′, it can be considered as an expression for calculating the feature of the content. Since each column of the matrix F ₁ had expressed the word is referred to as a characteristic word in the following what each row of D 2 _'represents.

Next, the compressed attribute calculating section 28 firstly, based on the word appearance rate data of each content stored as the attribute data 32, an attribute data matrix (in the same manner as above) in which the content is assigned to each row and the word is assigned to each column. This matrix is denoted as F). If the number of contents is m and there are p types in total, the attribute data matrix F is an mxp matrix.

Using the equation (8) stored as the attribute conversion method data 38, the number of columns of the attribute data matrix F is reduced from p to k ₂ (set as the degree after compression). ₂ is generated and stored in the storage unit 30 as compression attribute data 34.

[0154] Attribute data matrix F represents the word occurrence rate of each content, the matrix F ₂ of the matrix F compressed by Equation (8) represents the incidence of feature word of each content.

[0155] This feature word, since it is intended to represent the features of the characterizing content to express the characteristics of the evaluation tendency of users, matrix F _2, which represents the rate of appearance of feature words, the original word appearance rate From the information that the matrix F has,
The part that does not express the characteristic of the evaluation tendency is deleted, and only the part that expresses the characteristic of the evaluation tendency is included.

Next, the evaluation feature calculating unit 23 first evaluates the contents in the respective rows and the evaluation data matrix in which the users are allocated to the respective columns based on the evaluations in the evaluation value 33 (similar to the above, this matrix is , E). If the number of contents is m and the number of users who have evaluated the contents is n, the evaluation data matrix E becomes an m × n matrix.

The matrix F representing the appearance rate of the characteristic word of each content stored as the compressed attribute data 34
To generate a matrix _{E 2} that connect the ₂ and the matrix E. This matrix E
₂ means that the content is assigned to each row, and
Users are assigned up to the columns, and from the (n + 1) th column to the (n + k) th
Up to _two columns are generated as an m-row n + k _two- column matrix to which each feature word is assigned.

[0158] Next, singular value decomposition of this matrix _{E 2.} E ₂ = D ₃ · R ₃ · U ₃ ^┬ (9)

[0159] Next, decide the number of dimensions after compression (hereinafter referred to as k ₃ the number of dimensions after compression), by performing similar compression conversion processing described so far, D _3, R ₃ ,
And compressing each matrix of _{U 3, D 3 ', R} 3', calculates a matrix of _{U 3} '. Then, the matrix R ₃ ′ and the matrix U ₃ ′ are stored as the user evaluation feature data 35, and the matrix D ₃ ′ is stored as the content evaluation feature data 36.

[0160] Matrix E ₂ is obtained by coupling the evaluation data, the compressed attribute data is a portion affecting the evaluation of the attribute data. Therefore Customer rating characteristic data calculated by compressing convert E _2, the evaluation data and the attribute and the data that expresses the features of evaluation tendency of the user based on, and wherein the evaluation trend attribute data Is not included. This is the same for the content evaluation feature data.

Next, the evaluation predicting section 24 sets the request input section 11
When the requester's information filtering request is received from
Matrix R ₃ ′ stored as user evaluation feature data 35
And a matrix U ₃ ′ corresponding to the requester, and a matrix D ₃ ′ stored as the content evaluation feature data 36.
By calculating the column of the matrix E ₂ ′ corresponding to the requester, a predicted value of the evaluation given by the requester to each content is obtained.

Further, in this embodiment, the word appearance frequency of each content is used as the attribute data, but the TF / IDF value of the word may be used. Bibliographic data such as the author of the content can also be used.
Further, when the content has a reference relationship between other content such as a web page, a reference from the content to another page or information on a reference from another page can be used as attribute data.

As described above, according to the information filtering system of the present invention, based on both the attribute of the content and the evaluation value of the content of another user,
Highly accurate information filtering can be efficiently realized.

Next, a second embodiment of the present invention will be described.

FIG. 6 is a block diagram showing the configuration of the information filtering system according to the second embodiment of the present invention. In the first embodiment, attribute data is used as related data. However, in the present embodiment,
User data, which is information about the user such as age, gender, and hobby of each user, is used as related data.

That is, in the present embodiment, only the portion that affects the evaluation is extracted from this user data and filtering is performed in the same manner as in the first embodiment. The user data compression conversion unit 22a according to the present embodiment is configured so that the attribute compression conversion unit 22 according to the first embodiment extracts only a part that affects the evaluation based on the attribute data and the evaluation data. In the same manner as the calculation of the data, based on the user data and the evaluation data, only the part that affects the evaluation is extracted from the user data and stored as the compressed user data 34a.

The evaluation feature calculator 23a of the present embodiment
Based on the evaluation data and the compressed user data 34a, the first
The user evaluation characteristic data 35 and the content evaluation characteristic data 36 are calculated by the same method as the evaluation characteristic calculation unit 23 in the embodiment, and stored in the storage unit 30.

Next, the processing of the evaluation feature calculating section 23a will be described in detail. Compressed user data 34a as a stored matrix (hereinafter referred to as F ₃₎ is the number of users n, the type of compressed user data when it was k ₃ kinds, and k ₃ rows n columns of the matrix. First, the evaluation feature calculation unit 23a generates an evaluation data matrix in which the content is assigned to each row and the user is assigned to each column based on the evaluation in the evaluation value 33 (similarly, this matrix is described as E). Generate. The number of contents is m,
If n users rated your content,
The evaluation data matrix E is an m × n matrix.

[0169] Then, to produce a matrix _{E 3} obtained by connecting the stored matrix _{F 3} and the matrix E as compressed user data 34a. The matrix E _3, the user is assigned to each column, first
Content is allocated from the column to the m-th row, and the (m + 1) -th
From the column until the n + k ₃ rows to generate a compressed user data is assigned m + k ₃ row n matrix.

Thereafter, by performing the same processing as in the first embodiment, it is possible to perform filtering based on the evaluation data and the part of the user data that affects the evaluation.

The user data includes the user's age, gender, hobbies, and the like. However, the relationship between the user and other users (boss, subordinate, etc.), keywords indicating the user's interest, A user profile or the like calculated by the CBF method can also be used.

Also, the present embodiment and the first embodiment are combined,
It is also possible to perform filtering based on the evaluation data and the part of the user data that affects the evaluation, and the part of the attribute data that affects the evaluation.

Further, in each of the above embodiments, the method of processing the compression conversion by using the singular value decomposition has been described as an example. However, the compression conversion method of the present invention is not limited to this method. An embodiment using a dimension reduction method used in multivariate analysis, independent component analysis, a neural network, or the like is also possible.

It should be noted that the information filtering systems 100 and 100a of each of the above embodiments are
20a, an attribute extraction unit 21, an attribute compression conversion unit 22,
A user data compression conversion unit 22a, an evaluation feature calculation unit 23,
Of course, the functions of the evaluation prediction unit 24, the content selection unit 25, the evaluation compression conversion unit 26, the second attribute compression conversion unit 27, the compression attribute calculation unit 28, and other functions are realized by hardware. A computer program having each function can be realized by being loaded into a memory of a computer processing device. This computer program is stored on a magnetic disk, a semiconductor memory, or other recording media 90, 90a. Then, the functions described above are realized by being loaded from the recording medium into the computer processing device and controlling the operation of the computer processing device.

Although the present invention has been described with reference to the preferred embodiments and examples, the present invention is not necessarily limited to the above-described embodiments and examples, and various modifications may be made within the scope of the technical idea. Modifications can be made.

[0176]

As described above, according to the information filtering system of the present invention, the following effects can be achieved.

First, in the information filtering system of the present invention, a part that is not effective for evaluation prediction can be deleted from the attribute of the content by the attribute compression conversion unit.
For this reason, it is possible to reduce or eliminate the problem that the filtering accuracy is reduced due to the influence of a part not related to the evaluation in the conventional CBF method (or a method combined with the SIF method).

Second, in the information filtering system of the present invention, it is possible to predict the evaluation value by using the attribute of the content. For this reason, in the conventional SIF method, if the evaluation of a certain amount is not collected in advance, the filtering accuracy is poor, and the obtained information is only information that has already been evaluated and recommended by another user. Can be solved.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of an information filtering system according to a first embodiment of the present invention.

FIG. 2 is a flowchart for explaining an operation of the information filtering system according to the first embodiment of the present invention.

FIG. 3 is a flowchart for explaining the operation of the information filtering system according to the first embodiment of the present invention.

FIG. 4 is a block diagram illustrating an example of a configuration of an attribute compression conversion unit according to the first embodiment of this invention.

FIG. 5 is a flowchart illustrating an example of an operation of the attribute compression conversion unit according to the first embodiment of this invention.

FIG. 6 is a block diagram illustrating a configuration of an information filtering system according to a second embodiment of the present invention.

FIG. 7 is a block diagram showing a configuration of a conventional information filtering system.

FIG. 8 is a flowchart for explaining the operation of a conventional information filtering system.

FIG. 9 is a flowchart illustrating the operation of a conventional information filtering system.

[Explanation of symbols]

 100, 100a Information filtering system 10 Input / output unit 11 Request input unit 12 Content presentation unit 13 Evaluation input unit 20, 20a Data processing unit 21 Attribute extraction unit 22 Attribute compression conversion unit 23 Evaluation feature calculation unit 24 Evaluation prediction unit 25 Content selection unit 26 evaluation compression conversion unit 27 second attribute compression conversion unit 28 compression attribute calculation unit 30 storage unit 31 content data 32 attribute data 33 evaluation value 34 compression attribute data 35 user evaluation characteristic data 36 content evaluation characteristic data 37 evaluation conversion method data 38 Attribute conversion method data 22a User data compression conversion unit 23a Evaluation feature calculation unit 32a User data 34a Compressed user data 40 Data processing unit 41 Profile learning unit 42 Profile conversion unit 43 Attribute conversion unit 44 Evaluation prediction unit 50 Storage unit 1 profile data 52 conversion method data 53 converts attribute data 54 converted profile data 60 registered content 90,90a recording medium

Claims (30)

[Claims] 1. An information filtering system for predicting a user's evaluation of individual contents and detecting contents suitable for the user from among the contents, comprising: Data of an evaluation value indicating an evaluation of each of the contents of the user, and a storage unit provided with pre-recorded related data which is information related to the evaluation; and Compression conversion means for calculating compressed data that is data obtained by removing a portion that does not affect the evaluation from each of the related data, based on both information of the compressed data and the evaluation value, Evaluation feature calculating means for calculating user evaluation feature data indicating the evaluation feature indicated by the user and content evaluation feature data; Information filtering system, characterized in that it comprises an evaluation prediction means with the content rating characteristic data and the user evaluation feature data based on calculated predicted evaluations each user gives for each content. 2. The information filtering system according to claim 1, wherein the related data includes attribute data representing the content of each of the contents. 3. The information filtering system according to claim 1, wherein the related data includes user data indicating information on each of the users. 4. The compression conversion means compresses and converts the related data based on the related data and the evaluation value, thereby removing data that does not affect evaluation from each of the related data. 4. The information filtering system according to claim 1, wherein said certain compressed data is calculated. 5. The evaluation conversion method data, wherein the compression conversion means is data for designating a compression conversion method for performing compression based on the evaluation value so as to remove a portion that is not valid for evaluation prediction from the evaluation value. Compression conversion method data, which is a data designating a compression conversion method for calculating the compressed data from the relevant data based on the relevant data and the evaluation conversion method data. 5. A second compression conversion unit that calculates the compression data, and a compression calculation unit that calculates the compressed data from the related data based on a calculation method indicated by the compression conversion method data. 6. Information filtering system. 6. The information according to claim 4, wherein the compression conversion unit compresses and converts the related data and the evaluation value by singular value decomposition to calculate the compressed data. Filtering system. 7. The evaluation feature calculating means compresses and converts data indicated by using information of both the compressed data and the evaluation value to thereby obtain an evaluation feature indicated by each of the users. 7. The information filtering system according to claim 1, wherein user evaluation feature data and content evaluation feature data are calculated. 8. The evaluation feature calculating means compresses data represented by using both information of the compressed data and the evaluation value by singular value decomposition, and performs evaluation conversion of the evaluation shown by each of the users. The information filtering system according to claim 7, wherein user evaluation characteristic data indicating characteristics and content evaluation characteristic data are calculated. 9. An evaluation input means for inputting and recording an evaluation value indicating each user's evaluation of each of the contents, which has been shown by each user so far, and contents of each of the contents. The information filtering system according to any one of claims 1 to 8, further comprising an attribute extracting unit configured to extract attribute data that is data expressing the information. 10. The information filtering system according to claim 1, wherein the content published on the Internet is filtered. 11. An information filtering method for predicting a user's evaluation of individual contents and detecting contents suitable for the user from among the contents, comprising the steps of: Referring to pre-recorded data of evaluation value data indicating an evaluation of the user for each of the contents and related data that is information related to the evaluation; and the related data of each of the contents and the evaluation. Based on the value, a compression conversion step of calculating compressed data that is data obtained by removing a portion that does not affect the evaluation from each of the related data, based on both information of the compressed data and the evaluation value, Evaluation features for calculating user evaluation feature data indicating the evaluation features indicated by each of the users and content evaluation feature data An information filtering method, comprising: a calculating step; and an evaluation prediction step of predicting and calculating an evaluation given to each content by each user based on the user evaluation characteristic data and the content evaluation characteristic data. . 12. The information filtering method according to claim 11, wherein the related data includes attribute data representing the contents of each of the contents. 13. The information filtering method according to claim 11, wherein the related data includes user data indicating information on each of the users. 14. The compression conversion step is a step of compressing and converting the relevant data based on the relevant data and the evaluation value, thereby removing data that does not affect evaluation from each of the relevant data. 14. The information filtering method according to claim 11, wherein the certain compressed data is calculated. 15. The compression conversion step, wherein the compression conversion step is data for specifying a compression conversion method for compressing the evaluation data so as to remove a portion that is not valid for evaluation prediction from the evaluation data. Evaluation compression conversion step of calculating the compression conversion method data, which is data specifying a compression conversion method of calculating the compressed data from the relevant data based on the relevant data and the evaluation conversion method data. The method according to claim 14, further comprising: a second compression conversion step of calculating the compression data; and a compression calculation step of calculating the compressed data from the related data based on a calculation method indicated in the compression conversion method data. Information filtering method. 16. The information according to claim 14, wherein, in the compression conversion step, the related data and the evaluation value are compression-converted by singular value decomposition to calculate the compressed data. Filtering method. 17. The evaluation feature calculating step includes: performing compression conversion of data indicated by using information of both the compressed data and the evaluation value to thereby obtain an evaluation feature indicated by each of the users. 17. The information filtering method according to claim 11, wherein the user evaluation feature data and the content evaluation feature data are calculated. 18. The evaluation feature calculating step, wherein the data indicated by using both the information of the compressed data and the evaluation value is compression-converted by singular value decomposition, and the evaluation of the evaluation indicated by each user is performed. 18. The information filtering method according to claim 17, wherein user evaluation characteristic data indicating characteristics and content evaluation characteristic data are calculated. 19. An evaluation inputting step of inputting and recording an evaluation value, which has been shown by each user so far, indicating an evaluation of each of said users for each of said contents, and contents of each of said contents. The information filtering method according to any one of claims 11 to 18, further comprising an attribute extracting step of extracting attribute data that is data representing the information. 20. The information filtering method according to claim 11, wherein the content published on the Internet is filtered. 21. By controlling a computer,
In an information filtering program for predicting a user's evaluation of individual contents and detecting contents suitable for the user from among the contents, each of the users indicated so far by each user Data of an evaluation value indicating the evaluation of the content, and processing of referring to pre-recorded data of related data that is information related to the evaluation, based on the related data and the evaluation value of each of the contents A compression conversion process for calculating compressed data that is data obtained by removing a portion that does not affect evaluation from each of the related data; and, based on both information of the compressed data and the evaluation value, An evaluation feature calculation process for calculating user evaluation feature data indicating the feature of the indicated evaluation and content evaluation feature data; An information filtering program for executing, based on user evaluation characteristic data and the content evaluation characteristic data, an evaluation prediction process of predicting and calculating an evaluation given by each user to each content. 22. The information filtering program according to claim 21, wherein the related data includes attribute data representing the contents of each of the contents. 23. The information filtering program according to claim 21, wherein the related data includes user data indicating information on each of the users. 24. In the compression conversion processing, based on the relevant data and the evaluation value, the relevant data is subjected to compression conversion to remove data that does not affect evaluation from each of the relevant data. The information filtering program according to any one of claims 21 to 23, wherein a process for calculating the compressed data is performed. 25. The compression conversion processing, wherein the evaluation conversion method is data that specifies a compression conversion method for compressing the evaluation data based on the evaluation value so as to remove a portion that is not valid for evaluation prediction from the evaluation data. An evaluation compression conversion process for calculating data; and a compression conversion method for specifying a compression conversion method for calculating the compressed data from the relevant data based on the relevant data and the evaluation conversion method data. 25. A second compression conversion process for calculating data, and a compression calculation process for calculating the compressed data from the related data based on a calculation method indicated in the compression conversion method data. Information filtering program described. 26. The compression conversion process according to claim 24, wherein the related data and the evaluation value are subjected to compression conversion by singular value decomposition to execute a process of calculating the compressed data. 25. The information filtering program according to 25. 27. In the evaluation feature calculation processing, data represented by using both information of the compressed data and the evaluation value is subjected to compression conversion to thereby obtain evaluation features indicated by the respective users. The information filtering program according to any one of claims 21 to 26, wherein a process for calculating user evaluation characteristic data indicating content and content evaluation characteristic data is executed. 28. In the evaluation feature calculation process, data indicated using both information of the compressed data and the evaluation value is compression-converted by singular value decomposition, and the evaluation indicated by each of the users is performed. 28. The information filtering program according to claim 27, wherein the information filtering program is configured to execute a process of calculating user evaluation characteristic data indicating the characteristic of (1) and content evaluation characteristic data. 29. An evaluation input process for inputting and recording an evaluation value indicating an evaluation of each of the respective users for each of the contents, which has been shown by each of the users so far, and a content of each of the contents. The information filtering program according to any one of claims 21 to 28, wherein an attribute extraction process for extracting attribute data which is data expressing the attribute is executed. 30. The information filtering program according to claim 21, wherein the content published on the Internet is filtered.