JP3452308B2 - Data analyzer - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION This invention relates to a data analysis for analyzing the collection of data represented by a numerical value or symbol information stored in a storage device such as a database device, processed converted into useful representation for users on the equipment.

[0002]

2. Description of the Related Art Due to advances in computer technology, the amount of data stored in computers is increasing year by year. In particular, this trend is becoming more and more noticeable, especially in online systems, as the network is advanced. At present, it is not uncommon for the number of records to exceed 1 million and the data amount to exceed giga (= 10 9) bytes.

The data accumulated in the computer is often only a collection of numerical values and symbols. Therefore, a technique has been proposed in which this collection of data is converted into useful information for the user and the data is effectively utilized. The most widely known method is a statistical method such as correlation analysis or multiple regression analysis.

Further, as a relatively new method, a method of converting a rule format such as "if ... if it is easy for the user to understand" and outputting the result, that is, a knowledge acquisition method called rule induction is used. The method is known. For example, Hitachi Creative Workstation 2050 (product name) ES / TOOL /
Pages 23 to 31 of the W-RI commentary / operation manual describe a method of expressing the relationships existing between data in the form of rules.

This method was originally intended to generate rules that can be input to the expert system from given data. However, it is possible to utilize features such as causality and regularity included in the accumulated data for the purpose of discovering by a human user.

[0006]

SUMMARY OF THE INVENTION The above-mentioned prior art aims to generate rules that can be used by a computer. Therefore, although it is possible for a human user to interpret the rule, it does not necessarily generate the rule in a format that is easy for humans to understand, and thus it is difficult for humans to interpret and understand the characteristics of data. It was not possible to generate a valid rule. The method described in the above document will be explained using an example.

First, consider the data as a set of individual cases. For example, in a method of using a quality control database in a semiconductor manufacturing process to analyze semiconductor defect factors, each case is managed in a manufacturing unit called a wafer, and processing parameters in various manufacturing processes and various types are managed. A set of information such as inspection results can be treated as one case. FIG. 1 shows an example of such data.

[0008] In addition, in a usage method such as investigating a customer's purchase trend of a financial product from a customer database of a bank, an example is a set of information for each customer such as age, deposit balance, occupation, annual income, financial product purchase history. , The data to be analyzed can be regarded as a collection of this case.
This example will be described in detail in Examples.

An example of rule generation according to the above conventional technique will be described. As an example, consider examining common features of customers who have purchased a financial product (denoted as product A). In this case, in the prior art, the cases corresponding to the customers who have purchased the financial product A and the cases corresponding to the customers who have not purchased the financial product A are classified as accurately as possible based on the values of the respective items of data (age, deposit balance, etc.). The goal is to generate rules.

In the above-mentioned prior art, a set that most accurately classifies given data among the set of item values (for example, the age is 40 years or older and the deposit balance is 10 million or more) is generated. The accuracy in this case means that the characteristics of the customer who purchased the product are classified more accurately as the proportion of cases corresponding to the customer who purchased the financial product A in the subset of cases having a specific value is larger. I think. This set of values is "IF age 40 and over AND deposit balance 10
It can be expressed in the form of a rule "more than one million THEN purchase financial product A".

Next, the case explained by the generated rule is removed from the entire set of cases. In the above example, the case where the age is 40 or over and the deposit balance is 10 million or more is excluded. Then, with respect to the remaining set of cases, a set of values of items that are most accurately classified is obtained. By repeating the above processing, the financial product A
It is possible to obtain a rule group for distinguishing a customer who has purchased from a customer who has not purchased.

As can be seen from the above description, the rule group obtained by the above-mentioned prior art is that the IF age is 40 years old or more AND the deposit balance is 10 million yen or more
HEN Buy financial product A ELSE IF Occupation is self-employed AND Annual income is over 8 million THEN
Purchase financial instrument A ELSE IF…, such as IF… ELSE IF… ELSE IF….

When the computer classifies using this rule group, the processing can be executed mechanically only by sequentially checking from the first IF. However, as the number of rules increases, it becomes more difficult for humans to understand the characteristics of the customer who purchased the financial product A from this rule.
Furthermore, since the process of searching for a rule from the set of remaining cases is repeated every time one rule is generated,
It is also a problem that the required processing time increases rapidly as the number of cases increases.

More seriously, in the case of real-world data such as the above example, it must be considered that the data generally contains a very large amount of noise.
In other words, whether or not to purchase the financial product A should not be essentially determined from the items included in the database, so that it is not expected that a highly accurate classification rule is originally generated. Further, in the case of the above-described semiconductor failure factor analysis, since the occurrence of defects is influenced by the factors that randomly change, the data contains large noise. Even in such a case, it is often difficult to desire to generate a deterministic rule.

In the case of the above problems, an analysis method that expresses the general characteristics of the data is effective. However, in the method like the above-mentioned conventional technique, since the values of many items are combined to search for a rule with good classification accuracy, the number of conditions generally appearing in the IF part increases, but a case applicable to that rule. The phenomenon occurs that the number of is reduced. This makes it difficult to meet the goal of understanding the general characteristics of the data.

A great deal of information is stored in an actual database. In the case of the above semiconductor quality control data, the wafer number, the date of manufacturing start, etc., and in the case of the customer data, the name, telephone number, etc., which are clearly unrelated to the purpose of analysis, are included. ing. On the other hand, in the case of semiconductor quality control data, the product type code,
In the case of customer data, there is also information that is useful for analysis, such as an address.

Therefore, in the case of performing the analysis as in the above example using such data composed of many kinds of information, first, what item of data should be used for classification? The user must specify in advance. The more items you have, the more complex this task becomes. If an attempt is made to analyze all the items that may possibly be related, the number of items to be used will inevitably increase, resulting in an increase in processing time. Therefore, if you want to analyze efficiently, you must carefully select the items to be used, and depending on the degree of know-how of the user,
The quality of the analysis results will be greatly influenced.

Further, depending on the item, analysis from various viewpoints may be required. For example, in the case of customer data, if the trends of product purchase differ by region, whether the item of address should be considered by prefecture, whether it should be considered at the level of Tohoku region / Kanto region, or the two categories of East Japan / West Japan Basically, I don't know unless I analyze it. If all the cases like this are to be exhausted, it is necessary to repeat the analysis by the conventional technique over and over again, and there is a problem that the burden on the user increases.

In order to avoid this, there is also a method of adding all viewpoints at various levels to the data item.
In the case of addresses, it is possible to consider items such as prefecture / region / east / west data items to be analyzed.
However, in the analysis according to the conventional technique, no significant relation between items is taken into consideration, so that a great waste occurs in processing.

For example, while trying to classify by prefecture, it is not necessary to consider the values of the items of higher level, such as by region or by east and west, but nevertheless, it is nevertheless conventional. With technology, useless analysis is carried out. Furthermore, depending on the processing order of items, there is a possibility that rules that are clearly semantically redundant, such as “IF address is Kanto region AND address is Kanagawa prefecture THEN…” are generated. For computer classification purposes, such redundancy does not affect classification accuracy, but is obviously detrimental to humans' understanding of the characteristics of the data.

In addition, an actual database often contains unknown data items called missing values.
When performing analysis using statistical methods, etc., it is unavoidable to simply ignore missing values and regard them as having no data. Even in the above rule induction method, a data item having a missing value does not affect the classification accuracy and therefore does not appear in the classification rule.

However, there are cases where it is significant that the data values are missing. For example, when the item of address is a missing value, which means that a bank account is anonymously created, this may affect the purchase of the financial product A. In such a case, the rule that "IF address is missing value THEN ..." has meaning. In the conventional method, such a rule cannot be generally generated, and in order to generate such a rule, a manual process of explicitly converting a missing value into a specific value is required. It was

Further, in the above-mentioned method according to the prior art, since the rule is generated by giving priority to the classification accuracy, a general rule is not always generated first in the sense of explaining as many cases as possible. . On the other hand, in the data analysis method that we are currently thinking about, the user may interrupt during the long processing time. In such a case, if a general and simple rule that has high utility value for the user is generated first, it is possible to use the rule generated up to the time when the interrupt is issued. There is also a problem that such a method cannot be used by the method based on the technology.

The invention, in equipment you converts useful representation for the user to analyze the data stored in such a database, even massive data including noise considerable processing time It aims to be able to present the characteristics of the data in a format that is easy for humans to understand without requiring it. Moreover, if even a user having no special knowledge about the processing method, and an object thereof is to provide a data amount 析装 location that can produce results that represent accurately the characteristics of the data.

[0025]

The present invention provides a database
The data item based on a plurality of data items stored in
A data analyzer that generates rules that characterize eye-to-eye relationships.
Data stored in the database
IF clause and THEN of IF-THEN rule selected from the items
Means for receiving the data items used in the section and the
Received and received data item, item name and symbol value
IF-THEN format by combining data items consisting of
Generating means for generating a rule candidate of the
Coverage equivalent to the number of data that IF clause applies, and IF clause
And the THEN clause are both accurate enough to
Each value is counted, and each value is
And the higher the coverage rate and the better the accuracy, the higher
From the evaluation means for obtaining an evaluation scale and the rule candidates,
Determine and output at least one rule candidate with a high valuation scale
And an output means for

The present invention is also stored in a database.
The relationship between the data items based on multiple data items
A data analysis device for generating rules for characterizing,
Select from the data items stored in the database
The data used in the IF and THEN clauses of the IF-THEN rule
Means for receiving data items and said selected and receiving
If the data item contains a numeric value,
A conversion means for converting to a value and said selected and received
Data items (not including numerical values) and the conversion means
From the data items that are converted from numerical values to symbolic values,
IF by combining data items consisting of name and symbol value pairs
-Generating means for generating rule candidates in THEN format,
Corresponding to the number of data items that the IF clause of the candidate
Equivalent to the rate and the number of data that both IF clause and THEN clause apply
The accuracy values for each rule are counted to determine the rule candidates.
For each, the coverage rate is high and the accuracy is high.
Evaluation means for obtaining a higher evaluation scale and the rule
From the supplement, at least one rule candidate with a high evaluation scale was decided.
It is characterized in that it is provided with an output means for determining and outputting .

[0027]

[0028]

[0029]

[0030]

[0031]

According to the present invention, first, from the data items stored in the database, the IF clause of the IF-THEN rule and THE
The data items used in Section N are selected. If the selected data item contains a number, that number is converted to a symbolic value. Next, rule candidates representing the correlation between the selected data items in the rule format having at least one pair of item name and symbol value are generated, and for each of the rule candidates, the strong correlation between the data items is generated. Is evaluated.
Then, a rule candidate having a high evaluation value is determined and output.

According to the present invention, first, one data item used in the conclusion clause of the rule and a plurality of data items used in the conditional clause of the rule are selected from the data items stored in the database. If the selected data item contains a number, that number is converted to a symbolic value. Next, a conditional clause is generated by combining at least one pair of item name and symbol value of the condition item, and a conclusion clause is generated by combining at least one pair of item name and symbol value of the conclusion item, A plurality of rule candidates are generated by combining the generated conditional clause and conclusion clause. For each of the rule candidates, the strength of correlation between data items is evaluated. Then, a rule candidate having a high evaluation value is determined and output.

[0033]

Embodiments of the present invention will be described below with reference to the drawings. In the present embodiment, the above-described example of analysis of customers who purchase financial products is used.

FIG. 2 shows an example of data to be analyzed. Consider analyzing the purchase trends of a financial product. The data in Figure 2 includes customer number, name, age, gender, address, branch code, prefecture code, area code,
The purchase record of a certain financial product is stored together with information such as deposit balance, loan balance, and credit card type.

A user who wants to analyze this data
By analyzing what attributes of the customer influence the purchase of the financial product, strategies for expanding the sales of the financial product such as direct mail and door-to-door sales (what kind of customer is best to approach, etc.) ) Is intended to be considered.

FIG. 3 shows the overall flow of the analysis process. First, in process 301, analysis target data is defined and a table of analysis target data as shown in FIG. 2 is created. Next, process 302
Then, the data to be analyzed is read.

In process 303, first, the item "purchase of goods" is designated as a conclusion item. Next, each item of data is examined to see if the item has a symbolic or numeric attribute. It is determined whether or not each item is selected as a condition item, and if the selected item is an item having a numerical attribute, it is converted into a symbol value. Finally, the data that becomes the input of the rule generation processing is generated.

In process 304, parameters (maximum number of conditional clauses, minimum coverage, etc.) of the rule generation process are determined based on information such as the number of cases. In step 305, the determined parameters are displayed on the display device, and the user corrects the parameters if necessary. In process 306, the rule generation process is executed, and in process 307, the result is output to the output device.

Although the rule learning process is omitted in FIG. 3, after the process 307, the learning process can be executed at any time according to a user's instruction.

FIG. 4 shows a block diagram of a system for executing the above processing. The system consists of three processing modules. A user interface module 401 for controlling interaction with a user, a rule generation processing module 402 for generating rules from data to be analyzed, and a rule learning processing execution module for executing fine adjustment of rules generated by learning membership functions. 403.

The module 401 recognizes an instruction from the user as a command, inputs data from the analysis target database 404 based on the instructed method, and inputs the data to the rule generation processing execution module 402 as a rule generation work file. 405 is generated, and at the same time, information of parameters for controlling the rule generation processing is output to the rule generation parameter file 406. When the rule generation execution command from the user is input to the module 401,
The module 401 activates the module 402.

When the module 402 is activated, the information in the files 405 and 406 is read, the rule generation processing is executed on the data in the file 405 with the parameters specified in the file 406, and the result is stored in the rule file 407. Output. In the rule file 407,
Information on the generated rule and information on the membership function that defines the predicate included in the generated rule are stored.

When the execution of the module 402 is completed, the control is returned to the module 401, and the generated rule is displayed to the user in a predetermined format. The user is
Based on the display result, it is possible to continue the operations such as changing the parameters and executing the above processing again.

When the rule learning command generated here is input to the module 401, the module 401
Activates the module 403. When the module 403 is activated, first, the contents of the database 404 and the rule file 407 are read, and the membership function learning is executed for each rule contained in the rule file 407. After the processing is completed, the learned membership function information is stored in the rule file 407 again, and the control is returned to the module 401.

Of the processing shown in FIG. 3, the processing 306 is executed by the module 402, and the remaining processing is performed by the module 4.
It is executed at 01.

The details of each process will be described below in order according to the flow of FIG.

5 and 6 are views for explaining the contents of the analysis target data definition processing executed in the processing 301. The data to be analyzed is stored in a plurality of relation tables. From the customer information table of FIG. 5 (a), the branch information table of FIG. 5 (b), and the product purchase history table of FIG. 5 (c), respectively. Become. FIG. 6 is a diagram schematically showing the table structure definition of FIG. 2 to be analyzed by designating the mutual relation of the three tables of FIG. These definitions are defined by the relational database management system (RD
BMS) can be used.

In process 302 of FIG. 3, the analysis target data shown in FIG. 2 is read from the database and expanded in the memory in the computer. When reading data, determine whether the value of each item is a numerical value or a symbolic value, and if it is a numerical value, obtain the maximum / minimum value, and if it is a symbolic value, find the number of different symbols for each item. .

FIG. 7 shows the conditions / conditions in the process 303 of FIG.
A screen for specifying a conclusion item is shown. In the process 303, this FIG.
The screen is displayed as described above, and the user is prompted to specify the condition item and the conclusion item.

First, immediately after reading the data, no item is selected, and the names of the items of the analysis target data in FIG. 2 are displayed in the unused item name list box 701, and the condition item name list box is displayed. The contents of 702 and the conclusion item name text box 703 are empty.

Here, when the list box 701 is displayed, among the attributes having the symbol value obtained in the process 302, the number of different symbols is a predetermined number, for example, 20 or more, the condition item. / It is not possible to select as a conclusion item, and it is displayed that the item name is not selectable. This is to prevent the generated rule from being subdivided and lowering the utility value when an item having a larger number of symbol values than the number of cases is designated as a condition or a conclusion item. For example, the customer number, name, and address in FIG. 7 correspond. In FIG. 7, “*” is added to the head of the item name to indicate that the item cannot be selected, but of course the same effect can be achieved by using a shaded display / light color display. By doing so, it is possible to prevent the use of items that have no meaning even when used for rule generation.

In the list box 701 (selectable)
Select multiple items using an input device such as a mouse, and press the button 7
By clicking 06, the selected item is selected as a condition item, the item name is deleted from the list box 701, and the item name is added and displayed in the list box 702. Similarly, a conclusion item is selected by selecting one item in the list box 701 and clicking the button 709. Similarly, when deleting the items in the list boxes 702 and 703, the deleted items are selected and the buttons 707 and 710 are clicked.

By clicking the button 705,
All the items displayed in the list box 701 other than the non-selectable items can be added to the condition items. On the contrary, the button 708 is a button for deleting all the items in the list box 702 from the condition items.

After completing these selections, button 711
By clicking, the condition / conclusion item selection process ends. Further, when the button 712 is clicked, all item selection information at that time is discarded.

Next, the condition item automatic selection processing executed when the button 704 is clicked will be described. However, since this processing can be executed when the conclusion item is specified, the button 704 is valid only when the conclusion item is specified.

FIG. 8 is a processing flow of the condition item automatic selection processing. In FIG. 7, when the button 704 is clicked after the conclusion item (text box 703) is designated, the processing of FIG. 8 starts to be executed.

In the process 801, the processes 802 and 803 are repeated for all items except unusable items. In processing 802, it is checked whether the item currently processed has a numerical value attribute or a symbol value attribute, and if it has a numerical value attribute, processing 803 is executed.

In process 803, the value of the item being processed is converted into a predetermined number, for example, five symbol values. In the example of the item “age” in FIG. 2, the range of possible values of the item value is, for example, “less than 20” “20 or more and less than 30” “30 or more and less than 40” “40 or more and less than 50” “50”.
The above is divided into 5 ranges, and is classified into 5 categories. There are various methods for this classification, for example, the following three methods.

1) Equal range division: The range of possible values is equally divided. For example, if the minimum age value is 0 and the maximum age value is 75, 0-14, 15-29, 30-44, 45-59,
Classify as 60-75.

2) Mean / standard deviation division: The mean and standard deviation of the age value distribution are obtained, and division is performed based on these values. For example, if the average is μ and the standard deviation is σ, μ−0.84σ, μ−0.
The values of 25σ, μ + 0.25σ, μ + 0.84σ are set as the division points. This value is a value that equalizes the probabilities included in each classification, assuming that the distribution has a normal distribution.

3) Equal division: The distribution of actual age values is examined, and classification is performed so that the number of cases included in each classification is equal.

As to which method to use for classification, a default method is prepared in advance, and a means that can be designated by the user is prepared as needed.

The numerical data classified by the above method is converted into symbol value data having values of "category 1", "category 2" to "category 5" for convenience.

In process 804, all the similarity measures between two items are calculated for all combinations of any two items in the items other than the item designated as the conclusion item. The similarity measure is calculated as follows.

First, let two items be X and Y, respectively.
The possible values are xi (i = 1, Nx) and yj (j = 1, Ny). X,
Y may be an item originally having a symbol value or an item obtained by converting a numerical value into a symbol value by the above method.
Considering X and Y as random variables, mutual information is defined between them.

The mutual information amount is explained in many textbooks of information theory. The mutual information is, for example, X =
It is an amount related to how the information amount of the probability distribution of the value of Y changes when the information xi is given, that is, an amount indicating the degree of association between X and Y. The mutual information amount I (X; Y) is defined by the following equation. I (X; Y) = ΣiΣj P (xi, yj) log [P (xi, yj) / P (xi) P (yj)] (1)

Here, Σi represents the sum of the subscript i, P (xi) represents the probability that X = xi, and P (xi, yj) represents the probability that X = xi and Y = yj. The probability is calculated from the frequency of appearance of records in the data. For example, P (xi) = N (xi) / N, N is the total number of records, and N (xi) is the number of records where X = xi.

There is a property that I (X; Y) ≧ 0.
It can be considered that the larger the value of (X; Y), the more the relationship between X and Y is shown. In extreme cases, X
And Y are completely independent, P (xi, yj) = P (xi) P (yj)
And I (X; Y) = 0.

From this property, the similarity measure D (X; Y) between the item X and the item Y can be defined by the following formula. D (X; Y) = 1 / I (X; Y) (2) However, when I (X; Y) = 0, it is promised that the value of D will be a very large fixed value.

In process 805, items other than the conclusion items are clustered on the basis of the similarity scale D calculated for each combination of all items other than the conclusion items in process 804. That is, D (X; Y) is regarded as the distance between the item X and the item Y, and items with close distances are put together based on the distances between all items and classified as some clusters.

Various algorithms are known for the clustering method, and the details thereof are described in general statistical textbooks and literatures, and therefore the description thereof is omitted here. For example, a method called k-means method is used. As a result, among items that are candidates for condition items, items having a high degree of similarity (relevance) to each other can be put together in one cluster.

In many algorithms, the number of clusters is designated in advance. Here, it is assumed that the default value is 1/10 of the number of items of all data as the number of clusters.
That is, when the number of items of the analysis target data in FIG. 2 is 62, 62/10 = 6 (truncated) clusters are generated. Of course, a means for the user to directly specify the number of clusters should be prepared if necessary.

In process 806, a dependency measure is calculated for the combination of the conclusion item and the other items. The dependency measure is, for example, the mutual information I described in the processing 804.
(X; Y) can be used.

In process 807, for the clusters classified in process 805, items having the largest dependency scale with the conclusion item among the items belonging to each cluster are selected, and those items are adopted as condition items.

As described above, according to the processing of FIG. 8, the degree of relevance to the conclusion item is the highest (the degree of dependence is high) and the degree of relevance between the condition items is as low as possible from the data in which the items having numerical values / symbol values are mixed. It becomes possible to automatically select the condition item (greater independence). As a result, the accuracy of the rules generated by the following processing can be improved, and since similar condition items are not used, the independence between rules can be increased, and the readability / usefulness of the generated rules can be improved. There is an effect that can be.

In process 303, the missing value countermeasure process is executed for the item selected as the condition item in the above process. In the missing value countermeasure process, the following steps are performed for each item selected as the condition item.

Step 1: It is checked whether or not missing value data is included in the selected item. The missing value data is
It shall be represented by a special numerical value (eg negative maximum value) or symbolic value (eg empty string). If no missing value is included, the process ends.

Step 2: The mutual information amount I (X; Z) with the conclusion item Z when the value of the selected item is classified as a missing value or a value other than the missing value is obtained. X is a random variable having a value indicating whether the value of the item being selected is a missing value or a value other than the missing value.

Step 3: When the mutual information amount I (X; Z) obtained in step 2 is larger than a predetermined threshold value Ith, whether the value is a missing value or not and a conclusion item are predetermined. Since it is related to more than the degree of, the missing value of the selected item is added as a new symbol value.

By the above processing, when it is determined in step 3 that the missing value is meaningful, the special symbol value "missing value" is added to the item and the rule generation described later is performed. You will be able to use the symbolic value. In other cases, the missing value of the item is ignored in the rule generation processing described later (it does not belong to any symbol value and never appears in the generation rule).
Treated as a special value.

The numerical value → symbol value conversion process in the process 303 of FIG. 3 is performed as follows.

FIG. 9 is a list display of numerical value → symbol value conversion methods. The user refers to the display as shown in FIG. 9 on the display device and specifies the conversion method for each item using the mouse or the like.

In FIG. 9, in the item name display field 901,
The name of the item that originally has a numeric attribute and is converted to a symbol value is displayed. The symbol number display field 902 displays the number of different symbols after conversion into symbol values. Symbol name display column 9
In 03, the name of the symbol value is displayed.

In the example of FIG. 9, the age is less than 20, the twenties,
Different symbol values are assigned, such as: deposit balance is small, normal, large, loan balance is small, medium, and so on. In the conversion method display field 904, the numerical value → symbol value conversion method described in the processing 803 of FIG. 8 is displayed.
In the example of FIG. 9, the conversion method corresponding to the age is designated by the user, and this user designation indicates that the user directly designated the division method.

After the designation of the symbol value conversion method for each item is completed, the conversion method specification process is completed by clicking the button 906, and the specification contents up to that point are discarded by clicking the button 907. When the button 905 is clicked, the conversion method of all items is reset by the default symbol value conversion method as described in the processing 803 of FIG.

To specify the conversion method for each item, use the column 9
Double-click the 01 item name. As a result, the conversion method designation screen of FIG. 10 is displayed, and the conversion method of the item can be designated.

In FIG. 10, the option button 100
1 is a method of converting the symbolic value of the selected item such as even number division, equal range division, average / standard deviation division, and user specified division, and whether each division method is performed in a crisp manner. Specify whether to do it manually.

In the text box 1002, the number of symbols to be converted into a symbol value is input.

The current symbol name is displayed in the symbol name display field 1003. When the user wants to change the symbol name, he / she can directly input the symbol name by clicking the corresponding area of the column 1003. The same applies to the minimum value display column 1004 and the maximum value display column 1005 of each symbol value. However, when the user directly changes the minimum value or the maximum value, the option button 1001 automatically changes to a state in which the user-specified division method is specified.

When the value of the column 1004 or the column 1005 is changed, the content of the record number display column 1006 is automatically changed accordingly. This is done by using the data stored in the memory and counting the number of records corresponding to each symbol value.

By specifying either crisp division or fuzzy division among the option buttons 1001, whether the symbol value conversion processing is performed in a crisp or fuzzy manner within the range of the minimum / maximum values displayed in the columns 1004 and 1005. Can be specified. By performing the method in a fuzzy manner, the amount of information lost by the symbol value conversion can be minimized and the accuracy of the generated rule can be improved. Hereinafter, this process will be described.

Even when the fuzzy division is performed, the value range of each symbol value is displayed as the minimum value / maximum value in the columns 1004 and 1005 of FIG. 10, but it is interpreted as fuzzy. FIG. 11 is a diagram for explaining such fuzzy division using the example of symbol value conversion of the loan balance of FIG.

The horizontal axis of FIG. 11 shows the value of the loan balance in units of 10,000 yen, and the vertical axes are "small", "medium" and "large".
Represents the goodness of fit of the (fuzzy) symbolic value. "small"
The membership function 1101 corresponds to the symbol value of, the membership function 1102 corresponds to “medium”, and the membership function 1103 corresponds to “large”. The shape of the membership function is such that the absolute value of the slope is constant and the dividing points of each symbol value (in the figure, the dividing points are “small” and “medium”).
Membership function value becomes 0.5 at each of 0, "medium" and "large" dividing points of 1000), and a membership function having a value other than 0 for any value is The number is at most two, and the sum of membership function values can be set to 1.0.

For example, in FIG. 11, "loan balance = 600
The numerical value "(10,000 yen)" has a membership value of 0.8 for the symbol value of "medium" and 0.2 for the symbol value of "large". In other words, a record having a value of “loan balance = 600” behaves as if the record of “loan balance” is divided into 0.8 records for “medium” and 0.2 records for “large”. Can be regarded as a thing.

In the case of such fuzzy division, as shown in FIG.
The number of records displayed in the 0 column 1006 can also be considered as the sum of the membership values, and generally takes a real value.

Next, the rule generation parameter calculation processing in the processing 304 of FIG. 3 will be described. The following five types of parameters are considered as parameters for controlling the rule generation process.・ Number of generated rules ・ Overall coverage rate ・ Maximum number of clauses ・ Minimum coverage rate ・ Coverage priority coefficient

The number of generated rules is a parameter that specifies the number of rules to be generated in the rule generation processing. The default value is determined according to the following relational expression from the number N of condition items to be used.

(Default value of number of generation rules) = N *
(Maximum number of conditional clauses) This reflects that generally, the greater the number of conditional items, the greater the number of rules that should be generated to explain the characteristics of data.

The coverage is defined as the ratio of the number of cases satisfying the condition part of a rule to the total number of cases.
The overall coverage ratio is a parameter that specifies the ratio of the number of cases that satisfy the condition part of at least one rule to the total number of cases in the generated rules. The larger the number of generated rules, the larger the overall coverage. No default value is prepared for the overall coverage (that is, the value can be set only by the explicit specification of the user).

The maximum number of conditional clauses is a parameter that specifies the maximum value of the number of predicates included in the conditional part of the rule. For example, if you specify the maximum number of conditional clauses is 3, if X1 is A1 and X2 is A2 and X3 is A3.
Z can generate a rule called B, but in addition, "~ and X4
If is A4, the rule "~" will not be generated (because the number of conditional clauses is 4). The default value is 3.

The minimum coverage ratio is a parameter for designating the minimum value of the coverage ratio of the generated rule, and the rule having the coverage ratio less than the value specified here is not generated as a special rule. The default minimum coverage is 10
0 / (number of generation rules) (percentage). This is because in general, the larger the number of generated rules, the more detailed rules are required.

The coverage ratio priority coefficient specifies the relationship between the coverage ratio and accuracy in the evaluation scale for the rule.
Details will be described in the rule generation processing described later. The default value is 1.0 (coverage has the highest priority).

In process 305 of FIG. 3, the rule generation parameter calculated in process 304 is displayed on the display device so that the user can modify it as necessary. FIG. 12 shows an example of the display screen.

In the text boxes 1201 to 1205, the value of the rule generation parameter calculated in the process 304 of FIG. 3 is displayed. In the example of the figure, the number of generation rules = 2
4, overall coverage ratio = not specified, maximum condition clause number = 3, minimum coverage ratio = 4%, coverage ratio priority coefficient = 1.0.

The user can input the value of the displayed parameter by directly clicking the text box. On the display screen of FIG. 12, other information of the data to be analyzed (data file name, number of records, number of condition items, conclusion item and its symbol value) is displayed in the display area 1206. It is possible to correct the value of the rule generation parameter while referring to the information of.

After the parameter is modified, the button 1207 is clicked to make the modified value effective, and when the button 1208 is clicked, the modified value is discarded.

In the process 306 of FIG. 3, the rule generation process is executed according to the rule generation parameters designated in the processes 304 and 305.

In the rule generation processing, rules of the following format are generated.・ If X1 is A1 and X2 is A2 and ...
If Xn is An, then Y is B. Here, Xi is the condition item name, and Ai is the symbol value name of the item Xi. Similarly, Y is the conclusion item name, and B is the symbolic value name of the item Y.

The set "Xi is Ai" is called a conditional clause. The above rule is an example of a rule having n conditional clauses. The entire conditional clause is called the conditional part. Moreover, "Y is B" is called a conclusion clause. In this embodiment, the conclusion section is made up of one conclusion clause.

A real value called an evaluation scale is assigned to the rule. A rule with a larger evaluation scale is considered to be more valuable as a rule.

In the process 306 of FIG. 3, a process of selecting a rule having a large rule evaluation scale from the combination of the conditional clause and the conclusion clause is executed. This makes it possible to select rules with high value.

In this embodiment, "if A, then
The evaluation scale μ (A → B) of the rule “B” is defined by the following equation. μ (A → B) = P (A) ∧β * log [P (B | A) / P (B)] (3)

Here, a∧b means a to the b-th power. P
(A) represents the probability that the condition part A is satisfied in the data to be analyzed, that is, the proportion of cases satisfying the condition A in all cases. Similarly, P (B) represents the probability that the conclusion part B is satisfied, and P (B | A) represents the probability that the conclusion part B is satisfied on the assumption that the condition A is satisfied. β is the cover ratio priority coefficient described in the process 304 of FIG. Rewriting equation (3) with the number of cases, P (B | A) = P (A
& B) / P (A), μ (A → B) = [N (A) / N] ∧β * log [N ・ N (A & B) / N (A) N (B)] (4 ).

However, N: Total number of cases N (A): Number of cases that satisfy the condition part A N (B): Number of cases that satisfy conclusion B N (A & B): Number of cases that simultaneously satisfy the condition part A and the conclusion part B Is.

The first factor in the definition of equation (4) is the β power of the rule coverage, and the second factor corresponds to the accuracy. That is, the larger the coverage rate (explaining as wide a range of cases as possible) and the better the accuracy (the greater the probability that the entire rule will hold), the higher the evaluation scale. In general, the accuracy tends to decrease as a wide range is covered, so these two factors tend to conflict with each other.
Coverage priority coefficient β, which is one of the rule generation parameters
Determines how much importance is placed on the coverage ratio in the evaluation scale. When β = 1.0, the coverage ratio is given the highest priority, and when β = 0.0, the coverage ratio is not considered and only the accuracy is considered. And evaluate the rules.

Of course, the rule evaluation scale defined here is just one candidate, and it is possible to add the definitions of several evaluation scales, switch those definitions as user options, and use them. is there.

In the rule generation processing, a specified number of rules are generated from the rules having a large rule evaluation scale defined by the equation (4). The space to search is a combination of all possible conditional clauses.

Even with the problem of 10 condition items and the number of symbols of each is about 3, since this search space is enormous in size (on the order of about 30 to the 10th power), it is necessary to execute the exhaustive search in a realistic time. Is impossible. Therefore, the method described below is used to increase the efficiency of the search by limiting the search range and pruning.

First, the rule search range can be limited by using the maximum number of conditional clauses and the minimum coverage rate.

As a search method in the space after these restrictions, the depth-first search method is used in this embodiment.

FIG. 13 shows an image of the search space. Since it is basically a combination of conditional clause predicates, a tree structure search method can be used.

In FIG. 13, for example, the node 1301
The rule corresponding to is if X1 isA then…,
The rule corresponding to the node 1302 is if X1 is A a
ndX2 is B then…. That is, the conditional clause of the rule corresponding to the child node is the conditional clause of the rule corresponding to the parent node plus one predicate. The numerical value attached to the vicinity of each node indicates the predicate of the conditional clause of the rule corresponding to the node. For example, node 130
In the vicinity of 1, {1} is written, which indicates the predicate "X1 is A" in the corresponding rule. Also, {12} is written near the node 1302, which means that the predicate “X1 isA and X2 is” in the corresponding rule.
B ”is shown.

FIG. 13 shows the case where the maximum number of conditional clause predicates is three. The height of the tree (the maximum number of links from the root to the terminal node) is the maximum number of conditional clause predicates.

Since the order of the predicates appearing in the conditional clause has no meaning, for example, the child node of the node 1303 is not {21},
Note that it starts at {23}. Further, one conditional clause does not include two predicates having different symbol values in the same condition item. It is necessary to take this condition into consideration when adding a predicate to the conditional clause of the parent node to generate a child node.

In the depth-first search, all the nodes shown in FIG. 13 are generated and evaluated in the following order. 1 → 12 → 123 → 124 →… → 12n → 13 → 134 → 135 → …… → (n-
2) (n-1) n → (n-1) → (n-1) n → n

In the case of the depth-first search, the candidate rules may always be searched in the above order while holding N rules (N is the number of generated rules) having the largest evaluation scale until then.

Next, a pruning method for speeding up the above rule search will be described. Since the number of nodes in FIG. 13 is about the cube of 30 in the above example, it is not impossible to directly perform an exhaustive search, but several tens to several millions of data are targeted, and the number of items is 100. Since there are quite a few cases, it is necessary to adopt a search method that is as efficient as possible.

The basic idea of efficiency improvement is “pruning”. That is, by evaluating the state at a certain intermediate node, the evaluation of the child node can be skipped. In the case of the rule generation processing in the present embodiment, pruning based on the coverage ratio and pruning based on the rule evaluation measure are performed.

The basic idea of the pruning method based on the coverage ratio is that when a rule corresponding to a node is generated and the coverage ratio of the rule is below a user-specified lower limit of the coverage ratio, that node and all This means that it is not necessary to search the child nodes of. Since the conditional clause of the rule corresponding to the child node is the conditional clause of the rule corresponding to the parent node plus one predicate, if the rule coverage of the parent node falls below the lower limit, the rule of the child node The coverage of will always fall below the lower limit.

Pruning by the rule evaluation scale will be described. When N rules are generated in the search, the evaluation measure of the Nth candidate rule at a certain point in the search is μth. The basic idea of pruning is to find the maximum value of the rule evaluation scale in a subtree rooted at a node,
If the value is smaller than μth, the subtree node is not generated. Here, how to estimate the maximum value of the evaluation scale of the nodes included in the subtree is an important factor that determines the pruning efficiency.

The evaluation method will be described below. For simplification, the case where the coverage ratio priority coefficient β = 1.0 will be described.
Since it can be easily expanded in the case of general values, it is omitted here.

Assuming that the rule at a node is “if A, then B”, the evaluation scale is defined by the following equation. μ (A → B) ＝ [N (A) / N] ＊ log [N ・ N (A & B) / N (A) N (B)] (5)

Letting A ′ be a conditional clause in which some predicates are added to conditional clause A, the maximum value of μ (A ′ → B) is N (A ′) = N
Limited to (A & Bi). Because in the above equation, N, N (B)
Can be regarded as a constant, and since the above formula is monotonically increasing with respect to N (A & B), N (A & B) can also be regarded as a constant (that is, rules that dare to reduce precision never maximize the evaluation scale). ). Therefore, equation (5) can be written as follows, for which N (A & B) ≤ N (A ') ≤ N
In the range of (A), μ (A '→ B) = [N (A') / N] * log [N ・ N (A '& B) / N (A') N (B)] (5) Find the maximum value. This is because when the above equation is differentiated by N (A '), it always becomes <0. Therefore, when N (A') = N (A & Bi), the maximum value μmax ＝ [N (A & B) / N] * log [ N / N (B)].
If this value is smaller than μth, it is not necessary to search the descendant node of that node.

In an actual search, it is effective to find a rule having a high evaluation scale at an early stage in terms of efficiency improvement. The sooner it is found, the larger the value of μth and the more likely it is that pruning will be performed.

As one method of a search strategy for that purpose, FIG.
Instead of searching the 3 tree structure with depth first,
First, it is possible to examine all cases where the number of conditional clause predicates is 1, and to perform a depth-first search when the number of conditional clause predicates ≧ 2. That is, if β> 0, a rule with a larger coverage has a larger evaluation scale. Therefore, the value of μth can be increased faster by first examining the case of the conditional clause predicate number 1.

By performing such a search, not only the rule generation processing can be speeded up, but also when the user gives an instruction to interrupt the processing in the middle of the rule generation processing, the rule obtained as an intermediate result shows the characteristics of the data. There is also an effect that it becomes a rule expressed as a whole.

It is desirable that the generated rule be as easy as possible for the user to understand. In other words, if there is no difference in the rule evaluation scale, the number of predicates in the conditional clause should be small. As a method of reflecting this in the rule generation, the constraint that "the precision of the rule to which the conditional clause predicate is added must be greater than the precision of the rule before the addition" is imposed. In other words, since the rule was complicated by adding the conditional clause predicate, the precision of the rule must be improved accordingly. By adding a rule and generating a rule, it is possible to generate a rule that is easier to understand and has high utility value.

Furthermore, in the rule generation processing of this embodiment,
A rule is generated in consideration of a one-to-many dependency relationship between condition items that originally have a symbolic value attribute. For example, consider the prefecture code and branch code in FIG. Since the branch code is unique to each branch, if the branch code is decided, the prefecture code is uniquely determined. The opposite is not true when multiple sales offices exist in one prefecture. At this time, there is a one-to-many relationship between the prefecture code and the branch code.

If a rule is generated for data containing these two condition items, there is a possibility that a rule including a branch code and a prefecture code at the same time will be generated in the conditional clause predicate. At least in the tree structure search, the evaluation measure is calculated as a candidate rule. But this is a complete waste. If the branch code is included in the conditional clause predicate, the prefecture code is uniquely determined from that code, so there is no need to add a conditional clause predicate related to the prefecture code.

The one-to-many relationship between categorical data is
It can be regarded as a hierarchical relationship of classification concepts. That is,
The prefecture code is the upper classification concept, and the branch code is the lower classification concept.

This hierarchical relationship cannot be expressed by a tree structure.
As an example of a branch code, consider a local code and a branch scale code (large scale, medium scale, small scale, etc.). Since the local code is a higher-level concept of the prefecture code, there is a hierarchical relationship between the sales office code and the local code. further,
Specifying the branch code also determines the branch scale code, so there is also a hierarchical relationship here. That is, the prefecture code and the branch scale code are both superordinate concepts of the branch code. On the contrary, one categorical data may include a plurality of subordinate concepts (for example, branch code and supplier code for prefecture code, etc.).

In this embodiment, for each condition item, a list of items corresponding to its superordinate concept is created. This list can be automatically created from the definition of the analysis target data shown in FIG. 6, for example. That is, in a relational database, the primary key of each table must be able to uniquely specify a record in the table, so there is a one-to-many relationship between the primary key items in the table and other items. Always exists. In the example of FIG. 6, there is a one-to-many relationship between the branch code and the prefecture code and between the branch code and the area code. Of course, these relationships can be easily and automatically discovered when the mutual information amount between the condition items is calculated at the time of selecting the condition item described in the process 303 of FIG.

When searching the tree structure at the time of generating a rule using these superordinate concept lists, whether the superordinate or subordinate items of the item corresponding to the predicate to be added are already included in the conditional clause. (Check not only the direct upper / lower concept, but also the upper / lower lower items). If it is included, the predicate of that item is not added.

Next, the processing contents when the fuzzy division is used in the numerical value → symbol value conversion will be described.

In the rule generation processing of this embodiment, the feature of the fuzzy division method shown in FIG. 11 is used. In the fuzzy division of Fig. 11, one numerical value is distributed into at most two symbolic values and their sum is always 1.0 (eg "small" 0.3, "medium" 0.7, "large" 0.0). . In addition, the rule generation work file 405 in FIG.
In the medium expression, "small" is assigned to 0, "medium" is assigned to 1, and "large" is assigned to 2. At this time, the number x is converted into one real number y by the following rules.

Fuzzy symbol value conversion rule: If the ratio of "small" is z, the ratio of "medium" is
If (1-z), y = 1-z If the ratio of "medium" is z, the ratio of "large" is
If (1-z), y = 2-z, but 0 ≦ z ≦ 1.

That is, the real value y between 0 and 1 is
We promise to represent the distribution ratio of the symbols "small" and "medium". From the characteristics of fuzzy division in FIG. 11, one numerical value is 3
It is not distributed over one or more symbolic values, and the sum is 1.0
Therefore, there is no contradiction in this definition.

A case fuzzy symbolized by the above method is defined as (y1, y2, ..., Yn). The number of cases is calculated in the rule generation process when the rule evaluation measure is calculated. In the case of crisp division, the process counts how many cases satisfy a certain conditional clause. The fuzzy symbolization is as follows.

In the case where x1 is “small”, 0 ≦ y
In the case with a value of 1 ≤ 1, the weight is 1-y1.

The case where x1 is “medium” is 0.
In the case of ≤y1≤2, the weight is 1- | 1-y1 |
become.

In the case where x1 is “large”, 1 ≦ y
In the case with a value of 1 ≤ 2, the weight is y1-1.

Of course, when there are a plurality of conditional clause predicates, the weights of the respective cases are multiplied by the above weights for the respective predicates. N (A) and N (A & Bi) appearing in the definition of the rule evaluation scale are considered as the sum of these case weights and treated as real numbers.

If the items converted into the symbol values by the fuzzy division in the above processing are used in the rule generation processing, the data near the boundary in the division can be handled correctly.
The generated rule can be highly accurate.

In the data in the work file 405 of FIG. 4, the value -1 represents a missing value. The numerical attribute itself is not used in the rule generation process, and the symbol values are assigned sequentially from 0, so this value does not cause confusion.

Whether or not the missing value itself is used for the rule generation depends on the missing value handling process in the process 303 of FIG. 3 or the contents specified by the user. If you use it for rule generation, "If x is a missing value,
The rule "..." is allowed. If the missing value is not used for rule generation, then the missing value is simply ignored (ie treated as a symbol that does not match any symbol). Since the missing value is a special value, the fuzzy symbolization rules in the previous section do not apply.

In this embodiment, the progress status indicating how far the rule generation processing has progressed is displayed. In the rule generation process, since the number of rules to be originally examined can be calculated in advance, the figure indicating what percentage of the rules has been examined is displayed at an appropriate timing. Naturally, rules that have been skipped due to pruning are also counted.

In the rule generation processing, it is assumed that a user interrupt is accepted. A button 1401 for accepting an interruption command is prepared in the above-mentioned rule generation progress status display. FIG. 14 shows a display example of the rule generation processing progress status and the suspend button 1401.

When the button 1401 is clicked and the interruption of the processing is designated, the next operation is either "continue" or "end". If you want to continue, continue processing,
If it is finished, the rule generation process is terminated and the rules generated up to the present time are displayed.

FIG. 15 shows a processing flow chart of the rule generation processing described above. Clear candidate rule set and μ
After initialization (1501) such as zero clearing of th, steps 1503 to 1512 are repeated for all candidates. First, the evaluation scale μ is calculated, and μ and μth are compared (1504). If μ> μth, the candidate rule is added to the candidate rule set (1505), and the minimum evaluation scale in the candidate rule set is set to μth (1506). μ ≦ μ
If th, the maximum value μmax of the evaluation scale related to the subnode is calculated (1507), and μmax and μth are compared. μmax
If <μth, all subnodes have been evaluated (15
09).

Next, the progress status of the rule generation processing is displayed (1510) and it is judged whether or not the processing interruption is instructed (1511). If interruption is instructed, the loop is exited (1512). When the loop processing is completed for all the candidates or when the loop is exited in step 1512, the candidate rule set is output (1513) and the processing is completed.

Next, the processing executed by the rule learning processing execution module 403 of FIG. 4 will be described.

FIG. 16 shows the shape parameters of the membership function to be learned. In this embodiment, the parameters adjusted by learning are the division points α1 and α2. That is, the division points α1 and α2 having the largest rule evaluation scale are obtained for the rule generated by the rule generation processing.
However, in the rule generation process using fuzzy partitioning, it is assumed that one numerical value is distributed to at most two symbolic values and the sum of them is always 1.0, so only within the range that satisfies the assumption. It is assumed that the values of α1 and α2 are changed. FIG. 16 shows an example of the membership function shape after learning.

Specific learning is performed in the following steps. Step 1: Specify the rules to be learned. The user specifies which rule to learn. Step 2: Steps 3 to 8 are repeated until the learning end condition is satisfied. Step 3: Change amount of each parameter Δαi = 0 (i = 1,2,
…, M). m is the total number of split point parameters to be adjusted. Step 4: Repeat steps 5 to 7 for each of the specified rules.

Step 5: Evaluation scale μ of the rule being processed
At the same time as calculating, ∂N (A) / ∂αi, ∂N (B) / ∂αi,
Calculate ∂N (A & B) / ∂αi. N (A) is the number of cases (considering fuzzy division) that satisfy the condition part A, and the change amount ΔN (A) of N (A) when αi is changed by a small amount Δαi is
It can be easily obtained from the membership function shape of FIG. The values of ∂N (B) / ∂αi and ∂N (A & B) / ∂αi are calculated in the same way. Step 6: From the definition formula (4) of μ and the value obtained in Step 5, ∂μ / ∂αi = ∂N (A) / ∂αi ・ ∂μ / ∂N (A) + ∂N
(B) / ∂αi ・ ∂μ / ∂N (B) ＋ ∂N (A & B) / ∂αi ・ ∂μ /
Using the relational expression ∂N (A & B), ∂μ / ∂αi (i = 1,…,
m) is calculated.

Step 7: For all i, Δαi ← Δ
Let αi + ∂μ / ∂αi. Step 8: For all i, αi ← αi + λ · Δαi. λ is a predetermined learning coefficient. Step 9: The learned value of αi is output to the rule file 407 in FIG.

The termination conditions in step 2 are, for example, that steps 3 to 8 are repeated a specified number of times, and | Δαi | <ε holds for all i (ε is a predetermined positive constant). ) It is possible to use conditions such as a change in the rule evaluation scale μ | Δμ | <ε, or a combination of the above conditions.

Through the above processing, membership function parameters that maximize the evaluation scale of one or more rules can be automatically obtained by learning. This makes it possible to more accurately extract the regularity / causal relationship that exists between data, and to determine the optimal division point itself of the item value that is optimal for explaining regularity. Has the effect of becoming. In addition, according to this learning, the sum of the evaluation measures of the entire generation rule or a subset thereof can be optimized, so that only the rules that have meaning to the user can be optimized by learning, and the more useful value There is also an effect that it is possible to generate a high-rule.

[0168]

As described above, according to the present invention,
Since the regularity / causal relationship existing between data can be extracted in the form of a rule that is easy for the user to understand, there is an effect that it is possible to effectively use a large amount of data, which was difficult with the conventional method. In particular, there is an effect that it becomes possible to extract effective information from the database in the analysis of the defect generation factor by the analysis of the inspection data in the manufacturing process and the planning of the marketing strategy for the new product by the analysis of the customer database. . Also, create a combination of values between data items as a candidate rule,
The candidate rules are compared with a predetermined evaluation scale, and the candidate rules with high evaluation values are output, so there is no need to perform repeated calculation for rule groups or cluster sets as in the past, and a great deal of processing time is required for large amounts of data. I don't need it. In addition, even when the user does not have special knowledge about the processing method, it is possible to generate a result that accurately represents the characteristics of the data. In addition, since the numerical data contained in the database is converted to symbolic values and used for rules, it is possible to find regularity corresponding to macro structure even for data with large noise, which is highly useful for users. There is an effect that a result can be generated.

[Brief description of drawings]

FIG. 1 is a diagram showing an example of analysis target data in semiconductor failure analysis.

FIG. 2 is a diagram showing an example of analysis target data in customer analysis.

FIG. 3 is a process flow chart of the data analysis method according to the present embodiment.

[Figure 4] Overall configuration of the data analysis system

FIG. 5 is a diagram showing an example of a data table.

FIG. 6 is a diagram showing an example of a relationship definition between a plurality of data tables.

FIG. 7 is a diagram showing a rule generation processing specification item designation screen.

[Figure 8] Automatic item selection process flow chart

FIG. 9 is a diagram showing a symbol value conversion method list display screen.

FIG. 10 is a diagram showing a symbol value conversion method designation screen.

FIG. 11 is a diagram showing an example of the shape of a membership function.

FIG. 12 is a diagram showing a rule generation parameter specification screen.

FIG. 13 is a diagram illustrating a rule search method.

FIG. 14 is a diagram showing a rule generation processing progress display screen.

FIG. 15 is a rule generation processing flow chart.

FIG. 16 is a diagram showing an example of a membership function shape change due to learning.

[Explanation of symbols]

401 ... User interface module, 402 ... Rule generation processing execution module, 403 ... Rule learning processing execution module, 404 ... Analysis target database, 40
5 ... Rule generation work file, 406 ... Rule generation parameter file, 407 ... Rule file.

Front page continued (72) Inventor Yoji Taniguchi 1099 Ozenji, Aso-ku, Kawasaki-shi, Kanagawa Hitachi, Ltd. System Development Laboratory Software Development Headquarters (72) Inventor Yori Takahashi 5030 Totsuka-cho, Totsuka-ku, Yokohama-shi, Kanagawa Hitachi, Ltd. Software Development Headquarters (56) References JP-A-6-68066 (JP, A) Saito / Nakano Written, "Rule Extraction from Cases Containing Noise: RF3 Algorithm", Journal of Information Processing Society of Japan, May 15, 1992, Vol. 33, No. 5, p. 636-644 Mori and Tanaka, “Efficient inference method considering the order of application of rules”, IEICE Research Report, May 26, 1992, Vol. 92, No. 42, p. 31-40 (58) Fields investigated (Int.Cl. ⁷ , DB name) G06N 5/00 G06F 19/00

Claims (2)

(57) [Claims] 1. A data analysis apparatus for generating a rule characterizing a relationship between data items based on a plurality of data items stored in a database, wherein the data analysis device stores data items stored in the database. A combination of a means for receiving the selected data items used in the IF clause and the THEN clause of the IF-THEN rule, and a data item consisting of a pair of an item name and a symbol value from the selected and received data items. IF
Generation means for generating rules candidate -THEN format, before and coverage corresponding to the number of data IF clause is true kill Lumpur candidates, the accuracy of the value corresponding to the number of data both IF clause and THEN clauses apply the counts each, the rule climate
High coverage and high accuracy for each
A data analysis apparatus, comprising: an evaluation unit that obtains an evaluation scale that increases to a higher degree; and an output unit that determines and outputs at least one rule candidate with a high evaluation scale from the rule candidates. 2. A data analysis device for generating a rule characterizing a relationship between the data items based on a plurality of data items stored in a database, wherein the data analysis device stores data items stored in the database. A means for receiving the selected data item used in the IF clause and the THEN clause of the IF-THEN rule, and a conversion means for converting the numeric value into a symbol value if the selected and received data item includes a numeric value And the selected and received data items (not including numbers
And a numerical value is converted to a symbol value by the conversion means.
From the data item has a generation means for generating rules candidate IF-THEN format by combining the data item comprising a pair of an item name and symbols values, the number of data to IF clause is true before kill Lumpur candidate a cover ratio corresponding, by counting precision of the values respectively corresponding to the number of data both IF clause and tHEN clause is true, the rule climate
High coverage and high accuracy for each
A data analysis apparatus, comprising: an evaluation unit that obtains an evaluation scale that increases to a higher degree; and an output unit that determines and outputs at least one rule candidate with a high evaluation scale from the rule candidates.