JP4255779B2 - Data analysis apparatus, data analysis method, and data analysis program - Google Patents

Description

The present invention provides an output attribute (object attribute) to be analyzed, such as characteristics of a product manufactured in a manufacturing process, and an input attribute (description attribute) that is an attribute that affects the output attribute.
For example, the present invention relates to a data analysis apparatus, a data analysis method, and a data analysis program for analyzing a causal relationship with manufacturing process conditions and the like.

A decision tree technique is known as an effective technique for analyzing a causal relationship between an output attribute and an input attribute (see Patent Document 1). In this method, a tree structure is created in which the values of output attributes are well organized in leaf portions that are sequentially cut by the values of input attributes.

FIG. 10 shows a section of the prior art of Patent Document 1 (paragraphs [0002] to [0 of Patent Document 1].
005] and FIG. 22), and the data group of Table 1 is the analysis target. The data group in Table 1 is a set in which twelve pieces of data including a set of four input attribute values x1, x2, x3, and x4 and a value of an output attribute y corresponding to these input attributes are collected. In a decision tree created by this method (hereinafter referred to as “conventional decision tree-1”), as shown in FIG. 10, the values X, Y,
Z is well separated by each value of the input attributes x1, x2, and x3.

However, in the creation of the conventional decision tree-1 in FIG. 10, when data is classified, it is classified into data sets corresponding to the number of values that the input attribute takes (the number of attribute value types). For example, since the input attribute x2 takes four types of values (a, b, c, d), the input attribute x2
Classification into four sets by classification according to. Therefore, if the number of values that the input attribute takes increases, the decision tree may become complicated.

As a solution to this problem, Patent Document 1 proposes a decision tree in which attribute values to be grouped are represented by one label for each attribute, and data is classified by the label.

FIG. 11 shows an example of Patent Document 1 (paragraphs [0010] to [0028 of Patent Document 1).
And FIG. 13). In this embodiment, for example, 4
For the x3 attribute composed of the seed attribute values (1, 2, 3, 4), the x3 attribute value “1” “
“2” is labeled “2.5 or less”, and the x3 attribute values “3” and “4” are “
A hierarchical structure is expressed with a label of “2.5 or more”. FIG. 12 (paragraphs [0010] to [0028] in FIG. 12 and FIG. 12 shows a decision tree created using this label hierarchical structure (hereinafter, this decision tree is referred to as a conventional decision tree-2). 14
The reference decision tree-1 shown in FIG. 10 is much simpler.
JP-A-8-314725 (Publication date: November 29, 1996)

The problems of the prior art will be described using the case where the above-described conventional decision tree generation method is applied to cause analysis of product characteristic defects in the manufacturing process of products such as devices.

Now, the input attributes x1, x2, x3, and x4 in Table 1 are various process data and in-line inspection data in the product manufacturing process, and the output attribute y is the product characteristic data. The output attribute y = Y is the product characteristic. It shall correspond to a defect. Then, the process engineer uses the decision tree-1 (FIG. 10) generated by the technique described in the prior art of Patent Document 1 or the technique described in Patent Document 1 for the product characteristic defect y = Y. It is assumed that the cause of product characteristic failure is investigated using the generated conventional decision tree-2 (FIG. 12).

At this time, in the decision tree-1 generated by the method described in the prior art of Patent Document 1, y = Y to be noticed is distributed at a plurality of locations (four locations in the example of FIG. 10) in the tree shape. Therefore, it is complicated, and it is difficult for a process engineer to determine the cause of a product characteristic failure such as “Which input attribute is in which value range, so that the product characteristic is bad?”. FIG.
In this example, there are only 4 input attributes and only 4 types of attribute values.
Somehow, the process engineer can also determine the cause of the product characteristic failure. However, in manufacturing sites of products such as actual devices (especially semiconductor devices), there are process data and in-line inspection data of about 10 to 100 attributes per process, and the values are multivalued and in a very wide range. Distributed. Furthermore, due to the influence of disturbance (attributes that cannot be detected as input attributes), the values of output attributes often vary even if the values of the input attributes are the same. In such a case, if the method described in the prior art of Patent Document 1 is used, a precise analysis is aimed at, but the data is classified into an infinite number of data sets. It becomes impossible to specify the cause of the characteristic failure.

On the other hand, in the decision tree-2 (FIG. 12) generated by the method disclosed in Patent Document 1, the decision tree is simple because the classification is based on the label hierarchy. Therefore, it is easy for the process engineer to identify the cause of the product characteristic failure where y = Y.

However, in order to create the simple decision tree-2 shown in FIG. 12, it is necessary to previously define the label hierarchical structure shown in FIG. For this reason, the decision tree generation method of Patent Document 1 cannot be applied when there is no idea of the attribute values to be collected. As described above, at the manufacturing site of a product such as an actual device, 10 to 1 per process is required.
There are process data and inline inspection data of about 00 attributes, and the values are multivalued and distributed in a very wide range. Furthermore, due to the influence of disturbance (attributes that cannot be detected as input attributes), the values of output attributes often vary even if the values of the input attributes are the same. Under these circumstances, it is very difficult even for an experienced process engineer to find an attribute value that is collected as one label for each input attribute.

The present invention has been made in view of the above-described conventional problems, and an object of the present invention is to analyze data that can derive a causal relationship between an output attribute and an input attribute in a concise form without predefining a label hierarchical structure. An apparatus, a data analysis method, and a data analysis program are provided.

In order to solve the above problem, the data analysis apparatus according to the present invention stores a plurality of input attributes x _j (1 ≦ j ≦ N, N is the number of input attributes) stored in the analysis target data storage unit , one output is the attribute y with the basic data group DA analyte is a set of data composed in the causal relationship between the input attributes and output attributes a data analyzer you analyze, the basic data group DA Character-numeric data conversion means for generating numeric-type basic data group DA0, which is a set of numeric attribute data, by converting the included character attribute data into numeric attribute data according to a unique conversion rule; type base data group DA0, and numeric numeric output attribute y included in the basic data group DA0, based on a comparison of the magnitude relation between the predetermined threshold value of the output attribute y, the first data group DA1, second data group DA And classifying means for classifying the bets, for one input attribute x _j of the plurality of input attributes for each numerical value can be assumed by the said one input attributes x _j, among the data having the following values the number, the An operation is performed to obtain a first frequency (1-x _j frequency cumulative%), which is a ratio of the number of data belonging to one data group DA1 to the number of all data belonging to the first data group DA1. one of each numerical value can take input attributes x _j, among the data having the following values the numerical, the number of data belonging to the second data group DA2, a ratio to the number of all data belonging to the second data group DA2 It performs operation for obtaining a certain second frequency (2-x _j frequency cumulative%), and, for each numerical value can take of the one input attributes x _j, the first frequency and the difference between the second frequency (x calculation for obtaining the _j frequency cumulative difference%) A first evaluation means for, for each of the plurality of input attributes for a single input attribute x _j of the plurality of input attributes, the numerical values in the first evaluation means can take of the one input attributes x _j based on the calculated difference (x _j frequency cumulative difference%) for each, the numerical value determined maximum difference to be determined as a threshold value x _j-th of the input attributes x _j, of the plurality of input attributes a threshold value determining means for, for each, for a single input attribute x _j of the plurality of input attributes, the threshold x _j-th of the input attributes x _j which is determined by the threshold value determining means, the first frequency ( 1-x _j frequency cumulative%) of the second frequency (2-x _j frequency cumulative%) and the threshold value x _{j- of the} input attribute x _j determined by the threshold value determining means. in _th, (100% - the first frequency ( -X _j Frequency Cumulative%)) with respect to (100% - as well as calculating a second ratio is the ratio of the second frequency _{(2-x j} Frequency Cumulative%)), the first ratio and the second ratio Of the plurality of input attributes is selected for each of the plurality of input attributes, and the ratio selected for each input attribute by the second evaluation means is the largest. input attributes x _j with the _ratio, threshold x _j-th of the input attributes x _j, and the type of outermost even larger ratio indicates which of the first ratio and the second ratio, shows the input attribute conditions It is characterized by including factor extraction means for extracting the input attribute condition in an analysis result data storage unit as well as extracting it as data.

In order to solve the above-described problem, the data analysis method according to the present invention uses the data analysis apparatus described above to store a plurality of input attributes x _j (1 ≦ j ≦ N, N the number of input attributes), and one output attribute y analyzed basic data group DA is a set of data composed in, you analyze the causal relationship between input attributes and output attributes data analysis methods The character-numeric data conversion means converts the character attribute data included in the basic data group DA into numeric attribute data according to a unique conversion rule, thereby obtaining a set of numeric attribute data. character generating a numeric base data group DA0 - and numeric data conversion step, by the classifying means, a numeric base data group DA0, and numerical output attribute y included in the numeric base data group DA0 Based on the comparison of the magnitude relationship between the predetermined threshold value of the output attribute y, the first data group DA1, a classification step of classifying the second data group DA2, by the first evaluation means, the plurality of input attributes for one input attributes x _j of out, for each numerical value can take of the one input attributes x _j, among the data having the following values the numerical, the number of data belonging to the first data group DA1, the first data For each numerical value that can be taken by the one input attribute x _j , the first frequency (1-x _j frequency cumulative%) that is a ratio to the number of all data belonging to the group DA1 is calculated. A second frequency (2-x _j frequency cumulative%) that is a ratio of the number of data belonging to the second data group DA2 to the number of all data belonging to the second data group DA2 among the data having the following numerical values Performance A calculation for calculating a difference between the first frequency and the second frequency (x _j frequency cumulative difference%) for each numerical value that can be taken by the one input attribute x _{j is} the plurality of input attributes. a first evaluation step of performing for each, by the threshold value determining means, for one input attribute x _j of the plurality of input attributes, the numerical values can be assumed by the said one input attribute x _j in the first evaluation means based on the calculated difference (x _j frequency cumulative difference%) for each, the numerical value determined maximum difference to be determined as a threshold value x _j-th of the input attributes x _j, of the plurality of input attributes a threshold determination step of performing for each, by the second evaluation means, for one input attribute x _j of the plurality of input attributes, the input attributes x _j which is determined by the threshold determining unit threshold x _j- the first frequency in _th (1 -X _j frequency cumulative%) to the first ratio that is the ratio of the second frequency (2-x _j frequency cumulative%) and the threshold value x _j-th of the input attribute x _j determined by the threshold value determining means in, (100% - the first frequency _{(1-x j} frequency cumulative%)) with respect to (100% - the second frequency _{(2-x j} frequency cumulative%)) computing a second ratio is the ratio of In addition, the second evaluation step in which the larger one of the first ratio and the second ratio is selected for each of the plurality of input attributes, and the factor extraction means, the second extraction step . of the ratios selected for each input attribute in the evaluation means, the input attributes x _j having the largest _ratio, threshold x _j-th of the input attributes x _j, and outermost even larger ratio first ratio and the the type indicating which of 2 ratio, the input attribute condition Is extracted as to the data, is characterized by including the factor extraction step of storing the input attribute condition analysis result data storage unit.

In order to solve the above problems, a data analysis program according to the present invention includes a plurality of input attributes x _j (1 ≦ j ≦ N, where N is the number of input attributes) stored in the analysis target data storage unit , A data analysis program for causing a computer included in a data analysis apparatus that analyzes a causal relationship between an input attribute and an output attribute to analyze a basic data group DA that is a set of data composed of one output attribute y The data analysis apparatus converts the numerical attribute data included in the basic data group DA into numerical attribute data according to a unique conversion rule, thereby obtaining a numerical basic that is a set of numerical attribute data. character generating data group DA0 - and numeric data conversion means, a numeric base data group DA0, and numerical output attribute y included in the numeric base data group DA0, out Based on the comparison of the magnitude relationship between the predetermined threshold value attribute y, the first data group DA1, a classification means for classifying the second data group DA2, for one input attribute x _j of the plurality of input attributes , each numerical value can be assumed by the said one input attributes x _j, among the data having the following values the numerical, the number of data belonging to the first data group DA1, the number of all data belonging to the first data group DA1 Among the data having a numerical value equal to or lower than the numerical value for each numerical value that can be taken by the one input attribute x _j , and calculating the first frequency (1-x _j frequency cumulative%) that is a ratio to Performing an operation for obtaining a second frequency (2-x _j frequency cumulative%) that is a ratio of the number of data belonging to the second data group DA2 to the number of all data belonging to the second data group DA2. One input attribute x first evaluation means for performing an operation for obtaining a difference between the first frequency and the second frequency (x _j frequency cumulative difference%) for each of the plurality of input attributes for each numerical value that _j can take ; for one input attributes x _j of the plurality of input attributes, based on the single input attribute x _j of possible numerical each of the calculated difference by the first evaluation means (x _j frequency cumulative difference%), maximum to be determined as a threshold value x _j-th of the input attributes x _j a numerical value difference is determined, and the threshold value determining means for, for each of the plurality of input attributes, the one of the plurality of input attributes for input attribute _{x j,} at the threshold _{x j-th} of the input attributes _{x j} which is determined by the threshold value determining means, a second frequency for the first frequency _{(1-x j} frequency cumulative%) _{(2-x j} The first ratio that is the ratio of the frequency cumulative%) and threshold decision At the threshold _{x j-th} of the determined at means input attribute _{x j,} (100% - the first frequency _{(1-x j} Frequency Cumulative%)) with respect to (100% - the second frequency (2-x _j frequency cumulative%)) is calculated for each of the plurality of input attributes, and the larger one of the first ratio and the second ratio is selected. a second evaluation means for, among the ratios selected for each input attribute in the second evaluation means, the threshold value x _j-th input attributes x _j, the input attributes x _j having the largest _ratio, and Extraction of the factor indicating whether the largest ratio is the first ratio or the second ratio as data indicating the input attribute condition, and extracting the input attribute condition in the analysis result data storage unit Means including the computer It is a data analysis program for functioning as each means .

In order to solve the above problems, a computer-readable recording medium according to the present invention records the above data analysis program.

According to the apparatus, method, program, or recording medium, the factor of the output attribute condition (result) corresponding to the second data group can be extracted in a concise form without defining the label hierarchical structure in advance. Therefore, for example, if the second data group is a data group corresponding to a bad result (for example, occurrence of a defective product), the user can easily grasp the cause of the bad result. Conversely, if the second data group is a data group corresponding to a good result (for example, occurrence of a product having excellent characteristics), the user can easily grasp the factor of the good result.

In the data analysis method according to the present invention, based on the input attribute condition extracted by the factor extracting means, the numerical basic data group DA0 is classified into a factor data group that satisfies the input attribute condition and the input attribute condition that does not satisfy the input attribute condition. Further comprising a dividing unit that divides the data group into at least one of the classified data groups as a new numerical basic data group DA0 to the classifying unit, processing by the classifying unit, processing by the first evaluation unit, It is more preferable that a series of processing including processing by the threshold value determination unit, processing by the second evaluation unit, processing by the factor extraction unit, and processing by the dividing unit is repeatedly executed.

According to the above configuration, a tree structure can be created with a plurality of factors as nodes. therefore,
Even if the analysis target is intertwined with multiple factors that are difficult to express with a single association rule,
The factor can be investigated with sufficiently high accuracy.

The data analysis apparatus according to the present invention further includes an end condition determination unit that determines whether or not an end condition is satisfied. When the end condition determination unit determines that the end condition is satisfied, the series of processes is executed. It is more preferable to end the process. Thereby, it is possible to avoid performing unnecessary processing more than necessary.

The first evaluation means calculates, for all the numerical values of each input attribute, the ratio of data whose input attribute in the first data group is equal to or lower than the numerical value as the first frequency, and in the second data group Frequency calculation means for calculating the ratio of data whose input attribute is less than or equal to the numerical value as the second frequency, and for all the numerical values of each input attribute,
More preferably, difference calculation means for calculating a difference between the first frequency and the second frequency is included. Thereby, the threshold evaluation index can be easily calculated.

The second evaluation means uses, as the first rule evaluation value, the ratio of data whose input attribute in the second data group is equal to or smaller than the threshold to the ratio of data whose input attribute in the first data group is equal to or smaller than the threshold. As the first ratio,
As the second rule evaluation value, the ratio of the ratio of the data whose input attribute exceeds the threshold in the second data group to the ratio of the data whose input attribute exceeds the threshold in the first data group is calculated as the second ratio. , The larger ratio of both ratios is extracted, and the factor extracting means has the maximum value among the ratios of the input attributes extracted by the second evaluating means. More preferably, the input attribute, the threshold value of the input attribute, and the type of the extracted ratio are extracted as data indicating the input attribute condition. Thereby, the first and second rule evaluation values can be easily calculated.

According to the apparatus, method, program, and recording medium of the present invention, as described above, without defining the label hierarchical structure in advance, “input attribute is below threshold” or “input attribute exceeds threshold” is very simple. In this way, it is possible to derive a factor that causes a specific output attribute condition (problem event) that is a problem event. In addition, if multiple factors are derived, the problem event can be expressed as a very simple decision tree based on a combination of conditions such as “input attribute is below threshold” or “input attribute exceeds threshold” for each factor (input attribute). Causal relationships related to can be derived.

  One embodiment of the present invention will be described below.

  First, the data analysis apparatus of this embodiment is demonstrated based on FIG.

As shown in FIG. 1, the data analysis apparatus includes a character-numerical data conversion unit 1, an analysis target data storage unit 2, a threshold setting unit (threshold setting unit) 3, and a data classification unit (classification unit) 4.
, Data string extraction unit 5, frequency calculation unit (first evaluation unit, frequency calculation unit) 6, frequency cumulative difference calculation unit (first evaluation unit, difference calculation unit) 7, input attribute threshold determination unit (threshold determination unit) 8) Frequency cumulative ratio calculation unit (second evaluation unit) 16, factor extraction unit (factor extraction unit) 9, factor undiscovered data extraction unit (division unit) 10, end condition determination unit (end condition determination unit) 11 , An input attribute threshold value table creation unit 12, a contribution rate calculation unit 13, an analysis result data storage unit 14, and an output unit 15.

Next, the data analysis method of the present embodiment will be described with reference to FIG. 2, taking as an example the case where the data group DA in the following Table 1 is an analysis target. The data group DA in Table 1 is stored in the storage unit 2 such as a hard disk.

The data group DA in Table 1 is composed of 12 pieces of data having ids (identifiers) of 1 to 12. In Table 1, x1, x2, x3, and x4 are input attributes. The input attribute x1 is a character attribute that takes one of four characters A, B, C, and D. The input attribute x2 is a character attribute that takes one of the four characters a, b, c, and d. The input attribute x3 is a discrete attribute that takes one of four discrete values 1, 2, 3, and 4. The input attribute x4 is a discrete attribute taking any one of four discrete values 10, 20, 30, and 40. The input attribute may be a continuous attribute that takes a continuous numerical value.

In Table 1, y is an output attribute. The output attribute may be a character attribute, a discrete attribute, or a continuous attribute. Here, three characters X,
It is a character attribute that takes either Y or Z.

In the data analysis method of the present embodiment, the case where y = Y is regarded as a problem event, and the cause of the output attribute y being Y is analyzed.

As an example of the analysis target data, for example, the input attribute is the manufacturing process condition and / or in-line inspection result (inspection result during the manufacturing line) in the product manufacturing process, the output attribute is the product quality determination result, y Data in which the problem event = Y is a bad quality determination result can be cited. In this case, the causal relationship between the input attribute and the output attribute is analyzed by the data analysis method of the present embodiment, and the cause of the problem event y = Y is derived, thereby eliminating the occurrence of defective products such as device characteristic defects. Can be easily achieved. Therefore, it is possible to easily improve the manufacturing process such as improvement in yield.

As a more specific example of the analysis target data, for example, input attributes x1, x2, x
3 and x4 are process data such as gas flow rate, gas pressure, input power, and film formation time of the plasma CVD process, and the output attribute y is data such as the film thickness of the thin film formed by the plasma CVD process. Can be mentioned. The values of the input attribute and output attribute may be continuous attributes, discrete attributes, or character attributes. In the case of a character attribute, for example, the output attribute is an example of film thickness, and is expressed as “large”, “medium”, or “small”.
[Step 0]
First, the character-numeric data conversion unit 1 converts the character attributes in the data group DA of Table 1 stored in the analysis target data storage unit 2 such as a hard disk into numeric attributes (numeric data) according to the following conversion rule (S0). ). Thereby, each data is converted into numerical data. Then, the character-numeric data conversion unit 1 sends the converted data group to the data classification unit 4.
(X1) A → 1, B → 2, C → 3, D → 4
(X2) a → 1, b → 2, c → 3, d → 4
(X3) No conversion (x4) No conversion (y) X → 1, Y → 2, Z → 3
It is preferable that the conversion rule is set so that the numerical value of the output attribute increases as the numerical value of the input attribute after conversion increases, or vice versa. The conversion rule is not limited to the above example as long as it is unique.

The data group DA0 converted into numerical data by the conversion rule is as shown in Table 2.

By this conversion, the obtained data group DA0 has a plurality of input attributes (discrete values).
This is a set of data composed of description attributes) and output attributes (purpose attributes). Less than,
The data group DA0 is called a basic data group.
[Step 1]
The threshold value setting unit 3 responds to predetermined setting information or in response to the user inputting the problem event attribute value y = Y from an input unit such as a keyboard or a mouse (not shown). = Basic data group DA0 corresponding to problem event Y
The output attribute y threshold (output attribute threshold) y _th is set and output to the data classification unit 4 (S1). In this example, the threshold of the output attribute y of the basic data group DA0 corresponding to the problem event y = Y of the data group DA is y _th = 2.
[Step 2]
Next, the data classification unit 4 determines the basic data based on the comparison logic (1) (2) between the value of the output attribute y of the basic data group DA0 and the output attribute threshold y _th output from the threshold setting unit 3. The group DA0 is divided into two groups (classification) into a first data group DA1 and a second data group DA2 (S2).

(1) y> y _th or y <y _th → DA1
(2) y = y _th → DA2
In other words, the data classification unit 4 includes the basic data group DA0, the first data group DA1 whose output attribute does not match the output attribute threshold y _th (that is, 1 or 3), and the output attribute that is the output attribute threshold y _th ( = 2) and the second data group DA2 that coincides with the second data group DA2. The second data group DA2 is a data group of problem events (for example, defective device characteristics). That is, in the second data group DA2, the output attribute y is an attribute value (
2), and the first data group DA1 is a data group in which the output attribute y is an attribute value (1 or 3) that does not represent a problem event.

  Table 3 shows the first data group DA1 and Table 4 shows the second data group DA2.

In the following, the first data group DA1 is appropriately referred to as a non-defective product (OK product) data group, and the second
The data group DA2 is referred to as a defective product (NG product) data group.
[Step 3]
Next, the data string extraction unit 5 extracts the input attribute xj from the good product data group DA1 (Table 3).
Each data string of (1 ≦ j ≦ 4) is extracted (S3). This data string is 1-xj
This is called a data group.

Similarly, the data string extraction unit 5 extracts each data string of the input attribute xj (1 ≦ j ≦ 4) from the defective product data group DA2 (Table 4) (S3). This data string is 2
It will be called a -xj data group.

  The 1-xj data group is shown in Tables 5-8, and the 2-xj data group is shown in Tables 9-12.

[Step 4]
The frequency calculation unit 6 uses each of the 1-xj data groups extracted from the non-defective product data group DA1 in step 3 and 2-x extracted from the defective product data group DA2 in step 3.
Each of the j data groups is rearranged in ascending order by the value of the input attribute xj. Then, for each numerical value of the input attribute xj, 1-xj frequency cumulative% representing the ratio of the number of data less than that value in the first data group and 2 representing the ratio of the number of data less than that value in the second data group. -Xj Frequency cumulative% is calculated (S4).

Here, using Tables 13 to 16 in which Tables 5 to 8 are rearranged in ascending order by the value of the input attribute xj, the number of data at the position equal to or higher than the position of the data in the table for each row (id) data. The ratio of one data group to the total number of data (= 8) is calculated as 1-xj frequency cumulative%. Similarly, using Tables 17 to 20 in which Tables 9 to 12 are rearranged in ascending order by the value of the input attribute xj, the number of data at the position equal to or higher than the position of the data in the table for each row (id) is calculated. The ratio of 2 data groups to the total number of data (= 4) is calculated as 2-xj frequency cumulative%. The values of 1-xj frequency cumulative% and 2-xj frequency cumulative% calculated here are shown in Table 13 to
20 shows.

In Steps 3 and 4 described above, after extracting and rearranging the data strings, 1-xj frequency cumulative% and 2-xj frequency cumulative% are calculated. The 1-xj frequency accumulation% and the 2-xj frequency accumulation% may be directly calculated without performing the above.

Further, the frequency calculation unit 6 is a table of 1-xj data groups that are non-defective product data groups for which 1-xj frequency cumulative% is calculated, and 2-items that are defective product data groups for which 2-xj frequency cumulative% is calculated. The table of the xj data group is combined. Specifically, the input attribute x
For Table 1, Table 13 and Table 17 are combined to obtain the x1 frequency accumulation table in Table 21, and for the input attribute x2, Table 14 and Table 18 are combined to create the x2 frequency accumulation table in Table 22, and the input attribute x3. Table 15 and Table 19 are combined to obtain the x3 frequency accumulation table of Table 23, and for the input attribute x4, Table 16 and Table 20 are combined to generate x 24 of Table 24.
A 4-frequency accumulation table is created for each.

Furthermore, the frequency calculation unit 6 sorts the frequency accumulation tables in Tables 21 to 24 in ascending order by the value of the input attribute xj. At this time, 1-xj frequency accumulation% and 2-xj
The value immediately before is substituted into the blank for frequency accumulation%. When the same value continues in the input attribute xj, only the rearranged final data is adopted. Thus, in the frequency calculation unit 6, for each value of the input attribute xj, 1-xj frequency cumulative% (A; first) representing the ratio of the number of data less than or equal to that value in the first data group that is a non-defective data group. And 2-xj frequency cumulative% (B; second frequency) representing the ratio of the number of data less than or equal to that value in the second data group, which is a defective product data group, is calculated (S4). .
[Step 5]
Next, the frequency cumulative difference calculation unit 7 performs a non-defective 1-xj for each value of the input attribute xj.
A difference (= | A−B |) between the frequency accumulation (A) and the 2-xj frequency accumulation (B) of the defective product is calculated (S5). This difference value is referred to as an xj frequency cumulative difference (= | A−B |). x
Tables 25 to 28 show the calculation results of the j-frequency cumulative difference.

The relationship between the input attribute xj, 1-xj frequency accumulation (A) for non-defective products, 2-xj frequency accumulation (B) for defective products, and xj frequency accumulation difference | A-B | is shown in FIGS.

The xj frequency cumulative difference | A−B | for each numerical value is determined by dividing the input attribute xj into a first non-defective product by dividing it into a range where the input attribute xj is less than the numerical value and a range where the input attribute xj exceeds the numerical value.
This is an index indicating whether the data group DA1 and the defective second data group DA2 are well separated. In other words, the xj frequency cumulative difference | A−B | is a threshold evaluation index that represents the degree to which data whose input attribute is equal to or less than the numerical value is biased to one of the first data group and the second data group.

Here, the xj frequency cumulative difference | A−B | is calculated as the threshold evaluation index, but as a threshold evaluation index for each numerical value, an index for evaluating the degree of data bias, for example, information gain (gain) Information gain ratio, Gini index, mean square error, etc. may be used.
[Step 6]
The input attribute threshold value determination unit 8 for each input attribute xj, among the individual values of xj,
The value of the input attribute xj when the value of the xj frequency cumulative difference | A−B | is maximized is extracted (S6). This value is called an input attribute threshold value xj-th.

As can be seen with reference to FIGS. 3 to 6, the input attribute threshold value xj−th is xj ≦ xj.
The value of the input attribute xj that makes it easy to distinguish between the non-defective first data group DA1 and the defective second data group DA2 by dividing into the range of -th and the range of xj> xj-th. Is shown.

Here, the processes of the third step to the sixth step are collectively performed for a plurality of input attributes, but the value of j is sequentially increased from 1 to N to increase the third step to
The process of the sixth step may be repeated.
[Step 7]
Next, the frequency cumulative ratio calculation unit 16 determines that 1−x of the non-defective product when xj = xj−th.
The ratio of the 2-xj frequency accumulation (B) of defective products to the j frequency accumulation (A) is calculated. This ratio is called a 2-xjth lower ratio (= B / A). Also, 100
10 to the value obtained by subtracting 1-xj frequency accumulation (A) of non-defective products from (= 100-A)
A ratio of a value (= 100−B) obtained by subtracting 2-xj frequency accumulation (B) of defective products from 0 is calculated. This ratio is expressed as 2-xjth upper ratio (= (100-B) / (100-A)
). And 2-xj representing the larger value of the ratio of both
Extract th ratio.

Here, the 2-xjth lower ratio represents a ratio at which the defective second data group can be detected separately from the first non-defective data group based on the input attribute condition “xj ≦ xj−th”. The 2-xjth upper ratio represents a ratio at which the defective second data group can be detected separately from the first non-defective data group based on the input attribute condition “xj> xj-th”.

In other words, the 2-xjth lower ratio is “input attribute xj is input attribute threshold xj−
It represents an evaluation value (first rule evaluation value) representing the probability of the first association rule that “the data is included in the second data group if it is equal to or less than th”. In addition, 2-
The ratio on xjth is an evaluation value (second rule) indicating the probability of the second correlation rule that “if the input attribute xj exceeds the input attribute threshold value xj−th, it is data included in the second data group”. Evaluation value).

Input attribute threshold value xj-th extracted for each input attribute xj, xj = xj-t
1-xj frequency accumulation of good products (A), 2-xj frequency accumulation of defective products (B) in h
, Xj frequency cumulative difference | AB |, 2-xjth lower ratio B / A, 2-xjth upper ratio (100-B) / (100-A), and 2-xjth ratio.

[Step 8]
The factor extraction unit 9 selects 2-xjt in step 7 from the input attributes x1 to x4.
The input attribute that maximizes the h ratio is extracted. As a result, the input attribute that maximizes the 2-xjth ratio, the threshold value, and the type of the employed ratio (upper and lower) are data indicating the cause of the output attribute condition (input attribute condition) corresponding to the second data group. Will be extracted. This is the highest 2 of the association rules for all input attributes.
This corresponds to extracting data indicating an input attribute condition of an association rule having a -xjth lower ratio or a 2-xjth upper ratio.

Here, the 2-xjth ratio is calculated as an index for extracting the input attribute of the correlation rule having the maximum rule evaluation value, but the input attribute of the correlation rule having the maximum rule evaluation value is extracted. As an indicator for other evaluation indicators, for example,
Support rate (support), certainty factor (confidence), information gain (gain), information gain ratio, Gini index, mean square error and the like may be used.

Referring to Table 29, when input attribute x2 = x2-th = 2, 2-x2th ratio = 2-x2th upper ratio = ∞. This is completely separated from the non-defective first data group DA1 under the input attribute condition “x2> 2,” and the defective second data group DA2 is separated.
It can be easily understood with reference to FIG.

The extracted input attribute (= x2), the input attribute threshold value representing the value of the input attribute (
= 2), and the data of the adopted ratio type (= top) are analyzed result data storage unit 1
Save to 4.

As described above, the input attribute condition “x2> 2” is extracted as one factor of the problem event (the second data group DA2 of defective products).
[Step 9]
In step 8, the input attribute condition “x2> 2” is extracted as one factor of the problem event (second data group DA2 of defective products). Next, another factor is investigated. For this reason, the factor undiscovered data extraction unit 10 performs basic data group DA0 (Table 2).
And a data group (factor data group) satisfying the input attribute condition “x2> 2” and a data group (other data group) in which the cause of the problem event has not yet been found in the basic data group DA0 (Table 2), that is, Satisfy the input attribute condition “x2 ≦ 2” (input attribute condition “x2>
The data group in which the cause of the problem event has not yet been found is extracted (Table 30).

The factor undiscovered data extraction unit 10 sends the extracted data group to the data classification unit 4 as the next (new) basic data group DA0.
[Step 10]
Then, the data group extracted in step 9 is set as the next basic data group DA0.
Until the end condition determination unit 11 determines that the end condition is satisfied, the processes of step 2 to step 9 are repeated. The end condition determination unit 11 of the present embodiment
When the number of data in the second data group DA2 of the defective product becomes 0 in the above step 2 during the repeated processing, it is determined as the end condition. As described above, detailed factor analysis results can be obtained by repeatedly performing the process until the number of data in the second data group DA2 of defective products becomes zero.

The end condition is another end condition based on the number of data in the second data group DA2.
For example, (1) when the number of data in the second data group DA2 is equal to or less than a predetermined number in step 2 during the iterative process, (2) the number of data in the first data group DA1 in step 2 during the iterative process. When the ratio of the number of data in the second data group DA2 is equal to or less than a predetermined ratio, (3) When the rule evaluation value of the input attribute condition extracted in step 8 during the iterative processing falls below a predetermined value, etc. Also good. When such termination conditions are used, a simpler and sufficient factor analysis result can be obtained. Further, when priority is given to obtaining a concise factor analysis result, the end condition is simply a case where the iterative process is performed a predetermined number of times, or the end condition determining unit 11 is omitted and the iterative process is performed as much as possible. May be.

In this example, defective data (second data group DA2; y = 2) is not included in the data group of x1 ≦ 2 that has not been found and extracted in step 9 during the second iteration. For this reason, the iterative process was completed at the second time (at the time when the second factor extraction was performed).
[Step 11]
The input attribute threshold value table creation unit 12 creates an input attribute threshold value table that stores the input attribute xj extracted for each repetition process of step 10, the input attribute threshold value xj-th, and the type of ratio adopted ( Table 31).

The input attribute threshold value table creating unit 12 converts the numerical value of the input attribute threshold value xj-th in the input attribute threshold value table into character data as necessary. The conversion rule for character data is a rule that is the reverse conversion of the conversion in step 0, and is as follows.
(X1) 1 → A, 2 → B, 3 → C, 4 → D
(X2) 1 → a, 2 → b, 3 → c, 4 → d
(X3) Not converted (x4) Not converted Table 32 shows an input attribute threshold value table in which the input attribute threshold value xj-th in the input attribute threshold value table of Table 31 is converted into character data.

This input attribute threshold value table is the conventional decision tree-2 described in Patent Document 1 (FIG. 12).
) Corresponds to the classification condition of the decision tree when focusing on the output attribute y = Y (y = 2).
[Step 12]
Next, the contribution rate calculation unit 13 extracts the problem event (y = 2: defective product data group, original second data group DA of the input attribute extracted from the input attribute threshold value table of Table 31.
2) The contribution ratio (corresponding to the support that is an evaluation index of the association rule) is obtained.

Table 33 shows the input attribute condition “x2> 2” extracted as the first factor as the cause in the original second data group DA2 (Table 4) which is a problem event (defective product), or extracted the second time. The data corresponding to the input attribute condition “x1> 2”
".

From Table 33, the input attribute condition “x” for the problem event (original second data group DA2).
The contribution ratio of “1> 2” and “x2> 2” is obtained as shown in Table 34.

In Table 34, the contribution ratio shown at the intersection of “x1> 2” and “x1> 2” and the contribution ratio shown at the intersection of “x2> 2” and “x2> 2” are “x1> 2 "
The contribution ratio of the single factor and the contribution ratio of “x2> 2” are shown. In addition, the contribution rate indicated at the intersection of “x1> 2” and “x2> 2” is “x1”.
It represents the contribution ratio of the composite factor of the “> 2” factor and the “x2> 2” factor. Table 3
4 can also be expressed as shown in FIG.

From Table 34 or FIG. 7, priority (rank 1: x1) is assigned to the problem event (y = 2).
, Ranking 2: x2), and measures can be taken.
[Step 13]
The data analysis is thus completed, and the input attribute threshold value table created by the input attribute threshold value table creating unit 12 and the contribution rate data are stored as analysis result data in the analysis result data storage unit 14 such as a hard disk. This analysis result data is
It is sent from the analysis result data storage unit 14 to the output unit 15 such as a display device or a printing device, and can be displayed as a decision tree or table on the display device, or printed as a decision tree or table on the printing device. .

According to the present embodiment, as in the conventional decision tree-2 (FIG. 12) described in Patent Document 1, the label hierarchy structure (FIG. 11) is not defined in advance, but the table 32 (or table 31) can be used.
The cause of the problem event can be derived in a very simple form as shown in the input attribute threshold value table. Then, using this, the contribution rate of each factor (input attribute) to the problem phenomenon can be obtained.

Here, when the input attribute threshold value table of this embodiment shown in Table 32 (or Table 31) is expressed in the form of a decision tree, it is expressed as shown in FIG. Also, when the contribution rate of each factor to the problem event y = Y (= 2) is expressed in the same format as FIG. 7 using the conventional decision tree-2 (FIG. 12), it is as shown in FIG.

When the decision tree derived from the present embodiment (FIG. 8) is compared with the conventional decision tree-2 (FIG. 12), the contribution of the input attribute x3 is not expressed in the present embodiment.
As can be seen by comparing FIG. 7 and FIG. 9, the problem event y = Y (y = 2)
This is because it does not occur due to each single factor of the input attributes x1 and x3, and in step 9 during the second repetitive operation, step 10 was not executed for the data group of x1> 2. Due to

When pursuing factors in detail, it is also necessary to extract the contribution of the input attribute x3. However, for the purpose of removing (improving) the problem event y = Y (y = 2), only the input attribute x1 is required. Even this extraction can sufficiently achieve this purpose. In the present embodiment, paying attention to this point, the main factor that should be taken against the problem phenomenon is extracted, so the input attribute x3 is not extracted. If detailed analysis is required, 2 in step 9 above.
Step 10 may be executed for both of the differentiated data groups.

In the above-described embodiment, a plurality of factors are derived and a decision tree is generated. However, if only one factor is desired to be extracted, the process may end in step 8.

The data analysis method described above can be realized by the computer executing a data analysis program including processes corresponding to S0 to S12 (steps 0 to 13) in FIG. Therefore, in the data analysis apparatus of FIG. 1, the data analysis program converts the computer into a character-numeric data conversion unit 1, an analysis target data storage unit 2,
Threshold setting unit 3, data classification unit 4, data string extraction unit 5, frequency calculation unit 6, frequency cumulative difference calculation unit 7, input attribute threshold value determination unit 8, frequency cumulative ratio calculation unit 16, factor extraction unit 9, factor undiscovered This can be realized by functioning as the data extraction unit 10, end condition determination unit 11, input attribute threshold value table creation unit 12, and contribution rate calculation unit 13.

The program can be provided to the user by storing it in a computer-readable recording medium. The recording medium may be a built-in medium built in the computer main body, or a removable medium configured to be separable from the computer main body. Examples of the built-in medium include ROM; rewritable nonvolatile memory such as flash memory; and hard disk. The removable media includes optical recording media such as CD-ROM and DVD; magneto-optical recording media such as MO; magnetic recording media such as floppy (registered trademark) disks, cassette tapes and removable hard disks; memory cards and the like. And a medium having a built-in rewritable nonvolatile memory such as a medium having a built-in ROM such as a ROM cassette.

The program may be configured to be executed by CPU access, or after reading the program stored in the recording medium and transferring the read program to the program storage area of the built-in medium, the program on the built-in medium May be executed by CPU access. In addition, the program is not limited to be sold in a state where it is stored in a computer-readable recording medium, and is sold in a format that is transferred to a user's computer via a communication network such as the Internet. It may be a thing.

In this embodiment, the data classification unit 4 classifies the output attribute by comparing the output attribute with the output attribute threshold value. However, when the output attribute is a character attribute, the character-numeric data conversion unit 1 sets the output attribute to a numerical value. Instead of converting into attributes, the data classification unit 4 may perform classification by comparing the output attributes with the output attributes (characters; Y) to be analyzed.

As described above, the data analysis method according to the present embodiment includes N columns of input attribute data including N attributes (N is an integer of 2 or more) and one column of output attributes including one attribute. A data analysis method for analyzing a causal relationship between the output attribute and the input attribute, and a first step of setting an output attribute threshold; A second step of dividing the basic data group into a first data group and a second data group based on a comparison between the value and the output attribute threshold;
From each of the data group and the second data group, a 1-J data string and a 2-J data string representing a data string of the Jth input attribute (J is an integer having a relationship of 1 ≦ J ≦ N),
A third step of extracting each of the values, and for each value of the J-th input attribute of the 1-J data string, a 1-J frequency accumulation (%) representing a ratio of the number of data less than or equal to the value is calculated; A fourth step of calculating, for each value of the J-th input attribute of the 2-J data string, a 2-J frequency accumulation (%) representing a ratio of the number of data less than that value;
For each individual value of the Jth input attribute including both the data string and the 2-J data string, the 1-J frequency accumulation (%) and the 2-J frequency accumulation (%) A fifth step of calculating the J-th frequency cumulative difference, which represents the absolute value of the difference, and extracting the value of the J-th input attribute when the value of the J-th frequency cumulative difference is maximum as the J-th input attribute threshold 6-step, and when the J-th input attribute is the J-th input attribute threshold, a 2-J lower ratio representing a ratio of the 2-J frequency accumulation (%) to the 1-J frequency accumulation (%), and 1
Calculate the 2-J upper ratio, which represents the ratio of the value obtained by subtracting the 2-J frequency accumulation (%) from 100 to the value obtained by subtracting the 1-J frequency accumulation (%) from 00. The seventh step of extracting the 2-J ratio, which indicates the larger value of J, and the value of J from 1 to N
The operation of the third step to the seventh step is repeated, and the 2nd of the first to Nth input attributes extracted in the seventh step during the repetitive operation is repeated.
-The J attribute is extracted in the eighth step of extracting and storing the input attribute having the maximum value, the input attribute threshold representing the value of the input attribute, and the type of the adopted ratio, and the eighth step. Based on the input attribute, at least one of the ninth step of bisecting the basic data group and the data group bifurcated in the ninth step is set as a new basic data group, and a predetermined end condition is set. A tenth step that repeats the operations of the second step to the ninth step until it is satisfied.

According to the above method, it is possible to derive a plurality of factors of problem events in a very simple form without defining the label hierarchical structure in advance. Then, by using this, it is possible to create a decision tree representing a causal relationship, and obtain the contribution rate of each factor (input attribute) to the problem event (output attribute).
In order to solve the above-described problem, the data analysis apparatus according to the present invention analyzes a basic data group that is a set of data including a plurality of input attributes and output attributes. A data analysis device that analyzes a causal relationship with an attribute and extracts information indicating the causal relationship, and classifies a basic data group into a first data group and a second data group according to an output attribute; First evaluation means for calculating a threshold evaluation index representing the degree to which data whose input attribute is equal to or less than the numerical value of each input attribute is biased to one of the first data group and the second data group And a threshold value determining means for determining a numerical value having the maximum threshold value evaluation index for each input attribute as a threshold value of each input attribute based on the threshold value evaluation index calculated by the first evaluation means, and the threshold value determining means. Each Based on the threshold value of the force attribute, a first rule evaluation value indicating the probability of the association rule “if the input attribute is equal to or less than the threshold value, data included in the second data group”, A second evaluation means for calculating, for each input attribute, a second rule evaluation value representing the probability of the correlation rule that the data is included in the second data group if it exceeds, and correlation for all input attributes You may make it include the factor extraction means which extracts the data which show the input attribute condition of the correlation rule with the highest rule evaluation value among rules as information which shows the factor of the output attribute condition corresponding to a 2nd data group.
Further, in order to solve the above-described problem, the data analysis method according to the present invention uses the data analysis apparatus described above, and a basic data group that is a set of data composed of a plurality of input attributes and output attributes. Is a data analysis method for analyzing the causal relationship between the input attribute and the output attribute, and extracting information indicating the causal relationship, wherein the basic data group is output from the first data according to the output attribute by the classification means. A classification step of classifying the input attribute into a group and a second data group, and the first evaluation means, for all the numerical values of each input attribute, the data whose input attribute is equal to or less than the numerical value is the first data group and the second data group A first evaluation step for calculating a threshold evaluation index representing a degree of bias to one of the input values, and the threshold determination means based on the threshold evaluation index calculated in the first evaluation step. A threshold value determining step for determining a numerical value having the maximum threshold evaluation index for each input attribute as a threshold value for each input attribute, and “input attribute” based on the threshold value for each input attribute determined in the threshold value determining step by the second evaluation unit. The first rule evaluation value indicating the probability of the association rule that the data is included in the second data group if is less than or equal to the threshold value, and “if the input attribute exceeds the threshold value, it is included in the second data group. The second evaluation step for calculating the second rule evaluation value representing the certainty of the correlation rule that is “data” for each input attribute, and the above-described factor extraction means, among the correlation rules for all input attributes. A factor extraction process for extracting data indicating an input attribute condition of an association rule having a high rule evaluation value as information indicating a factor of an output attribute condition corresponding to the second data group It may be included and-up.
Further, in order to solve the above problems, the data analysis program according to the present invention includes a classification unit for classifying a basic data group into a first data group and a second data group according to output attributes, and each input. A first evaluation means for calculating a threshold evaluation index representing a degree that data having an input attribute equal to or less than the numerical value is biased to one of the first data group and the second data group for all the numerical values of the attribute; Based on the threshold evaluation index calculated by one evaluation means, each numerical value having the maximum threshold evaluation index for each input attribute is
Based on the threshold value determining means for determining the threshold value of the input attribute and the threshold value of each input attribute determined by the threshold value determining means, the correlation rule “if the input attribute is equal to or less than the threshold value, the data is included in the second data group” A first rule evaluation value that represents the certainty of the association rule, and a second rule evaluation value that represents the certainty of the association rule “if the input attribute exceeds the threshold, the data is included in the second data group”. The second evaluation means for calculating each input attribute, and the data indicating the input attribute condition of the correlation rule having the highest rule evaluation value among the correlation rules for all input attributes, the output attribute corresponding to the second data group It may be a data analysis program for functioning as a factor extracting means for extracting as information indicating the factor of the condition.
Further, the data analysis apparatus according to the present invention provides a basic data group based on the input attribute condition extracted by the factor extracting means, a factor data group that satisfies the input attribute condition, and other data that does not satisfy the input attribute condition. A dividing unit that divides the data into groups and sends at least one of the classified data groups to the classification unit as a new basic data group, and includes processing by the classification unit, processing by the first evaluation unit, and threshold determination unit A series of processes including the process, the process by the second evaluation unit, the process by the factor extraction unit, and the process by the division unit may be repeatedly executed.

It is a block diagram which shows the structure of the data analyzer which concerns on one Embodiment of this invention. It is a flowchart which shows the data analysis method which concerns on one Embodiment of this invention. An example of the output of the frequency accumulation difference calculation unit 7 (step 5) in the data analysis apparatus according to the embodiment of the present invention is represented by a graph, with an input attribute x1 and non-defective 1-x1 frequency accumulation (A), The relationship between 2-x1 frequency accumulation (B) of defective products and x1 frequency accumulation difference | AB | is shown. An example of the output of the frequency accumulation difference calculation unit 7 (step 5) in the data analysis apparatus according to the embodiment of the present invention is represented by a graph, with an input attribute x2 and a non-defective 1-x2 frequency accumulation (A), The relationship between 2-x2 frequency accumulation (B) of defective products and x2 frequency accumulation difference | AB | is shown. FIG. 7 is a graph showing an example of the output of the frequency cumulative difference calculation unit 7 (step 5) in the data analysis apparatus according to the embodiment of the present invention. The input attribute x3 and the non-defective 1-x3 frequency cumulative (A), The relationship between 2-x3 frequency accumulation (B) of defective products and x3 frequency accumulation difference | AB | is shown. An example of the output of the frequency accumulation difference calculation unit 7 (step 5) in the data analysis apparatus according to the embodiment of the present invention is represented by a graph, with an input attribute x4 and a non-defective 1-x4 frequency accumulation (A), The relationship between 2-x4 frequency accumulation (B) of defective products and x4 frequency accumulation difference | AB | is shown. It is an example of the data output by the contribution rate calculating part 13 (step 12) in the data analyzer which concerns on one Embodiment of this invention, and the input attribute condition "with respect to the output attribute condition y = 2 (= Y) which is a problem event" x1> 2 ”and the contribution ratio of the input attribute condition“ x2> 2 ”. It is the figure which expressed the input attribute threshold value table of the embodiment of the present invention in the form of a decision tree. It is the figure which expressed the conventional decision tree-2 in the same format as FIG. It is a figure showing the conventional decision tree-1. It is a figure showing the label hierarchical structure of the conventional decision tree-2, (a) shows x1 attribute, (b) shows x2 attribute, (c) shows x3 attribute, (d) shows x4 attribute. It is a figure showing the conventional decision tree-2.

Explanation of symbols

3 threshold setting unit (threshold setting means)
4 Data classification part (classification means)
6 Frequency calculator (first evaluation means, frequency calculation means)
7 Frequency cumulative difference calculation unit (first evaluation means, difference calculation means)
8 Input attribute threshold value determination unit (threshold value determination means)
9 Factor extraction unit (factor extraction means)
10 Factor undiscovered data extraction unit (division means)
11 End condition determination unit (end condition determination means)
16 Frequency cumulative ratio calculation unit (second evaluation means)

Claims (10)

A basic data group that is a set of data composed of a plurality of input attributes x _j (1 ≦ j ≦ N, N is the number of input attributes) and one output attribute y stored in the analysis target data storage unit the DA was analyzed, the causal relationship between the input attributes and output attributes a data analyzer you analysis,
Character-numeric data that generates a numeric basic data group DA0, which is a set of numeric attribute data, by converting character attribute data contained in the basic data group DA into numeric attribute data according to a unique conversion rule Conversion means;
The numeric base data group DA0, and numeric numeric output attribute y included in the basic data group DA0, based on a comparison of the magnitude relation between the predetermined threshold value of the output attribute y, the first data group DA1, second data Classification means for classifying into group DA2 ,
For one input attributes x _j of the plurality of input attributes for each numerical value can be assumed by the said one input attributes x _j, among the data having the following values the numerical data belonging to the first data group DA1 An operation for obtaining a first frequency (1-x _j frequency cumulative%), which is a ratio of the number to the number of all data belonging to the first data group DA1, can be taken by the one input attribute x _j For each numerical value, a second frequency (2-) is the ratio of the number of data belonging to the second data group DA2 to the number of all data belonging to the second data group DA2 among the data having a numerical value less than or equal to the numerical value. x _j frequency cumulative%) is calculated, and the difference between the first frequency and the second frequency (x _j frequency cumulative difference%) is obtained for each numerical value that the one input attribute x _j can take. For each of the multiple input attributes A first evaluation means for performing
For one input attributes x _j of the plurality of input attributes, based on the single input attribute x _j of possible numerical each of the calculated difference by the first evaluation means (x _j frequency cumulative difference%) , determining a numerical value maximum of the difference is determined as a threshold value x _j-th of the input attributes x _j, a threshold determination means for, for each of the plurality of input attributes,
For one input attributes x _j of the plurality of input attributes, the threshold x _j-th of the input attributes x _j which is determined by the threshold value determining means, first frequency (1-x _j Frequency Cumulative%) second and first ratio is the ratio of frequency _{(2-x j} frequency cumulative%), at the threshold _{x j-th} of the input attributes _{x j} which is determined by the threshold determination means, (100% of - the 1 on the frequency _{(1-x j} frequency cumulative%)) (100% - the second frequency _{(2-x j} frequency cumulative%) as well as calculating a second ratio is the ratio of) first ratio And a second evaluation means for selecting the larger one of the second ratios for each of the plurality of input attributes ;
Of the ratios selected for each input attribute in the second evaluation means, the input attributes x _j having the largest _ratio, threshold x _j-th of the input attributes x _j, and outermost even larger ratio first And a factor extracting means for extracting the type indicating the ratio or the second ratio as data indicating the input attribute condition and storing the input attribute condition in the analysis result data storage unit, Data analysis equipment. Based on the input attribute condition extracted by the factor extracting means, the numerical basic data group DA0 is divided into a factor data group that satisfies the input attribute condition and another data group that does not satisfy the input attribute condition, and is classified. Further comprising a dividing means for sending at least one of the data groups to the classification means as a new numerical basic data group DA0 ,
A series of processing consisting of processing by the classification means, processing by the first evaluation means, processing by the threshold determination means, processing by the second evaluation means, processing by the factor extraction means, and processing by the dividing means is repeatedly executed. The data analysis apparatus according to claim 1 , wherein 3. The data analysis apparatus according to claim 2, wherein the dividing means selects only another data group from the classified data group and sends it to the classification means as a new numerical basic data group DA0. . It further includes an end condition determining means for determining whether or not the end condition is satisfied, and when the end condition determining means determines that the end condition is satisfied, the execution of the series of processes is ended. data analyzer according to claim 2, wherein. 5. The data analysis apparatus according to claim 4 , wherein the end condition determination unit determines whether or not the number of data of the second data group classified by the classification unit is 0 as an end condition. 3. The data analysis apparatus according to claim 1, further comprising threshold setting means for setting the predetermined threshold of the output attribute in accordance with predetermined setting information or in response to an input from a user. . The input attribute is a manufacturing process condition and / or an in-line inspection result in the product manufacturing process,
The above output attribute is the product quality judgment result,
The data analysis apparatus according to claim 1, wherein the second data group is a data group having a poor quality determination result. A plurality of input attributes x stored in the analysis target data storage unit using the data analysis device according to claim 1. _j (1 ≦ j ≦ N, where N is the number of input attributes) and a basic data group DA that is a set of data composed of one output attribute y, and the causal relationship between the input attribute and the output attribute is A data analysis method for analyzing,
Numeric type basic data which is a set of numeric attribute data by converting character attribute data included in the basic data group DA into numeric attribute data according to a unique conversion rule by the character-numeric data conversion means. A character-numeric data conversion step for generating the group DA0;
Based on the comparison of the magnitude relationship between the numerical value of the output attribute y included in the numerical basic data group DA0 and the predetermined threshold value of the output attribute y by the classification means, the first data group DA1 And a classification step for classifying the data into the second data group DA2.
One input attribute x out of the plurality of input attributes by the first evaluation means. _j The one input attribute x _j The first frequency which is the ratio of the number of data belonging to the first data group DA1 to the number of all data belonging to the first data group DA1 among the data having a numerical value less than or equal to the numerical value that can be taken (1-x _j Frequency accumulation%), and the one input attribute x _j For each possible numerical value, a second frequency that is a ratio of the number of data belonging to the second data group DA2 to the number of all data belonging to the second data group DA2 among data having a numerical value equal to or lower than the numerical value. (2-x _j Frequency accumulation%), and the one input attribute x _j For each possible numerical value, the difference between the first frequency and the second frequency (x _j A first evaluation step of performing an operation for obtaining a frequency cumulative difference%) for each of the plurality of input attributes;
One input attribute x of the plurality of input attributes is obtained by the threshold value determining means. _j The one input attribute x in the first evaluation means _j The difference calculated for each possible numerical value (x _j Based on the cumulative frequency difference%), the numerical value for which the maximum difference is obtained is the input attribute x _j Threshold x _j-th Determining a threshold value for each of the plurality of input attributes; and
One input attribute x out of the plurality of input attributes by the second evaluation means. _j For the input attribute x determined by the threshold value determination means _j Threshold x _j-th The first frequency (1-x _j Frequency to the second frequency (2-x) _j The first ratio that is the ratio of the frequency cumulative%) and the input attribute x determined by the threshold value determination means _j Threshold x _j-th (100% -first frequency (1-x _j (Frequency cumulative%))) to (100%-second frequency (2-x _j The second ratio that is the ratio of the frequency cumulative%)) is calculated, and the larger ratio of the first ratio and the second ratio is selected for each of the plurality of input attributes. A second evaluation step;
The input attribute x having the largest ratio among the ratios selected for each input attribute by the second evaluation means by the factor extracting means. _j , The input attribute x _j Threshold x _j-th , And the type indicating whether the largest ratio is the first ratio or the second ratio (pre-correction claim 7, paragraph 0082) as the data indicating the input attribute condition, and the input attribute And a factor extracting step of storing the condition in the analysis result data storage unit. Multiple input attributes x stored in the analysis target data storage _j (1 ≦ j ≦ N, where N is the number of input attributes) and a basic data group DA that is a set of data composed of one output attribute y, and the causal relationship between the input attribute and the output attribute is A data analysis program for causing a computer included in a data analysis device to analyze to function,
The above data analyzer is
Character-numeric data that generates a numeric basic data group DA0, which is a set of numeric attribute data, by converting character attribute data contained in the basic data group DA into numeric attribute data according to a unique conversion rule Conversion means;
Based on the comparison of the magnitude relationship between the numerical value of the output attribute y included in the numerical basic data group DA0 and the predetermined threshold value of the output attribute y, the numerical basic data group DA0 is compared with the first data group DA1 and the second data. Classification means for classifying into group DA2,
One input attribute x of the plurality of input attributes _j The one input attribute x _j The first frequency that is the ratio of the number of data belonging to the first data group DA1 to the number of all data belonging to the first data group DA1 among the data having numerical values equal to or smaller than the numerical value (1-x _j Frequency accumulation%), and the one input attribute x _j The second frequency, which is the ratio of the number of data belonging to the second data group DA2 to the number of all the data belonging to the second data group DA2 among the data having numerical values equal to or smaller than the numerical value that can be taken (2-x _j Frequency accumulation%), and the one input attribute x _j For each possible numerical value, the difference between the first frequency and the second frequency (x _j A first evaluation unit that performs an operation for calculating a frequency cumulative difference%) for each of the plurality of input attributes;
One input attribute x of the plurality of input attributes _j The one input attribute x in the first evaluation means _j The difference calculated for each possible numerical value (x _j Based on the cumulative frequency difference%), the numerical value for which the maximum difference is obtained is the input attribute x _j Threshold x _j-th Threshold value determining means for determining each of the plurality of input attributes,
One input attribute x of the plurality of input attributes _j For the input attribute x determined by the threshold value determination means _j Threshold x _j-th The first frequency (1-x _j Frequency to the second frequency (2-x) _j The first ratio that is the ratio of the frequency cumulative%) and the input attribute x determined by the threshold value determination means _j Threshold x _j-th (100% -first frequency (1-x _j (Frequency cumulative%))) to (100%-second frequency (2-x _j The second ratio that is the ratio of the frequency cumulative%)) is calculated, and the larger one of the first ratio and the second ratio is selected for each of the plurality of input attributes. A second evaluation means;
The input attribute x having the largest ratio among the ratios selected for each input attribute by the second evaluation means _j , The input attribute x _j Threshold x _j-th , And the type indicating whether the largest ratio is the first ratio or the second ratio is extracted as data indicating the input attribute condition, and the input attribute condition is stored in the analysis result data storage unit Factor extraction means,
A data analysis program for causing a computer to function as each of the above means. A computer-readable recording medium on which the data analysis program according to claim 9 is recorded.

Images