JPH0261769A - Generating device for classification determining tree - Google Patents

Abstract

PURPOSE:To generate a classification determining tree to lead out a correcter inferred result by setting a suitable attribute threshold to an attribute to obtain plural values, to which an order can be attached, and suitably dividing an attribute area. CONSTITUTION:At first, an attribute value to obtain a continuous value to all the elements of a training example set is wholly extracted and arranged in small order in each attribute. An intermediate value thetai of the adjacent attribute values is obtained. This intermediate value thetai is defined as the attribute threshold and a threshold dividing table is prepared. Next, mutual information quantity I(thetai) is calculated by using the threshold dividing table. Then, an attribute threshold thetan' is obtained to lead out the maximum value of the mutual information quantity I(thetai) and by this attribute threshold thetan', a dividing table to describe the class condition of a training example to be sectioned is prepared. The mutual information quantity is calculated in each attribute and the attribute to show the largest mutual information quantity is allocated to a next low order node. Thus, the classifying tree is generated.

Description

[Detailed Description of the Invention] [Summary] Regarding a classification decision tree generation device for generating a classification decision tree from a set of training examples, the present invention aims to automatically generate a classification decision tree that leads to more correct inference results. When the value of an attribute of a training example takes on a plurality of values that can be ordered, attribute threshold selection means selects a plurality of attribute thresholds from the range of the attribute value, and the selected attribute threshold Threshold contingency table generation means that generates a threshold contingency table that describes the class state of training examples to be partitioned for each selected attribute threshold; Mutual information calculation means for calculating mutual information between an attribute and a class for each selected attribute threshold, and mutual information of a significant maximum value among the plurality of mutual information calculated. According to the attribute threshold that takes
an attribute partition determining means for partitioning the attributes, the attribute partition determining means calculates the degree of correlation between the partitioned attributes and the class, and classifies attributes that take a plurality of values that can be ordered. When the node is selected as a lower node of the decision tree, processing is performed to allocate the attribute partitioned by this attribute partition determining means as the lower node.

[Industrial Application Field] The present invention relates to a classification decision tree generation device for generating a classification decision tree that is the basis of inference processing of the classification decision device.

Examples include disease diagnosis, which identifies the disease a patient is suffering from based on the patient's illness case and test results, and failure diagnosis, which estimates the cause of machine abnormalities based on abnormal meter readings, sounds, movements, etc. In various diagnostic fields, classification knowledge is stored in a storage device called a knowledge base, and when a diagnostic case is input, the class to which the diagnostic case belongs is inferred based on the knowledge in this knowledge base. Classification determining devices that perform this process are becoming widely used.

One of the most widely used conventional classification determination devices is a knowledge-based reasoning device. As shown in Fig. 11, this knowledge-based reasoning device has two methods: ``If the starter does not rotate and the light does not come on, the battery is abnormal.'' and ``If the starter does not rotate and the light does not come on, then the battery is abnormal.'' Knowledge base 1 stores the experiential knowledge of experts in the form of inference rules, such as "If the cable is connected abnormally", and the attributes of the diagnostic cases to be classified into classes. A diagnostic case input device 2, a class inference device 3 that infers the class to which the input diagnostic case belongs based on the attributes of the input diagnostic case and the inference rules of the knowledge base 1, and an inference result. The classification result output device 4 outputs the classes as shown in FIG.

FIG. 12 shows an example of inference rules stored in knowledge base l. Classification knowledge having such a tree structure is called a classification decision tree. As shown in this figure, this classification decision tree consists of a tree structure of nodes that are associated with attributes, and branches from a root node that has no skills from other nodes to lower nodes according to attribute values. By continuing, a structure is adopted that branches to leaf nodes that ultimately specify the class of the diagnostic case. Then, the class inference device 3 starts from this root node and traces this tree structure according to the attributes of the diagnostic case to infer the class to which the diagnostic case belongs.

However, in a knowledge-based inference device with such a configuration, experiential knowledge regarding classification must be extracted from experts, which makes building the knowledge base 1 very time-consuming. Become. Therefore, the first
As shown in Figure 3, for example, medical records for disease diagnosis,
A training example input device 5 is prepared for inputting a set of training examples representing correlations between attributes and classes that have been determined in advance, and attributes of training examples to be input from this training example input device 5 are provided. By providing a classification decision tree generation device 6 that automatically generates a classification decision tree according to the strength of the correlation between the knowledge base 1 and the class, the device configuration is such that the knowledge base 1 can be automatically constructed.

Specifically, the set of training examples is "Case E has attribute A.
, has Vil as the value of , and V+Z as the value of attribute A,
... has vIh as the value of attribute A6, and belongs to class C4. It consists of a collection of training examples described as ``. FIG. 14 shows an example of a set of training examples. This training example represents whether or not a specific abnormal state exists, such as whether or not a person has heart disease, and tests that form attributes. represents various tests such as body temperature, blood, electrocardiogram, etc., and these test values become attribute values.The 0 classification decision tree generation device 6 uses a collection of such diagnostic cases to create a classification decision tree. is automatically generated.

As is clear from the previous explanation, in order for the classification decision device configured in this way to produce correct inference results,
It is necessary to configure the classification decision tree generation device 6 so that it can construct a more accurate classification decision tree.

[Conventional technology]

Next, the conventional technology of the classification decision tree generation process executed by the classification decision tree generation device 6 will be described.

The conventional classification decision tree generation device 6 generates a classification decision tree according to the following procedure when the training example set S is given from the training example input device 5. That is, (stl) Create a root node for the set S of training examples.

(st2) If S is an empty set, the corresponding node is set as a leaf node of undefined class.

(st3) If all the case sets assigned to a node belong to the same class, the class is determined based on the cases, a label is attached to the node, and the procedure is completed with the node as a leaf node.

(sL4) If not all belong to the same class,
One attribute with the highest degree of correlation between the attribute and the class is selected from the attribute set T of the training examples, and the attribute is stored with the node as a branch node.

(s t5) Divide the case set into a finite number of exclusive subsets according to the attribute values, and create and assign lower nodes to each divided subset.

(s t6) Repeat from st2 for each lower node.

In this way, a classification decision tree is automatically generated that has a structure that branches to the next lower node according to the degree of correlation between attributes and classes. For example, "mutual information" is used to calculate this degree of correlation.

In the above-mentioned process (S15), in order to branch from the upper node to the lower node, according to the attribute value,
The training example set belonging to the upper node will be divided. Next, a conventional technique for this division process will be explained.

When an attribute value takes a value that cannot be ordered, such as "YES" or "NO", there is no degree of freedom in this division process, so it is divided according to the number of attribute values. Become. On the other hand, for example, there are cases where the attribute value takes continuous values, such as ``weight,'' or items where the attribute value takes multiple discrete values, such as ``vision J,'' where the attribute value has multiple values that can be ordered. As shown in Figure 15, there is no problem when the attribute value has no overlap with the class, but as shown in Figure 16, when the attribute value When there is overlap between attributes and classes, the degree of correlation between attributes and classes will vary greatly depending on how the attribute area is divided. If this is not done properly, an attribute that should be located at a higher-level node may be set at a lower-level node, or the number of lower-level nodes that should be branched may be set to an inappropriate number, leading to problems in classification decisions. This will lead to a decline in the classification ability of the device.

Therefore, in the conventional classification decision tree generation device 6, the division process of attributes that take multiple values that can be ordered is performed based on the empirical judgment of the classification decision tree creator or the "boundary value division method" or "intermediate value division method". This was done using a division method such as "value division method". Here, as shown in Figure 17, this "boundary value division method" calculates the maximum and minimum values of the attribute values of the training examples belonging to each class, and arranges them in descending order to create an attribute area. On the other hand, the "median value division method", as shown in Figure 18, calculates the average value of the attribute values of the training examples belonging to each class, and This is a method of determining an attribute threshold value for dividing an attribute area using a median value of average values.

[Problem to be solved by the invention]

However, the conventional technique based on the empirical judgment of the classification decision tree creator has the problem that the classification decision tree cannot be automatically generated because it depends on the know-how of the classification decision tree creator. In the conventional technology based on the "boundary value splitting method," the maximum and minimum values that serve as attribute thresholds depend only on specific elements, so the attribute values of training examples belonging to that class are appropriately determined. There was a problem in that it could not be said that the results were reflected. In addition, although the conventional technology based on the "intermediate value division method" does not have the problem of dependence on specific elements like the "boundary value division method", the first
As shown in FIG. 9, when the number of elements of the training examples belonging to both classes for which the average value is provided is significantly different, there is a problem that the attribute threshold value is not appropriate. In other words, it cannot be said that the conventional technology generates a classification decision tree that can achieve a high classification accuracy rate.

The present invention has been made in view of the above circumstances, and it is possible to appropriately divide an attribute area by making it possible to set an appropriate attribute threshold value for an attribute that takes a plurality of values that can be ordered. It is an object of the present invention to provide a classification decision tree generation device that can generate a classification decision tree that leads to more accurate inference results.

[Means to solve the problem]

FIG. 1 is a diagram showing the basic configuration of the present invention.

In the figure, 1 is a knowledge base that stores classification decision trees, 5 is a training example input device that performs input processing on a set of training examples, and 6 is a knowledge base that stores classification decision trees.
is a classification decision tree generation device that automatically generates a classification tree according to the magnitude of the correlation between the attributes and classes of training examples input manually from the training example input device 5. This classification decision tree generating device 6 includes an attribute mode determining means 60, an attribute threshold selecting means 61, a threshold contingency table generating means 62, a mutual information calculating means 63, an attribute partition determining means 64, a contingency table generating means 65,
It includes correlation degree calculation means 66, maximum correlation degree detection means 67, and lower node allocation means 68.

The attribute mode determining means 60 determines whether the attribute of the training example input from the training example input device 5 corresponds to an attribute that takes a plurality of values that can be ordered or an attribute that takes values that cannot be ordered. The attribute threshold selection means 61 determines,
When the attribute mode determining means 60 determines that the attribute takes a plurality of values that can be ordered, a plurality of attribute threshold values are selected from the range of the attribute value, and the threshold contingency table generating means 6
2 generates, for each selected attribute threshold, a threshold contingency table that describes the class state of the training examples partitioned by the attribute threshold selected by the attribute threshold selection means 61, and the mutual information calculation means 63, Using the threshold contingency table generated by the threshold contingency table generation means 62, the mutual information between the attributes and the classes when partitioned by the attribute threshold is determined by each selected attribute threshold according to a predetermined mutual information arithmetic formula. The attribute partition determination means 64 partitions and divides the attributes according to an attribute threshold value that takes the mutual information of the significant maximum value among the plurality of mutual information calculated by the mutual information calculation means 63. The table generation means 65 generates a contingency table that describes the class states of the training examples partitioned by the attribute partition determination means 64, and the attribute mode judgment means 60 generates a contingency table that describes the class state of the training examples partitioned by the attribute partition determination means 64. When it is determined that , calculates the magnitude of the degree of correlation between the attribute and the class, and the maximum correlation degree detection means 67 detects the degree of correlation showing the largest value from among the plurality of degrees of correlation calculated by the degree of correlation calculation means 66. By doing so, the attribute to be assigned to a lower node is specified, and the lower node allocation means 68 selects an attribute that takes a plurality of values that can be ordered as a lower node of the classification decision tree depending on the degree of correlation. In this case, the attribute section determined by the attribute section determining means 64 is assigned as a lower node, and attributes that take values that cannot be ordered are placed at the bottom of the classification determination tree depending on the degree of correlation. When the node is to be selected as a lower node, processing is performed to allocate an attribute partitioned by its value as a lower node.

[Effect]

The classification decision tree generation device 6 selects attributes from a set of training examples representing correlations between attributes and classes according to the degree of correlation between attributes and classes, and selects them in this way. By assigning the attributes to the lower nodes, a classification decision tree is generated that branches from the upper nodes to the lower nodes. At this time, in the classification decision tree generation device 6 of the present invention, for attributes that take a plurality of values that can be ordered, the mutual information calculation means 63
The attributes are divided according to an attribute threshold that takes the mutual information of the significant maximum value among the mutual information calculated by, and
With the degree of correlation between this partitioned attribute and class,
Processing is performed so that the degree of correlation is the criterion for selecting attributes when generating a classification decision tree.

The attributes defined by the attribute threshold determined according to this amount of mutual information have the maximum amount of information regarding the class. Moreover, it is uniquely determined by calculation processing. From now on, according to the present invention, a classification decision tree with high classification ability will be objectively and automatically generated.

(Example〕 Hereinafter, the present invention will be explained in detail according to examples.

FIG. 2 shows a flowchart of the classification decision tree generation process executed by the classification decision tree generation device 6 of the present invention. As shown in this flowchart, the present invention is no different from the prior art; first, the attribute A that is manually input from the training example human-powered device 5 is a plurality of values that can be ordered (hereinafter referred to as continuous values). It is determined whether the attribute takes values that cannot be ordered or values that cannot be ordered (hereinafter referred to as discontinuous values). Then, when the attribute value takes discrete values, a contingency table is created that describes the class states of the training examples divided by that value, and when the attribute value takes continuous values, the attribute threshold for dividing the attributes is first set. It then processes to create a contingency table that describes the class states of the training examples partitioned by the attribute thresholds. When this contingency table creation process is completed for all attributes, the priority of the attributes is calculated by calculating the correlation between attributes and classes using the generated contingency table, and the highest priority A classification decision tree is generated by selecting an attribute, that is, an attribute with the highest degree of correlation, and assigning it to lower nodes.

The present invention proposes to realize generation of a classification decision tree with high classification ability by determining attribute threshold setting processing for attributes that take continuous values according to mutual information. .

FIG. 3 shows the steps executed by the classification decision tree generation device 6 of the present invention.
A flowchart for determining this attribute threshold is shown.

Next, this flowchart will be explained. In addition,〔
As explained in the ``Prior Art'' section, the classification decision tree generation algorithm is basically executed recursively. Therefore, excluding the difference in training examples to be classified, the process described below is not only the generation process when branching from the root node to the lower node, but also the generation process from the intermediate node located in the middle to the next lower node. The generation process when branching to a node is exactly the same.

First, as shown in step 1, all attribute values of attributes that take continuous values are extracted for all elements of the training example set, and the extracted attribute values are arranged in descending order for each attribute. For convenience of explanation, the attribute value of this arranged attribute is (V+≦■2≦■,≦, , , ,
, ≦Vk). If we assume a set of training cases as shown in Figure 4, which shows whether or not a diabetes test is necessary, this process will determine the "weight" and "height".

For each attribute that takes continuous values, such as visual acuity j, a row of attribute values arranged in ascending order is obtained.For example, for "weight", (54, 3, 54, 5,54
, 6, ... 102.1) can be obtained.

Next, in step 21 of step 2, first, V, +V
,,.

Calculate θ,−. Through this calculation process, the intermediate value of the adjacent attribute values found in the process of step l is found. To explain using the example of "body weight," it is determined to be 54.4 kg, which is the intermediate value between 54.3 kg and 54.5 kg. In step 21, a threshold contingency table is created using the thus calculated θ1 as the attribute threshold.

The second threshold contingency table describes the class states of training examples divided by attribute thresholds. FIG. 5 shows an example of the threshold contingency table created. FIG. 5(A) shows θ,=65
．． This is an example of a threshold contingency table with 4kg as the attribute threshold.1. FIG. 5(B) is an example of a threshold contingency table with θ1=89.8 kg as the attribute threshold. In this way, in step 21,
A threshold contingency table is created for each calculated attribute threshold. Note that when there are multiple classes to be classified, this threshold contingency table is expressed in a more general description format, as shown in FIG. In Fig. 6, N is the number of diagnostic cases belonging to class C, and Fj(θ) is the number of diagnostic cases belonging to class C4 whose attribute value is smaller than the attribute threshold θ. , [N, −F, (θ
)) is the number of diagnostic cases belonging to class C1 whose attribute value is larger than the attribute threshold value θ. however,
1515m.

When the process in step 21 is finished, the next step 22
Using the created threshold contingency table, the mutual information I(θL
) is calculated according to the formula below.

1 (θt)−x N.

(Σ(F,(θ)logzFj(θ))±Σ((N,-
F, (θ))・lOgz(N;-F 7(θ)))-Σ
(N7 ・logzN;) 10N++ ・IogzNo)
...Formula (1), where No is the total number of training examples, Q ・logtO=O.

The mutual information amount I(θi) defined by equation (1) represents the amount of information regarding the class to which the corresponding case belongs. Therefore, if we use attributes partitioned by the attribute threshold e, which gives a large value of mutual information I(θ,), a classification tree with high classification ability can be created because there is a large amount of information for inferring classes. It will be possible to build. Mutual information itself has been used in the prior art as one of the evaluation formulas for calculating the degree of correlation between attributes and classes, but in the present invention, unlike the prior art, this mutual information The aim is to use the amount of information to evaluate the process of partitioning sections of attributes that take continuous values.

By executing the processes of steps 21 and 22 for each attribute for all V obtained in step l, mutual information I(θ,
) is calculated, and in the subsequent step 3, the attribute threshold θ is calculated for each attribute, which satisfies the inequality: ■(θto,)≦I(e,) and ■(θI)≧■(θゑ,1) , to detect.

That is, the attribute threshold value at which the mutual information II (θ,) exhibits the maximum value is determined. The attribute threshold values that bring about this maximum value are expressed as [θ, ゛, θ2゛, θ3゛, . . . θ, °].

In the present invention, using 87' obtained in this way, the value range of the attribute [5inA, maxA] is divided into the limited exclusive intervals [m1nA, θ
, °) [θ1", θ+), ......[θ6°
, maxA]. The reason why attributes that take continuous values are divided according to the attribute threshold that gives the local maximum value is to effectively use a large amount of information for inferring a class. When generating this interval, if the distance between adjacent θ,゛ is not large enough to be a significant difference, the interval between (3n+) is not a significant interval in terms of classification ability. , it is also meaningful to adopt a method of rejecting one or the other.

In step 3, the attribute threshold value θ,゛ that brings about the maximum value of mutual information is determined, and the attribute partition is determined by this attribute threshold value θ7゛. Then, the classification decision tree generating device 6 of the present invention creates a contingency table that describes the class states of the training examples partitioned by this attribute threshold value θ7''. 4kg and θ1=89.8k
If the attribute threshold value g is the attribute threshold value e, l' that brings about the maximum value of mutual information regarding the attribute "weight", then this process creates a contingency table as shown in Figure 7. That will happen. The process of creating this contingency table is
It is performed for all attributes that take continuous values among the attributes of the training example. Therefore, if we look at the set of training examples shown in FIG. 4, this contingency table is created by executing the processes in steps 1 to 3 for "height" and "visual acuity." Note that when there are multiple classes to be classified, this contingency table will be expressed in a more general description format as shown in FIG.

On the other hand, the classification decision tree generation device 6 of the present invention is able to immediately generate discontinuous values for attributes that take discontinuous values, such as "urinalysis" in the training example shown in FIG. We will create a contingency table that describes the class states of training examples partitioned by discrete values. FIG. 9(A) shows a specific example of the contingency table for urinalysis test J. This contingency table creation process is also executed for all attributes that take discontinuous values among the attributes of the training example. Figure 9 (B) shows the general description format of a contingency table when there are multiple classes to be classified.As is clear from this figure, a contingency table for attributes that take continuous values is The contingency tables for attributes that take discrete values will have exactly the same format, except for the difference in the format of the attribute value description.

In this way, for all attributes of the training example
After creating the j table, the classification decision tree generation device 6 of the present invention:
Similar to the prior art, the created contingency table is used to determine the degree of correlation between each attribute and class. Various evaluation formulas have been proposed for calculating this degree of correlation, but here, the mutual information used to obtain a contingency table of attributes that take continuous values will be explained.

Once a contingency table as shown in FIG. 9(B) is obtained, mutual information 1(A) is calculated for each attribute according to the following formula.

1 (A) --X N.

-ΣNJ-1ogzN7+No ・IogzNol Here, N is the total number of diagnostic cases, and N is the class C5
, Xij is the number of diagnostic cases belonging to the class CJ that has the value V + (value range for attributes that take continuous values), and Xi is the number of diagnostic cases that belong to the class CJ that has the value V, (value range for attributes that take continuous values). For attributes, it is the number of diagnostic cases that have a value range).

The larger the mutual information 11 (A), the greater the degree of correlation with the class. From now on, the attribute indicating the largest mutual information 11(A) is assigned to the next lower node, thereby generating a classification tree.

Finally, the processing contents of the present invention will be specifically explained in accordance with the classification decision tree generation processing procedure described in [Prior Art].

By calculating the degree of correlation between attributes and classes according to step (s t4) of the classification decision tree generation process, it is determined that "weight", which is partitioned as shown in FIG. 7, is the attribute with the highest degree of correlation. If it is determined that, as shown in Fig. 10, "weight" is assigned to the root node, and the interval [54, 3, 65, 4) [65
, 4, 89, 8) Assign [89, 8, 102, 1] and label them respectively. And the weight value is
The training example set is divided into three parts depending on whether they belong to these three sections. The classification decision tree generation process is recursively applied to this divided set of three training examples. Note that 54.3 kg represents the minimum weight among the diagnosed cases, and 102.1 kg represents the maximum weight among the diagnosed cases.

According to (st3) of the classification decision tree generation process, all elements of the set of training examples whose weight values fall in the interval [54, 3, 65, 4) belong to the "test not required" class. When detected, the node assigned to this branch is set to the leaf node labeled "Unnecessary J" as shown in Figure 1O.Similarly, if the weight value is [89, 8, 10
2, 1], when it is detected that all elements of the set of training examples that fall in the interval belong to the “inspection required” class.
As shown in FIG. 1O, the node assigned to this branch is set to a leaf node labeled "required."

The elements of the training example set whose weight values fall in the interval [65, 4, 89, 8) do not belong to the same class, so
According to (st4) of the classification decision tree generation process, "weight"
Contingency tables are created by determining attribute thresholds as necessary for other attributes except for , and the degree of correlation between attributes and classes is calculated. According to this calculation, if "urine test" is the attribute with the highest degree of correlation, the value of weight is [65, 4, 89,
The set of training examples falling in the section 8) is divided according to the urine test values O'° and "x"'.

The present invention has been described above in detail according to the illustrated embodiments, but
The present invention is not limited to 6 degrees, for example,
It is not limited to two cases such as ``need for inspection'' and ``required heat for inspection,'' but can also be applied to multiple cases. Furthermore, the scope of the present invention includes processing for generating classification knowledge in an IF-THEN format that uniquely corresponds to the classification decision tree.

(Effects of the Invention) As described above, according to the present invention, it becomes possible to objectively determine a threshold value for dividing an attribute for an attribute that takes a plurality of values that can be ordered.

Moreover, since this required threshold value has more information than other threshold values when performing class inference, it becomes possible to generate a classification decision tree that can lead to more accurate 11E theory results.

[Brief explanation of the drawing]

Figure 1 is a diagram of the principle configuration of the present invention, Figures 2 and 3 are flowcharts of the execution of the present invention, Figure 4 is an example of a collection of training examples, and Figure 5 is the process of step 21 in Figure 3. An example of a threshold contingency table that is created. Figure 6 is an explanatory diagram of a threshold contingency table. Figure 7 is an example of a contingency table for attributes that take continuous values. Figure 8 is an explanatory diagram of a contingency table for attributes that take continuous values. , Fig. 9 is an explanatory diagram of a contingency table for attributes that take discontinuous values, Fig. 1O is an example of a classification decision tree that is generated, Fig. 11 is a configuration diagram of a classification decision device, and Fig. 12 is a diagram of a classification decision tree. Fig. 13 is a configuration diagram of the classification determination device; Fig. 14 is an explanatory diagram of a set of training examples; Figs. 15[m and 16 are illustrations of overlapping attributes and classes; Figs. The figure is an explanatory diagram of the prior art, and FIG. 19 is an explanatory diagram illustrating the problems of the prior art. In the figure, 1 is a knowledge base, 2 is a diagnosis case manual device, 3 is a class inference device, 4 is a classification result output device, 5 is a training example input device, 6 is a classification decision tree generation device, 60 is an attribute mode judgment means, 61 is an attribute threshold selection means, 62 is a threshold contingency table generation means, 63 is a mutual information calculation means, 64 is an attribute partition determination means, 65 is a contingency table generation means, 66 is a correlation degree calculation means, and 67 is a maximum correlation degree detection means. Means 68 is a lower node allocation means. 4 Fumei's roasted bamboo zuro frochi de oneto (1) Tea 2 Fig. The running flow of the present invention 1,000 f-t (mutual) $ 3 Fig. (A) Fig. 3 Stella 7'21/), < 几Moon jT "I keba diruru Q4 direct now IA's a), jki 5 口蓼...1.B east" 1 many, 1st 4th resultant force -4 n° 1$
4 Figure 04 Straight 41 Character 11 Table/) Sail 0 Month Figure Number Figure Renf Bald 1. 1st birthday, Q now 31 table ψ ita, j fig. (A) (B) Nodo lhii direct 1 Z, robu family! (The order of August 2 Tower 9 Figure C4 bald J beg Otoro A also life J right now Atsu theory a month whistle diagram first wide ゛σ offer now the koshi system construction part I ] Section ] The structural diagram of the classification legal device is) Fig. A° diagram of Imasaku Iwoemaki 1 (1) Fig. 13 A fixed tree A is also on the dividing line of Fig. $ 12 Fig. μm 1 bell "Toy j Ririden Itori S Men Hatred η Section, Horse tiI, Kukusu overlap/l Explanatory diagram (1) No. 15
Figure A and the 7th Class Master, Narifi Iji T3r3I Ward No. 16
Figure (It) (A) (B) Igami Raifu 4 Eigo no 1 light Ql 12] (IL) No. 78 [21 (A) (B) l [The theory of rice-like stock 1 θ Higuchi (1) I Chuhe machine 〃Yong Ding r, P + is negative door N, first explanation 2] Figure 9

Claims (1)

[Claims] An attribute is selected from a set of training examples representing a correlation between the attribute and the class according to the degree of correlation between the attribute and the class, and is selected in this way. In a classification decision tree generation device (6) that generates a classification decision tree configured to branch from higher nodes to lower nodes by assigning attributes to lower nodes, the value of the attribute of a training example is determined. an attribute threshold selection means (61) for selecting a plurality of attribute thresholds from the range of the attribute value when the attribute value takes on a plurality of values that can be ordered; a threshold contingency table generating means (62) that generates a threshold contingency table describing the class states of training examples partitioned by the selected attribute thresholds for each selected attribute threshold;
2) Using the threshold contingency table generated by step 2), calculate the mutual information between attributes and classes when partitioned by attribute threshold for each selected attribute threshold according to a predetermined mutual information arithmetic formula. Mutual information calculating means (63); dividing attributes according to an attribute threshold value that takes the mutual information of a significant maximum value among the plurality of mutual information calculated by the mutual information calculating means (63); Attribute section determination means (6
4), the attribute division determination means (64) calculates the degree of correlation between the divided attributes and classes, and determines the possible ordering based on the calculated degree of correlation. When an attribute that takes a plurality of values is selected as a lower node of the classification decision tree, processing is performed to assign the attribute partitioned by the attribute partition determining means (64) as the lower node. A classification decision tree generation device.