JP2001175640A - Cluster classification device and method, and recording medium recorded with program for cluster classification - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a cluster classifying device which is improved in performance by performing classification and analysis by taking a subset (area) of featured data from actual property data which have overlaps between different classes and are used for, e.g. credit examination, marketing, etc. SOLUTION: Past property data are stored in a past property data storage part 21, property data stored in the storage part 21 are received, and a cluster number determination part 31 determines the number of clusters corresponding to the received property data; and the number of clusters and the property data are received from the decision part 31 and clustered by a clustering part 32, the clustering result is received from the clustering part 32, and a result analysis part 33 analyzes the result.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a cluster classification device, a cluster classification method, and a recording medium storing a cluster classification program, and more particularly, to a data set (data space) including at least two or more types of data having different properties. The present invention relates to a cluster classification apparatus, a cluster classification method, and a recording medium that records a cluster classification program for extracting and classifying target data by extracting a subset (data area) of characteristic data.

[0002]

2. Description of the Related Art When performing credit screening, marketing, and the like using a large amount of attribute data such as customer data and product data, it is necessary to classify target data.

For example, in credit screening, it is necessary to classify customer data in order to determine whether a new customer can be repaid.

[0004] In marketing, very useful information can be obtained if customers can be classified according to purchase tendency or products that are likely to be of interest.

[0005] The data classification will be described below by taking a credit check as an example. Conventionally, various methods and apparatuses have been proposed for such data classification, and most of them perform class classification for constructing a classification boundary surface in a target data space.

However, in the classification of actual attribute data,
There is a large overlap between different classes, and in the method of constructing and classifying a boundary surface in the data space, no matter how the boundary surface is constructed, a limit may often occur in the classification performance. For example, in the credit check, a part of the customer data (defective data) for which the loan cannot be returned may have the same attribute as the data (normal data) of the customer who normally repaid without delay.

In the classification method for forcibly constructing a boundary surface, data having meaningful characteristics is often recognized only as a correct classification or a misclassification. For example,
In credit screening, in the past cases, fortunately, loans were repaid, but customers with attributes that make it difficult to repay loans should be recognized as those who have difficulty in repayment, Techniques such as constructing classification boundaries,
Such recognition is difficult.

As a method for solving the above problems,
There is known a method of extracting a characteristic region from a data space instead of separating the data space at a boundary surface.

[0009] For example, from the customer data space of the loan credit, "the area with many normal customers", "the area with many bad customers",
By extracting a characteristic area "area where both normal and defective customers are mixed", detailed analysis of past customer data becomes possible.

[0010] Further, for a new customer, not only the determination of “normal” and “bad” is made, but also the determination of “likely to be normal” is made based on past cases. It can be shown together with the probability value.

It is also possible to extract a customer list whose attributes are close to the new customer concerned from past cases.

A technique satisfying the above requirements can be said to be a clustering technique rather than a class classification. That is, as shown in the above-described example, the characteristic regions “region with many normal customers”, “region with many bad customers”, and “region where normal and bad customers are mixed” are extracted as clusters, and classified and analyzed. It does.

Typical clustering methods include a maximum likelihood estimation method, a K-means method, an LBG method, and a hierarchical method.

Among them, the maximum likelihood estimation method, the K-means method, and the LBG method assume the number of clusters, the position, and the distribution shape of data,
This is a method for performing classification and analysis. However, when dealing with actual data, the number of clusters, the position, and the distribution shape of the data are often unknown before classification, so there is a possibility that mistakes may be made or appropriate initial values cannot be set. is there. In such a case, excellent performance cannot be expected with these clustering methods (such as the maximum likelihood estimation method, the K-means method, the LBG method, and the hierarchical method).

The hierarchical method is a method of sequentially integrating and dividing data and performing clustering based on a distance between data and a cluster defined by a user. There is a problem that the result greatly changes depending on the setting of the initial state of the algorithm.

As a method for solving the above problem, a self-organizing feature map (SOM) is known.

The SOM can perform clustering without having to assume the number of clusters, the position, and the distribution shape of data before classification.

This SOM will be described below. FIG.
0 is a diagram for explaining SOM. Referring to FIG. 10, the SOM includes an input layer 51 and a competition layer 52.
This is a two-layer neural network. The number of units in the input layer 51 is equal to the number of dimensions of the target vector data. In the competitive layer 52, the competitive layer units are two-dimensionally arranged, and each competitive layer unit is fully connected to the input layer unit.

When clustering is performed using SOM,
Finally, one or more competitive layer units correspond to one cluster.

That is, past data belonging to each cluster is assigned to one of the competitive layer units corresponding to each cluster.

The clustering process by SOM will be described. In SOM, 1) a competition by a competition layer unit for input data, and 2) an update of a weight vector of a combination of a competition layer unit that has won the competition and a unit that is close to the competition layer unit on the competition layer. The clustering process is performed at the stage.

When the number of competitive layer units is M, each competitive layer unit is given a number i = 1,..., M. Let wi be the weight vector of the connection between the i-th competitive layer unit and the input layer.

When past data x is input, a weight vector wc of a competitive layer closest to x is found in the data space.

That is,

## EQU1 ## This wc is the competitor unit that won the competition.

At this time, the following update is performed on the weight vector of each competitive layer unit.

[0027]

Here, Nc is in the vicinity of the c-th competitive layer unit,
α (t) represents a learning coefficient, and t represents time.

While repeating the update, Nc and α (t) are gradually reduced.

By the above processing, the weight vector of the combination of the competitive layer units becomes a vector representing the target data space. Then, the weight vector of the combination of the respective competitive layer units becomes representative data of the subspace in which the weight vector is closest.

Further, past customer data with each weight vector being closest is assigned to a competitive layer unit corresponding to each weight vector, and clustering can be realized.

SOM is described in, for example, a document by T. Kohonen (“Self-Organizing Map”, Proc. IEEE, Vol.
78, No. 9, pp. 1464-1480 (1990)) (referred to as “Reference 1”).

[0033]

Although a method using the above-mentioned SOM property for clustering has been proposed, the number of competing layer units must be reduced in order to effectively extract the above-mentioned characteristic region. Must adapt to data.

Further, for example, JP-A-10-283336, JP-A-8-36557, JP-A-7-64948, JP-A 2-2115
The method described in Japanese Patent Publication No. 76, etc. merely uses the SOM simply by giving the number of competitive layer units at the initial stage. Therefore, an appropriate feature area is not necessarily obtained from a data set in which overlap exists between different classes. Can not be taken out.

In the clustering method described in, for example, Japanese Patent Application Laid-Open No. 5-205058, in order to obtain an optimal cluster, a method is employed in which the number of units in the competitive layer is sequentially changed and each case is examined. However, since the clustering is repeated from a case where the number of clusters is one mechanically irrespective of the target data to a predetermined maximum number of clusters without prior knowledge of the data, there are many useless processes.

Further, since the evaluation performed after checking the number of each cluster is based on the average value of the distance between each data and the representative data of each cluster and the average value of the data in the cluster, It is still unsuitable for dealing with overlapping data between disparate classes.

In the devices and the like described in the above patent publications, only the original algorithm of SOM is used as it is, and data is concentrated only in some clusters due to the characteristic of the SOM algorithm. In some cases, cluster trends cannot be clearly identified.

As described above, although attribute data to be classified in, for example, credit screening, marketing, and the like often overlaps between classes, there have been many conventionally proposed class classifications and classifications. In the clustering method, the classification is simply performed without considering the overlap between the classes, and it cannot be appropriately handled.

Accordingly, the present invention has been made in view of the above-mentioned problems, and has as its object to provide an overlap between different classes, for example, a real attribute used in credit screening, marketing and the like. It is an object of the present invention to provide an apparatus and a method and a recording medium that perform classification and analysis by extracting a subset (region) of characteristic data from data to improve classification accuracy and performance. Other objects, advantages, features, etc. of the present invention will be readily apparent to those skilled in the art from the following description.

[0040]

According to a first aspect of the present invention, there is provided an apparatus for receiving past attribute data storing means for storing attribute data in the past, and receiving attribute data stored in the past attribute data storing means. Cluster number determining means for determining the number of clusters corresponding to the attribute data; cluster number and attribute data from the cluster number determining means; clustering means for performing clustering; and clustering results from the clustering means, And a result analyzing means for analyzing the result.

The apparatus according to the present invention further comprises an output means for outputting a result, and an input means for designating the number of clusters, wherein the number of clusters is determined while observing the analysis result of the result analyzing means. , The user can specify the optimal number of clusters.

In the apparatus according to the present invention, the number-of-clusters determining means determines the number of clusters based on the class ratio of the target data. In the device according to the present invention,
The result analysis unit analyzes the class ratio of each cluster, the representative data, and the similarity between the data in the cluster. The attribute data is customer data.

The method according to the present invention further comprises the steps of: storing attribute data in the past in a storage means; determining the number of clusters corresponding to the attribute data stored in the storage means; Receiving the attribute data and performing the clustering, and analyzing the clustering result.

In the method according to the present invention, the user can specify the optimum number of clusters based on the analysis result, and can analyze the target data. In the method according to the present invention, in the step of determining the number of clusters, the number of clusters is determined based on a class ratio of the target data. Further, in the method according to the present invention, in the step of analyzing the clustering result, analyzing the class ratio of each cluster, representative data, and the similarity between the data in the cluster is one of the features. I have.

According to the present invention, a customer data classification for classifying customer data for judging whether or not a new customer can be repaid in a credit check, and a customer in a marketing for each product which is likely to have a purchasing tendency or interest. It is suitable for use in business such as data classification for classifying into.

[0046]

Embodiments of the present invention will be described below. In a preferred embodiment, the cluster classification device of the present invention stores attribute data in the past in past attribute data storage means (21), and receives attribute data stored in past attribute data storage means (21). The number of clusters corresponding to the received attribute data is determined by the number-of-clusters determining means (31), the number of clusters and the attribute data are received from the number-of-clusters determining means (31), and clustering is performed by the clustering means (32). The clustering result is received from the means (32), and the result is analyzed by the result analyzing means (33).

The clustering means (32) extracts a characteristic region in the data space by performing clustering without assuming data distribution or the like before classification.

The result analysis means (33) analyzes each cluster, and based on the analysis result, determines the number of clusters (3).
Using 1), the number of clusters suitable for the target data can be determined.

Further, the processing of the number-of-clusters determining means (31), the clustering means (32), and the result analyzing means (33) is repeated until an area having an overlap between classes and an area having no overlap can be extracted. By extracting an objective data area and analyzing each feature data set, highly accurate data classification and classification determination of new data become possible.

In a preferred embodiment of the method of the present invention, (a) storing attribute data in the past in a storage means; and (b) the number of clusters corresponding to the attribute data stored in the storage means. (C) receiving the determined number of clusters and attribute data and performing clustering, and (d) analyzing the clustering result.

According to the present invention, the user can specify the optimum number of clusters on the basis of the analysis result and analyze the target data.

In the present invention, the number of clusters is determined based on the class ratio of the target data.

In the present invention, the result analysis analyzes the similarity between the class ratio of each cluster, the representative data, and the data in the cluster.

In the present invention, a program executed by a data processing device such as a computer includes (a) a process of storing attribute data in the past in a past attribute data storage unit, and (b) a process of storing attribute data in the past in a past attribute data storage unit. Receiving the attribute data, and determining the number of clusters corresponding to the attribute data; and (c) receiving the cluster number and the attribute data from the cluster number determining process.
A clustering process for performing clustering; and (d) a result analysis process for receiving a clustering result from the clustering process and analyzing the result, and causing the computer to execute each of the processes (a) to (d). The present invention can be implemented by reading out the program from a recording medium storing the program or a wireless / wired communication medium carrying the program into a data processing device and executing the program.

[0055]

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram showing an embodiment of the present invention;

FIG. 1 is a diagram showing the configuration of one embodiment of the present invention. Referring to FIG. 1, an embodiment of the present invention includes an input device 1 such as a keyboard and a storage device 2 for storing information.
And a data processing device 3 operating under program control
And an output device 4 such as a display device or a printing device.

The data processing device 3 includes a cluster number determining unit 3
1, a clustering unit 32, and a result analysis unit 33.

The storage device 2 includes a past attribute data storage section 21 for storing past attribute data input from the input device 1.

The number-of-clusters determining unit 31 reads out past attribute data stored in the past attribute data storage unit 21, shapes the attribute data into a form that can be clustered, and totals.

The number-of-clusters determining unit 31 sets the initial number of clusters based on the result of the aggregation of the attribute data stored in the past attribute data storage unit 21.

The clustering section 32 receives the attribute data that has been formed into a form that can be clustered by the cluster number determining section 31, and performs clustering according to the SOM algorithm. For the SOM algorithm, reference is made to the description of the above-mentioned document 1 of, for example, T. Kohonen.

When the original algorithm of the SOM is used in the clustering unit 32, the victory of the competition is concentrated on only a part of the clusters due to the characteristic characteristic of the SOM algorithm. In some cases, it is not possible to clearly extract the tendency of each cluster.

In such a case, the above situation is avoided by performing correction using a conscience mechanism algorithm. The conscience mechanism algorithm is an algorithm that makes it possible to comprehend the competition based on the victory history of the competition in the clustering process and thereby grasp the tendency of the entire target data in the entire competition layer. For the above-mentioned problems that may occur in the original algorithm of SOM and the conscience mechanism algorithm, see the document by D. DeSieno (“Adding a Conscience to
Competitive Learning, ”Proc. Int. Conf. On Neural
Networks, I, pp. 117-124 (1988)) (referred to as “Reference 2”) and the like.

The result analyzing unit 33 includes the clustering unit 32
Then, an analysis is performed to determine whether to receive the clustering result and display the result on the output device 4 to end the processing, or to change the number of clusters again to perform clustering.

As a result of the result analysis, when the tendency of each cluster cannot be clarified, the number of clusters is changed and the above-described clustering process and result analysis are repeated until the tendency is clarified.

By the above processing, a characteristic data area in the target data space can be extracted as a cluster, and the result can be used to classify even data having an overlap between classes. For example, when certain data is assigned to a cluster to which many data of a certain class belongs, there is a high possibility that the data belongs to the class.

When certain data is assigned to a cluster to which a plurality of classes of data belong, the probability that the data belongs to a specific class is reduced. Further, the probability can be estimated by summing up the past data.

FIG. 2 is a flowchart for explaining the operation of one embodiment of the present invention. In the following, an example of performing a credit check at a financial institution using the customer data (attribute, age, gender, annual income, occupation, industry, marriage, repayment status, etc.) shown in FIG. The operation will be described.

When the operation is started (step 10 in FIG. 2)
1) The cluster number determination unit 31 shapes and totals the attribute data used for clustering, and sets the initial cluster number (step 102 in FIG. 2).

The shaping of the attribute data is performed using, for example, a table as shown in FIG. FIG. 5 is a diagram showing an example in which the attribute data is shaped for the customer data shown in FIG.

The setting of the initial number of clusters in step 102 of FIG. 2 is performed, for example, as follows. The number ratio of customer data that has been normally repaid without repayment delays (normal data) and customer data that has an abnormal occurrence such as repayment delays (defective data) is checked. In many cases, the number of special data (defective data in this example) is small.

Using the conscience mechanism algorithm,
Assuming that the same number of data is allocated to each cluster, it is ideal if a cluster with only bad data and a cluster with only normal data can be generated.

Accordingly, an example of the initial cluster number is an integer value close to a value obtained by dividing the total number of data by the number of defective data. For example, when clustering is performed on data having 900 normal data and 100 defective data, the initial number of clusters is assumed to be 10.

The number-of-clusters determining unit 31 passes the data for clustering and the initial number of clusters to the clustering unit 32.

Next, in the clustering unit 32, the initial cluster number received from the cluster number determination unit 31
Perform clustering according to SOM algorithm
(Step 103 in FIG. 2).

For the SOM algorithm, for example, the description of the above-mentioned document 1 by T. Kohonen is referred to.

The clustering section 32 sends the clustering result to the result analyzing section 33.

The result analyzing unit 33 includes the clustering unit 32
Receives and analyzes the clustering result (step 104 in FIG. 2).

The result analyzer 33 outputs the analysis result to the output device 4
And the output device 4 displays the received result (step 105 in FIG. 2).

As an analysis performed by the result analysis unit 33, for example, a normal / defective ratio of data belonging to each cluster is obtained as a result of clustering. As a result of examining the normal / defective ratio of the data belonging to each cluster, "a cluster where most of the belonging data is normal data", "a cluster where most of the belonging data is bad data", "almost half of the belonging data are normal, Clusters with almost half of which are defective, etc. are obtained. If the user can grasp the tendency of the customer data, the user's judgment (step 106 in FIG. 2) ends the clustering (step 108 in FIG. 2).

When the result analyzer 33 sends the above-mentioned normal / defective ratio to the output device 4, the graphical data as shown in FIG. 3 is also sent so that the user's understanding can be further assisted. Become.

FIG. 6 is a schematic diagram of a competitive layer of the SOM. As described above, each competitive layer unit corresponds to a separate cluster.

FIG. 6 is a pie chart showing the normal / defective ratio of customer data belonging to each cluster. In the example shown in FIG. 6, the competitive layer unit 1 is “a cluster in which most of the data to which the data belongs is normal data”; the competitive layer unit 4 is a “cluster in which most of the data to which the data belongs is bad data”; Further, it can be seen that the competitive layer unit 3 is a “cluster in which normal and defective data are mixed”.

By clicking each competitive layer unit (cluster) to display the customer data belonging to each cluster as shown in FIG. 7, it is possible to analyze the customer data belonging to each cluster in detail. I can do it.

In FIG. 7, “representative data” is the closest attribute value corresponding to the value of the weight vector of the competitive layer unit.

Here, it is also possible to arrange the data belonging to each cluster to the representative data in ascending order of distance such as Euclidean distance.

For the reason that the tendency has not been clarified yet,
If the user determines that further analysis is to be performed (step 106 in FIG. 2), the number of clusters is changed by the user's instruction to the cluster number determination unit 32 (step 107 in FIG. 2), and clustering is performed again (step 106 in FIG. 2). Step 1 of FIG.
03), The above processing is repeated until a satisfactory result is obtained by the user.

For example, as described above, if the tendency is not clarified in ten clusters as described above, there is a possibility that the tendency may appear by performing finer cluster division. The number of clusters is assumed to be 20
Increase to pieces.

After a satisfactory clustering result is obtained for the user, a new customer can be determined. The method will be described below.

When new customer data is input from the input device 1, the clustering unit 32 finds a competitive layer unit having a weight closest to the new customer data, that is, a cluster, and passes the result to the result analysis unit 33. .

New customer data can be determined by counting the classes of data belonging to the cluster determined to be closest.

The result analyzing section 34 passes the analyzed result to the output device 4, and the output device 4 displays the result. FIG. 8 shows an example of the display. In FIG. 8, the cluster corresponding to the competitive layer unit to which the newly input customer data belongs, that is, the cluster corresponding to the competitive layer unit 4 shines on the screen, and when the lit cluster is clicked with a mouse or the like, the new data input and the new data are input. Past customer data close to is arranged and displayed.

In the example shown in FIG. 8, the cluster closest to the new customer is a cluster to which many defective data belong.
The new customer data can be determined to have a high possibility of repayment delay or the like. Of course, it is also possible to calculate the probability that the new data is defective from the number of past data belonging to the cluster.

In the above embodiment, the past attribute data storage unit 21 is a writable and readable storage device such as a magnetic disk device and a semiconductor memory storage device.

The program-controlled data processing device 3 including the number-of-clusters determining unit 31, the clustering unit 32, and the result analyzing unit 33 comprises an information processing device such as a personal computer or a workstation.

Further, the input device 1 comprises an input device such as a keyboard and a mouse, and the output device 4 comprises a CRT display or a liquid crystal display.

Next, a second embodiment of the present invention will be described. FIG. 2 is a diagram showing the configuration of the second embodiment of the present invention.

FIG. 9 is a block diagram showing the configuration of the cluster classification device according to the second embodiment of the present invention. In FIG. 9, the same or equivalent elements as those in FIG. 1 are denoted by the same reference numerals. Referring to FIG. 9, the second embodiment of the present invention includes, in addition to the configuration of the above-described embodiment, a recording medium 5 on which a cluster classification program is recorded, which is composed of a magnetic disk, a semiconductor memory, or another recording medium. . The cluster classification program implements the functions and processes of the above-described cluster number determination unit 31, clustering unit 32, and result analysis unit 33 on the data processing device 3. The program is read from the recording medium 5 into the data processing device 3 via a reading device (not shown), and an executable program is loaded into the main memory of the data processing device 3 and the CPU of the data processing device 3 executes the program. Thus, the present invention can be implemented.

[0099]

As described above, according to the present invention,
In classifying the data, it is possible to extract a data area in which data of each class mainly exists or a data area in which data of a plurality of classes are mixed.

The reason is that, for example, in order to deal with the overlap existing between different classes, which is a property that attribute data to be classified in credit screening and marketing of financial institutions and the like generally has, the present invention In this case, the clustering by SOM is repeated while analyzing the target data.

According to the present invention, by using the data area obtained as a result of the classification, not only can the data be classified with high performance, but also the characteristics of the target data can be analyzed in detail. The effect is as follows.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating a configuration of a cluster classification device according to an embodiment of the present invention.

FIG. 2 is a flowchart for explaining the operation of one embodiment of the present invention.

FIG. 3 is a diagram for specifically describing one embodiment of the present invention, and is a diagram illustrating an example of attribute data.

FIG. 4 is a diagram illustrating an example of conversion of attribute data into numerical values according to an embodiment of the present invention.

FIG. 5 is a diagram illustrating an example of digitizing attribute data according to an embodiment of the present invention.

FIG. 6 is a diagram schematically illustrating a display example of a clustering result according to an embodiment of the present invention.

FIG. 7 is a diagram schematically showing a detailed display example of a clustering result in one embodiment of the present invention.

FIG. 8 is a diagram schematically illustrating a display example of a determination result when new customer data is applied in one embodiment of the present invention.

FIG. 9 is a diagram showing a configuration of another embodiment of the present invention.

FIG. 10 is a diagram schematically showing a configuration of an SOM.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 input device 2 storage device 3 data processing device 4 output device 5 recording medium 21 past attribute data storage unit 31 cluster number determination unit 32 clustering unit 33 result analysis unit

Claims (29)

[Claims] A storage means for storing attribute data in the past; a cluster number determination means for determining a cluster number corresponding to the attribute data from the attribute data stored in the storage means; A clustering unit that receives the determined number of clusters and the attribute data and performs clustering; and a result analysis unit that receives a clustering result output from the clustering unit and analyzes the clustering result. Cluster classifier. An output means for outputting a result; and an input means for designating the number of clusters, wherein an optimum value is provided to the number-of-clusters determining means based on an analysis result of the result analyzing means. 2. The cluster classification device according to claim 1, wherein the number of clusters can be set. 3. The cluster classification apparatus according to claim 1, wherein said cluster number determination means determines the number of clusters based on a class ratio of the target data. 4. The method according to claim 1, wherein said result analyzing means analyzes a class ratio of each cluster, representative data, and similarity between data in the cluster. Cluster classifier. 5. The cluster classification device according to claim 1, wherein the attribute data is customer data. 6. A customer data classification device for classifying customer data for judging whether a new customer can be repaid in a credit screening is provided for a credit screening system comprising the cluster classification device according to claim 5. Cluster classifier. 7. A cluster classification device for marketing customers, comprising a cluster classification device according to claim 5, wherein a customer data classification device for classifying customers into products that are likely to be purchased or interested in marketing. 8. A step of storing attribute data in the past in storage means, a step of determining the number of clusters corresponding to the attribute data stored in the storage means, and receiving the determined number of clusters and attribute data. A cluster classification method, comprising: performing a clustering; and analyzing a clustering result. 9. The cluster classification method according to claim 8, wherein a user designates an optimal number of clusters based on the analysis result, thereby enabling analysis of the target data. 10. The cluster classification method according to claim 8, wherein in the step of determining the number of clusters, the number of clusters is determined based on a class ratio of the target data. 11. In the step of analyzing the clustering result, the class ratio of each cluster, representative data,
The cluster classification method according to any one of claims 8 to 10, wherein similarity between data in the cluster is analyzed. 12. The attribute data is customer data.
The cluster classification method according to any one of claims 8 to 11, wherein: 13. The cluster classification method according to claim 12, wherein the cluster classification method according to claim 12 is used as a classification method of customer data for classifying customer data for determining whether a new customer can be repaid in a credit check. A cluster classification method characterized by the following. 14. A cluster classification method using the cluster classification method according to claim 12 as a method of classifying customer data for classifying customers into products that are likely to be purchased or interested in marketing. Method. 15. A process for storing (a) past attribute data in a storage unit; and (b) a cluster number for receiving attribute data stored in the storage unit and determining a cluster number corresponding to the attribute data. A determination process; (c) a clustering process of receiving the number of clusters and attribute data from the cluster number determination process and performing clustering; and (d) a result analysis process of receiving a clustering result from the clustering process and analyzing the result. And a recording medium on which a program for causing a computer to execute each of the processes (a) to (d) is recorded. 16. The recording medium according to claim 15, wherein: (f) outputting the analysis result so that the user can specify the optimum number of clusters while observing the analysis result and analyze the target data. A recording medium storing a program for causing the computer to execute an output process and (g) an input process for designating the number of clusters. 17. The recording medium according to claim 15, wherein said number-of-clusters determining process determines the number of clusters based on a class ratio of target data. A recording medium on which a program to be executed is recorded. 18. The recording medium according to claim 15, wherein the result analysis process analyzes a class ratio of each cluster, representative data, and similarity between data in the cluster. And the result analysis processing is
A recording medium recording a program to be executed by the computer. 19. The recording medium according to claim 15, wherein said attribute data is customer data. 20. (a) for the attribute data stored in the storage unit, shaping the attribute data used for clustering and totalizing the attribute data to determine the number of clusters; and (b) determining the number of clusters. Performing the clustering of the attribute data; (c) analyzing the clustering result and displaying and outputting the analysis result on a display device; and (d) calculating the data of the data belonging to each cluster from the display output of the result. A step of ending clustering when the tendency is grasped and the classification is completed; and (e) the number of clusters is changed and changed when further analysis is performed because the data tendency is not clarified from the display output of the result. Returning to step (b) and performing clustering again with the number of clusters. Star classification method. 21. The cluster classification method according to claim 20, wherein, in determining the cluster in the step (a), the number of clusters is determined based on the class ratio of the target data. 22. In the analysis of the step (c),
Find the ratio of normal to bad data belonging to each cluster.
21. The cluster classification method according to claim 20, wherein: 23. In the step (d), based on the analysis result, the user can see at a glance that a cluster is mostly normal data, a cluster is mostly defective data, and a cluster is a mixture of normal data and data. 21. The cluster classification method according to claim 20, wherein the information is displayed on the display device in a display format understandable by (1). 24. The method according to claim 20, wherein in the clustering step (b), clustering is performed using an SOM (Self Organizing Map) algorithm or an SOM algorithm and a conscience algorithm.
The described cluster classification method. 25. In the clustering step (b), clustering is performed using an SOM algorithm or an SOM algorithm and a conscience algorithm. In the step (d), S
For each of the competitive layers of the OM, the display device is displayed in a display format that can be understood by the user at a glance, such as a cluster where most of the data is normal data, a cluster where most of the data is bad, and a cluster where normal data and data are mixed. 21. The cluster classification method according to claim 20, further comprising displaying the attribute data belonging to a cluster designated by a click operation on the display device. 26. The cluster classification method according to claim 20, wherein in the step (c), a class ratio of each cluster, representative data, and similarity between data in the cluster are analyzed. 27. When new attribute data is input, a cluster having a weight closest to the new attribute data is detected,
21. The cluster classification method according to claim 20, wherein the classification and analysis of the new attribute data are performed by totalizing the classes of the data belonging to the cluster determined to be closest. 28. A storage unit for storing attribute data, means for reading attribute data from the storage unit, shaping attribute data used for clustering and totalizing attribute data to determine the number of clusters, and Means for clustering the attribute data, means for analyzing and displaying the clustering result, and display of the result. From the display output of the result, the number of clusters is determined when further analysis is to be performed. And a means for controlling to perform clustering again with the changed number of clusters. 29. When new attribute data that is not stored in the storage unit is input, a cluster having a weight closest to the new attribute data is detected, and data of the data belonging to the cluster determined to be the closest is detected. 29. The cluster classification apparatus according to claim 28, wherein new attribute data is classified and analyzed by totalizing the classes.