JP2003330583A - Cluster analysis system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a cluster analysis system that efficiently extracts a part of clusters, which are adequate results of classification, from the results of classification obtained by short time trial and error. <P>SOLUTION: The cluster analysis system is provided with a first means which inputs a plurality of groups of data to be classified; a second means which accumulates the groups of the partial data of the results of the classification executed to each of the plurality of groups of data inputted by the first means with a plurality of kinds of similarity arithmetics respectively different in parameter; and a third means that checks to which group of the partial data of the results of the similarity arithmetics, among the groups of the partial data of the results of the classification accumulated by the second means, a specific group of data among the plurality of groups of the data inputted by the first means belongs and displays the results of the check thus obtained in a visually recognizable form. <P>COPYRIGHT: (C)2004,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a cluster analysis system for classifying a large number of data according to the degree of similarity, and in particular, has a property that the classification result differs depending on the similarity calculation expression or the parameter value at the time of classification. The present invention relates to a cluster analysis system suitable for classifying held data.

[0002]

2. Description of the Related Art Conventionally, there is a cluster analysis system having a function of inputting a plurality of classification target data and converting them into partial data, that is, clustering according to the mutual similarity. In this cluster analysis system, the contents of the obtained cluster, that is, which data belongs to which cluster, depends on various calculation methods applied when clustering or initial parameters set before execution for each calculation method. It is known that there is a change. Of the changing clusters, there are often mixed clusters that are judged to be inappropriate according to the original nature of the original data itself. Therefore, the analyst must select a calculation method and an initial parameter to be set before the execution so that the classification that is considered appropriate can be executed by judging from the original property of the data itself which is the source of the classification.

However, it is necessary to select a proper calculation method and an initial parameter to be set before the execution, which makes it possible to carry out a classification which seems to be appropriate by judging from the original property of the data which is the source of the classification. Has become very difficult. This is because those calculation methods themselves have not been developed as methods that are guaranteed to always perform appropriate classification on specific data, and there are cases where empirical superior classification can be performed over other methods. This is because it is known that there is a certain degree, and there may be a case where a significantly different classification result is output due to the influence of the subtle nature, error, or tendency of the data that is actually input.

[0004]

On the other hand, when performing classification with a specific data group as an input, different initial parameters of different calculation methods are selected from random or empirical judgment, and this is selected by trial and error. Iteratively repeats the process until it obtains a classification result that is considered appropriate based on the original nature of the input data.

However, as a result of relatively short trial and error, an appropriate classification result is not always obtained, and trial and error must be stopped at a stage where an appropriate classification result cannot be obtained due to physical constraints such as time. Often it does not. In this case, all the classification results obtained up to that point are treated as inappropriate classification results, not only the expected result cannot be obtained, but also a great amount of work time is wasted. Therefore, even among the classification results obtained by short trial and error, only some of the clusters included in each of them that have appropriate classification results are extracted, and at least those clusters are classified appropriately. It is necessary to improve work efficiency by assuming that the result is obtained.

However, heretofore, no consideration has been given to means for efficiently extracting a part of clusters having an appropriate classification result.

An object of the present invention is to provide a cluster analysis system capable of efficiently extracting a part of clusters having an appropriate classification result from the classification results obtained by short trial and error. .

[0008]

In order to achieve the above object, the present invention has a structure in which a plurality of trial-and-error classification results are displayed for easy visual comparison. And That is, the first means for inputting a plurality of sets of data to be classified, and the classification results obtained by performing a plurality of types of similarity calculation with different parameters for each of the plurality of sets of data input by the first means. Second means for accumulating a partial data group of
The data of a specific set of the plurality of sets of data inputted by the means belong to any kind of the partial data group of the similarity calculation result in the partial data group of the classification result accumulated in the second means. And a third means for displaying the result in a visually recognizable format. Further, the third means displays a plurality of sets of data input by the first means on one axis of a two-dimensional table displayed on the screen in a format that can be individually distinguished,
The partial data group of the classification result is displayed on the other axis in a form that can be individually distinguished, and each set of data belongs to the corresponding partial data group on each grid point whose display position on each axis is two-dimensional coordinates. It is characterized in that it is provided with a means for displaying in a format in which it can be visually recognized. A format in which the third means can visually distinguish which kind of similarity calculation result is obtained when displaying the partial data group of the classification result on one axis of the two-dimensional table. Is further provided with means for displaying. Further, it is characterized by further comprising means for selecting a set of data to be collated in the third means and an execution unit of similarity calculation.

With such a configuration, the classification result is significantly different under the influence of the delicate property of the input data, the error, the tendency, etc. under the specific calculation method for executing the classification and its initial parameters. For each case, by performing classification using multiple calculation methods and initial parameters and comparing the results,
Among the clusters included in each of the execution results, it is possible to efficiently extract a part of the clusters having an appropriate classification result, which can contribute to reducing waste of analysis work time.

[0010]

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a system configuration diagram showing an embodiment of a cluster analysis system according to the present invention. In FIG. 1, 101 is a display device having a function of drawing characters and figures, 102 is a mouse device having a function of indicating one point on the screen of the display device 101, and 103 and 104 are numerical and character data. It is an auxiliary storage device that has the function of reading and writing existing files. The auxiliary storage device 103 stores the data to be classified, and the auxiliary storage device 104 stores the cluster data of the classification result. Reference numeral 105 denotes an arithmetic unit having a function of reading out numerical values and characters from the files on the auxiliary storage devices 103 and 104 and a function of instructing to draw a character or a figure at an arbitrary position on the screen of the display device 101. The arithmetic unit 105 performs a plurality of types of similarity arithmetic operations with different parameters for a plurality of sets of data to be classified, which are input from the attached input means and are stored in the auxiliary storage unit 103, and a portion of the arithmetic operation result. The data group is stored in the auxiliary storage device 104. In addition,
A publicly known one can be used for the similarity calculation, and thus detailed description thereof will be omitted.

FIG. 2 shows the structure of the input data file to be classified, which is stored in the auxiliary storage device 103. In FIG. 2, 201 is a data ID unique in this file. 202 is a unique data name in this file, and here, for 10 sets of classification target data, Da
The names ta1 to Data9 are given. 203-20
5 is an element of the classification target data of each set. 3 in this example
A set of data is composed of one numerical value.

FIG. 3 shows the structure of a cluster file stored in the auxiliary storage device 104. In the present embodiment, a plurality of types of similarity calculation with different parameters are executed a plurality of times, and the cluster data group (partial data group) generated in each execution unit is written in one file to perform a single operation. Although it is a method of inputting all of them, there is also a method of saving the execution results of each classification in separate files and reading them all together or in order. 301 is a classification execution unit ID that is unique in this file. 302 is a classification execution unit name unique in this file. 303 is an ID of the cluster data group generated as a result of execution of each execution unit, which is unique in this file. Three
Reference numeral 04 is a cluster name corresponding to the unique cluster ID 303 in this file. 305 is a data ID of the partial data belonging to each cluster data group.

FIG. 4 is a diagram showing an example of the structure of a menu displayed on the display device 101, and 401 is a menu for listing the data names of selectable data sets.
A menu 402 lists the classification execution unit names of the selectable classification execution units. 403 is an OK button for instructing the end of menu selection.

FIG. 5 is a diagram showing a configuration example of the cluster comparison screen. In FIG. 5, 501 is the menu 40 of FIG.
1 is an area for displaying a list of data names of the data set selected above. Reference numeral 502 is an area for displaying a list of classification execution unit names of the classification execution units selected on the menu 402 in FIG. 503 is a list display area 50 of classification execution unit names
This is an area for displaying a list of cluster names of clusters belonging to the classification execution unit displayed in 2. Reference numeral 504 is a cell area for displaying whether or not the data having the specific data name belongs to the cluster having the specific cluster name. When it is found that the data having the data name in the data name list display area 501 belongs to the cluster having any of the cluster names in the cluster name list display area 503, the name display line of the data and the cluster name The cell area at the position intersecting the display column is filled with a specific color (for example, black).

FIG. 6 is a flow chart showing details of the processing of this embodiment. The details of the processing of this embodiment will be described below with reference to this flowchart. First, in step 601, the arithmetic unit 105 causes the auxiliary storage device 10 to operate.
The classification target data (FIG. 2) of the input data file is read from No. 3. Next, in step 602, a plurality of types of similarity calculation with different parameters for the classification target data are executed a plurality of times, and the cluster data file in which the classification target data is classified as shown in FIG. To store. Next, in step 603, the arithmetic device 105 reads the cluster data file from the auxiliary storage device 104. Next, in step 604,
The arithmetic unit 105 has the data ID 201 and the data name 2
02, and creates the list.

Next, in step 605, the arithmetic unit 105 refers to the classification execution unit ID 301 and the classification execution unit name 302 to generate a list thereof. Next, in step 606, step 604 and step 60
Based on the list generated in step 5, a list of data names and classification execution unit names is displayed on the screen of the display device 101 in the format shown in FIG. Next, in step 607, the mouse device 102 designates one or more areas of the items of the menu 401 and the menu 402 to select a set of classification target data and a classification execution unit name of cluster data. . In the example of FIG. 4, as shown by the shaded display, three data of Data1, Data2 and Data4 and two classification execution unit names of Run2 and Run3 are selected. Next, in step 608, the area of the OK button 403 is instructed by the mouse device 102, thereby instructing the arithmetic device 105 that the menu selection is completed.

Next, in step 609, the data names of all the data selected in step 607 are displayed in the area 501 of FIG. In the example of FIG. 5, the menu 40 of FIG.
Assuming that 10 sets of data are selected in 1, the area 5
In 01, 10 sets of data names of Data1 to Data9 are displayed in a list as rows having the same height in the vertical direction. Next, in step 610, the cluster names of all clusters belonging to all classification execution units selected in step 607 are shown in FIG.
Is displayed as a continuous area for each classification execution unit. In the example of FIG. 5, cluster names Cluster11 to Cluster33 are displayed as a list in the area 503 as columns having the same width in the horizontal direction. Then, the execution unit names of all classification execution units are displayed in the area 502. In the example of FIG. 5, the cluster name display area 503 of the cluster included in each execution unit
Run 1 to Run 3 classification run unit names are displayed as columns with the same width as the total column width.

Next, in step 611, for each of all the data selected in the menu 401, the process of collating which cluster data group each data belongs to is repeated. Matching result, for example data Data1
Is found to belong to the cluster data group Cluster1 of the execution unit Run1, fill the cell area at the position where the display row of the data name Data1 and the display column of the cluster data name Cluster1 intersect with a specific color (for example, black) . In this case, if each piece of data belongs to a plurality of different cluster data groups in the cluster data file shown in FIG. 3, the corresponding area for each of the cluster groups is filled. When the process of step 611 is executed for all the data, the user visually refers to the screen shown in FIG. 5 and visually determines whether or not each cluster is appropriate.

The example of FIG. 5 shows that, as a result of classifying nine sets of data three times, each group is classified into three cluster groups. For example, three sets of data of Data1, Data4, and Data7 are classified into the same cluster data group in the classification execution unit Run1, but other classification execution units Run
2 and Run 3 are classified into different cluster data groups.
On the other hand, Data2, Data5, and Data8 are classified into the same cluster data group in the classification execution units Run1 and Run2. Therefore, since Data2, Data5, and Data8 are classified into the same cluster data group by the two types of classification execution units, it can be determined that Cluster12 or Cluster22 is a relatively appropriate cluster.

[0020]

As described above, according to the present invention, the classification is performed under the influence of the delicate property of the input data, the error, the tendency, etc. under the specific calculation method for executing the classification and its initial parameters. For cases where the results are very different, classification is performed using multiple types of calculation methods and initial parameters, and the results are compared to find the appropriate classification result among the clusters included in each of the execution results. It becomes possible to efficiently extract a part of the clusters that have become obsolete, which can contribute to reducing waste of analysis work time.

[Brief description of drawings]

FIG. 1 is a system configuration diagram of an embodiment of the present invention.

FIG. 2 is a configuration diagram of input data.

FIG. 3 is a configuration diagram of cluster data.

FIG. 4 is a diagram showing an example of a menu screen for selecting classification target data and classification execution units.

FIG. 5 is a diagram showing an example of a cluster comparison screen.

FIG. 6 is a flowchart showing the overall processing.

[Explanation of symbols]

101 ... Display device, 102 ... Mouse device, 10
3, 104 ... Auxiliary storage device, 105 ... Arithmetic device.

Continued front page    (72) Inventor Nobuyuki Yamamoto             6-81 Onoe-cho, Naka-ku, Yokohama-shi, Kanagawa             Hitachi Software Engineering Stock Association             In-house F-term (reference) 5E501 AA01 AC18 BA03 FA01 FA24                       FA47

Claims (4)

[Claims] 1. A first method for inputting a plurality of sets of data to be classified
Means, and second means for accumulating a partial data group of a classification result classified by a plurality of types of similarity operation with different parameters for each of the plurality of sets of data input by the first means, The data of a specific set of the plurality of sets of data input by the first means is the second data.
Of the partial data group of the classification result accumulated in the means for checking which kind of partial data group of the similarity calculation result belongs, and displaying the result in a visually recognizable format. A cluster analysis system comprising: 2. The second means for displaying on a screen 2
A plurality of sets of data input by the first means are displayed on one axis of the dimension table in a format that can be individually distinguished, and a partial data group of classification results is displayed on the other axis in a format that can be individually distinguished. , Means for displaying on a grid point having a display position on each axis as a two-dimensional coordinate in a format capable of visually recognizing whether or not each set of data belongs to a corresponding partial data group. The cluster analysis system according to claim 1, which is characterized in that. 3. When displaying the partial data group of the classification result on one axis of the two-dimensional table, the third means visually indicates which kind of similarity calculation result is obtained. The cluster analysis system according to claim 1 or 2, further comprising means for displaying in a distinguishable format. 4. The method according to claim 1, further comprising means for selecting a set of data to be collated by the third means and an execution unit of similarity calculation. Cluster analysis system.