JPH0540829A - Data clustering method - Google Patents

Abstract

PURPOSE:To obtain an equivalent result to that obtained with an operator's subjective clustering while seeing entire data by performing data conversion by means of a neural network as a preliminary processing of a clustering means. CONSTITUTION:A data input means 1, a data conversion neural network 2 inputting the data of the data input means 1 and a data inversion neural network 3 inputting the output of the data inversion neural network 2, are incorporated. A data comparison/learning control means 4 inputting the data of the data input means 1 as well as the output of the data inversion neural network 3 and controlling the learning of the network 2 and network 3, is also incorporated. The data conversion by means of the neural network is performed as the preliminary processing of the clustering means. Thus, the result equivalent to that obtained with the operator's subjective clustering while looking at the structure of the entire data, can be easily obtained.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a data clustering method used for information compression, pattern recognition and the like.

[0002]

2. Description of the Related Art As an example of a conventional method for clustering data, consider the LBG algorithm. This is a method of clustering the entire data by considering several set cluster center points, considering the distance between it and each data point as an error, and correcting the cluster center points so as to reduce the total error. is there.

Although this is a very effective method, it is basically a method that uses only information between two points.

As a problem with these methods, consider the case shown in FIG. 3, for example. In this figure, it is assumed that the data are on the semicircle indicated by number 9 and the semicircle indicated by 10 in the figure. Consider dividing this into two clusters.

It is very easy for a human to split this data in two. In other words, it is only divided into 9 semicircles and 10 semicircles.

However, in the conventional method of considering the distance between two points, this problem is not so easy. That is, for example, if the center of the cluster of 10 semicircles is found by a conventional method and the distance between the points 11 and 12 is considered, the point 11 should be closer to the cluster of 10 semicircles than the point 12. Become. In other words, the cluster obtained by the conventional method is completely different from the above-mentioned human-divided cluster.

[0007]

The problem to be solved by the present invention is that a conventional clustering that simply considers only the distance between two points gives humans a multifaceted view of the entire data, It is said that it will be completely different from a cluster divided subjectively, and an object of the present invention is to provide a clustering method that realizes clustering closer to human subjectivity.

[0008]

FIG. 1 is a schematic view of the constitution of the method of the present invention. When the configuration of the present invention is shown using this figure, the present invention is as follows: 1), a data input means 1 and 2), a data conversion neural network 2 which receives data from the data input means 1 and 3), 2 A data inverse transformation neural network 3 that receives the output of the data transformation neural network of 4), and 4) a data transformation neural network that receives the data from the 1 data input means and the output of the 3 data inverse transformation neural network as an input Data comparison that controls the learning of the data inversion neural network of 3
This is a data clustering method characterized by including learning control means and 4 in its configuration.

[0009]

The present invention will be described with reference to the example of FIG.

FIG. 1 is an overall schematic view of the present invention. FIG. 2 specifically shows the data conversion neural network and the data inverse conversion neural network in FIG. 1 as an example of processing the example of FIG.

As an example of the data shown in FIG. 3, consider the case where the 34 points shown as the input data in Table 1 are input to this network.

[0012]

[Table 1]

First, an appropriate initial value is set for each network, input data is given to the data conversion neural network 2, and the conversion output is calculated. Next, the converted output is given as input data to a data inverse conversion neural network, and the inverse converted output is calculated.

In this way, the data is successively supplied, and each neural network is trained by an appropriate learning algorithm until the input data and the inverse transform output of the data inverse transform neural network become the same.

Table 1 shows the conversion output and the inverse conversion output corresponding to each data after learning the neural network. In this example, the McClellan error was used as the error evaluation, and the error backpropagation algorithm was used as the learning algorithm. As is clear from Table 1, the input data and the inverse transform output almost match. FIG. 4 is a schematic diagram of the conversion output of Table 1.

The numeral 13 in FIG. 4 corresponds to the semicircle 9 in FIG. 3, 14 corresponds to the semicircle 10, point 15 corresponds to the point 11 and point 16 corresponds to the point 12.

As is clear from FIG. 4, even if the algorithm for clustering the data is the same as when humans are clustered, the result obtained by using the data thus converted is the same. give.

In this example, the input data and the converted output have the same dimension. However, depending on the structure of the data, if the converted output has a smaller dimension or vice versa, the dimension may be increased. It may be possible to express the data more naturally.

[0019]

As described above, according to the present invention, it is possible to easily obtain the same result as human being subjectively clusters from the structure of the entire data.

[Brief description of drawings]

FIG. 1 is an overall schematic diagram of a clustering method according to the present invention.

FIG. 2 is a configuration diagram of a neural network as one embodiment of the present invention.

FIG. 3 is a diagram showing an example of data for explaining the present invention.

FIG. 4 is a diagram showing the data of FIG. 3 converted according to the present invention.

[Explanation of symbols]

 1: Data input means 2: Data conversion neural network 3: Data inverse conversion neural network 4: Data comparison / learning control means 5: Clustering means 6: Data conversion network 7: Data inverse conversion network 8: Clustering means 9: Semicircular shape Data 10: semi-circular data 11: end point of data 12: end point of data 13: data corresponding to 9: data corresponding to 14:10 data corresponding to 15: 11 data corresponding to 16: 12

Claims (1)

[Claims] 1. A data inputting means, 2) a data conversion neural network which receives data from the 1 data inputting means, and 3) a data inverse which inputs the output of the 2 data conversion neural network. A conversion neural network, 4) data for controlling learning of the data conversion neural network 2 and the data inverse conversion neural network 3 with the data from the one data input means and the output of the data inverse conversion neural network 3 as input Comparison
A data clustering method characterized by including learning control means in its configuration.