Method of generating classification model and recording medium

Details Method of generating classification model and recording medium Method of generating classification model and recording medium

1999-07-19

2003-01-21

Werner, Brian (Department: 2621)

Image analysis

Pattern recognition

Classification

Reexamination Certificate

active

06510245

ABSTRACT:

1. TECHNICAL FIELD
The present invention relates to a classification model generating method for pattern recognition or situation classification, which is used for recognition of, e.g., speech or image patterns or classification of situations, and a recording medium on which a program for making a computer execute the classification model generating method is recorded.
2. BACKGROUND ART
Systems used in the field of process control and the like are required to perform situation classification to discriminate whether the current situation is an abnormal situation or demands a predetermined operation. Situation classification for abnormality determination or operation decision can be regarded as a problem for classifying situations by classifying them into abnormal and normal situations or operations A and B in a feature space defined by feature amounts (to be referred to as variates hereinafter) used for situation classification.
As a conventional method of implementing situation classification, a discriminant analysis method is known. According to the discriminant analysis method, when there are classes characterized by a plurality of types of variates, a specific class to which the situation to be classified belongs is discriminated on the basis of data belonging to the respective classes. This method is generally based on statistical techniques.
Assume that a class that has achieved a given object is defined as class A, and a class that has not achieved the object is defined as class B, and that a plurality of data characterized by variates x
1
, x
2
, . . . , xn (e.g., the numbers of times of visits to customers, telephone charges, and the numerical values obtained by quantifying enthusiasm) have been obtained for the respective classes. In this case, the discriminant analysis method uses a discrimination function Y that assigns weights to the respective variates to clarify the difference between classes A and B.
Y=a
1
×1+
a
2
×2+ . . . +
an×n
  (1)
where a
1
, a
2
, . . . , an are the weights for the respective variates. Note that equation (1) is written, as an example of discrimination functions, for a case in which the discrimination function Y is linear (variance-covariance matrixes for the respective classes are equal).
FIG. 21
shows how the discrimination function Y is determined when the space of class A as a set of data Da and the space of class B as a set of data Db are present in the two-dimensional feature space defined by the variates x
1
and x
2
. With this function, when a situation in which Y≧0 occurs, it can be determined that the situation belongs to class A. If a situation in which Y<0 occurs, it can be determined that the situation belongs to class B.
Another method of implementing situation classification, a pattern recognition method of recognizing an object on the basis of a form, mode, or pattern which characterizes the object is known. As this pattern recognition method, a method using a neural network has been proposed (Gail A. Carpenter and Stephen Grossberg, “PATTERN RECOGNITION BY SELF-ORGANIZING NEURAL NETWORKS”, A Bradford Book, 1991). As another pattern recognition method, a method using an RCE (Restricted Coulomb Energy) network has been proposed (D. L. Reilly, L. N. Cooper and C. Elbaum, “Self Organizing Pattern Class Separator and Identifier”, U.S. Pat. No. 4,326,259. Awarded Apr. 20, 1982).
A neural network is an attempt to implement a parallel information processing mechanism based on neurons as in the brain of a creature in terms of engineering. When a neural network is to be used for situation classification, variates included in several typical situations and discrimination results to be output from the neural network in accordance with the variates must be supplied to the neural network to make it learn to obtain desired discrimination results. As a method of making the neural network learn, a back propagation method is generally used.
An RCE network is used to classify a feature space by approximating classes occupying a linearly inseparable multi-dimensional space with a plurality of basic graphic patterns (e.g., multi-dimensional hyperspheres). In the case shown in
FIG. 22
, the spaces of linearly inseparable classes A and B are respectively approximated with basic graphic patterns Ca and Cb to classify the two-dimensional feature space defined by variates x
1
and x
2
.
3. DISCLOSURE OF INVENTION
[Problem to be Solved by the Invention]
According to the discrimination analysis method, however, when the spaces of the respective classes cannot be linearly separated, a discrimination function must be approximated with a higher order polynomial. If, therefore, many types of variates are required, and the space of each class is complicated, a discrimination function is difficult to derive.
In the method using the neural network, the learning speed of the neural network is low (in general, about 100 to 1,000 learning processes are required; it takes about one week in some cases). In addition, it is difficult to determine an optimal network configuration for classification. Furthermore, since it takes much time to perform classification processing, i.e., classifying situations on the basis of variates characterizing the situations, an expensive semiconductor chip is required to increase the processing speed.
In the method using the RCE network, the basic graphic patterns Ca and Cb respectively centered on the data Da and Db belonging to classes A and B are generated to have sizes that do not interfere with the remaining classes. However, the data Da and Db serving as the centers of the basic graphic patterns do not always exist at positions where the spaces of classes A and B can be properly approximated. For this reason, a situation that should not be included in a given class may be determined as a situation belonging to the class. That is, a recognition error may occur. For example, in the case shown in
FIG. 22
, the basic graphic patterns Cb properly approximate the space of class B, whereas some basic graphic patterns Ca protrude from the space of class A. In this case, therefore, a situation that should not be included in class A may be determined as a situation belonging to class A. In addition, according to the method using the RCE network, when a few data are remote from the data groups of the respective classes, classification is affected by those data.
The present invention has been made to solve the above problems, and has as its object to provide a classification model generating method in which the learning speed and the classification processing speed are high, and the spaces of classes can be properly approximated even if the spaces of the respective classes cannot be linearly separated, and a recording medium on which a program for making a computer execute the classification model generating method is recorded.
[Means of Solution to the Problem]
As described in claim
1
, according to the present invention, a classification model generating method of the present invention comprises the steps of, when n-dimensional data which belongs to one class in an n-dimensional feature space defined by n types of variates and whose position is specified by the variates is input, dividing the feature space into m
n
divided areas by performing m-part division for each of the variates, and determining a division number m on the basis of a statistical significance level in the division by regarding a degree of generation of a divided area containing one data as a degree following a probability distribution with respect to the division number m, setting a divided area containing n-dimensional data as a learning area belonging to the class, and associating each input data with a corresponding divided area, adding divided areas around the learning area as learning areas to expand a learning area group, and removing a learning area located on a boundary between the learning area and a divided area which is not a learning area from the learning area group to con

Affiliated with

Also associated with

Rating

Method of generating classification model and recording medium does not yet have a rating. At this time, there are no reviews or comments for this patent.

If you have personal experience with Method of generating classification model and recording medium, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method of generating classification model and recording medium will most certainly appreciate the feedback.

Rate now

Comments { 0 }

Profile ID: LFUS-PAI-O-3016526