Browse Prior Art Database

Algorithm for Concept Learning Utilizing Novel Attribute Space- Dividing Method

IP.com Disclosure Number: IPCOM000100965D
Original Publication Date: 1990-Jun-01
Included in the Prior Art Database: 2005-Mar-16
Document File: 3 page(s) / 100K

Publishing Venue

IBM

Related People

Nakasuka, S: AUTHOR [+2]

Abstract

Disclosed is an algorithm of "CONCEPT LEARNING" which utilizes a novel attribute space-dividing method. The following three techniques are invented for automatic generation of "A Binary Decision Tree" which is often utilized as a schema for concept learning: (1) automatic generation of new "USEFUL ATTRIBUTES",(2) automatic generation of "GENERALIZATION TREE", and (3) a new criterion for deciding the best decision formula.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 52% of the total text.

Algorithm for Concept Learning Utilizing Novel Attribute Space- Dividing Method

       Disclosed is an algorithm of "CONCEPT LEARNING" which
utilizes a novel attribute space-dividing method. The following three
techniques are invented for automatic generation of "A Binary
Decision Tree" which is often utilized as a schema for concept
learning: (1) automatic generation of new "USEFUL ATTRIBUTES",(2)
automatic generation of "GENERALIZATION TREE", and (3) a new
criterion for deciding the best decision formula.

      Concept learning is a well known technique in which, by
analyzing large number of data as to the relationships between a set
of attributes and the class to which the data belongs, the
generalized expression of relationships between each class and
attributes is derived.  This process is carried out mainly by
dividing the attribute space (multi-dimensional). The derived
expression is utilized for predicting which class a certain new data
belongs to.  A "BINARY DECISION TREE", such as shown in Fig. 1, is
often utilized as a scheme for attribute space division. A decision
formula is generated at each node, and by this formula, the data for
learning are divided into two groups which, respectively, formulate
another new node.  This division process is iterated and the tree is
extended until each tip node nearly consists of data belonging to
only one class.

      The conventional technique in this field can be used only under
quite ideal conditions and is of little use in the practical
application areas where empirical data are the mere source of
learning. The invention overcomes these deficiencies by employing a
new attribute space-dividing method and provides powerful concept
learning capability even in the practical application area.  The
detailed descriptions of the newly invented techniques are as
follows.

      Method for automatic generation of new USEFUL ATTRIBUTES: By
programming some basic mathematical operators (such as +, -, *, /,
etc.)  beforehand, various new attributes are defined automatically
during the learning phase in the form of m...