Advanced Decision Trees - Data Mining with Decision Trees: Theory and Applications - page 175

[ John (1996) ] and others ( [ Utgoff (1989a) ] ; [ Lubinsky (1993) ] ; [ Sethi and
Yoo (1994) ] ).
Growing of oblique decision trees was first proposed as a linear
combination extension to the CART algorithm. This extension is known
as the CART-LC [ Biermann et al . (1982) ] . oblique classifier 1 (OC1) is an
inducer of oblique decision trees designed for training sets with numeric
instances [ Murthy et al . (1994) ] . OC1 builds the oblique hyperplanes by
using a linear combinations of one or more numeric attributes at each
internal node; these trees then partition the space of examples with both
oblique and axis-parallel hyperplanes.

11.7 Incremental Learning of Decision Trees

To reflect new data that has become available, most decision trees inducers
must be rebuilt from scratch. This is time-consuming and expensive and
several researchers have addressed the issue of updating decision trees
incrementally. Utgoff [ Utgoff (1989b); Utgoff (1997) ] , for example, presents
several methods for incrementally updating decision trees while Crawford
(1989) describes an extension to the CART algorithm that is capable of
inducing incremental changes.

11.7.1 The Motives for Incremental Learning

In the ever-changing world of information technology there are two
fundamental problems to be addressed:

• Vast quantities of digital data continue to grow at staggering rates.

In organizations such as e-commerce sites, large retailers and telecom-
munication corporations, data increases of gigabytes per day are not
uncommon. While this data could be extremely valuable to these
organizations, the tremendous volume makes it virtually impossible to
extract useful information. This is due to the fact that KDD systems in
general, and traditional data mining algorithms in particular, are limited
by several crippling factors. These factors, referred to as computational
resources, are the size of the sample to be processed, running time
and memory. As a result, most of the available data is unused which
leads to underfitting. While there is enough data to model a compound
phenomenon, there is no capability for fully utilizing this data and
unsatisfactorily simple models are produced.