A comparative study of pruned decision trees and fuzzy decision trees

Dissertant

Bin Ibrahim, Huda

Thesis advisor

Bin Said, A.

University

Al Akhawayn University

Faculty

School of Science and Engineering

Department

Computer Science

University Country

Morocco

Degree

Master

Degree Date

2000

English Abstract

Decision trees are a widely used symbolic technique for classification tasks in machine learning and they successfully compete with other models such as neural networks.

However, they have often been criticized for some of their limitations such as problems of handling continuous attributes, missing data and performance in noisy domains.

Some of these limitations have been addressed by some models such as CART and C4.5.

Yet, the resulting knowledge exhibits lower comprehensibility and over speciahzation.

These problems were in turn addressed by tree pruning techniques and, more recently, by combining fuzzy representation with decision trees.

We propose in this thesis, a comparative study of pruned decision trees and fuzzy decision trees.

The focus is twofold.

First, for crisp decision tree experiments, we use C4.5 and apply and compare five post-pruning methods to perform pruning in C4.5 [6], Second, FID3.0 is used to build fuzzy decision trees.

The fuzzy decision tree algorithm requires data with predefined fuzzy sets for the continuous-valued attributes.

We have experimented with (1) the inductive reasoning algorithm, and (2) a fuzzy clustering algorithm (FCM) to construct those fuzzy sets.

Furthermore, we have compared three methods (random selection, heuristic-based selection and cluster validity-based selection) for selecting granularity, or the number of fuzzy sets to use for each continuous-valued input variable.

Numerical experiments are performed on 12 data sets to compare the performance of the different trees.

Among the five pruning methods, “Error Based Pruning” algorithm showed better performance stability over 12 data sets.

For fuzzy decision tree, using cluster validity produces the best results for deciding on the granularity of the input fuzzy sets.

Combining this choice of granularity with the use of FCM to build fuzzy membership functions produces superior results.

Cross validation experiments and statistical significance tests reveal that this combination produces fuzzy decision trees that are superior to those of the “Error Based Pruning” algorithm in 6 out of 12 data sets, are inferior for one data set and are equivalent for the remaining 5 data sets.

Main Subjects

Information Technology and Computer Science

No. of Pages

69

Table of Contents

• APA
• MLA
• AMA

Language

English

Data Type

Arab Theses

Record ID

BIM-629787