Authors:
Artur Ferreira
1
and
Mario Figueiredo
2
Affiliations:
1
Instituto Superior de Engenharia de Lisboa and Instituto Superior Técnico, Portugal
;
2
Instituto Superior Técnico and Instituto de Telecomunicações, Portugal
Keyword(s):
Classification, Feature Discretization, Mutual Information, Quantization, Supervised Learning.
Related
Ontology
Subjects/Areas/Topics:
Classification
;
Feature Selection and Extraction
;
Pattern Recognition
;
Theory and Methods
Abstract:
In many learning problems, feature discretization (FD) techniques yield compact data representations, which
often lead to shorter training time and higher classification accuracy. In this paper, we propose two new FD
techniques. The first method is based on the classical Linde-Buzo-Gray quantization algorithm, guided by a
relevance criterion, and is able to work in unsupervised, supervised, or semi-supervised scenarios, depending
on the adopted measure of relevance. The second method is a supervised technique based on the maximization
of the mutual information between each discrete feature and the class label. For both methods, our experiments
on standard benchmark datasets show their ability to scale up to high-dimensional data, attaining in many cases
better accuracy than other FD approaches, while using fewer discretization intervals.