Characterizing Forest of Fuzzy Decision Trees Errors

Christophe Marsala 1 Maria Rifqi 1
1 MALIRE - Machine Learning and Information Retrieval
LIP6 - Laboratoire d'Informatique de Paris 6
Abstract : Forest of fuzzy decision trees are known to be a powerful machine learning tool in terms of good classification rate. The idea of this paper is to focus on the misclassified examples of this method to improve it. To this aim, the preliminary step consists in understanding what are the sources of misclassifications. Concretely, we propose to characterize misclassified examples by finding different types of errors or different groups of errors. We modify J. Forest’s algorithm of supervised segmentation to find these groups. Our general process is the following: first, we learn a forest of fuzzy decision trees on a learning dataset and we validate it on a validation dataset. We class the examples on two categories: those who are well classified by the majority of the trees and those who are badly classified by the majority of the trees. Then, this new dataset is characterized by supervised segmentation algorithm. For instance, we can obtain that the badly classified examples can be distinguished in two groups: those who can be interpreted as outliers and those who are close to the frontiers of decision. We apply our general process on several datasets of UCI machine learning repository to show its benefits.
Document type :
Conference papers
Complete list of metadatas

https://hal.archives-ouvertes.fr/hal-01286230
Contributor : Lip6 Publications <>
Submitted on : Thursday, March 10, 2016 - 3:01:37 PM
Last modification on : Thursday, March 21, 2019 - 2:16:11 PM

Identifiers

  • HAL Id : hal-01286230, version 1

Citation

Christophe Marsala, Maria Rifqi. Characterizing Forest of Fuzzy Decision Trees Errors. The 4th International Conference of the ERCIM WG on COMPUTING & STATISTICS (ERCIM'11), Dec 2011, London, United Kingdom. ⟨hal-01286230⟩

Share

Metrics

Record views

142