Features Handling by Conformal Predictors. / Yang, Meng.

2015. 143 p.

Research output: ThesisDoctoral Thesis

Unpublished

Documents

Abstract

Unlike many conventional machine learning methods, conformal predictors
allow to supply individual predictions with valid measurement of confidence.
In this thesis we adapt conformal predictors to address three common problems
related to feature handling.
First of all, we consider the problem of feature selection in the context
of conformal predictors. The main idea of our method is to use confidence
measures as an indicator of usefulness of di↵erent feature subsets.
The second one is the problem of how to utilize the additional information
which is only available in training set. Recently, Vapnik proposed a
novel learning paradigm to incorporate additional information within SVM
algorithm. Inspired by Vapnik’s method, we propose an approach to deal
with additional information by conformal predictors.
The last problem is classification using features with missing information.
Conventionally, missing information is dealt with in pre-processing step, either
by ignoring it or imputing it. We suggest a method which embeds the
processing of missing information within conformal predictors.
Experiments have been carried out to evaluate the proposed methods
using public datasets. Results demonstrate the e↵ectiveness of these methods
for feature handling.
Original languageEnglish
QualificationPh.D.
Awarding Institution
Supervisors/Advisors
Award date1 Apr 2015
Publication statusUnpublished - 2015
This open access research output is licenced under a Creative Commons Attribution-NonCommercial-NoDerivs 3.0 Unported License.

ID: 24168436