A classifier fitness measure based on Bayesian likelihoods: An approach to the problem of learning from positives only

Andrew Skabar, Anthony Maeder, Binh Phanr

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

A classifier evaluation function based on Bayesian likelihoods of necessity and sufficiency is defined. This function can be used to measure the performance of an arbitrary classifier on a set of examples consisting of labeled positives together with a corpus of unlabeled data. A neural network system has been implemented in which the evaluation function is used as a heuristic to guide search through the space of network weight configurations. Results are presented from testing the system on three artificial datasets. The results are comparable to those that can be obtained using back-propagation, despite the fact that the latter method requires labeled counter-examples.

Original languageEnglish
Title of host publicationPRICAI 2000, Topics in Artificial Intelligence - 6th Pacific Rim International Conference on Artificial Intelligence, Proceedings
Pages177-187
Number of pages11
Publication statusPublished - 2000
Externally publishedYes
Event6th Pacific Rim International Conference on Artificial Intelligence, PRICAI 2000 - Melbourne, VIC, Australia
Duration: 28 Aug 20001 Sep 2000

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume1886 LNAI
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference6th Pacific Rim International Conference on Artificial Intelligence, PRICAI 2000
CountryAustralia
CityMelbourne, VIC
Period28/08/001/09/00

Fingerprint Dive into the research topics of 'A classifier fitness measure based on Bayesian likelihoods: An approach to the problem of learning from positives only'. Together they form a unique fingerprint.

Cite this