12,000

We have over 12,000 students, from over 100 countries, within one of the safest campuses in the UK

93%

93% of Lancaster students go into work or further study within six months of graduating

Home > Research > Publications & Outputs > Detecting and reacting on drifts and shifts in ...
View graph of relations

« Back

Detecting and reacting on drifts and shifts in on-line data streams with evolving fuzzy systems.

Research output: Contribution in Book/Report/ProceedingsPaper

Published

Publication date06/2009
Host publicationProceedings of the Joint 2009 International Fuzzy Systems Association World Congress and 2009 European Society of Fuzzy Logic and Technology Conference, Lisbon, Portugal, July 20-24, 2009.
EditorsJoão Paulo Carvalho, Didier Dubois, Uzay Kaymak, João Miguel da Costa Sousa
Place of publicationLisbon
PublisherIFSA
Pages931-937
Number of pages7
ISBN (Print)978-989-95079-6-8
Original languageEnglish

Conference

Conference2009 International Fuzzy Systems Association WORLD CONGRESS
CityLisbon, Portugal
Period20/07/0924/07/09

Conference

Conference2009 International Fuzzy Systems Association WORLD CONGRESS
CityLisbon, Portugal
Period20/07/0924/07/09

Abstract

In this paper, we present new approaches to handle drift and shift in on-line data streams using evolving fuzzy systems (EFS), which are characterized by the fact that their structure is not fixed and not pre-determined. When dealing with drifts and shifts in data streams one needs to take into account two major issues: a) automatic detection of, and b) automatic reaction to this. To address the first problem we propose an approach based on the concepts of age and utility of fuzzy rules/clusters. The second problem itself is composed of two sub-problems concerning the influence of the drifts and shifts on: 1) the antecedent parts (fuzzy set and rule structure) and 2) the consequent parts (parameters) of the fuzzy models. To address the latter sub-problem we propose an approach that introduces a gradual forgetting strategy in the local learning process. To address the former sub-problem we introduce two alternative methods: one that is based on the evolving density-based clustering, eClustering (used in eTS); and one that is based on the automatic adaptation of the learning rate of the evolving vector quantization approach (eVQ) (used in FLEXFIS). The paper is concluded with an empirical evaluation of the impact of the proposed approaches in (on-line) real-world data sets where drifts and shifts occur.