Method for extracting knowledge from data stream - algorithm k- nearest neighbor incremental anytime

• Grant number: 14/14174-3
• Support type: Scholarships abroad - Research Internship - Master's degree
• Effective date (Start): October 06, 2014
• Effective date (End): April 05, 2015
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computational Mathematics
• Principal researcher: Gustavo Enrique de Almeida Prado Alves Batista
• Grantee: Cristiano Inácio Lemes
• Supervisor abroad: Pedro Pereira Rodrigues
• Home Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
• Research place: Universidade do Porto (UP), Portugal
• Associated to the scholarship: 13/16081-0 - Anytime Algorithms for Data Stream Classification with Application on Insect Classification, BP.MS

Abstract

Learning from data stream is an important research area that has growing up in the last years. A many learning algorithm applied in data stream has been development. For a learning system succeeds in acquiring knowledge is necessary that it has some desirable properties: (i) update the decision model along the time; (ii) perform online learning, processing the sample as soon as it is available; (iii) perform the sample processing as fast as it possible using fixed memory, considering that in real problems exist computational constraints; (iv) build a decision model doing a single scan over the training data; (v) taking drift into account, once we can not garantee that the examples came from a stationary probability distribution. The k-nearest neighbor algorithm presents good results when applied in the stationary problems, however its performance is lower in the data stream problems. The anytime version of this algorithm presents a good solution for data stream problems where the time interval between two subsequent events is varied, however it do not take usually drift that happen on concept into account. Its incremental version fix the concept drift problem, but it is not able to determine a faster answer, desirable property in data stream problem where has constant events that happen in variable time. In this project intend research adapting methods for k-nearest neighbor algorithm that be able to have the anytime algorithm properties and be able to take the concept drift into account, thus searching for a k-nearest neighbor algorithm incremental anytime. (AU)