Providing theoretical guarantees to the detection of concept drift in data streams
| Grant number: | 14/13323-5 |
| Support type: | Regular Research Grants |
| Duration: | October 01, 2014 - September 30, 2016 |
| Field of knowledge: | Physical Sciences and Mathematics - Computer Science |
| Principal Investigator: | Rodrigo Fernandes de Mello |
| Grantee: | Rodrigo Fernandes de Mello |
| Home Institution: | Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil |
| Assoc. researchers: | Ricardo Araújo Rios |
Abstract
Several industrial, scientific and commercial processes produce data continuously over time, in large volumes and at high frequencies, called data streams. In the Machine Learning area, researchers have been modeling and analyzing the behaviour of these streams in attempt to understand the different phenomena. In many scenarios, data streams have their behaviour changed over time, which is called concept drift. The detection of such changes is an important task as it increases the information of the phenomenon under study. Several works accomplish such task assuming that (i) data is previously labeled and/or (ii) there is not temporal relations among observations produced -- those assumptions are hard to assert for data streams. This project proposes a method for detecting concept drift on unlabeled data streams assuming temporal dependencies among observations. Initially, consecutive windows of data will be extracted of the same stream. Then every window will be decomposed on the deterministic and stochastic components using the Empirical Mode Decomposition method and the Recurrence Quantification Analysis. Next, these components are modeled with the purpose of removing the dependencies among observations. In the next step, an stable clustering algorithm is applied on data, producing dendrograms. Finally, these models are then compared using the Gromov-Hausdorff distance which results in order to indicate concept drift. (AU)