COLLABORATIVE CONCEPT DRIFT DETECTION

Beverly Abadines Quon, Jean-Luc Gaudiot

Keywords: concept drift, heterogeneous ensemble, SVD, FCBF, correlation, entropy

TL;DR: Collaborative Concept Drift Detection (C2D2) combines Fast Correlated Based Filtering (FCBF) and Singular Value Decomposition (SVD) to detect concept drifts in 5 synthetic datasets.

Abstract:

Collaborative Concept Drift Detection (C2D2) combines Fast Correlated Based
Filtering (FCBF) and Singular Value Decomposition (SVD) to detect concept
drifts in 5 synthetic datasets. We compare our results against 6 diveregence tests
and introduce Performance Gain Update Cost Ratio (PGUCR). Post-hoc Tukey
HSD test confirmed that C2D2 outperformed the other tests in terms of PGUCR.
Much of C2D2’s improvement is based on its conservative signals for updates.