LDCnet: Minimizing the cost of supervision for various types of concept drift

Abstract

Supervision cost is often overlooked when designing decision systems to cope with concept drift. The solution presented in this article utilizes low supervision while achieving similar efficiency to the state-of-the-art methods. Algorithm bases on a net of classification models which cover the whole feature space, preparing the system for every possible concept. The experiments are performed in the simulated environment with four scenarios representing different types of concept drift.