Deep Architectures on Drifting Concepts: A Simple Approach

Abstract

Many real-world problems may vary over time. These non stationary problems have been widely studied in the literature, often called drifting concepts problems. Recently, deep architectures have drawn a growing attention, given that they can easily model functions that are hard to approximate with shallow ones and an effective way of training them have been discovered. In this work we adapt a deep architecture to problems that present concept drift. To this end we show a way of combining them with a widely known drifting concept technique, the Streaming Ensemble Algorithm. We evaluate the new method using appropriate drifting problems and compare its performance with a more traditional approach. The results obtained are promising and show that the proposed variation is effective at combining the expressive power of a deep architecture with the adaptability of SEA.