Entropy-based Concept Shift Detection

When monitoring sensory data (e.g., from a wearable device) the context oftentimes changes abruptly: people move from one situation (e.g., working quietly in their office) to another (e.g., being interrupted by one's manager). These context changes can be treated like concept shifts, since the underlying data generator (the concept) changes while moving from one context situation to another. We present an entropy based measure for data streams that is suitable to detect concept shifts in a reliable, noise-resistant, fast, and computationally efficient way. We assess the entropy measure under different concept shift conditions. To support our claims we illustrate the concept shift behavior of the stream entropy. We also present a simple algorithm control approach to show how useful and reliable the information obtained by the entropy measure is compared to a ensemble learner as well as an experimentally inferred upper limit. Our analysis is based on three large synthetic data sets representing real, virtual, and a combination of both concept drifts under different noise conditions (up to 50%). Last but not least, we demonstrate the usefulness of the entropy based measure context switch indication in a real world application in the context-awareness/wearable computing domain.