In this paper we present a general approach to learning in domains in which concepts change over time. We introduce a new metric, concept instability, that detects whether a concept is changing by examining changes in the concept representation formed by a supervised incremental learning algorithm. When a change is detected, we form the hypothesis that the change is due to a shift in the underlying concept. If the hypothesis is accepted, then a learning method that adapts the concept representation to the new concept is applied. We illustrate the approach in conjunction with an incremental decision tree algorithm. The approach is applicable to a wide class of learning systems because it is designed to be used in conjunction with any incremental learning algorithm that adjusts its concept representation in response to each observed positive or negative example of the target concept.