Learning stable concepts in a changing world

Concept drift due to hidden changes in context complicates learning in many domains including nancial prediction, medical diagnosis , and network performance. Existing machine learning approaches to this problem use an incremental learning, on-line paradigm. Batch, oo-line learners tend to be ineeective in domains with hidden changes in context as they assume that the training set is homogeneous. We present an oo-line method for identifying hidden context. This method uses an existing batch learner to identify likely context boundaries then performs a form of clustering called contextual clustering. The resulting data sets can then be used to produce context speciic, locally stable concepts. The method is evaluated in a simple domain with hidden changes in context.