Anticipative and Dynamic Adaptation to Concept Changes (Extended Abstract)
Ghazal Jaber, Philippe Tarroux · 2013
Recent years have witnessed the emergence of a whole new set of applications involving data streams made of pairs (xt, yt), where the “answer” or true label yt is revealed (sometimes long) after the input xt. When learning from data streams, it is necessary to rely on on-line learning with the capability to adapt to changing conditions a.k.a. concept drifts. Previous works have focused on means to detect changes and to adapt to them. Ensemble methods relying on committees of base learners have been among the most successful approaches. Most adaptive strategies operate either by passively tracking the evolving concept or by using an explicit detection mechanism of concept changes before launching an adaptation or relearning process. However, better learning strategies may take advantage of the examination of the history of past concepts in order to anticipate likely future changes or to recognize when a past concept recurs. We have developed ADACC (Anticipative Dynamic Adaptation to Concept Change), a system that uses this kind of second order learning to accelerate its adaptation to changing conditions in the environment.