Reacting to concept changes using a committee of experts
Résumé
We present a general framework to deal with concept changes in online machine learning. Our approach relies on a committee of experts where each expert is trained on a different history size. The experts change constantly based on their performance which creates a dynamic committee that can adapt to a large variety of concept changes. The experiments, based on synthetic data, simulate abrupt and global concept changes. We test different methods to weight the experts and to combine their predictions. The experimental results show that our ensemble algorithm learns a concept change better than when a single expert learns each concept separately. Different types of experts as neural networks, SVMs and others may coexist in the committee in order to increase the diversity and improve the overall performance. We show how our algorithm is robust to the existence of potentially “bad” types of experts in the committee.