A cellular automata approach to detecting concept drift and dealing with noise

Learning of drifting concepts has recently received great attention. This is mostly due to its capability of modeling natural phenomena more realistically, provided that it is done effectively. So far incremental and window based ensemble learning have been widely used as the two most effective methods for tracking concept changes. In windowing methods, the most recent samples are considered relevant as training examples to be fed to the underlying "base" learning algorithm, as well as for evaluating its accuracy. Here we present a cellular automata- (CA) based approach which improves the current widow- based relevance criterion by adding neighborhood distance as another relevance measure for data samples. Emergence of new samples in the stream affects their "nearby" samples\´ chance of being considered relevant for the learning task. Experiments show that a good choice of local rules for CA can reduce the concept convergence time considerably and increase model robustness to noise; thus presenting a more accurate stream-learning.