Adaptive online incremental learning for evolving data streams

Recent years have witnessed growing interests in online incremental learning. However, there are three major challenges this area. The first difficulty is concept drift, that is, the probability distribution streaming data would change as arrives. second catastrophic forgetting, forgetting what we learned before when learning new knowledge. last one often ignore of latent representation. Only good representation can improve prediction accuracy model. Our research builds on observation and attempts to overcome these difficulties. To end, propose an Adaptive Online Incremental Learning for evolving streams (AOIL). We use auto-encoder with memory module, hand, obtained features input, other according reconstruction loss could successfully detect existence drift trigger update mechanism, adjust model parameters time. In addition, divide features, which derived from activation hidden layers, into two parts, used extract common private respectively. By means approach, learn coming instances, but do not forget past (shared features), reduces occurrence forgetting. At same time, get fusion feature vector self-attention mechanism effectively fuse extracted further improved