MANIFOLD REGULARIZATION FOR SEMI-SUPERVISED SEQUENTIAL LEARNING

The sequential data flux in many time-series applications require that only a small fraction of the data are stored for future processing. Furthermore, labels for these data are possibly sparse and they might show some biases. To support learning under such restrictive constraints, we combine manifold regularization with sequential learning under a semi-supervised learning scenario. The online learning mechanism integrates a regularization based on the data smoothness assumptions. We present a proof-of-concept for illustrative toy problems, and we apply the algorithm to a real-world sparse online classification task for music categories.