Large Margin Dimensionality Reduction for Time Series

Dimensionality reduction techniques are widely used in time series data mining. Dimensionality reduction can not only speed up the computation but also lead to improved performance. Most available techniques implement the reduction process without supervised information. This operation can be used to de-noise the insignificance detail, or blur the discriminative information which is important for supervised learning. To solve the problem, we design a framework of dimensionality reduction method, called the Large Margin Dimensionality Reduction (LMDR), based on large margin criterion. It is shown empirically that the LMDR significantly improves the performance in terms of time series data mining.