Fast algorithms for time series mining

Abstract

In this paper, we present fast algorithms on mining coevolving time series, with or with out missing values. Our algorithms could mine meaningful patterns effectively and efficiently. With those patterns, our algorithms can do forecasting, compression, and segmentation. Furthermore, we apply our algorithm to solve practical problems including occlusions in motion capture, and generating natural human motions by stitching low-effort motions. We also propose a parallel learning algorithm for LDS to fully utilize the power of multicore/multiprocessors, which will serve as corner stone of many applications and algorithms for time series.