A Novel Fractal Representation for Dimensionality Reduction of Large Time Series Data

摘要

Recent research has attempted to speed up time series data mining tasks which focus on dimensionality reduction, indexing, and lower bounding function, among many others. For large time series data, current dimensionality reduction techniques cannot reduce the total dimensions of time series data by a large margin without losing their global characteristics. In this paper, we introduce a novel Fractal Representation which uses merely three real values to represent a whole time series data sequence. Moreover, our proposed representation can be efficiently used under Euclidean distance. We demonstrate effectiveness and utility of our novel Fractal Representation on classification problems and our proposed method outperforms existing methods in terms of speed performance and accuracy. Our results reconfirm that this representation can effectively represent global characteristics of the data, especially in larger time series data.