Efficient Time Series Mining Using Fractal Representation

As time series mining has become more prevalent and attracted much research interest, recent goals and efforts have been shifted toward scalability issue. One of the successful solutions is finding suitable representation of the data via dimensionality reduction. In this work, we introduce a novel fractal representation for time series data, which uses merely three real values to represent any time series. One of its unique advantages is that this representation does capture the pattern and self similarity within, thus representing the time series´ global structure using only a few values, producing incredible speedup in similarity search. We demonstrate the utility of our proposed method on clustering problems, which are shown to be scalable to much larger datasets.