Time Series Representation and Clustering with Directional Distance-Based Features (DDBF)

Time series clustering has many applications in numerous domains. For example, time series clustering can be applied to medical time series data such as EEG, ECG, FMRI, Laparoscopic videos and etc. The high dimensionality of time series data and the challenges about choosing a suitable similarity function for time series with different lengths, convert the time series clustering to a problem with high complexity. So the time series clustering is considered in this paper. The time series clustering performance is highly influenced by its representation schema. In this paper, new feature for time series representation and clustering is introduced which discretize time series data in a different manner. This feature is directional distance based feature (DDBF) and can be used for discretization and representation of univariate and multivariate time series. In this research, symbolic quantized time series are considered as documents with bag of words model. The combined features for document clustering are N-grams of DDBF symbols for l < N < 3. This approach is tested on several data sets. Results of these experiments show that this approach has better performance in comparison with some other state of the art time series symbolic representation and clustering methods.