Clustering of Time Series using Hybrid Symbolic Aggregate Approximation

Clustering of time series is one of the best-known grand challenges in time series analysis because of its application potentialities and difficulty. It is like data clustering and the task of partitioning time series into several groups based on their similarities, such that time series in a cluster are similar and they are not similar to other clusters. In the last decade, symbolic aggregate approximation (SAX), which is a high-level symbolic representation for time series, has attracted the attention of many data mining researchers. SAX enables time series analysis to be applied to sequence mining techniques. In this study, we propose a new approach for clustering time series that utilizes a moving average convergence divergence (MACD)-histogram-based SAX (MHSAX) and the k-medoids method. MHSAX is a hybrid symbolic aggregate approximation combining the SAX strings of a time series and its MACD histogram. By utilizing MHSAX, we can calculate the more accurate distance between time series compared with other approaches. This improves the affinity with the k-medoids method and improves the accuracy of clustering. We actually implemented the proposed clustering method and conducted experiments using the whole UCR Time Series Archive data sets. The experimental results show that the proposed method is superior to other state-of-the-art methods.