Graph-based Clustering for Time Series Data
Peiyu Li, Soukaïna Filali Boubrahimi, Shah Muhammad Hamdi
Abstract
Clustering is an important data mining task that consists of grouping data without a prior knowledge of classes. In the recent years, with the rise of big data and its various applications, unsupervised learning such as clustering algorithms have attracted massive interest in the data mining field. For the case of time series data, temporal changes are difficult to detect, which makes the clustering for temporal sequence data more complex than traditional vector data. In this paper, we propose a new model that generates graphs from time series data to preserve important relations between different data points. In particular, every time series data will be considered as a node, then edges will be added between nodes if the dynamic time warping distances of time series data achieve a specific threshold, then spectral clustering algorithm is applied to the generated graph. Our results shows that our new proposal time series graph representation outperforms state-of-the-art clustering algorithms.