DFECTS: A Deep Fuzzy Ensemble Clusterer for Time Series

Time series clustering plays an important role in various fields such as anomaly detection and resource scheduling. With the increase of complexity and scale of time series datasets, many deep-learning-based time series clustering methods have emerged and achieved great success. These methods mostly use a deep neural network to extract features which are fed into classic clustering methods. The focus of such methods is to enhance the neural network to extract more meaningful features. Differently, in this paper, we propose the first ensemble clusterer, which uses time series splitting method to construct different sub-series representing various aspects of the original time series. For each group of sub-series, a clusterer is generated, which consists of an autoencoder and a fuzzy C-means algorithm. After collecting the results from each clusterer, a consensus is reached by voting. However, traditional voting algorithm widely-used in ensemble classifier does not work in ensemble clusterer because labels mismatch among clusterers. To deal with this problem, we propose an efficient label aligning algorithm. Then, a weighted voting algorithm is applied and a new mechanism with fitness matrix is utilized to further improve the performance. Extensive experiments show that our method outperforms seven representative algorithms under three metrics.