Two-stage time-series clustering approach under reducing time cost requirement

Abstract

Clustering is an essential task of unsupervised learning, which is valuable as a specific data mining tool and as an auxiliary stage of numerous highly demanded tasks, including recognizing structures, tuning of forecast parameters, detecting anomalies, and others. Significantly data-driven, especially of specific data such as time-series considered here, as well as with an impressive growth of the volume data, the computational cost becomes a vital critical issue. In the research presented, the authors developed a two-step approach to clustering based on the split of a massive dataset into two unequal parts under the control of the clusterability metric through the instance-based and feature-based combination of time-series clustering. The conducted experimental study on the well-known test data set confirmed the competitiveness of the proposed method under the conditions of the requirement to reduce time costs.