An ensemble solution for multivariate time series clustering

Abstract

Technologies such as Big Data and IoT have shown the need for intelligent unsupervised processing of Multivariate Time Series (MTS), MTS clustering among them. The challenges in MTS clustering includes not only the selection of the algorithm but also the MTS representation and the similarity measurement among the instances. This study proposes an ensemble of MTS clustering methods that merges different MTS representations and distance functions, aggregating them to obtain a similarity measurement. Furthermore, a proposal for prior knowledge representation is propose to balance the aggregation of the distances. The final clustering is performed either using k-means or hierarchical clustering. The experimentation set up includes the implementation of the ensemble with either 4 or 5 different methods, including an MTS extension of k-Shape. The results show that the ensemble is biased towards the best methods, which helps the clustering practitioner in the selection of the most suitable prototypes. Moreover, the evaluation of the ensemble with the number of clusters set to the number of labels shows that metrics, such as the sensitivity and specificity, must drive the rule of the elbow; alternatively, this value represents the most interesting prior knowledge bit in MTS clustering. Further work includes the study of digital markers to compare MTS representations and distance functions and the use of external metrics to balance the aggregation of the methods.