Evaluation of clustering methods for time series

The main goal of this paper is the evaluation of selected clustering methods on thirteen randomly selected data files from the UCR archive. Clustering itself is performed by using the TSclust and dtwclust packages in R. The clustering results vary significantly across all thirteen datasets. Based on the most successful procedures, it is possible to guarantee at least a third of correctly classified objects, where for half of the datasets the success rate increases to around one half. The most successful hierarchical clustering method was by far the farthest neighbor method (highest achieved similarity coefficient 0,8655), while the most successful distance measure is DTW distance (highest achieved ARI 0,6135). Second relatively successful distance is the ACF measure. For non-hierarchical methods, it is not possible to unequivocally determine a method that would significantly outweigh the quality of the results for the other investigated non-hierarchical methods. In general, the results of the non-hierarchical methods were very comparable. In terms of quality, hierarchical methods provide slightly better results (ex. dataset Yoga hierarchical methods achieved ARI 0,0266 while non-hierarchical ARI 0,0031). A significant visual difference between groups and an overall lower number of categories led to better clustering results (ex. the lowest similarity coefficient 0,2928 was found for the dataset Phoneme with 39 categories). The investigated clustering methods were not able to reliably distinguish between the more subtle differences of time series. Quality coefficients often provide rather questionable conclusions, leading to the necessity of verifying their results via closer examination of the resulting clusters (ex. DistalPhalanxOutlineCorrect similarity coefficient 0,5513 and ARI 0,0037).