Random-shapelet: An algorithm for fast shapelet discovery

Abstract

Time series shapelets proposes an approach to extract subsequences most suitable to discriminate time series belonging to distinct classes.Computational complexity is the major issue with shapelets: the time required to identify interesting subsequences can be intractable for large cases. In fact, it is required to evaluate all the subsequences of all the time series of the training dataset. In the literature, improvements have been proposed to accelerate the process, but few provide a solution that dramatically reduces the time required to find a solution.We propose a random-based approach that reduces the time necessary to find a solution, in our experimentation until 3 orders of magnitude compared to the original method.Based on extensive experimentations on several data sets from the literature, we show that even with a few time available, random-shapelet algorithm is able to find very competitive shapelets.

J48SS: A Novel Decision Tree Approach for the Handling of Sequential and Time Series Data

“…The results of the experimentation of single J48SS trees are reported in Table 8, together with the accuracies originally presented in [45] (Random Shapelet). Both performances have been obtained by averaging over 100 executions of the algorithms.…”

“…. Searching such a huge space for the best shapelets in an exhaustive manner is unfeasible and, for this reason, several studies have focused on how to speed up the shapelet extraction process [40][41][42][43][44][45][46], by means, for instance, of heuristics-driven search, or random sampling of the shapelet candidates. Tipically, shapelets are used as follows.…”

“…All the following experiments have been performed on a selection of 16 UCR [10] datasets given the fact that, as witnessed by the literature, it has been broadly used in the machine learning community [14,42,43,45,46]. In addition, such a decision allows us to compare our results to the ones presented in [45]. The details concerning the dataset selection are reported in Table 5.…”

“…The performance of J48SS has then been evaluated with respect to the pre-processed datasets, and against the results presented in [45] (Random Shapelet), that has been chosen for comparison since it is relatively recent approach that also embeds the shapelet generation phase in a decision tree inducing algorithm. Moreover, the paper thorougly describes all aspects of the experimentation phase, which makes our comparison task easier.…”

“…Despite the fact that J48SS has shown an inferior classification performance on some datasets than Random Shapelet, it should be observed that the latter requires very long computation times, especially when the larger sampling values are used [45]: typically hundreds of seconds are needed in order to build a single model on the smaller datasets, and thousands of seconds (or even worse) are required on the larger ones. On the contrary, while J48SS training times tend to increase with the number of evaluations, they are comparatively much lower, as shown in Table 9.…”