LMDS-SEA: Upgrading the Shape Exchange Algorithm (SEA) to handle general time series classification by local matching and distance selection

Abstract

In this paper, we propose an enhanced version of the Shape Exchange Algorithm (SEA) for the purpose of general Time Series Classification (TSC). Indeed, the SEA method is very effective in quasi-periodic time series matching. However, for general time series, it has been found that this method needs some kind of enhancement to meet state-of-the-art TSC techniques. The proposed LMDS-SEA method (Local Matching with Distance Selection Shape Exchange Algorithm) is based on two ideas: a) Distance Selection as an adaptation paradigm for classification method and b) Local Matching. The distance selection contribution consists in dynamically selecting the best distance to apply as a function of the data to classify from a pool of Euclidian Distance (ED), SEA and LM-SEA (Local Matching SEA). The Local Matching (SEA) contribution consists in matching equal and corresponding sub-sequences in the two time series to match, X and Y, using the SEA method locally. The newly proposed LMDS-SEA method was applied to classification of the public UCR benchmark (UCR: General Time Series Classification/Clustering dataset of the University of California at Riverside) based on 1NN classifier. Results show significant enhancements in the classification accuracy of the proposed 1NN LMDS-SEA classifier over the 1NN SEA classifier (1NN LMDS-SEA: 64 WINs, 1NN SEA: 00 WINs and 21 Ties). Results show also that the proposed 1NN LMDS-SEA classifier is more effective than the very effective 1NN DTW classifier (1NN LMDS-SEA: 50 WINs, 1NN DTW: 28 WINs and 7 Ties). Note that a win is credited to a classifier A over a classifier B if A is more accurate than classifier B, class-wise, over the 85 classes of the UCR benchmark.