A Fast Shapelet Extraction Algorithm Based on Trend Distance
Aiming at the problem that the existing Shapelet extraction method cannot reflect the trend characteristics and the extraction result deviates slightly from the original data,an improved fast Shapelet selection algorithm was proposed.A distance calculation method considering the relative trend of time series was proposed,which could measure the similarity of time series more accurately.Secondly,the Shapelet features were combined with the ensemble network to enable the classifier to benefit from the residual linear connection and attention mechanism,which enhanced the generalization ability of the algo-rithm.Finally,controlled trials were conducted on 12 datasets.Experimental results show that the pro-posed method can obtain an average accuracy of 88.0%,which is 2.9%higher than the fast Shapelet algorithm,especially on the ChlorineConcentration dataset,and the accuracy is increased by 13.3%.In terms of acceleration rate,the method can extract faster than the original algorithm on all 10 datasets,so it can extract Shapelet in time series data more efficiently.
Shapelettrend characteristicsShapelet transformsubclass divisiontime series classification
国家科技期刊平台
NSTL
万方数据
