Short Isometric Shapelet Transform For Binary Time Series Classification

Time series is prevalent in today 's application.It is a special sort of high-dimensional data.And Time Series Classification is the most important task in the analysis of time series.The results of comprehensive comparison experiments carried out by a flock of outstanding researchers show some notable defects in this research area.On the one hand,those algorithms can not perform well on all the data sets,but none of them can distinguish whether a data is suitable to be dealt with by it.On the other hand,the state-of-the-art in this field have achieved a great accuracy of classification.However,the time consumption of them is too considerable to make them be used in practical application.Our work aims to improve the abovementioned defects in the field of time series classification.For the challenge of the high time complexity of the state-of-the-art,we analyze the ensemble shapelet transform algorithm which is one of the state-of-the-art and improve it from the following 4 aspects:features selection,features evaluation,construction of feature space and selection of classifier in feature space.And we fi-nally propose the short isometric shapelet transform algorithm for binary time series classification problems.Furthermore,we give strict theoretical supports under each improvement.And experimental results also show that our proposed algorithm can largely reduce the time consumption without a loss of the accuracy achieved by the state-of-the-art.For the challenge of the adaptive capacity of the algorithm to data sets,we analyze the key point of our algorithm to find the crucial dependable feature of data set,then we quantify these features so that we can judge whether a data set is suitable to be dealt with by our proposed algorithm by the quantified features.And the experimental results show that our proposed algorithm can distinguish whether a given data set is appropriate to be dealt with by itself.The main contributions of this dissertation are listed as follows:·A novel and fast algorithm is proposed to reduce the time complexity of the state of the art without a loss of accuracy.·We design a primary Fitness Index to evaluate whether our algorithm is appropri-ate to deal with a given data set.·We find theoretical evidences of feasibility of some general optimization strate-gies in time series classification problems.