Research On Computation Of Uncertain Time Series Similarity

Time series is a sequence of data that are ranked according to chronological order.Time series has been widely applied in various practical applications such as speech recognition,astronomy,medicine,machine learning and pattern recognition.For time series processing and analysis,time series similarity calculation is a primary issue and some approaches have been proposed(e.g.,Minkowski distance,edit distance and DTW distance).In real-world applications,information is often uncertain.As a result,modeling and handling uncertain time series are receiving more attention from both academic community and industrial community.Being different from tranditinal time series,the research on calculating similarity of uncertain time series is still in its infancy there are few approaches for the similarity of uncertain time series(e.g.,PROUD distance,MUNISH distance and DUST distance).In this thesis,the methods of similarity computation for uncertain time series is investigated.First,we aim at solving the problem of similarity computation of massive discrete uncertain time series.By analyzing,the concept of discrete expected distance is introduced to solve the computation of discrete uncertain time series data.On the basis of the similairity calculation methods of traditional deterministic time series,MapReduce framework for large scale data computing is introduced and then a new similarity calculation method named MR-FastDTW algorithm is proposed,which can be used for maasive discrete uncertain time series.When calculating the similarity of uncertain time series,the MRFastDTW algorithm divides the calculation matrix of the recursive return phase into multiple submatrices and then calculates the submatrix with the MapReduce framework.Finally,the path of the submatrix is summed up and the merging path is obtained.Second,A CRV-EV algorithm is proposed to calculate the similarity of uncertain time series to solve the problem of similarity calculation of continuous uncertain time series.According to different error distribution functions,a specific formula for different distributions is developed.To accelerate the execution efficiency of the proposed CRV-EV algorithm,an accelerated computation method is given as well.The experimental results show that the proposed MR-FastDTW algorithm has good computational accuracy and low time complexity,and the proposed CRV-EV algorithm has good accuracy and the acceleration method of CRV-EV(hash)reduces the time consumption.