Research On Uncertain Time Series Classification Based On DTW

As a special form of data,time series is widely used in economics,biology,medicine,astronomy and other fields,and is an important research object in data mining.With the emergence of uncertain time series,the analysis of uncertain time series data has become a research hotspot.This paper focuses on the classification analysis of uncertain time series,proposes a similarity matching algorithm suitable for uncertain time series,improves the traditional shapelet classification algorithm,and seeks a classification algorithm more suitable for uncertain time series.According to the basic concept of uncertain time series,a continuous uncertain time series representation model expressed in probability density is used.Analyze and compare the matching algorithm of deterministic and uncertain time series.Considering the advantages and disadvantages of various algorithms,the DTW(Dynamic Time Warping)based distance algorithm is adopted for the orderly,high-dimensional and uncertain features of uncertain time series.Based on the probabilistic data model used,the error function represented by the probability density is used to represent the data uncertainty.The distance between the corresponding points of the uncertain time series is calculated using the expected distance combined with the weight function method.The probability distance method is used to calculate the distance between two points in the expected distance.Considering the error function,the weight function further ensures the accuracy by weighting.UWDTW(Uncertain Weighted Dynamic Time Warping)distance calculation method is proposed by using the idea of DTW algorithm.This algorithm is applied to the nearest neighbor classification of uncertain time series for evaluation.For the problem of large time complexity of classification algorithm,LB_Keogh(Lower Bound by Keogh),a lower bound function filtering algorithm,is used to reduce computational cost and improve the performance of the classification algorithm.Experiments show that the similarity matching algorithm proposed in this paper can obtain the results closer to the real value when dealing with uncertain time series,and provide better classification accuracy.Aiming at the fact that the nearest neighbor classification algorithm is not explanatory and the complexity of the algorithm is too large,the shapelet is used to represent the most representative subsequence in the uncertain time series,and the shapelet transformation classification algorithm is proposed.Firstly,the key point-based PLR(Piecewise Linear Representation)method is used to reduce the dimensionality of uncertain time series to solve the data high dimensional problem.Then,for the problem of a large number of similarity elements in the shapelet selection process,the previous shapelet selection algorithm is improved,and the similar elements in the collection are removed by the shapelet pruning strategy to obtain a simple but accurate shapelet.Finally,a shapelet transformation classification algorithm is proposed.The UWDTW distance algorithm is used in the classification process.Experiments show that the shapelet-based transformation classification algorithm can be used to classify uncertain time series.Moreover,the algorithm provides classification interpretability,which makes up for the shortcomings of the nearest neighbor classification algorithm and helps to better analyze the data.