Research On Key Techniques Of Discriminative Patterns Discovery And Classification Methods Of Time Series

Time series are a series of data recorded in chronological order.In real life,there are a large number of data acquired based on a series of time observation points.The basic characteristics of time series data include orderliness,numerical value,high dimensionality,large amount of data,and constant updating.The problem of time series classification is to mine the patterns or rules that can be used for classification prediction from a large amount of time series.The discriminative patterns of time series are implied with the attributes,and the dependencies among attributes will affect the accuracy of classification.However,the attributes of time series are numerical,which makes their dependencies difficult to analyze.It has become one of the main challenges faced by current researchers that how to quickly discover the discriminative characteristics or patterns of time series while improving the performance and interpretability of classifiers.Although many studies have been carried out on time series classification methods based on discriminative patterns,there are still problems to be solved.First of all,for most classifiers such as decision trees or neural networks,the input data are feature vectors,but time series do not have clear features.Second,since the time series are numerical data,it is very difficult to describe the dependencies among attributes.Third,the patterns formed by time series' attributes and attributes' values play a very critical role in classification problems,while time series have high dimensionality,thus the amount of discriminative patterns is often considerable.It is still a difficult problem to balance the number of discriminative patterns and classification accuracy.The dissertation revolves around the above issues and deeply studies how to complete the time series classification based on the discriminative patterns.The main contributions are as follows:(1)A Bayesian classification method based on selective patterns is proposed.The dependencies among discriminative patterns are built based on Bayesian networks,and two Bayesian classification algorithms based on selective patterns' discrimination ability are proposed.The algorithms combine the selective patterns with the naive Bayes classifier and the aggregated one-dependence estimators(AODE)respectively.The AODE is used to deal with the dependencies among attributes.While reducing the limitation of the attribute conditional independence assumption,it further balances the dependencies among the internal and external attributes of the selective patterns.The experimental results show that,through fully mining the classification ability of the selective pattens,analyzing the dependencies among the internal and external attributes of the selective patterns,the proposed classifers can achieve the highest accuracy on some datasets.Compared with the bench mark classifiers naive Bayes classifer and aggregating one-dependence estimators,the classification accuracies of the proposed classifiers are significantly higher,which verifies their correctness and rationality.(2)A shapelets-based Bayesian network time series classification method is proposed.The method of mining time series discriminative patterns is studied,and time series are transformed based on the time series discriminative patterns-shapelets.Then the transformed numerical time series are discretized into nominal data,and classified by the weighted aggregating one-dependence estimator(WAODE).Experimental analysis shows that the average accuracy of the proposed classification method is better than the ensemble classifiers-canonical time-series characteristics(Catch22),and has significant difference.Compared with the ensemble classifier-the hierarchical vote collective of transformation-based ensembles(HIVE?COTE),there isn't significant difference between HIVE?COTE and the proposed method.(3)A naive Bayesian time series classification method based on discrete Fourier transform is proposed.The method of symbolizing time series is analyzed,and the discriminative features are extracted by discrete Fourier transform,then the time series are transformed from the time domain to the frequency domain and discretized into nominal data.After that,the final classification of the time series is completed by a simple and effective naive Bayes classifier.The influence of the the sliding window length and the alphabet size on the classification accuracy is analyzed.The experiment result indicates that,compared with the benchmark algorithms naive Bayes classifier,the nearest neighbor classifier based on Euclidean distance,and the nearest neighbor classifier based on dynamic time warping,the proposed method has the highest average accuracy and significantly improves the accuracy of the naive Bayes classifier.(4)A weighted Bayesian network time series classification method based on the interest attenuation policy is proposed.A method to control structured peer-to-peer network flow based on interest attenuation policy is proposed.According to the principle of early classification of time series,the interest attenuation policy is used to analyze and characterize the attributes' weights in Bayesian networks.A cosine-weighted aggregation one-dependence estimator is proposed.The time series are transformed into nominal data by Fourier transform,and the classification is completed by the weighted aggregating one-dependence estimator and cosine-weighted aggregation one-dependence estimator respectively.Experimental results show that the proposed method has outstanding classification accuracy.Compared with the weighted aggregation one-dependence estimator based on discrete Fourier transform,the nearest neighbor classifier based on Euclidean distance and the nearest neighbor classifier based on dynamic time warping,the proposed method has significant difference in classification performance.(5)The military-oriented research on time series classification methods based on discriminative patterns is carried out.The important role of open source intelligence in military is clarified,and the characteristics of open source data in military are analyzed.According to communication,early classification and other principles,through the time series classification methods based on discriminative patterns,the application dimension of time series analysis of open source data in military is studied.The open source intelligence application scenarios in military are enriched by the time series classification methods based on discriminative patterns,and it lays a foundation for the further analysis of open source intelligence in military.The above methods expound the time series classification methods based on the discriminative patterns from multiple aspects.Through a large number of experiments and comparative analysis,the high performance of each method in the time series classification is shown.The correctness and rationality of the proposed methods are veryfied,and a firmer theoretical basis is founded.