Research On Feature Extraction And Classifier Ensemble Method Of Time Series

Time series data exists in many real world applications,which reflects unobservable inner state of things.Unlike traditional classification task,time series classification task is more challenging because of order relation between attributes.In time series classification task,extracting features effectively and building corresponding classifiers can improve classification accuracy.Thus how to extract features and design classification algorithm become critical issue in time series classification.In this paper,we firstly make a brief introduction to the related work of time series classification,then analysis disadvantages of current feature extraction of classification algorithms.At present,instanced-based algorithms are time-consuming,which are not suitable for massive data processing.Model-based algorithms require prior knowledge of time series data,while ensemble algorithms suffer from high time and space complexity which imposes restrictions on application environments.To overcome previous mentioned problems,we focus on feature extraction and classification algorithm of time series,the main work includes:(1)We propose a feature extraction algorithm based on feature points detection,transforming time series data from original high dimensional time space into low dimensional feature space with main characteristics kept.This algorithm applies local feature algorithm in image process domain to time series data to extract feature points and generates feature vectors utilizing key sub-series around feature points.Moreover,it assigns weight to feature vectors and filteres them by weights to achieve feature extraction and dimension reduction.Comparing with traditional algorithms,the feature point detection algorithm we proposed in this paper not only extracts discriminative features,but also has linear time complexity and supports incremental update which is suitable to be applied to massive data process.In addition,stages of the algorithm can be easily controlled.Accuracy can be improved in the experiments in which this feature extraction algorithm and classification algorithm are combined.(2)We design a new classification ensemble algorithm based on scale space theory.This algorithm increases diversity of training dataset in which tendency information are added,while original information are kept at the same time.In the perspective of structure,it has multiple levels.Classifiers are trained level by level in a specified sequence with the information passing from one level to another while extending the feature space of dataset,the last level ouputs the final prediction result.Comparing with traditional algorithms,levels are not independent between each other.Types of classifiers can be replaced by others flexibly.Furthermore time and space complexity can be controlled by limiting number of level,which widens the application fields.In experiments,this ensemble algorithm combined with feature extraction algorithm proposed in this paper improves the accuracy of classifiers.