| Time series data exist widely in various fields,such as electrocardiograms in medicine,traffic flow data on traffic networks and electricity consumption data of residents in national power grids.Time series data mining aims to analyze and mine the potential patterns behind the data from time series data,and help people make useful decisions.Time series data mining covers several research directions such as prediction,classification,anomaly detection,clustering,similarity measurement and visualization,among which the prediction and classification directions have broad application prospects and can bring great help to people’s production and life.For example,the prediction of traffic flow data can predict the congestion of road network in the future period in advance;the classification of medical patient attendance data in medical treatment can predict the mortality rate of critically ill patients.Therefore,this dissertation studys the problem of time series prediction and classifi cation in time series data mining.Time series data are dynamic,complex,high-dimensional and easy to reach large scale,which makes it difficult for traditional machine learning algorithms to automatically extract the feature information from time series data.In contrast,deep learning methods have powerful feature learning capability,which can automatically capture the complex features of time series and show better performance in the case of a large amount of sample data.Therefore,this dissertation proposes two deep neural network models to solve the time series prediction problem and the time series classification problem.For the time series prediction problem,this dissertation proposes a time series prediction model based on graph convolutional neural networks.For the complex dependencies between different variables of multivariate time series,the model designs an adaptive variable relationship graph construction network module,which can get different adjacency graphs according to the different inputs of multivariate time series,so as to learn the relationship between different variables.In addition to extracting the features of the interaction of different variables on the generated graph structure,the model designs the coding and decoding network module of the graph convolution,which can capture the features of the sequence time using both the gate recurrent unit component and the attention component.Experiments on power,energy,financial exchange rate,and traffic datasets show that the proposed prediction model outperforms the vast majority of baseline models.The effectiveness of the proposed model is validated by the fact that it outperforms the vast majority of the baseline models.For the classification time series problem,this dissertation proposes an time series classification model based on a variational auto-encoder,and the model is suitable for complex time data with irregular sampling.For the problem of irregularly sampled time series data with a large number of missing data,the model uses a variational autoencoder to directly learn the distribution of samples with missing data to avoid preprocessing the samples and use the information of hidden variables generated during the learning process to perform classification.To address the problem of irregular sampling intervals,a continuous convolution layer with adjustable kernel width is designed to convolve irregularly sampled time series data to obtain a fixed dimensional feature representation.The binary and multi-classification experiments done on medical health dataset and human activity dataset show that the proposed classification model outperforms most of the baseline models,and in addition,the training time of the model is faster than other baseline models. |
