Time Series Classification,Retrieval Methods And Applications

Time series is a group of data points that were collected according to fixed or non-fixed timestamps. As the development of information technology, sensors become cheaper and more prevalent in recent years. Hence, a large amount of time series data (e.g., electrocardiograph, human activity records and so forth) can be collected easily from different domains such as bioinformatics and finance. During last two decades, many researchers tried to explore the time series data mining and achieved many sig-nificant outcomes. Most of previous studies on time series focused on proposing new theoretical approaches but not being driven by real world applications. With the devel-opment of time series data mining technologies, until now, the research of time series data mining includes a wide range of topics, e.g., financial prediction, gene expression analysis, scientific data analysis, earthquake data mining and so forth. The complexity of time series data mining has many facets due to the nature of time series data, which includes three parts, i.e., large in data size, high dimensionality and necessary to update continuously. The most prominent factors are the challenges of its high dimensions and the difficult of defining similarity metric. More digital devices make the collec-tion of time series data increasing rapidly. Moreover, for different scenarios, different representations and different similarity or distance metrics of time series are heavily af-fecting the performance of time series data mining tasks. In this thesis, I focus on time series data mining and its business applications. Specifically, for several time series data mining tasks, i.e., time series classification, time series representation, similarity or distance metric, time series retrieval and time series segmentation, driven by the real world applications, I investigate these problems and propose several novel models to solve the specific issues and fulfill the business demands. Moreover, extensive experi-mental results provide the evidence that these methods are effective and superior to some of previous methods to some extent. The objective of this thesis is through proposing novel data mining techniques to fulfill the demands of real world applications and also lead more future work. The research efforts of this thesis can be summarized as follows.Time series classification has attracted amount of interest during last decade. Ac-tually, many studies on time series classification methods have been proposed and it is suggested that Nearest Neighbor classifier (especially,1-NN) is difficult to beat. Since the performance of1-NN algorithm depends critically on the distance metric given for specific tasks, the subsequent question then becomes how to determine the distance metric for so many applications. Many distance metrics have been proposed and among them, two of the most widely used are Euclidean distance and Dynamic Time Warping (DTW). Even though they could achieve a good performance for some applications, both of them have their own deficiencies, either in efficiency or in accuracy. Moreover, current evidence shows that there is no distance metric that is best for all time series data. Hence, inspired by the learning perspective, we investigate to use distance metric learning to obtain better distance metric and further to improve the classification perfor-mance for time series data. Specifically, we propose a novel Convolutional Nonlinear Neighborhood Components Analysis (CNNCA) model, which could learn a suitable distance metric from the time series data automatically. Both of the effectiveness and efficiency are evaluated. We summarize the contributions of this paper in these parts:· Though there are several studies that have explored the distance metric learning for time series data, to the best of our knowledge, we are the first to consider the time shift property when learning distance metric for the time series classification task.· Along this line, we propose a novel distance metric learning method CNNCA for time series data, which can obtain combined feature representation by concate-nating CNN and Multiple Layers Perceptron (MLP), and then learn a distance metric based on the scheme of stochastic neighbor assignments.· We conduct comprehensive experiments on amount of public data sets, then com-pare the performance of CNNCA with other distance metrics, including not only three conventional distance metrics, but also two learnt by Linear NCA and Non-linear NCA. The results prove that CNNCA can improve classification accuracy to some extent, especially for the relatively large-scale data sets.Compared to univariate time series, multivariate time series can provide more pat-terns and views of the same underlying phenomena, and help improve the classifica-tion performance. In this thesis, we focus on the classification of multivariate time se-ries. The distance-based method k-Nearest Neighbor (k-NN) combined with Dynamic Time Warping (DTW) could reach the best classification performance in most scenar-ios. However, both of k-NN and DTW needs expensive computation for large data set and long time series. The performance of traditional feature-based methods depends on the quality of hand-crafted features heavily, due to the difficulty of manually design-ing good features for time series data, most of time, feature-based methods are inferior to distance-based ones. Inspired by the feature learning, in this thesis, we propose a deep learning framework named Multi-Channels Deep Convolutional neural Networks (MC-DCNN) for multivariate time series classification. The experimental results on real world data sets reveal that our MC-DCNN model outperforms the baseline meth- ods with significant margins and has a good generalization, especially for weakly la-beled data. Besides, novel activation function, pooling strategy and visualization of learnt features are introduced. To further improve the performance, we also apply an unsupervised initialization to pretrain the convolutional neural networks and propose the pretrained version of MC-DCNN model.It is important to derive the source focal mechanism for earthquakes in real time, in addition to location and magnitude. Moreover, real-time estimation of source focal mechanism can be useful for building early-warning systems and monitoring fault ac-tivities. The challenge is the automatic and fast estimation of the earthquake source mechanism in a few seconds after receiving the seismic data at a few stations. In this thesis, we develop an image-based earthquake search engine (SeisE), similar to web search engines, to estimate earthquake parameters within a second by searching for sim-ilar seismograms from a large database. Similar to voice recording or a one-dimensional image, a seismogram is a graph record of the ground motion at a recording station as a function of time. It contains information about both the earthquake source and the earth medium through which the waves propagated. By assuming that the earth veloc-ity model is known, we apply a forward modeling approach to build a database wave-form for scenario earthquakes over a grid. Our objective is to find the best matches to any new earthquake record from the database. This approach is fully automatic with-out parameter input or human interference. Therefore, it could be applied for routinely reporting earthquake parameters.Rapid development of data mining and machine learning makes it possible that people could investigate and construct models for the stock prediction. Most of previ-ous studies on stock prediction are based on machine learning models (e.g., neural net-works, support vector machines, fuzzy methods and case based reasoning) or statistical methods (e.g. Generalized AutoRegressive Conditional Heteroskedasticity, GARCH) to predict the stock price or stock price trend in next days. From the view of fundamen-tal analysis, many researchers proposed stock prediction methods based on the analysis of public information such as news, twitter mood and stock articles. However, these methods are doubtable of their reliability and hardly used for decision-marking of in-vestors. In this thesis, we aim to identify the turning points of historical stock price data and then construct the technical indicators as the features of turning points for fur-ther turning point prediction. Two effective classifiers, Random Forests and Gradient Boosting Decision Tree, are used to evaluate the classification performance, and the experimental results reveal that technical indicators can provide useful information to improve the performance of turning point prediction, even though the improvement is not that remarkable. The predicted turning points can be used for decision-marking of investors.