Dissimilarity-Preserving Representation Learning for One-Class Time Series Classification.

A deep multi-task representation learning method for time series classification and retrieval

Throughout recent years, dynamic time warping (DTW) has remained as a robust similarity measure in time series classification (TSC). 1-nearest neighbor (1-NN) algorithm with DTW is the most widely used classification method on time series serving as a benchmark. With the increasing demand for TSC on low-resource devices and the widespread of wearable devices, the need for a efficient and accurate time series classifier has never been higher. Although 1-NN DTW attains accurate results, it highly falls back on efficiency due to its quadratic complexity in the length of time series. In this paper, we propose a new approximation method for reducing the length of the time series as the input of DTW. We call it control chart approximation (CCA), after a similar concept used in statistical quality control processing. CCA representation approximates raw time series by transforming them into a set of segments with aggregated values and durations forming a reduced 3-D vector. We also propose an adaptation of DTW in 3-D space as a distance measure for 1-NN classifier, and denote the method as 1-NN 3-D DTW. Our experiments on 85 datasets from UCR archive—including 28 long-length (>500 points) time series datasets—show up to two orders of magnitude performance gain in running time compared to the state-of-the-art 1-NN DTW implementation. Moreover, it shows similar or better accuracy on the long time series in the experiment.

Time series classification using local distance-based features in multi-modal fusion networks

ABSTRACTA classification of the behavior of smart card users is important in the field of public transit demand analysis. It provides an understanding of people’s sequence of activities within a period of time. However, classical metrics such as Euclidean distance is not appropriate when dealing with time-series classification. To solve this problem, in this article a method for the classification of public transit smart card users’ daily transactions, which are represented in time series, is presented. The chosen approach uses cross-correlation distance (CCD), hierarchical clustering, and subgroups by metric parameter to understand the users’ temporal patterns. The clustering results are compared with dynamic time warping (DTW) distance (a common method to measure time-series distance). After a brief pedagogical example to explain the DTW and CCD concepts, a program is developed in R to validate the method on a real dataset of smart card data transactions. The dataset concerns the use of the public transit system in the city of Gatineau in September 2013. The results demonstrate that CCD performs better than DTW to classify the time series, and that the classification method identifies different public transit users’ daily behaviors. The results will help transit authorities to offer better services for smart card users from diverse groups.

Time-series classification is one of the most important machine learning tasks related to time series. It is the common denominator in various recognition tasks, such as signature verification, person identification based on keystroke dynamics, detection of cardiovascular diseases and brain disorders (e.g. early stage of Alzheimer disease or dementia). This tutorial aims to give an introduction to the most prominent challenges, methods, evaluation protocols and biomedical applications related to time series classification. Besides the "conventional" time series classification task, early classification and semi-supervised classification will be considered. Both preprocessing techniques -- FFT, SAX, etc. -- and wide-spread classifiers -- such as similarity-based, feature-based, motif/shaplet-based classifiers and convolutional neural networks -- will be covered. As dynamic time warping (DTW) is the one of the key components of many time series classifiers, including recent ones based on deep learning, we will describe this technique in detail. Slides: http://www.biointelligence.hu/pdf/timeseriestutorial.pdf

Classification of time series has attracted substantial interest over past decades. Methods based on Dynamic Time Warping (DTW), Symbolic Aggregate approXimation (SAX) and Shapelets are widely used and have achieved success in various real-world scenarios. However most existing time series classification methods either focus on global variation (e.g. DTW, SAX) or local variation (e.g. Shapelets). In this paper, we propose a Multi-Grained Ensemble Method for time series classification (MEGoT), which can make use of the variation of multi-grained data at the same time. In MEGoT, unstable base learners (Neural Networks) are assigned different weights to combine the ensemble. Different learners represent the learning features of different subsequences in time series, which can discover the discriminative regions, providing interpretability for classification. The training process of MGEoT is simpler and apt to parallel implementation. In the experiments, we conduct empirical evaluations and comparisons with various existing methods on 25 benchmark datasets. The final results show that dividing samples into smaller granularity is able to improve the diversity of ensemble, and MGEoT is competitive in accuracy under the Nemenyi test. Furthermore, MGEoT can discover the discriminative regions in time series, which may be neglected in the global methods.

Traditional works on time series classification usually use all of data in time series without distinction. However, that will swamp the discriminative information and decrease the correctness of classification. In this paper, a feature segment based time series classification algorithm was proposed, which only selects some highly discriminative time series data for classification. Firstly, an adaptive time series segmentation method was proposed. Then, a large margin based feature segment selection method was given. Based on these two methods, a time series classification framework was established after representing the time series with the optimal segments. By exploring the discriminative temporal patterns hidden in subsequences of time series and giving them more emphasize, the algorithm proposed in this paper can improve the time series classification performance greatly. Extensive experimental results showed that the proposed algorithm can achieve a good classification performance.

In the domain of time-series classification, one simple but persistently successful method is the 1-nearest neighbour (1NN) classifier coupled with an elastic distance measure such as Dynamic Time Warping (DTW). In this paper we evaluate the performance of DTW when constrained using the Itakura parallelogram, and compare it with the more commonly used Sakoe-Chiba band, as well as with the unconstrained DTW. Results show that although the Itakura parallelogram is generally inferior to the Sakoe-Chiba band, it is still superior to unconstrained DTW. Furthermore, on individual data sets the Itakura parallelogram can produce superior results, warranting further investigation into the merits of its use with DTW and other elastic distance measures for time-series classification.

Time Series Classification (TSC) is an important and challenging problem in data mining. With the increase of time series data availability, hundreds of TSC algorithms have been proposed. Among these methods, only a few have considered Deep Neural Networks (DNNs) to perform this task. This is surprising as deep learning has seen very successful applications in the last years. DNNs have indeed revolutionized the field of computer vision especially with the advent of novel deeper architectures such as Residual and Convolutional Neural Networks. Apart from images, sequential data such as text and audio can also be processed with DNNs to reach state-of-the-art performance for document classification and speech recognition. In this article, we study the current state-of-the-art performance of deep learning algorithms for TSC by presenting an empirical study of the most recent DNN architectures for TSC. We give an overview of the most successful deep learning applications in various time series domains under a unified taxonomy of DNNs for TSC. We also provide an open source deep learning framework to the TSC community where we implemented each of the compared approaches and evaluated them on a univariate TSC benchmark (the UCR/UEA archive) and 12 multivariate time series datasets. By training 8,730 deep learning models on 97 time series datasets, we propose the most exhaustive study of DNNs for TSC to date.

Classification of time series is an essential requirement of various applications that demand continuous monitoring of dynamical systems such as industrial process and health care monitoring. Feature extraction plays a vital role in deciding performance of the time series models. In recent years deep learning techniques have shown an excellent performance to extract highly discriminating features for the classification of the time series. A Convolutional Neural Network (CNN) is a unified framework that performs the feature learning and classification tasks simultaneously. Using the CNN to perform the classification of the multivariate time series is still a challenging task. In this paper, we propose an attention-based multivariate convolutional neural network (AT-MVCNN) that consists of the attention feature-based input tensor scheme to encode informations across the multiple time stamps. The method is capable of learning the temporal characteristics of the multivariate time series. The efficacy of the proposed method is tested on Human Activity Recognition (HAR) and Occupancy Detection datasets. The experiments and results show the proposed method outperforms the other deep learning and traditional machine learning models.

Dynamic Time Warping (DTW) is a well known algorithm for measuring similarity of two time series and widely used in classification, clustering or regression problems related to time series data. Unlike simple Euclid distance measure, DTW can handle time series of unequal lengths and is able to find an optimal alignment between two time sequences. Though very efficient, the computational cost of DTW is very high. There are several suboptimal variants of DTW for lowering computation, none of them is perfect. In this work, an approach to reduce computational burden of DTW has been proposed from the perspective of removing unimportant portion of the time series from computing, selected by a mask generated by edge detection algorithm commonly used in image processing or computer vision. The proposed Edge-Detectional Dynamic Time Warping (EDDTW) has been compared with original DTW by simulation experiments with 43 publicly available benchmark data sets. The simulation results show that EDDTW outperforms DTW regarding classification accuracy in more than half of the data sets, while reducing on the average 60% of the original time series leading to reduction in computational time.

The study is concerned with a concept and a design of granular time series and granular classifiers. In contrast to the plethora of various models of time series, which are predominantly numeric, we propose to effectively exploit the idea of information granules in the description and classification of time series. The numeric (optimization-oriented) and interpretation abilities of granular time series and their classifiers are highlighted and quantified. A general topology of the granular classifier involving a formation of a granular feature space and the usage of the framework of relational structures (relational equations) in the realization of the classifiers is presented. A detailed design process is elaborated on along with a discussion of the pertinent optimization mechanisms. A series of experiments is covered leading to a quantitative assessment of the granular classifiers and their parametric analysis.

A nearest neighbor approach to the classification of non-stationary time series is considered. A metric or measure of dissimilarity is computed between a new-to-be classified time series and each of a set of labeled sample time series. The new time series is classified by nearest neighbor rules. The metric is related to the criterion functional used in prediction error time series modeling methods. Engine fault time series data is considered. That data appears to be locally stationary. A Householder transformation - Akaike AIC criterion method for modeling time series by locally stationary AR models is applied to classify the data.

One of the advantages of the kernel methods is that they can deal with various kinds of objects, not necessarily vectorial data with a fixed number of attributes. In this paper, we develop kernels for time series data using dynamic time warping (DTW) distances. Since DTW distances are pseudo distances that do not satisfy the triangle inequality, a kernel matrix based on them is not positive semidefinite, in general. We use semidefinite programming (SDP) to guarantee the positive definiteness of a kernel matrix. We present neighborhood preserving embedding (NPE), an SDP formulation to obtain a kernel matrix that best preserves the local geometry of time series data. We also present an out-of-sample extension (OSE) for NPE. We use two applications, time series classification and time series embedding for similarity search, to validate our approach.

In time series classification the most commonly used approach is k Nearest Neighbor classification, where k = 1, coupled with Dynamic Time Warping (DTW) similarity checking. A challenge is that the DTW process is computationally expensive. This paper presents a new approach for speeding-up the DTW process, Sub-Sequence-Based DTW, which offers the additional benefit of improving accuracy. This paper also presents an analysis of the impact of the Sub-Sequence-Based method in terms of efficiency and effectiveness in comparison with standard DTW and the Sakoe-Chiba Band technique.