Imaging Time-Series to Improve Classification and Imputation

1 Imaging Time-Series to Improve Classification and ImputationZhiguang WangandTim OatesDepartment of Computer Science and Electric EngineeringUniversity of Maryland, Baltimore County{ , by recent successes of deep learningin computer vision, we propose a novel frame-work for encoding time series as different typesof images, namely, Gramian Angular Summa-tion/Difference Fields (GASF/GADF) and MarkovTransition Fields (MTF). This enables the use oftechniques from computer vision for time seriesclassification and Imputation . We used Tiled Con-volutional Neural Networks (tiled CNNs) on 20standard datasets to learn high-level features fromthe individual and compound GASF-GADF-MTFimages. Our approaches achieve highly competi-tive results when compared to nine of the currentbest time series Classification approaches.}

2 Inspiredby the bijection property of GASF on 0/1 rescaleddata, we train Denoised Auto-encoders (DA) on theGASF images of four standard and one synthesizedcompound dataset. The Imputation MSE on testdata is reduced by when comparedto using the raw data. An analysis of the featuresand weights learned via tiled CNNs and DAs ex-plains why the approaches IntroductionSince 2006, the techniques developed from deep neural net-works (or, deep learning) have greatly impacted natural lan-guage processing, speech recognition and computer visionresearch[Bengio, 2009; Deng and Yu, 2014]. One suc-cessful deep learning architecture used in computer vision isconvolutional neural networks (CNN)[LeCunet al., 1998].CNNs exploit translational invariance by extracting featuresthrough receptive fields[Hubel and Wiesel, 1962]and learn-ing with weight sharing, becoming the state-of-the-art ap-proach in various image recognition and computer visiontasks[Krizhevskyet al.]

3 , 2012]. Since unsupervised pretrain-ing has been shown to Improve performance[Erhanet al.,2010], sparse coding and Topographic Independent Compo-nent Analysis (TICA) are integrated as unsupervised pretrain-ing approaches to learn more diverse features with complexinvariances[Kavukcuogluet al., 2010; Ngiamet al., 2010].Along with the success of unsupervised pretraining appliedin deep learning, others are studying unsupervised learningalgorithms for generative models, such as Deep Belief Net-works (DBN) and Denoised Auto-encoders (DA)[Hintonet al., 2006; Vincentet al., 2008]. Many deep generativemodels are developed based on energy-based model or auto-encoders. Temporal autoencoding is integrated with RestrictBoltzmann Machines (RBMs) to Improve generative mod-els[H ausleret al.

4 , 2013]. A training strategy inspired byrecent work on optimization-based learning is proposed totrain complex neural networks for Imputation tasks[Brakelet al., 2013].A generalized Denoised Auto-encoder ex-tends the theoretical framework and is applied to Deep Gen-erative Stochastic Networks (DGSN)[Bengioet al., 2013;Bengio and Thibodeau-Laufer, 2013].Inspired by recent successes of supervised and unsuper-vised learning techniques in computer vision, we consider theproblem of encoding time series as images to allow machinesto visually recognize, classify and learn structures and pat-terns. Reformulating features of time series as visual clueshas raised much attention in computer science and physics. Inspeech recognition systems, acoustic/speech data input is typ-ically represented by concatenating Mel-frequency cepstralcoefficients (MFCCs) or perceptual linear predictive coeffi-cient (PLPs)[Hermansky, 1990].

5 Recently, researchers aretrying to build different network structures from time seriesfor visual inspection or designing distance measures. Re-currence Networks were proposed to analyze the structuralproperties of time series from complex systems[Donneretal., 2010; 2011]. They build adjacency matrices from thepredefined recurrence functions to interpret the time series ascomplex networks. Silvaet the recurrence plotparadigm for time series Classification using compression dis-tance[Silvaet al., 2013]. Another way to build a weightedadjacency matrix is extracting transition dynamics from thefirst order Markov matrix[Campanharoet al., 2011]. Al-though these maps demonstrate distinct topological proper-ties among different time series , it remains unclear how thesetopological properties relate to the original time series sincethey have no exact inverse present three novel representations for encoding timeseries as images that we call the Gramian Angular Summa-tion/Difference Field (GASF/GADF) and the Markov Transi-tion Field (MTF).

6 We applied deep Tiled Convolutional Neu-ral Networks (Tiled CNN)[Ngiamet al., 2010]to classifytime series images on 20 standard datasets. Our experimentalProceedings of the Twenty-Fourth International Joint Conference on Artificial Intelligence (IJCAI 2015)3939 time series xGASFP olar CoordinateGADFF igure 1: Illustration of the proposed encoding map ofGramian Angular a sequence of rescaled time se-ries in the Fish dataset. We transformXinto a polar coordi-nate system by eq. (3) and finally calculate its GASF/GADF images with eqs. (5) and (7). In this example, we build GAFswithout PAA smoothing, so the GAFs both have high demonstrate our approaches achieve the best perfor-mance on 9 of 20 standard dataset compared with 9 previousand current best Classification methods. Inspired by the bi-jection property of GASF on0/1rescaled data, we train theDenoised Auto-encoder (DA) on the GASF images of 4 stan-dard and a synthesized compound dataset.

7 The imputationMSE on test data is reduced by compared tousing the raw data. An analysis of the features and weightslearned via tiled CNNs and DA explains why the Imaging time SeriesWe first introduce our two frameworks for encoding time se-ries as images. The first type of image is a Gramian AngularField (GAF), in which we represent time series in a polar co-ordinate system instead of the typical Cartesian the Gramian matrix, each element is actually the cosine ofthe summation of angles. Inspired by previous work on theduality between time series and complex networks[Campan-haroet al., 2011], the main idea of the second framework,the Markov Transition Field (MTF), is to build the Markovmatrix of quantile bins after discretization and encode the dy-namic transition probability in a quasi-Gramian Gramian Angular FieldGiven a time seriesX={x1,x2.}

8 ,xn}ofnreal-valued ob-servations, we rescaleXso that all values fall in the interval[ 1,1]or[0,1]by: xi 1=(xi max(X)+(xi min(X))max(X) min(X)(1)or xi0=xi min(X)max(X) min(X)(2)Thus we can represent the rescaled time series Xin polarcoordinates by encoding the value as the angular cosine andthe time stamp as the radius with the equation below:{ = arccos ( xi), 1 xi 1, xi Xr=tiN,ti N(3)In the equation above,tiis the time stamp andNis a con-stant factor to regularize the span of the polar coordinate sys-tem. This polar coordinate based representation is a novelway to understand time series . As time increases, correspond-ing values warp among different angular points on the span-ning circles, like water rippling. The encoding map of equa-tion 3 has two important properties. First, it is bijective ascos( )is monotonic when [0, ].)

9 Given a time series ,the proposed map produces one and only one result in the po-lar coordinate system with a unique inverse map. Second, asopposed to Cartesian coordinates, polar coordinates preserveabsolute temporal relations. We will discuss this in more de-tail in future data in different intervals have different angularbounds.[0,1]corresponds to the cosine function in[0, 2],while cosine values in the interval[ 1,1]fall into the angu-lar bounds[0, ]. As we will discuss later, they provide dif-ferent information granularity in the Gramian Angular Fieldfor Classification tasks, and the Gramian Angular DifferenceField (GADF) of[0,1]rescaled data has the accurate inversemap. This property actually lays the foundation for imputingmissing value of time series by recovering the transforming the rescaled time series into the polarcoordinate system, we can easily exploit the angular perspec-tive by considering the trigonometric sum/difference betweeneach point to identify the temporal correlation within differ-ent time intervals.

10 The Gramian Summation Angular Field(GASF) and Gramian Difference Angular Field (GADF) aredefined as follows:GASF= [cos( i+ j)](4)= X X I X2 I X2(5)GADF= [sin( i j)](6)= I X2 X X I X2(7)Iis the unit row vector[1,1,..,1]. After transforming tothe polar coordinate system, we take time series at each timestep as a 1-D metric space. By defining the inner product<x,y >=x y 1 x2 1 y2and< x,y >= 1 x2 y x 1 y2, two types of Gramian Angular Fields (GAFs)are actually quasi-Gramian matrices[< x1, x1>].1 The GAFs have several advantages. First, they provide away to preserve temporal dependency, since time increases asthe position moves from top-left to bottom-right. The GAFscontain temporal correlations becauseG(i,j||i j|=k)repre-sents the relative correlation by superposition/difference ofdirections with respect to time intervalk.