Capturing theDynamics ofMultivariate Time Series Through Visualization UsingGenerative Topographic MappingThrough Time
Alfredo Vellido · 2006
Mostoftheexisting research ontimeseries concernslack ofadaptive optimization ofthemodelparameters. Onthe supervised forecasting problems. Incomparison, little researchcontrary, theGenerative Topographic Mapping (GTM:Bishop hasbeendevoted tounsupervised methods forthevisual explo-etal. (7)) isastochastic modelthatwasoriginally devised ration ofmultivariate timeseries. Inthispaper, thecapabili- ' a ties oftheGenerative Topographic MappingThrough Time, a asaprobabilistic alternative toSOM,aiming toovercome modelwithsolid foundations inprobability theory that performsitsaforementioned limitations. TheGTM,whichcanalsobe simultaneous timeseries dataclustering andvisualization, are understood asaconstrained mixture model, issuited fordata assessed indetail inseveral experiments. Thefocus isplaced on clustering butalso, asalatent variable model, isembodied thedetection ofatypical data, thevisualization oftheevolution withvisualization capabilities thatareakintothoseofthe ofsignal regimes, andtheexploration ofsuddentransitions, for SOM,whichhavebeenextensively studied (Vesanto, (8)). which anovelidentification index isdefined. TheGTM Through Time(henceforth referred toasGTM- IndexTerms-Generative Topographic Mapping; Topology- TT:Bishop etal, (9)) isoneofthemanypossible extensions constrained hiddenMarkovmodels; Multivariate timeseriesofthestandard GTM allowed byitsprobabilistic definition. analysis; Datavisualization; Clustering o h tnadGMalwdb