Transcription
DimStiller: Workflows forDimensional Analysis and ReductionStephen Ingram, Tamara MunznerVeronika Irvine, Melanie TorySteven Bergner, Torsten Möller1
Overview Dimensionality Reduction Users Related Work Guidance DimStiller2
Dimension(ality)Reduction3
Dimension ReductionFilterCullSynthesizeCollectcomplex combinationsof input dimensions(nuttiness, fruitiness)dimensions areuninteresting(weight of spoon)dimensions areredundant(caffeine s/2009/03/coffee tongue.jpghttp://commons.wikimedia.org/wiki/File:A small cup of 4
Dimension ReductionFilterCullSynthesizeCollectcomplex combinationsof input dimensions(nuttiness, fruitiness)dimensions areuninteresting(weight of spoon)dimensions areredundant(caffeine s/2009/03/coffee tongue.jpghttp://commons.wikimedia.org/wiki/File:A small cup of 5
Dimension ReductionFilterCullSynthesizeCollectcomplex combinationsof input dimensions(nuttiness, fruitiness)dimensions areuninteresting(weight of spoon)dimensions areredundant(caffeine s/2009/03/coffee tongue.jpghttp://commons.wikimedia.org/wiki/File:A small cup of 6
Dimension ReductionFilterCullSynthesizeCollectcomplex combinationsof input dimensions(nuttiness, fruitiness)dimensions areuninteresting(weight of spoon)dimensions areredundant(caffeine s/2009/03/coffee tongue.jpghttp://commons.wikimedia.org/wiki/File:A small cup of 7
Dimension ReductionFilterCullSynthesizeCollectcomplex combinationsof input dimensions(nuttiness, fruitiness)dimensions areuninteresting(weight of spoon)dimensions areredundant(caffeine s/2009/03/coffee tongue.jpghttp://commons.wikimedia.org/wiki/File:A small cup of 8
Synthetic DR Example123Face Image Dataset:700 Faces35x35 1225 Dimensions700 x 1225 Dataset700http://isomap.stanford.edu/web3.jpg9
Synthetic DR ExampleNew Dataset:700 Faces2 Dimensions700 x 2 Dataset10http://isomap.stanford.edu/web3.jpg
USERS11
Visual High Dimensional Analysis(VHDA) User MapMath / StatsData Knowledge12
VHDA User MapBest Paper at NIPSMath / StatsTook Stats in UndergradWhat’s a mean?Data Knowledge13
VHDA User MapTotal Information AwarenessMath / StatsDropped in lapData Knowledge14
VHDA User MapMath / StatsPedagogicalData Knowledge15
VHDA User MapMath / StatsDon’t NeedAnalysisData Knowledge16
VHDA User MapWell DefinedTasksMath / StatsData Knowledge17
VHDA User MapMath / StatsMiddle Ground UsersData Knowledge18
RELATED WORK19
Other SystemsToolTarget UsersLimitationsMatlab, R, etc.Needs PowerUsersDR ToolkitsOnly LessProgrammingXMDVTool, GGobiNo GuidanceBeyond VisJohansson &Johansson 2009No Synthetic DR20
Hole In Prev Work Access To Range Of DR Algos Guidance For Middle Ground Users
Contributions22
Design andImplementation ofDimStiller23
Global and LocalGuidanceGlobal : : Operators24View:SPLOM
GUIDANCE25
Sloppy,MisunderstoodCompact,EvocativeOperator Space26
Which Operations and What courses/cs322/2008sp/schedule.htmlOperator e plot for the initial dataset Figure 36.jpghttp://www.scielo.cl/scielo.php?pid S0716-078X2001000200019&script sci /400/data filter icon?r 1http://www.personality-project.org/R/SPLOM27
Global GuidanceWhich Operations and What courses/cs322/2008sp/schedule.htmlOperator e plot for the initial dataset Figure 36.jpghttp://www.scielo.cl/scielo.php?pid S0716-078X2001000200019&script sci /400/data filter icon?r 1http://www.personality-project.org/R/SPLOM28
Local GuidanceWhat to do with a given operator?FilterHow many principal components?Sloppy,MisunderstoodPCAPCAWhat do they du/courses/cs322/2008sp/schedule.htmlOperator e plot for the initial dataset Figure 36.jpghttp://www.scielo.cl/scielo.php?pid S0716-078X2001000200019&script sci /400/data filter icon?r ct,Evocative
DIMSTILLER30
DimStiller31
DimStillerWorkflowSelector32
DimStillerExpressionTree33
DimStillerOperatorControl34
DimStillerOperatorViews35
EXAMPLE36
5000 pts294 dim294 DIMS37
SelectReduce:PCAWorkflow294 DIMSViewOperatorList Here38
Cull:VarianceOperatorScree Plotof Variances294 DIMS39
Log-scalefor betterVisibility294 DIMS40
List ofCulled DimsChoose firstnonzerodimension(31)264 DIMS41
Data:NormOperator264 DIMS42
Correlation Sliderset to 1.0146 DIMS43
Correlation Sliderset to 0.937 DIMS44
Eigenvalue ScreePlot : values die offaround 1616 DIMS45
Manageable SPLOM16 DIMS46
Operators &Workflows47
Operator FamiliesFamily NameOperatorsCullVariance, NameCollectPearson’sReducePCA, MDSViewSPLOM, HistoAttribColor, ClusterFilterValue48
Custom Workflows Three Workflows Given Freeform Experimenting With Operators Custom Workflows after Success49
Conclusions Presented the design and implementationof the DimStiller software Provided Global and Local guidance toopen up dimensionality reduction formiddle ground users beyond experts in math AND data
Thanks! Download DimStiller at .http://www.cs.ubc.ca/ sfingram/dimstiller Doing Dim Reduction?sfingram@cs.ubc.ca Funded By NSERC51Let me know!
DimStiller: Workflows for Dimensional Analysis and Reduction Stephen Ingram, Tamara Munzner Ver
