Graphics Reference
In-Depth Information
berg (
), together with students T. Chomut (Chomut,
) (who implemented
the first
-coordsEDA
sotware), andthe ubiquitous M.Boz(Inselbergetal.,
)-
wasvery fruitful,as indicated bythe partial list (Inselberg,
,
,
;Inselberg
etal.,
)(applicationstostatistics)andInselbergandDimsdale(
).Itpavedthe
wayfornewcontributors andusers:R.P.Burton'sgroup at Brigham Young University
(since
)Cluffetal.(
),CohanandYang(
),Hinterberger(firststudieddata
densities using
-coords,
)SchmidandHinterberger(
),Helly(
),Fiorini
and Inselberg (
), Gennings et al., contributed a sophisticated statistical applica-
tion (Gennings et al.,
)(response surfaces based on
-coords), Wegman (greatly
promoted EDA applications) (Wegman,
),and Desai and Walters (
).he re-
sultsobtainedbyEickemeyer(
),HungandInselberg(
)andChatterjee(
)
wereseminal.Progresscontinued throughChatterjee etal.(
),Ward(
),Jones
(
), to the most recent work of Yang (
) and Hauser (
), which increased
the versatility of
-coords. At the current time of writing, a query for “parallel coor-
dinates” on Google returned more than
,
“hits.”
The Case for Visualization
14.1.2
Searching a dataset with M items for “interesting” (a term that depends on the ob-
jectives) properties is inherently a di
cult task. here are
M
possible subsets, any
one of which could most closely satisfy the objectives. he visual cues that our eyes
can pick out from a decent data display can enable us to quickly sort through this
combinatoric explosion. How this is done is only part of the story here. Clearly, if
the transformation data
picture loses information, a great deal is lost right at
the start. When exploring adatasetwithN variables rather than presenting it (as pie
charts, histograms, etc.), use a good display to:
.
preserve information
- it should be possible to reconstruct the dataset com-
pletely from the picture,
.
have low representational complexity
-thecomputational costofconstructing
the display should be low,
.
work for any
N
-itshouldnotbelimitedbythedimensionsofthedata,
.
treat each variable uniformly
,
.
exhibit invariance under projective transformations
- the dataset should be
recognizable ater rotations, translations, scalings and perspective transforma-
tions,
.
reveal multivariate relations in the dataset
-thisisthemostimportantand
controversial single criterion,
.
be based on a rigorous mathematical and algorithmic methodology
-thus
eliminating ambiguity in the results.
Neither completeness noruniqueness isclaimedfor this list,whichshouldinvite crit-
icism and changes. Further commentary on each item, illustrated by examples and
comparisons, is listed below in the same order.
abbreviation of exploratory data analysis
