<s>
Projection	B-General_Concept
pursuit	I-General_Concept
(	O
PP	O
)	O
is	O
a	O
type	O
of	O
statistical	O
technique	O
which	O
involves	O
finding	O
the	O
most	O
"	O
interesting	O
"	O
possible	O
projections	B-Algorithm
in	O
multidimensional	O
data	O
.	O
</s>
<s>
Often	O
,	O
projections	B-Algorithm
which	O
deviate	O
more	O
from	O
a	O
normal	O
distribution	O
are	O
considered	O
to	O
be	O
more	O
interesting	O
.	O
</s>
<s>
As	O
each	O
projection	B-Algorithm
is	O
found	O
,	O
the	O
data	O
are	O
reduced	O
by	O
removing	O
the	O
component	O
along	O
that	O
projection	B-Algorithm
,	O
and	O
the	O
process	O
is	O
repeated	O
to	O
find	O
new	O
projections	B-Algorithm
;	O
this	O
is	O
the	O
"	O
pursuit	O
"	O
aspect	O
that	O
motivated	O
the	O
technique	O
known	O
as	O
matching	B-General_Concept
pursuit	I-General_Concept
.	O
</s>
<s>
The	O
idea	O
of	O
projection	B-General_Concept
pursuit	I-General_Concept
is	O
to	O
locate	O
the	O
projection	B-Algorithm
or	O
projections	B-Algorithm
from	O
high-dimensional	O
space	O
to	O
low-dimensional	O
space	O
that	O
reveal	O
the	O
most	O
details	O
about	O
the	O
structure	O
of	O
the	O
data	O
set	O
.	O
</s>
<s>
Once	O
an	O
interesting	O
set	O
of	O
projections	B-Algorithm
has	O
been	O
found	O
,	O
existing	O
structures	O
(	O
clusters	O
,	O
surfaces	O
,	O
etc	O
.	O
)	O
</s>
<s>
Projection	B-General_Concept
pursuit	I-General_Concept
has	O
been	O
widely	O
used	O
for	O
blind	B-Application
source	I-Application
separation	I-Application
,	O
so	O
it	O
is	O
very	O
important	O
in	O
independent	B-Algorithm
component	I-Algorithm
analysis	I-Algorithm
.	O
</s>
<s>
Projection	B-General_Concept
pursuit	I-General_Concept
seeks	O
one	O
projection	B-Algorithm
at	O
a	O
time	O
such	O
that	O
the	O
extracted	O
signal	O
is	O
as	O
non-Gaussian	O
as	O
possible	O
.	O
</s>
<s>
Projection	B-General_Concept
pursuit	I-General_Concept
technique	O
were	O
originally	O
proposed	O
and	O
experimented	O
by	O
Kruskal	O
.	O
</s>
<s>
The	O
first	O
successful	O
implementation	O
is	O
due	O
to	O
Jerome	O
H	O
.	O
Friedman	O
and	O
John	O
Tukey	O
(	O
1974	O
)	O
,	O
who	O
named	O
projection	B-General_Concept
pursuit	I-General_Concept
.	O
</s>
<s>
The	O
original	O
purpose	O
of	O
projection	B-General_Concept
pursuit	I-General_Concept
was	O
to	O
machine-pick	O
"	O
interesting	O
"	O
low-dimensional	O
projections	B-Algorithm
of	O
a	O
high-dimensional	O
point	O
cloud	O
by	O
numerically	O
maximizing	O
a	O
certain	O
objective	O
function	O
or	O
projection	B-Algorithm
index	O
.	O
</s>
<s>
Several	O
years	O
later	O
,	O
Friedman	O
and	O
Stuetzle	O
extended	O
the	O
idea	O
behind	O
projection	B-General_Concept
pursuit	I-General_Concept
and	O
added	O
projection	B-General_Concept
pursuit	I-General_Concept
regression	O
(	O
PPR	O
)	O
,	O
projection	B-General_Concept
pursuit	I-General_Concept
classification	O
(	O
PPC	O
)	O
,	O
and	O
projection	B-General_Concept
pursuit	I-General_Concept
density	O
estimation	O
(	O
PPDE	O
)	O
.	O
</s>
<s>
The	O
most	O
exciting	O
feature	O
of	O
projection	B-General_Concept
pursuit	I-General_Concept
is	O
that	O
it	O
is	O
one	O
of	O
the	O
very	O
few	O
multivariate	O
methods	O
able	O
to	O
bypass	O
the	O
"	O
curse	O
of	O
dimensionality	O
"	O
caused	O
by	O
the	O
fact	O
that	O
high-dimensional	O
space	O
is	O
mostly	O
empty	O
.	O
</s>
<s>
In	O
addition	O
,	O
projection	B-General_Concept
pursuit	I-General_Concept
is	O
able	O
to	O
ignore	O
irrelevant	O
(	O
i.e.	O
</s>
<s>
Many	O
of	O
the	O
methods	O
of	O
classical	O
multivariate	O
analysis	O
turn	O
out	O
to	O
be	O
special	O
cases	O
of	O
projection	B-General_Concept
pursuit	I-General_Concept
.	O
</s>
<s>
Examples	O
are	O
principal	B-Application
component	I-Application
analysis	I-Application
and	O
discriminant	B-General_Concept
analysis	I-General_Concept
,	O
and	O
the	O
quartimax	O
and	O
oblimax	O
methods	O
in	O
factor	O
analysis	O
.	O
</s>
<s>
One	O
serious	O
drawback	O
of	O
projection	B-General_Concept
pursuit	I-General_Concept
methods	O
is	O
their	O
high	O
demand	O
on	O
computer	O
time	O
.	O
</s>
