<s>
In	O
machine	O
learning	O
,	O
multiple-instance	B-General_Concept
learning	I-General_Concept
(	O
MIL	O
)	O
is	O
a	O
type	O
of	O
supervised	B-General_Concept
learning	I-General_Concept
.	O
</s>
<s>
In	O
the	O
simple	O
case	O
of	O
multiple-instance	O
binary	B-General_Concept
classification	I-General_Concept
,	O
a	O
bag	O
may	O
be	O
labeled	O
negative	O
if	O
all	O
the	O
instances	O
in	O
it	O
are	O
negative	O
.	O
</s>
<s>
Depending	O
on	O
the	O
type	O
and	O
variation	O
in	O
training	O
data	O
,	O
machine	O
learning	O
can	O
be	O
roughly	O
categorized	O
into	O
three	O
frameworks	O
:	O
supervised	B-General_Concept
learning	I-General_Concept
,	O
unsupervised	O
learning	O
,	O
and	O
reinforcement	O
learning	O
.	O
</s>
<s>
Multiple	O
instance	O
learning	O
(	O
MIL	O
)	O
falls	O
under	O
the	O
supervised	B-General_Concept
learning	I-General_Concept
framework	O
,	O
where	O
every	O
training	O
instance	O
has	O
a	O
label	O
,	O
either	O
discrete	O
or	O
real	O
valued	O
.	O
</s>
<s>
More	O
precisely	O
,	O
in	O
multiple-instance	B-General_Concept
learning	I-General_Concept
,	O
the	O
training	O
set	O
consists	O
of	O
labeled	O
“	O
bags	O
”	O
,	O
each	O
of	O
which	O
is	O
a	O
collection	O
of	O
unlabeled	O
instances	O
.	O
</s>
<s>
One	O
of	O
the	O
proposed	O
ways	O
to	O
solve	O
this	O
problem	O
was	O
to	O
use	O
supervised	B-General_Concept
learning	I-General_Concept
,	O
and	O
regard	O
all	O
the	O
low-energy	O
shapes	O
of	O
the	O
qualified	O
molecule	O
as	O
positive	O
training	O
instances	O
,	O
while	O
all	O
of	O
the	O
low-energy	O
shapes	O
of	O
unqualified	O
molecules	O
as	O
negative	O
instances	O
.	O
</s>
<s>
Thus	O
formulating	O
multiple-instance	B-General_Concept
learning	I-General_Concept
.	O
</s>
<s>
They	O
tested	O
the	O
algorithm	O
on	O
Musk	O
dataset	O
,	O
which	O
is	O
a	O
concrete	O
test	O
data	O
of	O
drug	O
activity	O
prediction	O
and	O
the	O
most	O
popularly	O
used	O
benchmark	O
in	O
multiple-instance	B-General_Concept
learning	I-General_Concept
.	O
</s>
<s>
In	O
MIL	O
terms	O
,	O
the	O
image	O
is	O
described	O
as	O
a	O
bag	O
,	O
where	O
each	O
is	O
the	O
feature	B-Algorithm
vector	I-Algorithm
(	O
called	O
instance	O
)	O
extracted	O
from	O
the	O
corresponding	O
-th	O
region	O
in	O
the	O
image	O
and	O
is	O
the	O
total	O
regions	O
(	O
instances	O
)	O
partitioning	O
the	O
image	O
.	O
</s>
<s>
Numerous	O
researchers	O
have	O
worked	O
on	O
adapting	O
classical	O
classification	O
techniques	O
,	O
such	O
as	O
support	B-Algorithm
vector	I-Algorithm
machines	I-Algorithm
or	O
boosting	B-Algorithm
,	O
to	O
work	O
within	O
the	O
context	O
of	O
multiple-instance	B-General_Concept
learning	I-General_Concept
.	O
</s>
<s>
The	O
remainder	O
of	O
the	O
article	O
will	O
focus	O
on	O
binary	B-General_Concept
classification	I-General_Concept
,	O
where	O
.	O
</s>
<s>
The	O
first	O
step	O
tries	O
to	O
learn	O
instance-level	O
concepts	O
by	O
building	O
a	O
decision	B-Algorithm
tree	I-Algorithm
from	O
each	O
instance	O
in	O
each	O
bag	O
of	O
the	O
training	O
set	O
.	O
</s>
<s>
Each	O
bag	O
is	O
then	O
mapped	O
to	O
a	O
feature	B-Algorithm
vector	I-Algorithm
based	O
on	O
the	O
counts	O
in	O
the	O
decision	B-Algorithm
tree	I-Algorithm
.	O
</s>
<s>
GMIL-1	O
enumerates	O
all	O
axis-parallel	O
rectangles	O
in	O
the	O
original	O
space	O
of	O
instances	O
,	O
and	O
defines	O
a	O
new	O
feature	B-Algorithm
space	I-Algorithm
of	O
Boolean	O
vectors	O
.	O
</s>
<s>
A	O
bag	O
is	O
mapped	O
to	O
a	O
vector	O
in	O
this	O
new	O
feature	B-Algorithm
space	I-Algorithm
,	O
where	O
if	O
APR	O
covers	O
,	O
and	O
otherwise	O
.	O
</s>
<s>
A	O
single-instance	O
algorithm	O
can	O
then	O
be	O
applied	O
to	O
learn	O
the	O
concept	O
in	O
this	O
new	O
feature	B-Algorithm
space	I-Algorithm
.	O
</s>
<s>
Because	O
of	O
the	O
high	O
dimensionality	O
of	O
the	O
new	O
feature	B-Algorithm
space	I-Algorithm
and	O
the	O
cost	O
of	O
explicitly	O
enumerating	O
all	O
APRs	O
of	O
the	O
original	O
instance	O
space	O
,	O
GMIL-1	O
is	O
inefficient	O
both	O
in	O
terms	O
of	O
computation	O
and	O
memory	O
.	O
</s>
<s>
Xu	O
(	O
2003	O
)	O
proposed	O
several	O
algorithms	O
based	O
on	O
logistic	O
regression	O
and	O
boosting	B-Algorithm
methods	O
to	O
learn	O
concepts	O
under	O
the	O
collective	O
assumption	O
.	O
</s>
<s>
By	O
mapping	O
each	O
bag	O
to	O
a	O
feature	B-Algorithm
vector	I-Algorithm
of	O
metadata	O
,	O
metadata-based	O
algorithms	O
allow	O
the	O
flexibility	O
of	O
using	O
an	O
arbitrary	O
single-instance	O
algorithm	O
to	O
perform	O
the	O
actual	O
classification	O
task	O
.	O
</s>
<s>
Future	O
bags	O
are	O
simply	O
mapped	O
(	O
embedded	O
)	O
into	O
the	O
feature	B-Algorithm
space	I-Algorithm
of	O
metadata	O
and	O
labeled	O
by	O
the	O
chosen	O
classifier	O
.	O
</s>
<s>
Classification	O
is	O
done	O
via	O
an	O
SVM	B-Algorithm
with	O
a	O
graph	O
kernel	O
(	O
MIGraph	O
and	O
miGraph	O
only	O
differ	O
in	O
their	O
choice	O
of	O
kernel	O
)	O
.	O
</s>
<s>
So	O
far	O
this	O
article	O
has	O
considered	O
multiple	O
instance	O
learning	O
exclusively	O
in	O
the	O
context	O
of	O
binary	B-General_Concept
classifiers	I-General_Concept
.	O
</s>
<s>
However	O
,	O
the	O
generalizations	O
of	O
single-instance	O
binary	B-General_Concept
classifiers	I-General_Concept
can	O
carry	O
over	O
to	O
the	O
multiple-instance	O
case	O
.	O
</s>
