<s>
Apache	B-Application
Mahout	I-Application
is	O
a	O
project	O
of	O
the	O
Apache	O
Software	O
Foundation	O
to	O
produce	O
free	B-Application
implementations	O
of	O
distributed	B-Architecture
or	O
otherwise	O
scalable	B-Architecture
machine	O
learning	O
algorithms	O
focused	O
primarily	O
on	O
linear	B-Language
algebra	I-Language
.	O
</s>
<s>
In	O
the	O
past	O
,	O
many	O
of	O
the	O
implementations	O
use	O
the	O
Apache	B-Application
Hadoop	I-Application
platform	O
,	O
however	O
today	O
it	O
is	O
primarily	O
focused	O
on	O
Apache	B-Language
Spark	I-Language
.	O
</s>
<s>
Mahout	O
also	O
provides	O
Java/Scala	O
libraries	O
for	O
common	O
math	O
operations	O
(	O
focused	O
on	O
linear	B-Language
algebra	I-Language
and	O
statistics	O
)	O
and	O
primitive	O
Java	B-Language
collections	O
.	O
</s>
<s>
Apache	O
Mahout-Samsara	O
refers	O
to	O
a	O
Scala	B-Application
domain	O
specific	O
language	O
(	O
DSL	O
)	O
that	O
allows	O
users	O
to	O
use	O
R-Like	O
syntax	O
as	O
opposed	O
to	O
traditional	O
Scala-like	O
syntax	O
.	O
</s>
<s>
Apache	B-Application
Mahout	I-Application
's	O
code	O
abstracts	O
the	O
domain	O
specific	O
language	O
from	O
the	O
engine	O
where	O
the	O
code	O
is	O
run	O
.	O
</s>
<s>
While	O
active	O
development	O
is	O
done	O
with	O
the	O
Apache	B-Language
Spark	I-Language
engine	O
,	O
users	O
are	O
free	B-Application
to	O
implement	O
any	O
engine	O
they	O
choose	O
-	O
and	O
Apache	B-Application
Flink	I-Application
have	O
been	O
implemented	O
in	O
the	O
past	O
and	O
examples	O
exist	O
in	O
the	O
code	O
base	O
.	O
</s>
<s>
To	O
improve	O
speed	O
,	O
“	O
native	O
solvers	O
”	O
were	O
added	O
which	O
move	O
in-core	O
,	O
and	O
by	O
extension	O
,	O
distributed	B-Architecture
BLAS	O
operations	O
out	O
of	O
the	O
JVM	O
,	O
offloading	O
to	O
off-heap	O
or	O
GPU	O
memory	O
for	O
processing	O
via	O
multiple	O
CPUs	O
and/or	O
CPU	O
cores	O
,	O
or	O
GPUs	O
when	O
built	O
against	O
the	O
ViennaCL	O
library	O
.	O
</s>
<s>
Apache	B-Application
Mahout	I-Application
features	O
implementations	O
of	O
Alternating	O
Least	O
Squares	O
,	O
Co-Occurrence	O
,	O
and	O
Correlated	O
Co-Occurrence	O
,	O
a	O
unique-to-Mahout	O
recommender	O
algorithm	O
that	O
extends	O
co-occurrence	O
to	O
be	O
used	O
on	O
multiple	O
dimensions	O
of	O
data	O
.	O
</s>
<s>
While	O
Mahout	O
's	O
core	O
algorithms	O
for	O
clustering	B-Algorithm
,	O
classification	O
and	O
batch	O
based	O
collaborative	O
filtering	O
were	O
implemented	O
on	O
top	O
of	O
Apache	B-Application
Hadoop	I-Application
using	O
the	O
map/reduce	B-Operating_System
paradigm	O
,	O
it	O
did	O
not	O
restrict	O
contributions	O
to	O
Hadoop-based	O
implementations	O
.	O
</s>
<s>
Contributions	O
that	O
run	O
on	O
a	O
single	O
node	O
or	O
on	O
a	O
non-Hadoop	O
cluster	O
were	O
also	O
welcomed	O
.	O
</s>
<s>
For	O
example	O
,	O
the	O
'	O
Taste	O
 '	O
collaborative-filtering	O
recommender	O
component	O
of	O
Mahout	O
was	O
originally	O
a	O
separate	O
project	O
and	O
can	O
run	O
stand-alone	O
without	O
Hadoop	B-Application
.	O
</s>
<s>
The	O
environment	O
consists	O
of	O
an	O
algebraic	O
backend-independent	O
optimizer	O
and	O
an	O
algebraic	O
Scala	B-Application
DSL	O
unifying	O
in-memory	O
and	O
distributed	B-Architecture
algebraic	O
operators	O
.	O
</s>
<s>
Supported	O
algebraic	O
platforms	O
are	O
Apache	B-Language
Spark	I-Language
,	O
,	O
and	O
Apache	B-Application
Flink	I-Application
.	O
</s>
<s>
Support	O
for	O
MapReduce	B-Operating_System
algorithms	O
started	O
being	O
gradually	O
phased	O
out	O
in	O
2014	O
.	O
</s>
<s>
Apache	B-Application
Mahout	I-Application
is	O
developed	O
by	O
a	O
community	O
.	O
</s>
