<s>
Collaborative	B-Algorithm
filtering	I-Algorithm
(	O
CF	O
)	O
is	O
a	O
technique	O
used	O
by	O
recommender	B-Application
systems	I-Application
.	O
</s>
<s>
Collaborative	B-Algorithm
filtering	I-Algorithm
has	O
two	O
senses	O
,	O
a	O
narrow	O
one	O
and	O
a	O
more	O
general	O
one	O
.	O
</s>
<s>
In	O
the	O
newer	O
,	O
narrower	O
sense	O
,	O
collaborative	B-Algorithm
filtering	I-Algorithm
is	O
a	O
method	O
of	O
making	O
automatic	O
predictions	O
(	O
filtering	O
)	O
about	O
the	O
interests	O
of	O
a	O
user	O
by	O
collecting	O
preferences	O
or	O
taste	O
information	O
from	O
many	O
users	O
(	O
collaborating	O
)	O
.	O
</s>
<s>
The	O
underlying	O
assumption	O
of	O
the	O
collaborative	B-Algorithm
filtering	I-Algorithm
approach	O
is	O
that	O
if	O
a	O
person	O
A	O
has	O
the	O
same	O
opinion	O
as	O
a	O
person	O
B	O
on	O
an	O
issue	O
,	O
A	O
is	O
more	O
likely	O
to	O
have	O
B	O
's	O
opinion	O
on	O
a	O
different	O
issue	O
than	O
that	O
of	O
a	O
randomly	O
chosen	O
person	O
.	O
</s>
<s>
For	O
example	O
,	O
a	O
collaborative	B-Algorithm
filtering	I-Algorithm
recommendation	B-Application
system	I-Application
for	O
preferences	O
in	O
television	O
programming	O
could	O
make	O
predictions	O
about	O
which	O
television	O
show	O
a	O
user	O
should	O
like	O
given	O
a	O
partial	O
list	O
of	O
that	O
user	O
's	O
tastes	O
(	O
likes	O
or	O
dislikes	O
)	O
.	O
</s>
<s>
In	O
the	O
more	O
general	O
sense	O
,	O
collaborative	B-Algorithm
filtering	I-Algorithm
is	O
the	O
process	O
of	O
filtering	O
for	O
information	O
or	O
patterns	O
using	O
techniques	O
involving	O
collaboration	O
among	O
multiple	O
agents	O
,	O
viewpoints	O
,	O
data	O
sources	O
,	O
etc	O
.	O
</s>
<s>
Applications	O
of	O
collaborative	B-Algorithm
filtering	I-Algorithm
typically	O
involve	O
very	O
large	O
data	O
sets	O
.	O
</s>
<s>
Collaborative	B-Algorithm
filtering	I-Algorithm
methods	O
have	O
been	O
applied	O
to	O
many	O
different	O
kinds	O
of	O
data	O
including	O
:	O
sensing	O
and	O
monitoring	O
data	O
,	O
such	O
as	O
in	O
mineral	O
exploration	O
,	O
environmental	O
sensing	O
over	O
large	O
areas	O
or	O
multiple	O
sensors	O
;	O
financial	O
data	O
,	O
such	O
as	O
financial	O
service	O
institutions	O
that	O
integrate	O
many	O
financial	O
sources	O
;	O
or	O
in	O
electronic	O
commerce	O
and	O
web	O
applications	O
where	O
the	O
focus	O
is	O
on	O
user	O
data	O
,	O
etc	O
.	O
</s>
<s>
The	O
remainder	O
of	O
this	O
discussion	O
focuses	O
on	O
collaborative	B-Algorithm
filtering	I-Algorithm
for	O
user	O
data	O
,	O
although	O
some	O
of	O
the	O
methods	O
and	O
approaches	O
may	O
apply	O
to	O
the	O
other	O
major	O
applications	O
as	O
well	O
.	O
</s>
<s>
The	O
growth	O
of	O
the	O
Internet	O
has	O
made	O
it	O
much	O
more	O
difficult	O
to	O
effectively	O
extract	B-General_Concept
useful	I-General_Concept
information	I-General_Concept
from	O
all	O
the	O
available	O
online	O
information	O
.	O
</s>
<s>
The	O
overwhelming	O
amount	O
of	O
data	O
necessitates	O
mechanisms	O
for	O
efficient	O
information	B-Application
filtering	I-Application
.	O
</s>
<s>
Collaborative	B-Algorithm
filtering	I-Algorithm
is	O
one	O
of	O
the	O
techniques	O
used	O
for	O
dealing	O
with	O
this	O
problem	O
.	O
</s>
<s>
The	O
motivation	O
for	O
collaborative	B-Algorithm
filtering	I-Algorithm
comes	O
from	O
the	O
idea	O
that	O
people	O
often	O
get	O
the	O
best	O
recommendations	B-Application
from	O
someone	O
with	O
tastes	O
similar	O
to	O
themselves	O
.	O
</s>
<s>
Collaborative	B-Algorithm
filtering	I-Algorithm
encompasses	O
techniques	O
for	O
matching	O
people	O
with	O
similar	O
interests	O
and	O
making	O
recommendations	B-Application
on	O
this	O
basis	O
.	O
</s>
<s>
Collaborative	B-Algorithm
filtering	I-Algorithm
algorithms	O
often	O
require	O
(	O
1	O
)	O
users	O
 '	O
active	O
participation	O
,	O
(	O
2	O
)	O
an	O
easy	O
way	O
to	O
represent	O
users	O
 '	O
interests	O
,	O
and	O
(	O
3	O
)	O
algorithms	O
that	O
are	O
able	O
to	O
match	O
people	O
with	O
similar	O
interests	O
.	O
</s>
<s>
Typically	O
,	O
the	O
workflow	O
of	O
a	O
collaborative	B-Algorithm
filtering	I-Algorithm
system	O
is	O
:	O
</s>
<s>
A	O
key	O
problem	O
of	O
collaborative	B-Algorithm
filtering	I-Algorithm
is	O
how	O
to	O
combine	O
and	O
weight	O
the	O
preferences	O
of	O
user	O
neighbors	O
.	O
</s>
<s>
Collaborative	B-Algorithm
filtering	I-Algorithm
systems	O
have	O
many	O
forms	O
,	O
but	O
many	O
common	O
systems	O
can	O
be	O
reduced	O
to	O
two	O
steps	O
:	O
</s>
<s>
This	O
falls	O
under	O
the	O
category	O
of	O
user-based	O
collaborative	B-Algorithm
filtering	I-Algorithm
.	O
</s>
<s>
A	O
specific	O
application	O
of	O
this	O
is	O
the	O
user-based	O
Nearest	B-General_Concept
Neighbor	I-General_Concept
algorithm	I-General_Concept
.	O
</s>
<s>
Alternatively	O
,	O
item-based	B-Application
collaborative	I-Application
filtering	I-Application
(	O
users	O
who	O
bought	O
x	O
also	O
bought	O
y	O
)	O
,	O
proceeds	O
in	O
an	O
item-centric	O
manner	O
:	O
</s>
<s>
See	O
,	O
for	O
example	O
,	O
the	O
Slope	O
One	O
item-based	B-Application
collaborative	I-Application
filtering	I-Application
family	O
.	O
</s>
<s>
Another	O
form	O
of	O
collaborative	B-Algorithm
filtering	I-Algorithm
can	O
be	O
based	O
on	O
implicit	O
observations	O
of	O
normal	O
user	O
behavior	O
(	O
as	O
opposed	O
to	O
the	O
artificial	O
behavior	O
imposed	O
by	O
a	O
rating	O
task	O
)	O
.	O
</s>
<s>
These	O
predictions	O
then	O
have	O
to	O
be	O
filtered	O
through	O
business	B-Architecture
logic	I-Architecture
to	O
determine	O
how	O
they	O
might	O
affect	O
the	O
actions	O
of	O
a	O
business	O
system	O
.	O
</s>
<s>
However	O
,	O
there	O
are	O
other	O
methods	O
to	O
combat	O
information	O
explosion	O
,	O
such	O
as	O
web	O
search	O
and	O
data	B-Algorithm
clustering	I-Algorithm
.	O
</s>
<s>
Typical	O
examples	O
of	O
this	O
approach	O
are	O
neighbourhood-based	O
CF	O
and	O
item-based/user	O
-based	O
top-N	O
recommendations	B-Application
.	O
</s>
<s>
The	O
user	O
based	O
top-N	O
recommendation	B-Application
algorithm	I-Application
uses	O
a	O
similarity-based	O
vector	O
model	O
to	O
identify	O
the	O
k	O
most	O
similar	O
users	O
to	O
an	O
active	O
user	O
.	O
</s>
<s>
A	O
popular	O
method	O
to	O
find	O
the	O
similar	O
users	O
is	O
the	O
Locality-sensitive	B-Algorithm
hashing	I-Algorithm
,	O
which	O
implements	O
the	O
nearest	B-Algorithm
neighbor	I-Algorithm
mechanism	I-Algorithm
in	O
linear	O
time	O
.	O
</s>
<s>
The	O
advantages	O
with	O
this	O
approach	O
include	O
:	O
the	O
explainability	O
of	O
the	O
results	O
,	O
which	O
is	O
an	O
important	O
aspect	O
of	O
recommendation	B-Application
systems	I-Application
;	O
easy	O
creation	O
and	O
use	O
;	O
easy	O
facilitation	O
of	O
new	O
data	O
;	O
content-independence	O
of	O
the	O
items	O
being	O
recommended	O
;	O
good	O
scaling	B-Architecture
with	O
co-rated	O
items	O
.	O
</s>
<s>
Its	O
performance	O
decreases	O
when	O
data	B-Algorithm
gets	I-Algorithm
sparse	I-Algorithm
,	O
which	O
occurs	O
frequently	O
with	O
web-related	O
items	O
.	O
</s>
<s>
This	O
hinders	O
the	O
scalability	B-Architecture
of	O
this	O
approach	O
and	O
creates	O
problems	O
with	O
large	O
datasets	O
.	O
</s>
<s>
Although	O
it	O
can	O
efficiently	O
handle	O
new	O
users	O
because	O
it	O
relies	O
on	O
a	O
data	B-General_Concept
structure	I-General_Concept
,	O
adding	O
new	O
items	O
becomes	O
more	O
complicated	O
since	O
that	O
representation	O
usually	O
relies	O
on	O
a	O
specific	O
vector	O
space	O
.	O
</s>
<s>
In	O
this	O
approach	O
,	O
models	O
are	O
developed	O
using	O
different	O
data	B-Application
mining	I-Application
,	O
machine	O
learning	O
algorithms	O
to	O
predict	O
users	O
 '	O
rating	O
of	O
unrated	O
items	O
.	O
</s>
<s>
Bayesian	O
networks	O
,	O
clustering	B-Algorithm
models	I-Algorithm
,	O
latent	O
semantic	O
models	O
such	O
as	O
singular	O
value	O
decomposition	O
,	O
probabilistic	B-General_Concept
latent	I-General_Concept
semantic	I-General_Concept
analysis	I-General_Concept
,	O
multiple	O
multiplicative	O
factor	O
,	O
latent	O
Dirichlet	O
allocation	O
and	O
Markov	O
decision	O
process	O
based	O
models	O
.	O
</s>
<s>
Through	O
this	O
approach	O
,	O
dimensionality	B-Algorithm
reduction	I-Algorithm
methods	O
are	O
mostly	O
being	O
used	O
as	O
complementary	O
technique	O
to	O
improve	O
robustness	O
and	O
accuracy	O
of	O
memory-based	O
approach	O
.	O
</s>
<s>
In	O
this	O
sense	O
,	O
methods	O
like	O
singular	O
value	O
decomposition	O
,	O
principal	B-Application
component	I-Application
analysis	I-Application
,	O
known	O
as	O
latent	O
factor	O
models	O
,	O
compress	O
user-item	O
matrix	O
into	O
a	O
low-dimensional	O
representation	O
in	O
terms	O
of	O
latent	O
factors	O
.	O
</s>
<s>
It	O
handles	O
the	O
sparsity	B-Algorithm
of	O
the	O
original	O
matrix	O
better	O
than	O
memory	O
based	O
ones	O
.	O
</s>
<s>
Also	O
comparing	O
similarity	O
on	O
the	O
resulting	O
matrix	O
is	O
much	O
more	O
scalable	B-Architecture
especially	O
in	O
dealing	O
with	O
large	O
sparse	O
datasets	O
.	O
</s>
<s>
Importantly	O
,	O
they	O
overcome	O
the	O
CF	O
problems	O
such	O
as	O
sparsity	B-Algorithm
and	O
loss	O
of	O
information	O
.	O
</s>
<s>
Usually	O
most	O
commercial	O
recommender	B-Application
systems	I-Application
are	O
hybrid	O
,	O
for	O
example	O
,	O
the	O
Google	O
news	O
recommender	B-Application
system	I-Application
.	O
</s>
<s>
Some	O
generalize	O
traditional	O
Matrix	B-Application
factorization	I-Application
algorithms	O
via	O
a	O
non-linear	O
neural	O
architecture	O
,	O
or	O
leverage	O
new	O
model	O
types	O
like	O
Variational	O
Autoencoders	B-Algorithm
.	O
</s>
<s>
its	O
real	O
effectiveness	O
when	O
used	O
in	O
a	O
simple	O
collaborative	B-Application
recommendation	O
scenario	O
has	O
been	O
put	O
into	O
question	O
.	O
</s>
<s>
Similar	O
issues	O
have	O
been	O
spotted	O
by	O
others	O
and	O
also	O
in	O
sequence-aware	O
recommender	B-Application
systems	I-Application
.	O
</s>
<s>
Many	O
recommender	B-Application
systems	I-Application
simply	O
ignore	O
other	O
contextual	O
information	O
existing	O
alongside	O
user	O
's	O
rating	O
in	O
providing	O
item	O
recommendation	O
.	O
</s>
<s>
However	O
,	O
by	O
pervasive	O
availability	O
of	O
contextual	O
information	O
such	O
as	O
time	O
,	O
location	O
,	O
social	O
information	O
,	O
and	O
type	O
of	O
the	O
device	O
that	O
user	O
is	O
using	O
,	O
it	O
is	O
becoming	O
more	O
important	O
than	O
ever	O
for	O
a	O
successful	O
recommender	B-Application
system	I-Application
to	O
provide	O
a	O
context-sensitive	O
recommendation	O
.	O
</s>
<s>
According	O
to	O
Charu	O
Aggrawal	O
,	O
"	O
Context-sensitive	O
recommender	B-Application
systems	I-Application
tailor	O
their	O
recommendations	B-Application
to	O
additional	O
information	O
that	O
defines	O
the	O
specific	O
situation	O
under	O
which	O
recommendations	B-Application
are	O
made	O
.	O
</s>
<s>
As	O
an	O
instance	O
,	O
assume	O
a	O
music	O
recommender	B-Application
system	I-Application
which	O
provide	O
different	O
recommendations	B-Application
in	O
corresponding	O
to	O
time	O
of	O
the	O
day	O
.	O
</s>
<s>
Thus	O
,	O
instead	O
of	O
using	O
user-item	O
matrix	O
,	O
we	O
may	O
use	O
tensor	B-Device
of	O
order	O
3	O
(	O
or	O
higher	O
for	O
considering	O
other	O
contexts	O
)	O
to	O
represent	O
context-sensitive	O
users	O
 '	O
preferences	O
.	O
</s>
<s>
In	O
order	O
to	O
take	O
advantage	O
of	O
collaborative	B-Algorithm
filtering	I-Algorithm
and	O
particularly	O
neighborhood-based	O
methods	O
,	O
approaches	O
can	O
be	O
extended	O
from	O
a	O
two-dimensional	O
rating	O
matrix	O
into	O
a	O
tensor	B-Device
of	O
higher	O
order	O
.	O
</s>
<s>
Therefore	O
,	O
similar	O
to	O
matrix	B-Application
factorization	I-Application
methods	O
,	O
tensor	B-Device
factorization	O
techniques	O
can	O
be	O
used	O
to	O
reduce	O
dimensionality	O
of	O
original	O
data	O
before	O
using	O
any	O
neighborhood-based	O
methods	O
.	O
</s>
<s>
Services	O
like	O
Reddit	B-Application
,	O
YouTube	B-General_Concept
,	O
and	O
Last.fm	B-Application
are	O
typical	O
examples	O
of	O
collaborative	B-Algorithm
filtering	I-Algorithm
based	O
media	O
.	O
</s>
<s>
One	O
scenario	O
of	O
collaborative	B-Algorithm
filtering	I-Algorithm
application	O
is	O
to	O
recommend	O
interesting	O
or	O
popular	O
information	O
as	O
judged	O
by	O
the	O
community	O
.	O
</s>
<s>
As	O
a	O
typical	O
example	O
,	O
stories	O
appear	O
in	O
the	O
front	O
page	O
of	O
Reddit	B-Application
as	O
they	O
are	O
"	O
voted	O
up	O
"	O
(	O
rated	O
positively	O
)	O
by	O
the	O
community	O
.	O
</s>
<s>
Wikipedia	O
is	O
another	O
application	O
of	O
collaborative	B-Algorithm
filtering	I-Algorithm
.	O
</s>
<s>
Another	O
aspect	O
of	O
collaborative	B-Algorithm
filtering	I-Algorithm
systems	O
is	O
the	O
ability	O
to	O
generate	O
more	O
personalized	O
recommendations	B-Application
by	O
analyzing	O
information	O
from	O
the	O
past	O
activity	O
of	O
a	O
specific	O
user	O
,	O
or	O
the	O
history	O
of	O
other	O
users	O
deemed	O
to	O
be	O
of	O
similar	O
taste	O
to	O
a	O
given	O
user	O
.	O
</s>
<s>
The	O
more	O
a	O
given	O
user	O
makes	O
use	O
of	O
the	O
system	O
,	O
the	O
better	O
the	O
recommendations	B-Application
become	O
,	O
as	O
the	O
system	O
gains	O
data	O
to	O
improve	O
its	O
model	O
of	O
that	O
user	O
.	O
</s>
<s>
A	O
collaborative	B-Algorithm
filtering	I-Algorithm
system	O
does	O
not	O
necessarily	O
succeed	O
in	O
automatically	O
matching	O
content	O
to	O
one	O
's	O
preferences	O
.	O
</s>
<s>
As	O
in	O
the	O
personalized	O
recommendation	O
scenario	O
,	O
the	O
introduction	O
of	O
new	O
users	O
or	O
new	O
items	O
can	O
cause	O
the	O
cold	B-Application
start	I-Application
problem	O
,	O
as	O
there	O
will	O
be	O
insufficient	O
data	O
on	O
these	O
new	O
entries	O
for	O
the	O
collaborative	B-Algorithm
filtering	I-Algorithm
to	O
work	O
accurately	O
.	O
</s>
<s>
In	O
order	O
to	O
make	O
appropriate	O
recommendations	B-Application
for	O
a	O
new	O
user	O
,	O
the	O
system	O
must	O
first	O
learn	O
the	O
user	O
's	O
preferences	O
by	O
analysing	O
past	O
voting	O
or	O
rating	O
activities	O
.	O
</s>
<s>
The	O
collaborative	B-Algorithm
filtering	I-Algorithm
system	O
requires	O
a	O
substantial	O
number	O
of	O
users	O
to	O
rate	O
a	O
new	O
item	O
before	O
that	O
item	O
can	O
be	O
recommended	O
.	O
</s>
<s>
In	O
practice	O
,	O
many	O
commercial	O
recommender	B-Application
systems	I-Application
are	O
based	O
on	O
large	O
datasets	O
.	O
</s>
<s>
As	O
a	O
result	O
,	O
the	O
user-item	O
matrix	O
used	O
for	O
collaborative	B-Algorithm
filtering	I-Algorithm
could	O
be	O
extremely	O
large	O
and	O
sparse	O
,	O
which	O
brings	O
about	O
challenges	O
in	O
the	O
performance	O
of	O
the	O
recommendation	O
.	O
</s>
<s>
One	O
typical	O
problem	O
caused	O
by	O
the	O
data	O
sparsity	B-Algorithm
is	O
the	O
cold	B-Application
start	I-Application
problem	O
.	O
</s>
<s>
As	O
collaborative	B-Algorithm
filtering	I-Algorithm
methods	O
recommend	O
items	O
based	O
on	O
users	O
 '	O
past	O
preferences	O
,	O
new	O
users	O
will	O
need	O
to	O
rate	O
a	O
sufficient	O
number	O
of	O
items	O
to	O
enable	O
the	O
system	O
to	O
capture	O
their	O
preferences	O
accurately	O
and	O
thus	O
provides	O
reliable	O
recommendations	B-Application
.	O
</s>
<s>
The	O
new	O
item	O
problem	O
does	O
not	O
affect	O
content-based	O
recommendations	B-Application
,	O
because	O
the	O
recommendation	O
of	O
an	O
item	O
is	O
based	O
on	O
its	O
discrete	O
set	O
of	O
descriptive	O
qualities	O
rather	O
than	O
its	O
ratings	O
.	O
</s>
<s>
As	O
the	O
numbers	O
of	O
users	O
and	O
items	O
grow	O
,	O
traditional	O
CF	O
algorithms	O
will	O
suffer	O
serious	O
scalability	B-Architecture
problems	O
.	O
</s>
<s>
As	O
well	O
,	O
many	O
systems	O
need	O
to	O
react	O
immediately	O
to	O
online	O
requirements	O
and	O
make	O
recommendations	B-Application
for	O
all	O
users	O
regardless	O
of	O
their	O
millions	O
of	O
users	O
,	O
with	O
most	O
computations	O
happening	O
in	O
very	O
large	O
memory	O
machines	O
.	O
</s>
<s>
Synonyms	B-Application
refers	O
to	O
the	O
tendency	O
of	O
a	O
number	O
of	O
the	O
same	O
or	O
very	O
similar	O
items	O
to	O
have	O
different	O
names	O
or	O
entries	O
.	O
</s>
<s>
Most	O
recommender	B-Application
systems	I-Application
are	O
unable	O
to	O
discover	O
this	O
latent	O
association	O
and	O
thus	O
treat	O
these	O
products	O
differently	O
.	O
</s>
<s>
The	O
prevalence	O
of	O
synonyms	B-Application
decreases	O
the	O
recommendation	O
performance	O
of	O
CF	O
systems	O
.	O
</s>
<s>
Gray	O
sheep	O
refers	O
to	O
the	O
users	O
whose	O
opinions	O
do	O
not	O
consistently	O
agree	O
or	O
disagree	O
with	O
any	O
group	O
of	O
people	O
and	O
thus	O
do	O
not	O
benefit	O
from	O
collaborative	B-Algorithm
filtering	I-Algorithm
.	O
</s>
<s>
Black	O
sheep	O
are	O
a	O
group	O
whose	O
idiosyncratic	O
tastes	O
make	O
recommendations	B-Application
nearly	O
impossible	O
.	O
</s>
<s>
Although	O
this	O
is	O
a	O
failure	O
of	O
the	O
recommender	B-Application
system	I-Application
,	O
non-electronic	O
recommenders	B-Application
also	O
have	O
great	O
problems	O
in	O
these	O
cases	O
,	O
so	O
having	O
black	O
sheep	O
is	O
an	O
acceptable	O
failure	O
.	O
</s>
<s>
In	O
a	O
recommendation	B-Application
system	I-Application
where	O
everyone	O
can	O
give	O
the	O
ratings	O
,	O
people	O
may	O
give	O
many	O
positive	O
ratings	O
for	O
their	O
own	O
items	O
and	O
negative	O
ratings	O
for	O
their	O
competitors	O
 '	O
.	O
</s>
<s>
It	O
is	O
often	O
necessary	O
for	O
the	O
collaborative	B-Algorithm
filtering	I-Algorithm
systems	O
to	O
introduce	O
precautions	O
to	O
discourage	O
such	O
manipulations	O
.	O
</s>
<s>
Collaborative	B-Algorithm
filters	I-Algorithm
are	O
expected	O
to	O
increase	O
diversity	O
because	O
they	O
help	O
us	O
discover	O
new	O
products	O
.	O
</s>
<s>
Because	O
collaborative	B-Algorithm
filters	I-Algorithm
recommend	O
products	O
based	O
on	O
past	O
sales	O
or	O
ratings	O
,	O
they	O
cannot	O
usually	O
recommend	O
products	O
with	O
limited	O
historical	O
data	O
.	O
</s>
<s>
Several	O
collaborative	B-Algorithm
filtering	I-Algorithm
algorithms	O
have	O
been	O
developed	O
to	O
promote	O
diversity	O
and	O
the	O
"	O
long	O
tail	O
"	O
by	O
recommending	O
novel	O
,	O
unexpected	O
,	O
and	O
serendipitous	O
items	O
.	O
</s>
<s>
Cross-System	O
Collaborative	B-Algorithm
Filtering	I-Algorithm
where	O
user	O
profiles	O
across	O
multiple	O
recommender	B-Application
systems	I-Application
are	O
combined	O
in	O
a	O
multitask	O
manner	O
;	O
this	O
way	O
,	O
preference	O
pattern	O
sharing	O
is	O
achieved	O
across	O
models	O
.	O
</s>
<s>
Robust	B-Application
collaborative	I-Application
filtering	I-Application
,	O
where	O
recommendation	O
is	O
stable	O
towards	O
efforts	O
of	O
manipulation	O
.	O
</s>
<s>
User-item	O
matrix	O
is	O
a	O
basic	O
foundation	O
of	O
traditional	O
collaborative	B-Algorithm
filtering	I-Algorithm
techniques	O
,	O
and	O
it	O
suffers	O
from	O
data	O
sparsity	B-Algorithm
problem	O
(	O
i.e.	O
</s>
<s>
cold	B-Application
start	I-Application
)	O
.	O
</s>
<s>
As	O
a	O
consequence	O
,	O
except	O
for	O
user-item	O
matrix	O
,	O
researchers	O
are	O
trying	O
to	O
gather	O
more	O
auxiliary	O
information	O
to	O
help	O
boost	O
recommendation	O
performance	O
and	O
develop	O
personalized	O
recommender	B-Application
systems	I-Application
.	O
</s>
<s>
The	O
interaction-associated	O
information	O
–	O
tags	O
–	O
is	O
taken	O
as	O
a	O
third	O
dimension	O
(	O
in	O
addition	O
to	O
user	O
and	O
item	O
)	O
in	O
advanced	O
collaborative	B-Algorithm
filtering	I-Algorithm
to	O
construct	O
a	O
3-dimensional	O
tensor	B-Device
structure	O
for	O
exploration	O
of	O
recommendation	O
.	O
</s>
