<s>
Robust	B-Application
collaborative	I-Application
filtering	I-Application
,	O
or	O
attack-resistant	O
collaborative	B-Algorithm
filtering	I-Algorithm
,	O
refers	O
to	O
algorithms	O
or	O
techniques	O
that	O
aim	O
to	O
make	O
collaborative	B-Algorithm
filtering	I-Algorithm
more	O
robust	O
against	O
efforts	O
of	O
manipulation	O
,	O
while	O
hopefully	O
maintaining	O
recommendation	O
quality	O
.	O
</s>
<s>
Collaborative	B-Algorithm
filtering	I-Algorithm
predicts	O
a	O
user	O
's	O
rating	O
to	O
items	O
by	O
finding	O
similar	O
users	O
and	O
looking	O
at	O
their	O
ratings	O
,	O
and	O
because	O
it	O
is	O
possible	O
to	O
create	O
nearly	O
indefinite	O
copies	O
of	O
user	O
profiles	O
in	O
an	O
online	O
system	O
,	O
collaborative	B-Algorithm
filtering	I-Algorithm
becomes	O
vulnerable	O
when	O
multiple	O
copies	O
of	O
fake	O
profiles	O
are	O
introduced	O
to	O
the	O
system	O
.	O
</s>
<s>
There	O
are	O
several	O
different	O
approaches	O
suggested	O
to	O
improve	O
robustness	O
of	O
both	O
model-based	O
and	O
memory-based	O
collaborative	B-Algorithm
filtering	I-Algorithm
.	O
</s>
<s>
However	O
,	O
robust	B-Application
collaborative	I-Application
filtering	I-Application
techniques	O
are	O
still	O
an	O
active	O
research	O
field	O
,	O
and	O
major	O
applications	O
of	O
them	O
are	O
yet	O
to	O
come	O
.	O
</s>
<s>
One	O
of	O
the	O
biggest	O
challenges	O
to	O
collaborative	B-Algorithm
filtering	I-Algorithm
is	O
shilling	O
attacks	O
.	O
</s>
<s>
In	O
general	O
,	O
item-based	O
collaborative	B-Algorithm
filtering	I-Algorithm
is	O
known	O
to	O
be	O
more	O
robust	O
than	O
user-based	O
collaborative	B-Algorithm
filtering	I-Algorithm
.	O
</s>
<s>
However	O
,	O
item-based	O
collaborative	B-Algorithm
filtering	I-Algorithm
are	O
still	O
not	O
completely	O
immune	O
to	O
bandwagon	O
and	O
segment	O
attacks	O
.	O
</s>
<s>
Robust	B-Application
collaborative	I-Application
filtering	I-Application
typically	O
works	O
as	O
follows	O
:	O
</s>
<s>
Follow	O
the	O
workflow	O
of	O
regular	O
collaborative	B-Algorithm
filtering	I-Algorithm
system	O
,	O
but	O
only	O
using	O
rating	O
data	O
of	O
non-spam	O
users	O
.	O
</s>
<s>
to	O
make	O
memory-based	O
collaborative	B-Algorithm
filtering	I-Algorithm
more	O
robust	O
.	O
</s>
<s>
Some	O
popular	O
metrics	O
used	O
in	O
collaborative	B-Algorithm
filtering	I-Algorithm
to	O
measure	O
user	O
similarity	O
are	O
Pearson	O
correlation	O
coefficient	O
,	O
interest	O
similarity	O
,	O
and	O
cosine	O
distance	O
.	O
</s>
<s>
The	O
rest	O
of	O
the	O
algorithm	O
works	O
exactly	O
same	O
as	O
normal	O
item-based	O
collaborative	B-Algorithm
filtering	I-Algorithm
.	O
</s>
