<s>
A	O
recommender	B-Application
system	I-Application
,	O
or	O
a	O
recommendation	B-Application
system	I-Application
(	O
sometimes	O
replacing	O
'	O
system	O
 '	O
with	O
a	O
synonym	O
such	O
as	O
platform	O
or	O
engine	O
)	O
,	O
is	O
a	O
subclass	O
of	O
information	B-Application
filtering	I-Application
system	I-Application
that	O
provide	O
suggestions	O
for	O
items	O
that	O
are	O
most	O
pertinent	O
to	O
a	O
particular	O
user	O
.	O
</s>
<s>
Recommender	B-Application
systems	I-Application
are	O
particularly	O
useful	O
when	O
an	O
individual	O
needs	O
to	O
choose	O
an	O
item	O
from	O
a	O
potentially	O
overwhelming	O
number	O
of	O
items	O
that	O
a	O
service	O
may	O
offer	O
.	O
</s>
<s>
Recommender	B-Application
systems	I-Application
are	O
used	O
in	O
a	O
variety	O
of	O
areas	O
,	O
with	O
commonly	O
recognised	O
examples	O
taking	O
the	O
form	O
of	O
playlist	O
generators	O
for	O
video	O
and	O
music	O
services	O
,	O
product	O
recommenders	B-Application
for	O
online	O
stores	O
,	O
or	O
content	O
recommenders	B-Application
for	O
social	O
media	O
platforms	O
and	O
open	O
web	O
content	O
recommenders	B-Application
.	O
</s>
<s>
These	O
systems	O
can	O
operate	O
using	O
a	O
single	O
input	O
,	O
like	O
music	O
,	O
or	O
multiple	O
inputs	O
within	O
and	O
across	O
platforms	O
like	O
news	O
,	O
books	O
and	O
search	O
queries	B-Library
.	O
</s>
<s>
There	O
are	O
also	O
popular	O
recommender	B-Application
systems	I-Application
for	O
specific	O
topics	O
like	O
restaurants	O
and	O
online	O
dating	O
.	O
</s>
<s>
Recommender	B-Application
systems	I-Application
have	O
also	O
been	O
developed	O
to	O
explore	O
research	O
articles	O
and	O
experts	O
,	O
collaborators	O
,	O
and	O
financial	O
services	O
.	O
</s>
<s>
Recommender	B-Application
systems	I-Application
usually	O
make	O
use	O
of	O
either	O
or	O
both	O
collaborative	B-Algorithm
filtering	I-Algorithm
and	O
content-based	O
filtering	O
(	O
also	O
known	O
as	O
the	O
personality-based	O
approach	O
)	O
,	O
as	O
well	O
as	O
other	O
systems	O
such	O
as	O
knowledge-based	B-General_Concept
systems	I-General_Concept
.	O
</s>
<s>
Collaborative	B-Algorithm
filtering	I-Algorithm
approaches	O
build	O
a	O
model	O
from	O
a	O
user	O
's	O
past	O
behavior	O
(	O
items	O
previously	O
purchased	O
or	O
selected	O
and/or	O
numerical	O
ratings	O
given	O
to	O
those	O
items	O
)	O
as	O
well	O
as	O
similar	O
decisions	O
made	O
by	O
other	O
users	O
.	O
</s>
<s>
We	O
can	O
demonstrate	O
the	O
differences	O
between	O
collaborative	O
and	O
content-based	O
filtering	O
by	O
comparing	O
two	O
early	O
music	O
recommender	B-Application
systems	I-Application
–	O
Last.fm	B-Application
and	O
Pandora	B-Application
Radio	I-Application
.	O
</s>
<s>
Last.fm	B-Application
creates	O
a	O
"	O
station	O
"	O
of	O
recommended	O
songs	O
by	O
observing	O
what	O
bands	O
and	O
individual	O
tracks	O
the	O
user	O
has	O
listened	O
to	O
on	O
a	O
regular	O
basis	O
and	O
comparing	O
those	O
against	O
the	O
listening	O
behavior	O
of	O
other	O
users	O
.	O
</s>
<s>
Last.fm	B-Application
will	O
play	O
tracks	O
that	O
do	O
not	O
appear	O
in	O
the	O
user	O
's	O
library	O
,	O
but	O
are	O
often	O
played	O
by	O
other	O
users	O
with	O
similar	O
interests	O
.	O
</s>
<s>
As	O
this	O
approach	O
leverages	O
the	O
behavior	O
of	O
users	O
,	O
it	O
is	O
an	O
example	O
of	O
a	O
collaborative	B-Algorithm
filtering	I-Algorithm
technique	O
.	O
</s>
<s>
Pandora	B-Application
uses	O
the	O
properties	O
of	O
a	O
song	O
or	O
artist	O
(	O
a	O
subset	O
of	O
the	O
400	O
attributes	O
provided	O
by	O
the	O
Music	O
Genome	O
Project	O
)	O
to	O
seed	O
a	O
"	O
station	O
"	O
that	O
plays	O
music	O
with	O
similar	O
properties	O
.	O
</s>
<s>
In	O
the	O
above	O
example	O
,	O
Last.fm	B-Application
requires	O
a	O
large	O
amount	O
of	O
information	O
about	O
a	O
user	O
to	O
make	O
accurate	O
recommendations	O
.	O
</s>
<s>
This	O
is	O
an	O
example	O
of	O
the	O
cold	B-General_Concept
start	I-General_Concept
problem	O
,	O
and	O
is	O
common	O
in	O
collaborative	B-Algorithm
filtering	I-Algorithm
systems	O
.	O
</s>
<s>
Whereas	O
Pandora	B-Application
needs	O
very	O
little	O
information	O
to	O
start	O
,	O
it	O
is	O
far	O
more	O
limited	O
in	O
scope	O
(	O
for	O
example	O
,	O
it	O
can	O
only	O
make	O
recommendations	O
that	O
are	O
similar	O
to	O
the	O
original	O
seed	O
)	O
.	O
</s>
<s>
Recommender	B-Application
systems	I-Application
are	O
a	O
useful	O
alternative	O
to	O
search	B-Application
algorithms	I-Application
since	O
they	O
help	O
users	O
discover	O
items	O
they	O
might	O
not	O
have	O
found	O
otherwise	O
.	O
</s>
<s>
Of	O
note	O
,	O
recommender	B-Application
systems	I-Application
are	O
often	O
implemented	O
using	O
search	O
engines	O
indexing	O
non-traditional	O
data	O
.	O
</s>
<s>
Recommender	B-Application
systems	I-Application
have	O
been	O
the	O
focus	O
of	O
several	O
granted	O
patents	O
.	O
</s>
<s>
Elaine	O
Rich	O
created	O
1979	O
unknowingly	O
the	O
first	O
recommender	B-Application
systems	I-Application
Grundy	O
.	O
</s>
<s>
Montaner	O
provided	O
the	O
first	O
overview	O
of	O
recommender	B-Application
systems	I-Application
from	O
an	O
intelligent	O
agent	O
perspective	O
.	O
</s>
<s>
Adomavicius	O
provided	O
a	O
new	O
,	O
alternate	O
overview	O
of	O
recommender	B-Application
systems	I-Application
.	O
</s>
<s>
Herlocker	O
provides	O
an	O
additional	O
overview	O
of	O
evaluation	O
techniques	O
for	O
recommender	B-Application
systems	I-Application
,	O
and	O
Beel	O
et	O
al	O
.	O
</s>
<s>
have	O
also	O
provided	O
literature	O
surveys	O
on	O
available	O
research	O
paper	O
recommender	B-Application
systems	I-Application
and	O
existing	O
challenges	O
.	O
</s>
<s>
One	O
approach	O
to	O
the	O
design	O
of	O
recommender	B-Application
systems	I-Application
that	O
has	O
wide	O
use	O
is	O
collaborative	B-Algorithm
filtering	I-Algorithm
.	O
</s>
<s>
Collaborative	B-Algorithm
filtering	I-Algorithm
is	O
based	O
on	O
the	O
assumption	O
that	O
people	O
who	O
agreed	O
in	O
the	O
past	O
will	O
agree	O
in	O
the	O
future	O
,	O
and	O
that	O
they	O
will	O
like	O
similar	O
kinds	O
of	O
items	O
as	O
they	O
liked	O
in	O
the	O
past	O
.	O
</s>
<s>
Collaborative	B-Algorithm
filtering	I-Algorithm
methods	O
are	O
classified	O
as	O
memory-based	O
and	O
model-based	O
.	O
</s>
<s>
A	O
well-known	O
example	O
of	O
memory-based	O
approaches	O
is	O
the	O
user-based	O
algorithm	O
,	O
while	O
that	O
of	O
model-based	O
approaches	O
is	O
Matrix	B-Application
factorization	I-Application
(	O
recommender	B-Application
systems	I-Application
)	O
.	O
</s>
<s>
A	O
key	O
advantage	O
of	O
the	O
collaborative	B-Algorithm
filtering	I-Algorithm
approach	O
is	O
that	O
it	O
does	O
not	O
rely	O
on	O
machine	O
analyzable	O
content	O
and	O
therefore	O
it	O
is	O
capable	O
of	O
accurately	O
recommending	O
complex	O
items	O
such	O
as	O
movies	O
without	O
requiring	O
an	O
"	O
understanding	O
"	O
of	O
the	O
item	O
itself	O
.	O
</s>
<s>
Many	O
algorithms	O
have	O
been	O
used	O
in	O
measuring	O
user	O
similarity	O
or	O
item	O
similarity	O
in	O
recommender	B-Application
systems	I-Application
.	O
</s>
<s>
For	O
example	O
,	O
the	O
k-nearest	B-General_Concept
neighbor	I-General_Concept
(	O
k-NN	B-General_Concept
)	O
approach	O
and	O
the	O
Pearson	O
Correlation	O
as	O
first	O
implemented	O
by	O
Allen	O
.	O
</s>
<s>
When	O
building	O
a	O
model	O
from	O
a	O
user	O
's	O
behavior	O
,	O
a	O
distinction	O
is	O
often	O
made	O
between	O
explicit	O
and	O
implicit	B-General_Concept
forms	O
of	O
data	O
collection	O
.	O
</s>
<s>
Asking	O
a	O
user	O
to	O
create	O
a	O
list	O
of	O
items	O
that	O
he/she	O
likes	O
(	O
see	O
Rocchio	B-Algorithm
classification	I-Algorithm
or	O
other	O
similar	O
techniques	O
)	O
.	O
</s>
<s>
Examples	O
of	O
implicit	B-General_Concept
data	I-General_Concept
collection	I-General_Concept
include	O
the	O
following	O
:	O
</s>
<s>
Collaborative	B-Algorithm
filtering	I-Algorithm
approaches	O
often	O
suffer	O
from	O
three	O
problems	O
:	O
cold	B-General_Concept
start	I-General_Concept
,	O
scalability	O
,	O
and	O
sparsity	O
.	O
</s>
<s>
Cold	B-General_Concept
start	I-General_Concept
:	O
For	O
a	O
new	O
user	O
or	O
item	O
,	O
there	O
is	O
n't	O
enough	O
data	O
to	O
make	O
accurate	O
recommendations	O
.	O
</s>
<s>
One	O
of	O
the	O
most	O
famous	O
examples	O
of	O
collaborative	B-Algorithm
filtering	I-Algorithm
is	O
item-to-item	O
collaborative	B-Algorithm
filtering	I-Algorithm
(	O
people	O
who	O
buy	O
x	O
also	O
buy	O
y	O
)	O
,	O
an	O
algorithm	O
popularized	O
by	O
com	O
's	O
recommender	B-Application
system	I-Application
.	O
</s>
<s>
Many	O
social	O
networks	O
originally	O
used	O
collaborative	B-Algorithm
filtering	I-Algorithm
to	O
recommend	O
new	O
friends	O
,	O
groups	O
,	O
and	O
other	O
social	O
connections	O
by	O
examining	O
the	O
network	O
of	O
connections	O
between	O
a	O
user	O
and	O
their	O
friends	O
.	O
</s>
<s>
Collaborative	B-Algorithm
filtering	I-Algorithm
is	O
still	O
used	O
as	O
part	O
of	O
hybrid	O
systems	O
.	O
</s>
<s>
Another	O
common	O
approach	O
when	O
designing	O
recommender	B-Application
systems	I-Application
is	O
content-based	O
filtering	O
.	O
</s>
<s>
Content-based	O
recommenders	B-Application
treat	O
recommendation	O
as	O
a	O
user-specific	O
classification	O
problem	O
and	O
learn	O
a	O
classifier	O
for	O
the	O
user	O
's	O
likes	O
and	O
dislikes	O
based	O
on	O
an	O
item	O
's	O
features	O
.	O
</s>
<s>
This	O
approach	O
has	O
its	O
roots	O
in	O
information	B-Library
retrieval	I-Library
and	O
information	B-Application
filtering	I-Application
research	O
.	O
</s>
<s>
A	O
history	O
of	O
the	O
user	O
's	O
interaction	O
with	O
the	O
recommender	B-Application
system	I-Application
.	O
</s>
<s>
Simple	O
approaches	O
use	O
the	O
average	O
values	O
of	O
the	O
rated	O
item	O
vector	O
while	O
other	O
sophisticated	O
methods	O
use	O
machine	O
learning	O
techniques	O
such	O
as	O
Bayesian	B-General_Concept
Classifiers	I-General_Concept
,	O
cluster	B-Algorithm
analysis	I-Algorithm
,	O
decision	B-Algorithm
trees	I-Algorithm
,	O
and	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
in	O
order	O
to	O
estimate	O
the	O
probability	O
that	O
the	O
user	O
is	O
going	O
to	O
like	O
the	O
item	O
.	O
</s>
<s>
When	O
the	O
system	O
is	O
limited	O
to	O
recommending	O
content	O
of	O
the	O
same	O
type	O
as	O
the	O
user	O
is	O
already	O
using	O
,	O
the	O
value	O
from	O
the	O
recommendation	B-Application
system	I-Application
is	O
significantly	O
less	O
than	O
when	O
other	O
content	O
types	O
from	O
other	O
services	O
can	O
be	O
recommended	O
.	O
</s>
<s>
To	O
overcome	O
this	O
,	O
most	O
content-based	O
recommender	B-Application
systems	I-Application
now	O
use	O
some	O
form	O
of	O
the	O
hybrid	O
system	O
.	O
</s>
<s>
Content-based	O
recommender	B-Application
systems	I-Application
can	O
also	O
include	O
opinion-based	O
recommender	B-Application
systems	I-Application
.	O
</s>
<s>
These	O
user-generated	O
texts	O
are	O
implicit	B-General_Concept
data	O
for	O
the	O
recommender	B-Application
system	I-Application
because	O
they	O
are	O
potentially	O
rich	O
resources	O
of	O
both	O
feature/aspects	O
of	O
the	O
item	O
and	O
users	O
 '	O
evaluation/sentiment	O
to	O
the	O
item	O
.	O
</s>
<s>
Popular	O
approaches	O
of	O
opinion-based	O
recommender	B-Application
system	I-Application
utilize	O
various	O
techniques	O
including	O
text	B-Algorithm
mining	I-Algorithm
,	O
information	B-Library
retrieval	I-Library
,	O
sentiment	O
analysis	O
(	O
see	O
also	O
Multimodal	B-General_Concept
sentiment	I-General_Concept
analysis	I-General_Concept
)	O
and	O
deep	O
learning	O
.	O
</s>
<s>
Most	O
recommender	B-Application
systems	I-Application
now	O
use	O
a	O
hybrid	O
approach	O
,	O
combining	O
collaborative	B-Algorithm
filtering	I-Algorithm
,	O
content-based	O
filtering	O
,	O
and	O
other	O
approaches	O
.	O
</s>
<s>
Hybrid	O
approaches	O
can	O
be	O
implemented	O
in	O
several	O
ways	O
:	O
by	O
making	O
content-based	O
and	O
collaborative-based	O
predictions	O
separately	O
and	O
then	O
combining	O
them	O
;	O
by	O
adding	O
content-based	O
capabilities	O
to	O
a	O
collaborative-based	O
approach	O
(	O
and	O
vice	O
versa	O
)	O
;	O
or	O
by	O
unifying	O
the	O
approaches	O
into	O
one	O
model	O
(	O
see	O
for	O
a	O
complete	O
review	O
of	O
recommender	B-Application
systems	I-Application
)	O
.	O
</s>
<s>
These	O
methods	O
can	O
also	O
be	O
used	O
to	O
overcome	O
some	O
of	O
the	O
common	O
problems	O
in	O
recommender	B-Application
systems	I-Application
such	O
as	O
cold	B-General_Concept
start	I-General_Concept
and	O
the	O
sparsity	O
problem	O
,	O
as	O
well	O
as	O
the	O
knowledge	O
engineering	O
bottleneck	O
in	O
knowledge-based	O
approaches	O
.	O
</s>
<s>
Netflix	O
is	O
a	O
good	O
example	O
of	O
the	O
use	O
of	O
hybrid	O
recommender	B-Application
systems	I-Application
.	O
</s>
<s>
The	O
website	O
makes	O
recommendations	O
by	O
comparing	O
the	O
watching	O
and	O
searching	O
habits	O
of	O
similar	O
users	O
(	O
i.e.	O
,	O
collaborative	B-Algorithm
filtering	I-Algorithm
)	O
as	O
well	O
as	O
by	O
offering	O
movies	O
that	O
share	O
characteristics	O
with	O
films	O
that	O
a	O
user	O
has	O
rated	O
highly	O
(	O
content-based	O
filtering	O
)	O
.	O
</s>
<s>
Mixed	O
:	O
Recommendations	O
from	O
different	O
recommenders	B-Application
are	O
presented	O
together	O
to	O
give	O
the	O
recommendation	O
.	O
</s>
<s>
Feature	O
Combination	O
:	O
Features	O
derived	O
from	O
different	O
knowledge	O
sources	O
are	O
combined	O
together	O
and	O
given	O
to	O
a	O
single	O
recommendation	B-Application
algorithm	I-Application
.	O
</s>
<s>
Cascade	O
:	O
Recommenders	B-Application
are	O
given	O
strict	O
priority	O
,	O
with	O
the	O
lower	O
priority	O
ones	O
breaking	O
ties	O
in	O
the	O
scoring	O
of	O
the	O
higher	O
ones	O
.	O
</s>
<s>
These	O
recommender	B-Application
systems	I-Application
use	O
the	O
interactions	O
of	O
a	O
user	O
within	O
a	O
session	O
to	O
generate	O
recommendations	O
.	O
</s>
<s>
Session-based	O
recommender	B-Application
systems	I-Application
are	O
used	O
at	O
Youtube	O
and	O
Amazon	B-Application
.	O
</s>
<s>
Most	O
instances	O
of	O
session-based	O
recommender	B-Application
systems	I-Application
rely	O
on	O
the	O
sequence	O
of	O
recent	O
interactions	O
within	O
a	O
session	O
without	O
requiring	O
any	O
additional	O
details	O
(	O
historical	O
,	O
demographic	O
)	O
of	O
the	O
user	O
.	O
</s>
<s>
The	O
recommendation	O
problem	O
can	O
be	O
seen	O
as	O
a	O
special	O
instance	O
of	O
a	O
reinforcement	O
learning	O
problem	O
whereby	O
the	O
user	O
is	O
the	O
environment	O
upon	O
which	O
the	O
agent	O
,	O
the	O
recommendation	B-Application
system	I-Application
acts	O
upon	O
in	O
order	O
to	O
receive	O
a	O
reward	O
,	O
for	O
instance	O
,	O
a	O
click	O
or	O
engagement	O
by	O
the	O
user	O
.	O
</s>
<s>
One	O
aspect	O
of	O
reinforcement	O
learning	O
that	O
is	O
of	O
particular	O
use	O
in	O
the	O
area	O
of	O
recommender	B-Application
systems	I-Application
is	O
the	O
fact	O
that	O
the	O
models	O
or	O
policies	O
can	O
be	O
learned	O
by	O
providing	O
a	O
reward	O
to	O
the	O
recommendation	O
agent	O
.	O
</s>
<s>
Multi-criteria	O
recommender	B-Application
systems	I-Application
(	O
MCRS	O
)	O
can	O
be	O
defined	O
as	O
recommender	B-Application
systems	I-Application
that	O
incorporate	O
preference	O
information	O
upon	O
multiple	O
criteria	O
.	O
</s>
<s>
The	O
majority	O
of	O
existing	O
approaches	O
to	O
recommender	B-Application
systems	I-Application
focus	O
on	O
recommending	O
the	O
most	O
relevant	O
content	O
to	O
users	O
using	O
contextual	O
information	O
,	O
yet	O
do	O
not	O
take	O
into	O
account	O
the	O
risk	O
of	O
disturbing	O
the	O
user	O
with	O
unwanted	O
notifications	O
.	O
</s>
<s>
Therefore	O
,	O
the	O
performance	O
of	O
the	O
recommender	B-Application
system	I-Application
depends	O
in	O
part	O
on	O
the	O
degree	O
to	O
which	O
it	O
has	O
incorporated	O
the	O
risk	O
into	O
the	O
recommendation	O
process	O
.	O
</s>
<s>
Mobile	O
recommender	B-Application
systems	I-Application
make	O
use	O
of	O
internet-accessing	O
smart	B-Application
phones	I-Application
to	O
offer	O
personalized	O
,	O
context-sensitive	O
recommendations	O
.	O
</s>
<s>
This	O
is	O
a	O
particularly	O
difficult	O
area	O
of	O
research	O
as	O
mobile	O
data	O
is	O
more	O
complex	O
than	O
data	O
that	O
recommender	B-Application
systems	I-Application
often	O
have	O
to	O
deal	O
with	O
.	O
</s>
<s>
There	O
are	O
three	O
factors	O
that	O
could	O
affect	O
the	O
mobile	O
recommender	B-Application
systems	I-Application
and	O
the	O
accuracy	O
of	O
prediction	O
results	O
:	O
the	O
context	O
,	O
the	O
recommendation	O
method	O
and	O
privacy	O
.	O
</s>
<s>
Additionally	O
,	O
mobile	O
recommender	B-Application
systems	I-Application
suffer	O
from	O
a	O
transplantation	O
problem	O
–	O
recommendations	O
may	O
not	O
apply	O
in	O
all	O
regions	O
(	O
for	O
instance	O
,	O
it	O
would	O
be	O
unwise	O
to	O
recommend	O
a	O
recipe	O
in	O
an	O
area	O
where	O
all	O
of	O
the	O
ingredients	O
may	O
not	O
be	O
available	O
)	O
.	O
</s>
<s>
One	O
example	O
of	O
a	O
mobile	O
recommender	B-Application
system	I-Application
are	O
the	O
approaches	O
taken	O
by	O
companies	O
such	O
as	O
Uber	B-Application
and	O
Lyft	B-Application
to	O
generate	O
driving	O
routes	O
for	O
taxi	O
drivers	O
in	O
a	O
city	O
.	O
</s>
<s>
One	O
of	O
the	O
events	O
that	O
energized	O
research	O
in	O
recommender	B-Application
systems	I-Application
was	O
the	O
Netflix	O
Prize	O
.	O
</s>
<s>
From	O
2006	O
to	O
2009	O
,	O
Netflix	O
sponsored	O
a	O
competition	O
,	O
offering	O
a	O
grand	O
prize	O
of	O
$	O
1,000,000	O
to	O
the	O
team	O
that	O
could	O
take	O
an	O
offered	O
dataset	O
of	O
over	O
100	O
million	O
movie	O
ratings	O
and	O
return	O
recommendations	O
that	O
were	O
10%	O
more	O
accurate	O
than	O
those	O
offered	O
by	O
the	O
company	O
's	O
existing	O
recommender	B-Application
system	I-Application
.	O
</s>
<s>
Some	O
members	O
from	O
the	O
team	O
that	O
finished	O
second	O
place	O
founded	O
Gravity	O
R&D	O
,	O
a	O
recommendation	B-Application
engine	I-Application
that	O
's	O
active	O
in	O
the	O
RecSys	O
community	O
.	O
</s>
<s>
Evaluation	O
is	O
important	O
in	O
assessing	O
the	O
effectiveness	O
of	O
recommendation	B-Application
algorithms	I-Application
.	O
</s>
<s>
To	O
measure	O
the	O
effectiveness	O
of	O
recommender	B-Application
systems	I-Application
,	O
and	O
compare	O
different	O
approaches	O
,	O
three	O
types	O
of	O
evaluations	O
are	O
available	O
:	O
user	O
studies	O
,	O
online	O
evaluations	O
(	O
A/B	O
tests	O
)	O
,	O
and	O
offline	O
evaluations	O
.	O
</s>
<s>
The	O
commonly	O
used	O
metrics	O
are	O
the	O
mean	B-Algorithm
squared	I-Algorithm
error	I-Algorithm
and	O
root	B-General_Concept
mean	I-General_Concept
squared	I-General_Concept
error	I-General_Concept
,	O
the	O
latter	O
having	O
been	O
used	O
in	O
the	O
Netflix	O
Prize	O
.	O
</s>
<s>
The	O
information	B-Library
retrieval	I-Library
metrics	O
such	O
as	O
precision	O
and	O
recall	O
or	O
DCG	O
are	O
useful	O
to	O
assess	O
the	O
quality	O
of	O
a	O
recommendation	O
method	O
.	O
</s>
<s>
Evaluating	O
the	O
performance	O
of	O
a	O
recommendation	B-Application
algorithm	I-Application
on	O
a	O
fixed	O
test	O
dataset	O
will	O
always	O
be	O
extremely	O
challenging	O
as	O
it	O
is	O
impossible	O
to	O
accurately	O
predict	O
the	O
reactions	O
of	O
real	O
users	O
to	O
the	O
recommendations	O
.	O
</s>
<s>
In	O
A/B	O
tests	O
,	O
recommendations	O
are	O
shown	O
to	O
typically	O
thousands	O
of	O
users	O
of	O
a	O
real	O
product	O
,	O
and	O
the	O
recommender	B-Application
system	I-Application
randomly	O
picks	O
at	O
least	O
two	O
different	O
recommendation	O
approaches	O
to	O
generate	O
recommendations	O
.	O
</s>
<s>
The	O
effectiveness	O
is	O
measured	O
with	O
implicit	B-General_Concept
measures	O
of	O
effectiveness	O
such	O
as	O
conversion	O
rate	O
or	O
click-through	O
rate	O
.	O
</s>
<s>
For	O
instance	O
,	O
in	O
the	O
domain	O
of	O
citation	O
recommender	B-Application
systems	I-Application
,	O
users	O
typically	O
do	O
not	O
rate	O
a	O
citation	O
or	O
recommended	O
article	O
.	O
</s>
<s>
In	O
such	O
cases	O
,	O
offline	O
evaluations	O
may	O
use	O
implicit	B-General_Concept
measures	O
of	O
effectiveness	O
.	O
</s>
<s>
For	O
instance	O
,	O
it	O
may	O
be	O
assumed	O
that	O
a	O
recommender	B-Application
system	I-Application
is	O
effective	O
that	O
is	O
able	O
to	O
recommend	O
as	O
many	O
articles	O
as	O
possible	O
that	O
are	O
contained	O
in	O
a	O
research	O
article	O
's	O
reference	O
list	O
.	O
</s>
<s>
Typically	O
,	O
research	O
on	O
recommender	B-Application
systems	I-Application
is	O
concerned	O
with	O
finding	O
the	O
most	O
accurate	O
recommendation	B-Application
algorithms	I-Application
.	O
</s>
<s>
Recommender	B-Application
persistence	O
–	O
In	O
some	O
situations	O
,	O
it	O
is	O
more	O
effective	O
to	O
re-show	O
recommendations	O
,	O
or	O
let	O
users	O
re-rate	O
items	O
,	O
than	O
showing	O
new	O
items	O
.	O
</s>
<s>
Privacy	O
–	O
Recommender	B-Application
systems	I-Application
usually	O
have	O
to	O
deal	O
with	O
privacy	O
concerns	O
because	O
users	O
have	O
to	O
reveal	O
sensitive	O
information	O
.	O
</s>
<s>
Building	O
user	O
profiles	O
using	O
collaborative	B-Algorithm
filtering	I-Algorithm
can	O
be	O
problematic	O
from	O
a	O
privacy	O
point	O
of	O
view	O
.	O
</s>
<s>
Many	O
European	O
countries	O
have	O
a	O
strong	O
culture	O
of	O
data	O
privacy	O
,	O
and	O
every	O
attempt	O
to	O
introduce	O
any	O
level	O
of	O
user	O
profiling	B-Algorithm
can	O
result	O
in	O
a	O
negative	O
customer	O
response	O
.	O
</s>
<s>
Robustness	O
–	O
When	O
users	O
can	O
participate	O
in	O
the	O
recommender	B-Application
system	I-Application
,	O
the	O
issue	O
of	O
fraud	O
must	O
be	O
addressed	O
.	O
</s>
<s>
Serendipity	B-Protocol
–	O
Serendipity	B-Protocol
is	O
a	O
measure	O
of	O
"	O
how	O
surprising	O
the	O
recommendations	O
are	O
"	O
.	O
</s>
<s>
For	O
instance	O
,	O
a	O
recommender	B-Application
system	I-Application
that	O
recommends	O
milk	O
to	O
a	O
customer	O
in	O
a	O
grocery	O
store	O
might	O
be	O
perfectly	O
accurate	O
,	O
but	O
it	O
is	O
not	O
a	O
good	O
recommendation	O
because	O
it	O
is	O
an	O
obvious	O
item	O
for	O
the	O
customer	O
to	O
buy	O
.	O
</s>
<s>
"[Serendipity]	O
serves	O
two	O
purposes	O
:	O
First	O
,	O
the	O
chance	O
that	O
users	O
lose	O
interest	O
because	O
the	O
choice	O
set	O
is	O
too	O
uniform	O
decreases	O
.	O
</s>
<s>
Trust	O
–	O
A	O
recommender	B-Application
system	I-Application
is	O
of	O
little	O
value	O
for	O
a	O
user	O
if	O
the	O
user	O
does	O
not	O
trust	O
the	O
system	O
.	O
</s>
<s>
Trust	O
can	O
be	O
built	O
by	O
a	O
recommender	B-Application
system	I-Application
by	O
explaining	O
how	O
it	O
generates	O
recommendations	O
,	O
and	O
why	O
it	O
recommends	O
an	O
item	O
.	O
</s>
<s>
Recommender	B-Application
systems	I-Application
are	O
notoriously	O
difficult	O
to	O
evaluate	O
offline	O
,	O
with	O
some	O
researchers	O
claiming	O
that	O
this	O
has	O
led	O
to	O
a	O
reproducibility	O
crisis	O
in	O
recommender	B-Application
systems	I-Application
publications	O
.	O
</s>
<s>
In	O
the	O
context	O
of	O
recommender	B-Application
systems	I-Application
a	O
2019	O
paper	O
surveyed	O
a	O
small	O
number	O
of	O
hand-picked	O
publications	O
applying	O
deep	O
learning	O
or	O
neural	O
methods	O
to	O
the	O
top-k	O
recommendation	O
problem	O
,	O
published	O
in	O
top	O
conferences	O
(	O
SIGIR	O
,	O
KDD	O
,	O
WWW	O
,	O
RecSys	O
,	O
IJCAI	O
)	O
,	O
has	O
shown	O
that	O
on	O
average	O
less	O
than	O
40%	O
of	O
articles	O
could	O
be	O
reproduced	O
by	O
the	O
authors	O
of	O
the	O
survey	O
,	O
with	O
as	O
little	O
as	O
14%	O
in	O
some	O
conferences	O
.	O
</s>
<s>
Deep	O
learning	O
and	O
neural	O
methods	O
for	O
recommender	B-Application
systems	I-Application
have	O
been	O
used	O
in	O
the	O
winning	O
solutions	O
in	O
several	O
recent	O
recommender	B-Application
system	I-Application
challenges	O
,	O
WSDM	O
,	O
RecSys	O
Challenge	O
.	O
</s>
<s>
The	O
topic	O
of	O
reproducibility	O
is	O
not	O
new	O
in	O
recommender	B-Application
systems	I-Application
.	O
</s>
<s>
criticized	O
that	O
"	O
it	O
is	O
currently	O
difficult	O
to	O
reproduce	O
and	O
extend	O
recommender	B-Application
systems	I-Application
research	O
results	O
,	O
”	O
and	O
that	O
evaluations	O
are	O
“	O
not	O
handled	O
consistently	O
"	O
.	O
</s>
<s>
Konstan	O
and	O
Adomavicius	O
conclude	O
that	O
"	O
the	O
Recommender	B-Application
Systems	I-Application
research	O
community	O
is	O
facing	O
a	O
crisis	O
where	O
a	O
significant	O
number	O
of	O
papers	O
present	O
results	O
that	O
contribute	O
little	O
to	O
collective	O
knowledge	O
 [ … ] 	O
often	O
because	O
the	O
research	O
lacks	O
the	O
 [ … ] 	O
evaluation	O
to	O
be	O
properly	O
judged	O
and	O
,	O
hence	O
,	O
to	O
provide	O
meaningful	O
contributions.	O
"	O
</s>
<s>
As	O
a	O
consequence	O
,	O
much	O
research	O
about	O
recommender	B-Application
systems	I-Application
can	O
be	O
considered	O
as	O
not	O
reproducible	O
.	O
</s>
<s>
Hence	O
,	O
operators	O
of	O
recommender	B-Application
systems	I-Application
find	O
little	O
guidance	O
in	O
the	O
current	O
research	O
for	O
answering	O
the	O
question	O
,	O
which	O
recommendation	O
approaches	O
to	O
use	O
in	O
a	O
recommender	B-Application
systems	I-Application
.	O
</s>
<s>
Some	O
researchers	O
demonstrated	O
that	O
minor	O
variations	O
in	O
the	O
recommendation	B-Application
algorithms	I-Application
or	O
scenarios	O
led	O
to	O
strong	O
changes	O
in	O
the	O
effectiveness	O
of	O
a	O
recommender	B-Application
system	I-Application
.	O
</s>
<s>
They	O
conclude	O
that	O
seven	O
actions	O
are	O
necessary	O
to	O
improve	O
the	O
current	O
situation	O
:	O
"	O
(	O
1	O
)	O
survey	O
other	O
research	O
fields	O
and	O
learn	O
from	O
them	O
,	O
(	O
2	O
)	O
find	O
a	O
common	O
understanding	O
of	O
reproducibility	O
,	O
(	O
3	O
)	O
identify	O
and	O
understand	O
the	O
determinants	O
that	O
affect	O
reproducibility	O
,	O
(	O
4	O
)	O
conduct	O
more	O
comprehensive	O
experiments	O
(	O
5	O
)	O
modernize	O
publication	O
practices	O
,	O
(	O
6	O
)	O
foster	O
the	O
development	O
and	O
use	O
of	O
recommendation	O
frameworks	O
,	O
and	O
(	O
7	O
)	O
establish	O
best-practice	O
guidelines	O
for	O
recommender-systems	O
research.	O
"	O
</s>
