<s>
In	O
predictive	B-General_Concept
analytics	I-General_Concept
,	O
data	O
science	O
,	O
machine	O
learning	O
and	O
related	O
fields	O
,	O
concept	B-Algorithm
drift	I-Algorithm
or	O
drift	O
is	O
an	O
evolution	O
of	O
data	O
that	O
invalidates	O
the	O
data	B-Application
model	I-Application
.	O
</s>
<s>
Drift	O
detection	O
and	O
drift	O
adaptation	O
are	O
of	O
paramount	O
importance	O
in	O
the	O
fields	O
that	O
involve	O
dynamically	O
changing	O
data	O
and	O
data	B-Application
models	I-Application
.	O
</s>
<s>
In	O
machine	O
learning	O
and	O
predictive	B-General_Concept
analytics	I-General_Concept
this	O
drift	O
phenomenon	O
is	O
called	O
concept	B-Algorithm
drift	I-Algorithm
.	O
</s>
<s>
In	O
machine	O
learning	O
,	O
a	O
common	O
element	O
of	O
a	O
data	B-Application
model	I-Application
are	O
the	O
statistical	O
properties	O
,	O
such	O
as	O
probability	O
distribution	O
of	O
the	O
actual	O
data	O
.	O
</s>
<s>
Another	O
important	O
area	O
is	O
software	B-General_Concept
engineering	I-General_Concept
,	O
where	O
three	O
types	O
of	O
data	O
drift	O
affecting	O
data	O
fidelity	O
may	O
be	O
recognized	O
.	O
</s>
<s>
"	O
Structural	O
drift	O
"	O
happens	O
when	O
the	O
data	O
schema	B-Application
changes	O
,	O
which	O
may	O
invalidate	O
databases	O
.	O
</s>
<s>
In	O
the	O
case	O
of	O
cloud	B-Architecture
computing	I-Architecture
,	O
infrastructure	O
drift	O
that	O
may	O
affect	O
the	O
applications	O
running	O
on	O
cloud	O
may	O
be	O
caused	O
by	O
the	O
updates	O
of	O
cloud	O
software	O
.	O
</s>
<s>
Data	O
loss	O
happens	O
when	O
valid	O
data	O
are	O
ignored	O
due	O
to	O
non-conformance	O
with	O
the	O
applied	O
schema	B-Application
.	O
</s>
<s>
A	O
simple	O
drift	O
detection	O
is	O
to	O
run	O
checksum	B-Algorithm
regularly	O
.	O
</s>
<s>
For	O
example	O
,	O
if	O
weekly	O
merchandise	O
sales	O
are	O
to	O
be	O
predicted	O
,	O
and	O
a	O
predictive	B-General_Concept
model	I-General_Concept
has	O
been	O
developed	O
that	O
works	O
satisfactorily	O
.	O
</s>
<s>
The	O
model	O
is	O
likely	O
to	O
become	O
less	O
and	O
less	O
accurate	O
over	O
time	O
–	O
this	O
is	O
concept	B-Algorithm
drift	I-Algorithm
.	O
</s>
<s>
In	O
the	O
merchandise	O
sales	O
application	O
,	O
one	O
reason	O
for	O
concept	B-Algorithm
drift	I-Algorithm
may	O
be	O
seasonality	O
,	O
which	O
means	O
that	O
shopping	O
behavior	O
changes	O
seasonally	O
.	O
</s>
<s>
Concept	B-Algorithm
drift	I-Algorithm
generally	O
occurs	O
when	O
the	O
covariates	O
that	O
comprise	O
the	O
data	O
set	O
begin	O
to	O
explain	O
the	O
variation	O
of	O
your	O
target	O
set	O
less	O
accurately	O
—	O
there	O
may	O
be	O
some	O
confounding	O
variables	O
that	O
have	O
emerged	O
,	O
and	O
that	O
one	O
simply	O
cannot	O
account	O
for	O
,	O
which	O
renders	O
the	O
model	O
accuracy	O
to	O
progressively	O
decrease	O
with	O
time	O
.	O
</s>
<s>
Generally	O
,	O
it	O
is	O
advised	O
to	O
perform	O
health	O
checks	O
as	O
part	O
of	O
the	O
post-production	O
analysis	O
and	O
to	O
re-train	O
the	O
model	O
with	O
new	O
assumptions	O
upon	O
signs	O
of	O
concept	B-Algorithm
drift	I-Algorithm
.	O
</s>
<s>
To	O
prevent	O
deterioration	O
in	O
prediction	O
accuracy	O
because	O
of	O
concept	B-Algorithm
drift	I-Algorithm
,	O
reactive	O
and	O
tracking	O
solutions	O
can	O
be	O
adopted	O
.	O
</s>
<s>
Reactive	O
solutions	O
retrain	O
the	O
model	O
in	O
reaction	O
to	O
a	O
triggering	O
mechanism	O
,	O
such	O
as	O
a	O
change-detection	O
test	O
,	O
to	O
explicitly	O
detect	O
concept	B-Algorithm
drift	I-Algorithm
as	O
a	O
change	O
in	O
the	O
statistics	O
of	O
the	O
data-generating	O
process	O
.	O
</s>
<s>
When	O
concept	B-Algorithm
drift	I-Algorithm
is	O
detected	O
,	O
the	O
current	O
model	O
is	O
no	O
longer	O
up-to-date	O
and	O
must	O
be	O
replaced	O
by	O
a	O
new	O
one	O
to	O
restore	O
prediction	O
accuracy	O
.	O
</s>
<s>
Methods	O
for	O
achieving	O
this	O
include	O
online	B-Algorithm
machine	I-Algorithm
learning	I-Algorithm
,	O
frequent	O
retraining	O
on	O
the	O
most	O
recently	O
observed	O
samples	O
,	O
and	O
maintaining	O
an	O
ensemble	O
of	O
classifiers	O
where	O
one	O
new	O
classifier	O
is	O
trained	O
on	O
the	O
most	O
recent	O
batch	O
of	O
examples	O
and	O
replaces	O
the	O
oldest	O
classifier	O
in	O
the	O
ensemble	O
.	O
</s>
<s>
Contextual	O
information	O
,	O
when	O
available	O
,	O
can	O
be	O
used	O
to	O
better	O
explain	O
the	O
causes	O
of	O
the	O
concept	B-Algorithm
drift	I-Algorithm
:	O
for	O
instance	O
,	O
in	O
the	O
sales	O
prediction	O
application	O
,	O
concept	B-Algorithm
drift	I-Algorithm
might	O
be	O
compensated	O
by	O
adding	O
information	O
about	O
the	O
season	O
to	O
the	O
model	O
.	O
</s>
<s>
By	O
providing	O
information	O
about	O
the	O
time	O
of	O
the	O
year	O
,	O
the	O
rate	O
of	O
deterioration	O
of	O
your	O
model	O
is	O
likely	O
to	O
decrease	O
,	O
concept	B-Algorithm
drift	I-Algorithm
is	O
unlikely	O
to	O
be	O
eliminated	O
altogether	O
.	O
</s>
<s>
Concept	B-Algorithm
drift	I-Algorithm
cannot	O
be	O
avoided	O
for	O
complex	O
phenomena	O
that	O
are	O
not	O
governed	O
by	O
fixed	O
laws	O
of	O
nature	O
.	O
</s>
<s>
All	O
processes	O
that	O
arise	O
from	O
human	O
activity	O
,	O
such	O
as	O
socioeconomic	O
processes	O
,	O
and	O
biological	O
processes	O
are	O
likely	O
to	O
experience	O
concept	B-Algorithm
drift	I-Algorithm
.	O
</s>
<s>
:	O
An	O
open-source	O
Python	B-Language
library	O
for	O
detecting	O
univariate	B-General_Concept
and	O
multivariate	O
distribution	O
drift	O
and	O
estimating	O
machine	O
learning	O
model	O
performance	O
without	O
ground	O
truth	O
labels	O
.	O
</s>
<s>
RapidMiner	B-Algorithm
:	O
Formerly	O
Yet	O
Another	O
Learning	O
Environment	O
(	O
YALE	O
)	O
:	O
free	O
open-source	O
software	O
for	O
knowledge	O
discovery	O
,	O
data	B-Application
mining	I-Application
,	O
and	O
machine	O
learning	O
also	O
featuring	O
data	B-Algorithm
stream	I-Algorithm
mining	I-Algorithm
,	O
learning	O
time-varying	O
concepts	O
,	O
and	O
tracking	O
drifting	O
concept	O
.	O
</s>
<s>
It	O
is	O
used	O
in	O
combination	O
with	O
its	O
data	B-Algorithm
stream	I-Algorithm
mining	I-Algorithm
plugin	O
(	O
formerly	O
concept	B-Algorithm
drift	I-Algorithm
plugin	O
)	O
.	O
</s>
<s>
EDDM	O
(	O
)	O
:	O
free	O
open-source	O
implementation	O
of	O
drift	O
detection	O
methods	O
in	O
Weka	B-Language
.	O
</s>
<s>
MOA	B-Language
(	O
Massive	B-Language
Online	I-Language
Analysis	I-Language
)	O
:	O
free	O
open-source	O
software	O
specific	O
for	O
mining	O
data	O
streams	O
with	O
concept	B-Algorithm
drift	I-Algorithm
.	O
</s>
<s>
It	O
contains	O
a	O
prequential	O
evaluation	O
method	O
,	O
the	O
EDDM	O
concept	B-Algorithm
drift	I-Algorithm
methods	O
,	O
a	O
reader	O
of	O
ARFF	O
real	O
datasets	O
,	O
and	O
artificial	O
stream	O
generators	O
as	O
SEA	O
concepts	O
,	O
STAGGER	O
,	O
rotating	O
hyperplane	O
,	O
random	O
tree	O
,	O
and	O
random	O
radius	O
based	O
functions	O
.	O
</s>
<s>
MOA	B-Language
supports	O
bi-directional	O
interaction	O
with	O
Weka	B-Language
.	O
</s>
<s>
USP	O
Data	O
Stream	O
Repository	O
,	O
27	O
real-world	O
stream	O
datasets	O
with	O
concept	B-Algorithm
drift	I-Algorithm
compiled	O
by	O
Souza	O
et	O
al	O
.	O
</s>
<s>
Sensor	O
stream	O
and	O
Power	O
supply	O
stream	O
datasets	O
are	O
available	O
from	O
X	O
.	O
Zhu	O
's	O
Stream	O
Data	B-Application
Mining	I-Application
Repository	O
.	O
</s>
<s>
Text	B-Algorithm
mining	I-Algorithm
,	O
a	O
collection	O
of	O
text	B-Algorithm
mining	I-Algorithm
datasets	O
with	O
concept	B-Algorithm
drift	I-Algorithm
,	O
maintained	O
by	O
I	O
.	O
Katakis	O
.	O
</s>
<s>
It	O
is	O
often	O
used	O
as	O
a	O
benchmark	O
to	O
evaluate	O
handling	O
concept	B-Algorithm
drift	I-Algorithm
.	O
</s>
<s>
Real-World	O
Challenges	O
for	O
Data	B-Algorithm
Stream	I-Algorithm
Mining	I-Algorithm
Workshop-Discussion	O
at	O
the	O
ECML	O
PKDD	O
2013	O
,	O
Prague	O
,	O
Czech	O
Republic	O
.	O
</s>
