►
Description
Analyzing the locations of crimes in San Francisco using temporal crime data since 2003.
http://nupic2015spring.challengepost.com/submissions/37838-sf-crimes
A
So
if
you
feel
bad
about
your
preparedness,
just
watch
okay,
so
what
I
did
was
I
took
a
hack
that
Austin
Marshall
did
at
the
last
hackathon,
where
he
platted
he
plotted
earthquake
historical
earthquake
data
and
showed
anomaly.
Score
is
based
on
their
geographical
coordinates
and
the
strength
of
the
earthquake
and
I
did
something
somewhat
similar
with
crime
data
from
San
Francisco.
B
B
A
Sure,
there's
a
there's
local
correlations
to
these
crimes,
but
but
it,
but
it
is
at
least
it's
an
example
of
taking
a
large
data
set.
This
is
over
a
million
1,700,000
crimes
over
the
past
15
12
years,
or
something
San
Francisco
and
encoding,
the
the
latitude
and
longitude
and
the
time
so
I've
got
time
of
day
day
of
week
weekend
vs
weekday.
A
So
but
it's
not
that
impressive,
because
I
don't
think
the
anomaly
scores
mean
a
whole
lot
in
this
case,
just
because
there's
not
really
a
decent
association,
because
the
sea
I've
thought
about
that
using
the
seriousness
of
the
crime,
because
if
T,
that's
true
well,
the
data
set
I'm
going
to
jump
them
to
give
you
a
mic.
Because
you're
going
to
have
a
lot
of
comments,
the
the
dataset
had
quite
a
bit
of
information.
It
had
the
area
of
the
town.
So
if
I
had
more
time,
there's
some
other
interesting
things.
I
could
do
it.
A
Had
the
police
precinct
the
area
of
the
town
that
the
crime
occurred,
the
police
precinct
that
was
dispatched
to
deal
with
the
crime,
the
categorization
categorization
of
the
crime
and
then
a
short
description.
So
I
could
separate
this
out.
You
know
petty
larceny
versus
murderer
versus
drug
offenses.
That
sort
of
thing
it
also
segregated
out
so
that
each
difference
like
area
of
the
city
might
have
its
own
model,
so
pass
one
to
each
each
model.
But
I
didn't
have
a
lot
of
time,
so
this
is
basically
it.
A
A
So
it's
a
it's
an
interesting,
an
interesting
data
set
to
do
something
with
it.
This
is
about
as
far
as
I
got
with
it.
So,
but
it's
something
I
did
something.
B
A
C
B
C
A
C
C
A
couple
of
really
simple
things,
but
changing
the
learning
rate.
Might
you
know
we
get
over
this
fact
that
after
a
while,
it's
seeing
everything-
because
you
really
don't
want
to
know
the
history
10
years
ago-
probably
doesn't
really
relate
to
what's
going
on.
You
know
the
last
few
months
right
agreed.
D
B
A
D
Time
you
might
want
to
I
think
this
could
be
a
really
fun
data
set
to
play
with
in
the
future
as
well.
But
one
thing
that
would
be
really
cool
is
not
an
anomaly
model,
but
a
prediction
model
yeah
and
do
it
/
neighborhood,
maybe-
and
if
you
you
know,
I
would
do
it
as
like
the
number
of
crimes
in
this
region
per
hour
or
per
day
or
something
you
have
a
lot
of
data
and
see.
If
you
can
predict
when
the
next
crime
is
going
to
happen
in
that
area
and
yeah.
B
C
That's
a
great
idea,
of
course.
You
know
what
you
wouldn't
expect
to
predict
where
the
next
crime
can
happen,
but
you
could
say:
there's
a
you:
have
an
average
an
expected
average
for
a
region,
and
then
you
could
say,
could
we
do
better
than
that?
Let's
say
the
marina
district
has
on
average
thirty
percent
of
the
crime,
something
like
that.
C
What
you'd
want
to
do
is
say
as
it
predicting
it
would
give
you
a
likelihood
score
for
you
know
like
a
thirty
percent
per
chance
in
a
crime
in
that
area
or
forty
percent
chance,
and
you
could
score
it
over
time
and
see
if
you're
beat
how
much
you
beat
the
the
just.
You
know
the
regular
average
type
of
thing
you
know,
yeah.
E
A
Yeah,
it
didn't
have
anything
like
that.
All
it
had
was
like
neighborhood
area
of
town,
in
addition
to
like
they
had
a
street
area
latitude
and
longitude,
which
is
what
I
used
in
categories.
So
there's
more
information,
you
can
do
more
interesting
stuff
with
I.
Just
didn't
have
time
to
do
it.
So,
okay,
any
other
questions
and
I'll
move
on.
If
not
all
right.
The
next
one
is
a.