►
From YouTube: AutoDJ [DEMO #12] (2014 Fall NuPIC Hackathon)
Description
By George London
"I'm extracting my track-by-track music listening history from Last.fm and then seeing if NUPIC can predict what artist I'm going to listen to next given a sequence of my previous listens."
A
All
right,
so
I'm
george
in
my.
A
Music
recommendation,
one
of
the
problems
of
music
recommendation-
is
that
almost
all
existing
systems
work
by
trying
to
figure
out
your
preference
for
a
bunch
of
different
songs.
Assuming
you
have
the
same
preference,
no
matter
kind
of
what
time
it
is
what
you're
feeling?
What
you're
doing
so.
One
of
the
things
I'm
trying
to
do
is
to
make
music
recommendation
a
bit
more
context
sensitive.
A
A
B
A
The
artist
the
name
of
the
track,
my
name
and
then
a
timestamp.
A
A
Evaluate
quite
yet,
if
that's
a
reasonable
prediction,
but
that's
what
it
says,
I
think
that
for
actually
making
this
more
reasonable,
it
would
help
to
provide
it
with
better
context.
Data
like
perhaps
instead
of
feeding
in
artist
names
to
feed
in
some
kind
of
vector
of
mood
of
the
songs
or
possibly
just
to
feed
in
spectrograms.
For
each
thing,
I
think,
if
I
fit
in
the
spectrogram
of
the
time,
stamp
could
probably
come
up
with
something
pretty
interesting.
C
Actually,
if
anybody's
interested
in
this
program,
pandora's
internal
data
structure
resembles
sdrs
greatly.
Basically,
the
way
they
classify
songs
is
they
have
a
musician
sitting
listening
to
a
song
and
he
checks
off
different
qualities
about
this
song.
Whether
this
is
true
so
with
the
data
set
like
that
that,
like.
C
C
C
C
B
I
think
that
would
be
the
trick
right.
The
trick.
Is
you
really
you're
just
facing
these
categories?
It's
got
to
be
exactly
that
name
or
that
artist
or
whatever,
and
so
it's
not
being
able
to
pick
up
any
kind
of
other
meaning.
Besides,
you
know,
we
know
the
two
artists
are
almost
identical
or
similar
in
their
apologies.
You
know,
so
you
have
to
get
to
that
next
level
of
encoding
next
time.
Next
time,
that's
good
great
thanks,
george.