C[omp]ute

From: andrew cooke <andrew@...>

Date: Wed, 20 Feb 2013 18:29:27 -0300

The problem: if you have a large music collection you probably don't have a
good idea of all the artists and styles available.  Particularly if it
includes many compilations with obscure artists (which tends to be what you
download from the 'net when looking for new music...).


The solution: I've made various attempts at solving this problem over the
years (programs called UYKFD, UYKFE, UYKFF and UYKFG).  After that many
iterations, I think I've found a good solution: a playlist generator and,
separately, targetted selection based on tags.

The most important tool is the playlist generator.  This checks whatever music
player you use and, if there is a single track remaining, adds one that is
"similar".  If there are no tracks it adds one at random.

In UYKFG I don't generate the relationships between artists myself - I use
EchoNest's API to find related artists.  But I make two important adjustments:

  First, I traverse the graph backwards.  So if artist A is related to B then,
  when playing a track from B, I consider A's tracks as possible candidates to
  be played next.  This helps emphasise obscure artists (if A is obscure, then
  there may be no C such that C->A, but we *do* have a A->B, which we can
  invert).

  Second, I store only the top 4 links per artist.  If this isn't sufficient
  to give 4 backlinks then I consider forward links and artists in
  compilations.  This both keeps links relevant and avoids, as much as
  possible, having no links at all from an artist.

The above gives a system that slowly explores the space of available music.
With time it drifts into related styles.  If you get bored, clearing the
playlist gets you to a new, random point.


The playlist generator works well, and is what I use 90% of the time - it's
great "in the background".  But sometimes I want to be more active in my
exploration.  To support this I've added a number of tools based on EchoNest's
tags (called "terms" in their API).

In particular, I can show the tags defined for an artist, and add tracks to
the playlist that match certain tags (and/or explicitly avoid other tags).

The first of these is useful to get an idea of what tags are available - you
can always remember the name of a popular artists in a certain style.  So you
can see what tags are associated with them.  Then you can listen to tracks
from similar artists using the second command.

Unlike with the playlist generator, tracks added for a given set of tags don't
"wander off".  They stay centred around the given tags.  This works better
when you're curious about, say, "the blues".


So there are two approaches above: a "random walk" of related artists and a
"targeted list" based on tags.  In earlier iterations of the software I had
tried to combine these, but in practice it's simpler and more convenient to
leave them separate.


All the above assumes that you are exploring music you have already chosen.
That is important because it excludes music that you don't like (and so don't
own) - a "random walk" isn't going to go somewhere too unpleasant.  Services
like Pandora face a harder problem, because they don't know the boundaries of
a given user's taste.  They may need to combine the two approaches outlined.

Andrew