improved representation of bias

[Daniel Jones]

If two disjoint proportional biases are set, add up to greater than
100%, they are essentially normalized.

So (100% Artist A),(100% Artist B), is equivalent to (50% Artist
A),(50% Artist B).

So why not make this fact apparant in some way? Because it isn't so simple.

Suppose you had two biases.
1. 100% Artist A
2. 100% Jazz

If "Jazz" and "Artist A" overlap, then the generated playlist will be
100% Jazz and 100% Artist A.

If they are disjoint (don't overlap at all), it will look like 50%
Jazz, 50% Artist A.

So what the pie chart looks like depends on determining every
combination of intersections between n biases. Thats 2^n
intersections. Its possible to do for a small number biases, but it
certainly won't scale say, 20 biases. Which would be shame, because
the algorithm that actually generates the playlists works effeciently
(linearly) for any number of biases.

Which brings up another point: is a pie chart with 32 slices really
easier to understand than 5 sliders?

This also ignores the interaction with other biases (I know there's
only one other bias type now, but there will be more) which only makes
things more complicated. The point is, if we were to display a pie
chart, it would be, in general, an intractable (very much in NP-hard)
problem to determine what it actually looks like.


There is the idea of using points rather than percentages. It seemed
to me easier to reason about what the playlist would look like with
hard percentages, rather than points. But this is a discussion worth
having. Maybe both a point and percentage system could coexist.


Also, I do think that the description for fuzzy biases vs proportional
needs to be more clear. "Fuzzy bias" is pretty ambigous, but I thought
it sounded better than "gaussian distribution goodness of fit
constraint".


Anyway, thanks for your idea. I apologize for bombarding you with a
stream of conciousness technical discussion.

-Daniel


On Fri, Dec 12, 2008 at 1:04 PM, Caleb Cushing <xenoterracide at gmail.com> wrote:
> https://bugs.kde.org/show_bug.cgi?id=177627
>
> I've just filed this bug, and am bring it to the mailing list because
> I think it needs some serious discussion. I really can't think of a
> perfect solution, but it definitely needs work.
>
> main text of the bug is:
>
>
> problem is that I can't really tell how amarok calculates my playlist
> making it near impossible to tweak. I really don't know what it's
> doing.
>
> what does it do when I have proportional bias 75% genre rock and 60%
> genre pop or even more confusing what if I have a 80% fuzzy bias on
> year 2008 and a proportional and an 80% proportional bias on classic
> rock (assuming all tags are correct).
>
> in the spirit of the first example lets say 60/60 does that end up
> being 60/40 using the first value it comes across? 40/60 or 50/50?
>
> one suggestion was offer a pie chart... some kind of chart might be nice.
>
> another would be that in the case of proportional don't allow the
> sliders to add up to more than 100% I'm not sure how that affects
> fuzzy though.
>
> --
> Caleb Cushing
>
> http://xenoterracide.blogspot.com
> _______________________________________________
> Amarok mailing list
> Amarok at kde.org
> https://mail.kde.org/mailman/listinfo/amarok
>