Over the years, I've heard complaints from musicians that their pages on magnatune point to other artists with the "customers who bought this also bought" but that they're not seeing links from other musicians to them.
This happens because of the algorithm I use to find the purchasing relationships. Simply, what I do is find all the buyers of one artist, collect a big list of everything those buyers have bought, and then find the top 10 things those buyers bought. Simple enough, and it has historically produced results that made sense.
However, it's possible that an artist sells much less than another, so that they point to that other artists (their customers buy the other one) but that because they don't sell that much, they don't make it to the top 10 of the other artists. Hence the unfairness in the system: you pass sales to other artists but don't receive any in reverse, which only contributes to the lopsidedness where they sell much more than you.
Another problem with my algorithm is that the artists who have been on Magnatune the longest get the majority of the "also bought" recommendations, since their sales are the greatest (since they've been selling for a longer time).
I've decided to try tweaking the algorithm to see if I can alleviate these behaviors.
I've now changed the "also bought" algorithm so that it looks only at the past 100 days of purchasing. That will fix the "earliest artists to magnatune get the most recommendations" problem, and also means that artists who sell well in a brief recent period (such as after their new album launch) have more of a chance of getting onto the recommendations.
However, the downside is that I'm looking at a lot less data, so the recommendations will be less accurate.
Below is a chart of old vs new (before the 100-day tweak, and after) for a few related artists. You can see that the music genre hopping is more severe, but some more recent artists (such as Laurel Zucker) get featured in the new one and not in the older one, and some more obscure artists (ie Falik) show up.
I'll be interested to see what Magnatune's visitors think of this...
One other idea I've had is making an m3u playlist of "also bought" titles, so that you can listen to a few hours of music that people also liked who liked this artist. Does this seem like a *really* useful feature, or just nice to have?
> Does this seem like a *really* useful feature, or just nice to have?
Maybe not a *really* useful feature, but I would *really* like it. That would be a great and easy way for listener to discover new artists.
Posted by: JMG | January 28, 2006 at 07:31 PM
Divide by total sales to find items that correlate better, not just "popular" items? example
Posted by: Don Marti | January 28, 2006 at 09:46 PM
>>... You can see that the music genre hopping is more severe, but some more recent artists (such as Laurel Zucker) get featured in the new one and not in the older one, and some more obscure artists (ie Falik) show up.
Hey! I'm an obscure reference!!!!
But sereiously, folks, I think it's an improvement. Maybe there should be an "off the beaten path" link. Some randomly selected cross- genre link to those "obscure" artists?
Posted by: Charlie | January 29, 2006 at 11:04 AM
Really what I think you should be aiming for are ratings-based recommendations. (Netflix has an excellent implimentation that's worth checking out.) I am far less interested in the typical "Others also bought..." lists than I am in "We think you might like..." lists. The latter is always much more accurate and rarely lets me down, whereas the former is often just a popularity contest -- and what's popular is rarely what's best.
My two cents.
Posted by: Daniel Kinney | January 29, 2006 at 11:21 AM
Don, I need a little clarity on your algorithm suggestion. To use your example, suppose :
3 people who bought A also bought X
2 people who bought A also bought Y
1 person who bought A also bought Z
And if X has sold 300, Y has sold 200 and Z and has sold 50.
In your math:
X would be 1x correlated
Y would be 1x correlated
Z would be 2x correlated, and ranked top
is that right?
-john
Posted by: John Buckman | January 30, 2006 at 04:10 PM
John, yes, that's right.
Posted by: Don Marti | January 30, 2006 at 05:29 PM