onsdag 8 Nov 2006, 14:08

Hi everybody!

The new taste-o-meter is online. And the new neighborhood too!
Now the two things are directly related, so it should not happen that you get "medium" or "low" on your neighbor anymore.

How it works? Well, the whole neighbor computation is derived from a rather standard user-based collaborative filtering technique, with some additional "tricks". Bwahahahahah! ;)

This technique returns a (real) value between 0 and 1, that represent the "compatibility" between two users. Since it is rather uncommon that two users share exactly the same set of artists, that number is generally between 0.4 and 0, with a lot of variations for each user.

If we order these values for each user, we get the neighbors. To make an example, here is the list of my neighbors with the computed value:
athleston 0.299406
Splinter77 0.294277
Bowley 0.289653
ezzvaldez 0.289625
..etc..

The value is used as the basis for the taste-o-meter.
Since the top value varies a lot depending on how obscure your tastes are, we cannot use an absolute value to calibrate the taste-o-meter. Rather I coded a simple function that takes into consideration your first neighbor as base, and scale the other values accordingly.

The drawback of this technique is that the resulting match is not symmetric, that is we could have the case in which:
A->B : super
B->A : medium

This follows the principle of: "he is my best friend, but for him, I am only a good friend".

From a logical point of view this sounds correct, but some good hearted people here at Last.fm believes that we should make all friendships strong, and go for a symmetric version that return max(A->B, B->A) for both people and make everybody (?) happy. Tell me what you think… :)

One last - but very important - note:
Both neighborhood and the taste-o-meter are based on a 6 months dataset. If you (or your neighbor) did not scrobble nor listen to radio during that period.. well, no cookie! ;)

We will switch to 12 months if we see this is not working.

For now it's all, and sorry for my Swiss English! ;)