About build interaction matrix about lightfm

[springcoil]

Your interaction matrix needs to be positive and negative. The value 4.0 is arbitrary you could pick 3.0 for example.

[springcoil]

Is this correct @maciejkula ?

[maciejkula]

Yes, you need to do this for logistic loss.

For WARP loss you only need positives, as the negatives are sampled from other users' interactions. So in that sense this is not necessary. I guess the interpretation here is that only values above some X are actual positives.

[maciejkula]

@xiaop1987 are you happy about this answer? Or can I expand more?

[xiaop1987]

@maciejkula surely I'm happy, thanks a lot :).

I take this step as making explicit feedback training data, but when I reading the example/movielens/readme.md which said that warp loss is used in implicit feedback, did I miss anything?

btw: I'm implementing warp loss in spark as recommending algorithm, but the result is not so good, would you mind if I posted some problems about warp loss(wsabie) by email?

[maciejkula]

Yes, WARP loss is for implicit data. So this step basically drops anything less than 4 (but you could equally well leave everything in).

Sure, should I drop you an email at [email protected]?

[xiaop1987]

@maciejkula yes, that would be very appreciated!

[maciejkula]

Done, let me know if you didn't get it!