Domain Specific Biases(Geographical influence)Updated Aug,18st

data used :foursquare NYC data,

this code has not been finished yet..

here  is the link: https://github.com/FassyGit/LightFM_liu/blob/master/DomainBiases.py

I have succeeded in calculating the distannces, there are some functions in the codes.

I will briefly talk about the problems I have right now...

The major problem is how to preprocess the data. The concept is simple, we treat some negative as non-negative. 

But it is really a hassle doing this.

I have not come up with a way that do not touch the warp theory yet ...

---

After we discussed, we decided to take every poi as center and use fit_partial to train the model.

As for the nonneg porblem, I decided to change 0 to some small number like 0.01 or smaller.

But after I running the codes, it shows great time complexity.... because there are 38333 pois, and we need to preprocess every center and that means 38333 * 38333,

running the codes on cluster for one night, about 16 hours, and it only goes to 4044.....

that is a really big problem anyway...