BPR performance

[baltrunasl]

The performance of BPR is close to random.
Alex, could you try to fix it. If you run examples/models_example.py the performance should be better similar to popularity or at least better than random :)

BTW, I use 1M movielens to check the performance using this data loading:

prepare the data

df = pd.read_csv(DATAPATH+"/1M_movielens/ratings.dat",
                 sep=" ", header=None, names=["user", "item", "rating", "date"])
print df.head()
training, testing = testfm.split.holdoutByRandom(df, 0.8)