[BUG] ZeroDivisionError when fitting BPR on small samples

[jonibr22]

Description

This problem occurs when i use a small dataset to fit my BPR model. For example, I build a dataset of 3 users, 3 items, where each user has interacted with at least 2 items. The matrix is kinda like this:

   [[1, 1, 0],
    [1, 0, 1],
    [0, 1, 1]]

When i fit the BPR model using that dataset, it throws a ZeroDivisionError exception.

I tried to analyze the cause, and seems that in the fit method, self._fit_sgd returns skipped that also equals to len(user_ids), which makes the calculation (100.0 * correct / (len(user_ids) - skipped) be a zero division operation.

(See the source in /cornac/models/bpr/recom_bpr.pyx).

In which platform does it happen?

Any platform, but in my case, i use cornac-2.3.5 on Google Colab (Python 3.12)

How do we replicate the issue?

Run this code, it throws ZeroDivisionError: float division by zero

from cornac.models import BPR
from cornac.data import Dataset
model = BPR()
mat = Dataset.from_uir([(1,1,1),(1,2,1),(2,1,1),(2,3,1),(3,2,1),(3,3,1)])
model.fit(mat)

Expected behavior (i.e. solution)

If my dataset were valid, it would be nice to have a logic that handles len(user_ids) == skipped in the fit method. If not, an additional note on the documentation would be helpful.

[qtuantruong]

Thanks for flagging the issue. We should have added a small $\epsilon$ value to avoid zero division.