continue training for the ranking module?

[everdark]

It seems that the fit method does not support a n_more_iter for the bpr ranking module. Does it mean that if we'd like to iterate over different set of sample pairs the only way is to generate lots of such pairs at once?

For example, for each user-item interaction we'd like to sample one negative sample, so N interactions implies N such pairs. But we'd also like to iterate over each interaction by multiple different negative samples, so the only way is to sample, say, M times for N interactions so the training pair matrix becomes ((M*N), 2) which could be very large if N is large.