You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Option to redraw negatives that coincide with the target in CatalogUniformSampler #330
Add an opt-in parameter to CatalogUniformSampler (for example avoid_positive_collisions: bool = False) that redraws sampled negatives equal to the target item at the same position. It would be enabled through the existing negative_sampler_kwargs:
The default would keep the current behaviour, so existing configs and seeded results do not change.
Why this feature?
CatalogUniformSampler.get_negatives (rectools/models/nn/transformers/negative_sampler.py) samples negatives uniformly from all catalog items (as its docstring says) using torch.randint, without excluding the target. With BCE, gBCE or sampled_softmax, the same item can then be scored as both the positive and a negative in one loss term.
The UniSRec model added in 0.19.0 already handles this: _NegativeSampler in rectools/fast_transformers/unisrec/model.py redraws colliding negatives once. The transformer models in rectools.models.nn (SASRec, BERT4Rec, HSTU) have no such option, although users can subclass TransformerNegativeSamplerBase to get it.
The collision probability per negative is 1 / n_items. With n_negatives=256:
catalog size
share of negatives equal to the target
share of positions with at least one such negative
3,416 (ML-1M as used in the SASRec/gSASRec papers)
0.03%
7.2%
500
0.2%
40%
100
1.0%
92%
On a synthetic dataset with 100 items (2,000 users, 22 interactions each), SASRecModel with loss="sampled_softmax" and n_negatives=256 reached NDCG@10 of about 0.127 (RecTools NDCG(k=10), default settings). Redrawing colliding negatives until none were left changed this by -0.0009 on average over 3 seeds, which is within run-to-run noise. So I don't expect a measurable metric gain; the aim is consistency with UniSRec and sampling negatives the way the gBCE sampling rate alpha = n_negatives / (n_items - 1) already assumes.
Additional context
I have a small patch ready (one redraw pass, same as _NegativeSampler, plus unit tests and a changelog entry) and can open a PR if this is welcome.
Activity