Anchor size sensitivity
Each user has a fixed anchor — a slow-memory vector built from the positives in their earliest training impression. We asked whether it matters how many of those early clicks are used: 10, 25, or 50. We ran three keep-the-pad variants on MIND-small (seed 42, validation only). Val NDCG@10 rises gently from 0.390 to 0.395, a spread of 0.005. Cap-50 essentially matches uncapped keep-the-pad on this split. The model is not sensitive to this choice, which strengthens the anchor mechanism claim.
Anchor quality scales softly with N. The gap from 10 to 50 is 0.005 NDCG@10. Cap-50 ≈ uncapped. Ten early clicks already build a useful anchor.
Val NDCG@10 by anchor cap N
Full results · seed 42
| Variant | N | NDCG@10 | NDCG@5 | MRR | AUC | Best epoch |
|---|---|---|---|---|---|---|
| keep_anchor10 | 10 | 0.3902 | 0.3269 | 0.3524 | 0.6296 | 7 / 10 |
| keep_anchor25 | 25 | 0.3927 | 0.3280 | 0.3554 | 0.6285 | 6 / 9 |
| keep_anchor50 | 50 | 0.3951 | 0.3303 | 0.3573 | 0.6303 | 4 / 7 |
What the anchor cap controls
The anchor is the mean embedding of the user's earliest positives in the training set —
it acts as the slow component of their taste vector and is frozen at test time.
Cap N is not the notepad window (which stays at 20 for all runs).
A smaller cap means fewer early clicks go into the anchor; the rest of the model and training are identical.
The default uncapped keep-the-pad on this split scored ~0.395 val NDCG@10.
keep_anchor50 is an explicit cap of 50, not the same run, but the numbers land at the same place.
The result suggests that 50 early positives is enough to saturate anchor quality on MIND-small,
and that even 10 gives a competitive anchor.
Isolated suite on damlr-w03 (RTX 4080 SUPER). MIND-small, dataset seed 42, val fraction 0.1. Keep-the-pad only; no baselines retrained. Patience 3, max 10 epochs. Validation only. Test split sealed. Anchor = mean of the first N positives in the user's earliest training impression. Do not mix with published test-set tables.