Keep the notepad vs sequence models
Keep-the-pad stores a small notepad of recent news taste instead of re-reading the whole click history. Sequence models read that history with a GRU or with attention. We asked which ranks better when everything else is the same: same MIND-small split, same article vectors, same candidate lists, three training seeds. We trained SASRec, GRU4Rec, and HSTU. We did not retrain keep-the-pad; those scores come from the earlier keep-the-pad runs on this split. Keep-the-pad beat all three models on every seed. The gap is small but consistent. This is validation only. Published test-set tables use a different split. Do not mix the two.
Keep-the-pad wins every seed. Mean val NDCG@10: 0.395 vs SASRec 0.372, GRU 0.369, HSTU 0.365.
Mean NDCG@10 · three seeds
NDCG@10 by seed
| Model | 42 | 43 | 44 | Mean |
|---|---|---|---|---|
| Keep the pad | 0.3951 | 0.3962 | 0.3939 | 0.3951 |
| SASRec | 0.3712 | 0.3723 | 0.3715 | 0.3717 |
| GRU4Rec | 0.3699 | 0.3689 | 0.3691 | 0.3693 |
| HSTU | 0.3580 | 0.3708 | 0.3666 | 0.3651 |
Sequence models · mean AUC and MRR
| Model | AUC | MRR |
|---|---|---|
| SASRec | 0.6042 | 0.3355 |
| GRU4Rec | 0.6019 | 0.3329 |
| HSTU | 0.5962 | 0.3285 |
Isolated suite on damlr-w04. Same MIND-small split as keep-the-pad (dataset seed 42, val fraction 0.1). Sequence models trained here; keep-the-pad scores reused from the dense lab runs (seed 42 lab4080, seeds 43/44 confirm lab), not retrained. Config is SASRecProfile. Validation only. Test split sealed. Do not mix with published test-set tables.