All experiments
MIND-small · seed 42 · RTX 4080 SUPER
Rank-16 notepad on MIND
The full notepad stores a dense matrix per user. Factoring it to rank 16 cuts the per-user memory in half. We ran the same snapshot plus keep-the-pad setup on MIND-small with a rank-16 notepad and compared it to the dense version. Ranking stayed nearly equal. The memory saving is real; the speed is not improved yet because the rank-16 path is not optimised.
Takeaway
Rank-16 notepad cuts memory per user from 33 KB to 16.7 KB (−50%) at a cost of 0.001 NDCG@10. Agreement with full rebuild is 85.5%.
Keep NDCG@10 · dense vs rank-16
Scores · seed 42
| Row | AUC | MRR | NDCG@10 |
|---|
| Wipe, rank-16 | 0.6253 | 0.3496 | 0.3877 |
| Keep, rank-16 | 0.6291 | 0.3565 | 0.3944 |
| Wipe, dense | 0.6252 | 0.3509 | 0.3885 |
| Keep, dense | 0.6303 | 0.3579 | 0.3951 |
Memory per user
| Notepad | Bytes per user | vs dense |
|---|
| Rank-16 | 16,656 | −50% |
| Dense | 33,024 | — |
Rank-16 keep · NDCG@10 by epoch
| Epoch | Keep the pad |
|---|
| 1 | 0.3887 |
| 2 | 0.3903 |
| 3 | 0.3935 |
| 4 | 0.3943 |
| 5 | 0.3944 |
| 6 | 0.3929 |
| 7 | 0.3927 |
| 8 | 0.3906 |
Agreement · keep vs fresh rebuild
| Notepad | Pairs | Top-1 match | Largest score gap |
|---|
| Rank-16 | 200 | 85.5% | 2.77 |
| Dense | 200 | 90.5% | 2.55 |
Isolated trainer copy; dense keep-the-pad files untouched. Seed 42, MIND-small, RTX 4080 SUPER, torch 2.13, CUDA 13. Rank-16 path is not kernel-optimised; wall-clock is slower than dense. Wipe ran 10 epochs (best at epoch 8). Keep early-stopped after epoch 8 (patience 3, best at epoch 5). Validation only. Test split sealed.