How does the notepad compare to the field?
Every model in this project is our own rewrite of someone else's, so a comparison is only worth anything if all of them get the same deal. We gave nineteen models one identical training budget, ran each twice with different random starts, and scored them the same way on three very different datasets: movie ratings, Amazon game purchases, and short-video watch logs. Nothing was tuned for our own method. The point was not to win — the notepad's case rests on being cheap and inspectable — but to find out honestly where it sits among models people actually publish.
It depends on the dataset, and we are reporting both halves. On MovieLens-1M the notepad is genuinely strong at finding the right film — 2nd of 17 on hit rate at 0.8269, behind only HSTU — though it orders its top ten less sharply, landing 9th on nDCG@10. On Amazon it drops to 11th of 17. The efficiency claim is the clean part: the bounded rank-16 notepad matches the full one on every dataset, so the smaller memory costs nothing.
MovieLens-1M
Ranked by nDCG@10 on the held-out test set. All runs finished.
| # | Model | nDCG@10 | HR@10 |
|---|---|---|---|
| 1 | HSTU | 0.6419 | 0.8359 |
| 2 | Mamba4Rec | 0.6336 | 0.8240 |
| 3 | LRURec | 0.6310 | 0.8247 |
| 4 | SASRec | 0.6259 | 0.8219 |
| 5 | BSARec | 0.6235 | 0.8212 |
| 6 | TiSASRec | 0.6224 | 0.8215 |
| 7 | DuoRec | 0.6211 | 0.8198 |
| 8 | GRU4Rec | 0.6179 | 0.8161 |
| 9 | FWPRec (dense) | 0.6148 | 0.8269 |
| 10 | FWPRec (rank-16) | 0.6120 | 0.8248 |
| 11 | LinRec | 0.5859 | 0.7982 |
| 12 | FPMC | 0.5470 | 0.7797 |
| 13 | SeqRules | 0.5400 | 0.7439 |
| 14 | BERT4Rec | 0.4918 | 0.6858 |
| 15 | BPRMF | 0.3913 | 0.6682 |
| 16 | ItemKNN | 0.3282 | 0.5780 |
| 17 | Popularity | 0.2386 | 0.4363 |
| — | NRMS | n/a | n/a |
| — | LSTUR | n/a | n/a |
Amazon Video Games
Ranked by nDCG@10 on the held-out test set. All runs finished.
| # | Model | nDCG@10 | HR@10 |
|---|---|---|---|
| 1 | BSARec | 0.1412 | 0.2159 |
| 2 | BERT4Rec | 0.1411 | 0.2191 |
| 3 | SASRec | 0.1383 | 0.2162 |
| 4 | TiSASRec | 0.1379 | 0.2157 |
| 5 | GRU4Rec | 0.1364 | 0.2092 |
| 6 | HSTU | 0.1362 | 0.2115 |
| 7 | LRURec | 0.1353 | 0.2093 |
| 8 | Mamba4Rec | 0.1329 | 0.2070 |
| 9 | FPMC | 0.1123 | 0.2044 |
| 10 | SeqRules | 0.1097 | 0.1801 |
| 11 | FWPRec (dense) | 0.1061 | 0.1699 |
| 12 | FWPRec (rank-16) | 0.1056 | 0.1698 |
| 13 | DuoRec | 0.0989 | 0.1668 |
| 14 | ItemKNN | 0.0953 | 0.1607 |
| 15 | BPRMF | 0.0875 | 0.1469 |
| 16 | Popularity | 0.0745 | 0.1354 |
| 17 | LinRec | 0.0744 | 0.1352 |
| — | NRMS | n/a | n/a |
| — | LSTUR | n/a | n/a |
KuaiRand-1K
Ranked by nDCG@10 on the held-out test set. 13 of 28 runs finished so far.
| # | Model | nDCG@10 | HR@10 |
|---|---|---|---|
| 1 | BPRMF | 0.4077 | 0.8329 |
| 2 | Mamba4Rec | 0.3631 | 0.7613 |
| 3 | GRU4Rec | 0.3569 | 0.7523 |
| 4 | LinRec | 0.3547 | 0.6993 |
| 5 | FWPRec (rank-16) | 0.3531 | 0.7231 |
| 6 | FWPRec (dense) | 0.3527 | 0.7218 |
| 7 | HSTU | 0.3517 | 0.7505 |
| 8 | SeqRules | 0.0914 | 0.2004 |
| 9 | Popularity | 0.0426 | 0.0931 |
| — | ItemKNN | n/a | n/a |
| — | FPMC | n/a | n/a |
| — | NRMS | n/a | n/a |
| — | LSTUR | n/a | n/a |
How it was run
75 of 90 runs finished so far. Each one records the code version, a checksum of the data it read, the exact settings, the GPU it used and how long it took, and each re-checks its own scores from its saved predictions before being counted.
Two seeds per model, reported as the mean; we do not quote a standard deviation from two numbers. The three datasets are scored against different-sized candidate lists, so a column is only comparable down its own length, never across. Rows marked n/a could not be run and say why in the record: NRMS and LSTUR need item text none of these datasets carry, ItemKNN's memory grows with the square of a user's history, and FPMC's three embedding tables over 4.4 million videos exceed the cards we have. KuaiRand-1K is still running.