FWPRec Experiments
All experiments
All experiments
Amazon reviews · seed 42

Amazon snapshot and notepad

Same four-row design as MIND, now on Amazon reviews with frozen product vectors. Fixing the earliest-product mean helps when there is no notepad. Once the notepad is on, a moving profile ranks slightly better than that snapshot. The notepad itself is still the large gain. This is one seed, validation only, from the earlier published screen — not the Modal rerun. Treat it as a warning that a stale shopping mean may be the wrong slow state, not as a multi-seed claim.

Takeaway

The notepad still helps a lot. The earliest-product snapshot is not the best slow state once memory is on.

NDCG@10 · seed 42

Moving history, no notepad
0.06788
Snapshot, no notepad
0.10015
Moving history + notepad
0.26700
Snapshot + notepad
0.26309

Published seed-42 screen

RowAUCNDCG@10
Moving history, no notepad 0.51374 0.06788
Snapshot, no notepad 0.57726 0.10015
Moving history + notepad 0.79952 0.26700
Snapshot + notepad 0.79406 0.26309

Numbers from evaluation/FIXED_ANCHOR_CROSS_DATASET_SCREEN.md. Not rerun on Modal this week. One seed. Validation only. Coverage of the snapshot proxy is 58.4% of Amazon validation impressions.