Important things to note
Mixed results on real data. Report the losses. Do not treat one run as proof.
Read this first
Mixed. Tested on real data. Reported as-is. Not a clean win.
Beats other models on Amazon — explicit ratings. Strongest evidence.
Loses to standard models on MIND (news) and KuaiRand (video). Do not claim overall superiority.
No history re-read. True. Still uses more memory and compute than a tuned GRU. Not a clean efficiency win either.
Splitting stable taste (anchor) from fast recent mood (memory) looks real — especially a fixed anchor on MIND.
That fixed-anchor MIND result is one run. More seeds before it counts as solid.
No cherry-picking. No mixing incompatible setups. No reusing the test set to keep tuning. No treating one-off runs as proof.
In order
Next steps
Do these in this order. Then update the paper.
- Confirm the fixed-anchor result
- Align training with real-world use
- Recheck smaller, compact versions
- Test on a dataset with real explicit intent
- Redo the efficiency test on Amazon
- Update the paper
Baselines
Models we compare against. Hybrids check whether combining ideas helps.
StaticProfile
Average of past likes. No updating.
PositiveEMA / SignedEMA
Recent-average trackers. Newer events count more.
GRU4Rec
Classic recurrent sequence model.
LRURec
Another recurrent-style sequence model.
SASRecProfile
Attention-based sequence model.
HSTU
Newer, stronger Transformer-style model.
Mamba4Rec
Newer state-space sequence model.
SASRecSignedEMA / SASRecFWP
Transformer plus EMA, or plus this paper’s fast memory. Tests whether combining helps.
Datasets
What each corpus is for. Side checks and synthetic data do not support claims.
MIND
Real news clicks. Main test.
EB-NeRD
Another real news set. Double-check only. Not a final test.
Amazon Reviews
Video Games ratings. Explicit likes and dislikes. Strongest result.
KuaiRand
Real short-video watches.
KuaiRec
Related video set. Backup only. Not for main claims.
MovieLens
Old movie ratings. Code check only. Not for real claims.
Synthetic
Made-up data. Testing only. Never use it to support a claim.
This page is the guardrail. The argument is on The idea. Layout is on the method page. Scores live on the experiment pages.