- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes FWPRec, a recommender that splits the user state into a fixed “anchor” u₀ (e.g., a mean-content profile, query, or cached long-term representation) and two d×d fast-weight matrices, M⁺ and M⁻, for positive and negative feedback. Each interaction (item, signed reward) is passed through small MLP “writers” that produce a key, a correction vector, and a write gate. The authors claim a competitive ranking quality against static, lightweight-update, and sequential baselines on six datasets. Ablation studies support learned writes, separate positive and negative memories, and the anchor.
Critical strengths
A practical design with a clear deployment story. User-side-only adaptation over frozen item vectors leaves the item index and ANN stage untouched. This addresses a real industrial constraint.
Critical weaknesses
- Comparison fairness and strength of the empirical claims. FWPRec and the lightweight baselines receive signed feedback. The sequential baselines receive clicked-only histories. The headline wins on MIND and EB-NeRD may therefore come from extra negative-feedback information rather than the architecture. The paper itself says gains are strongest when negative feedback is informative. A signed-input HSTU/SASRec is the necessary control.
- Many comparisons (e.g., MIND, where Mamba4Rec, HSTU, LRURec, and GRU4Rec are within ~0.001–0.005 nDCG@10) are within plausible noise.
- Mixed outcomes compared to the “competitive” baselines. FWPRec is worse than HSTU, LRURec, and GRU4Rec on Amazon (MRR/nDCG). On ML-1M it trails ItemKNN and even Popularity. It is second on KuaiRand and loses on KuaiRec MRR. Absolute MIND AUC (~0.63) is far below tuned news models, and NRMS/LSTUR are described but not reported.
- Efficiency claims are thin. No latency or throughput is reported. The datasets use histories of only 50–100, so the “O(1) vs growing” advantage is not demonstrated in the tested regime. Recurrent baselines are also O(1).
Questions for authors
- How exactly is u₀ built in each dataset? If it is a mean of history items, are those the same interactions then written into memory? If so, is history processed twice, and how does this fit the “decoupled from history” claim?
- How were training instances and the negatives N_t built (impression non-clicks vs. random)? What is the training cost of backpropagating through 50–100 memory updates?
- KuaiRand “uses learned ID embeddings in each model’s native configuration”. How does this reconcile with the frozen-embedding protocol?
- Report scaling with much longer histories and wall-clock/QPS.
2 comments
Responses to Questions for Authors
Q1. How is constructed, and is history processed twice?
As clarified in Appendix D (Detailed Analysis of User Anchor), is the normalized mean of earlier positive interaction embeddings, using frozen content representations or learned ID embeddings on KuaiRand. These interactions precede the adaptation window and are not replayed when writing memory.
Once constructed, remains fixed, and only the latest 20 interactions contribute to memory updates (Eq. 9). Appendix D, Table 7 further evaluates alternative anchor constructions, including user ID embeddings and profile features. Thus, the history-independence claim concerns online adaptation, not the initial construction of .
Q2. How are training instances and negatives constructed, and what is the training cost?
As described in Section 3.3 and Appendix A.1–A.2, each training instance contains a causal history, a positive target, and a fixed candidate set. The target interaction is excluded from its prediction history. Please see the negative-sampling protocol below.
Table 3: Construction of training negatives.
| Dataset | Negative candidates |
|---|---|
| MIND / EB-NeRD | Unclicked items from the same logged impression |
| Amazon / MovieLens-1M | Random catalog items, excluding the target and previously positive items |
| KuaiRand / KuaiRec | Previously observed non-positive interactions, with catalog fallback |
Candidate sets are constructed using fixed random seeds and are not resampled across training epochs. Sampled candidate negatives are not automatically treated as negative feedback in the user history.
We also added Appendix A.8 (Training Cost and the Memory Write Window). FWPRec backpropagates through only 20 memory updates, rather than the full 50–100-event history. Please see the measured training costs below.
Table 4: Training time per epoch, including validation (seconds, RTX 4080 SUPER).
| Method | Amazon | MovieLens-1M | KuaiRand | KuaiRec |
|---|---|---|---|---|
| GRU4Rec | 8 | 19 | 20 | 59 |
| SASRec | 9 | 24 | 22 | 79 |
| HSTU | 14 | 21 | 34 | 81 |
| Mamba4Rec | 37 | 49 | 80 | 222 |
| FWPRec | 17 | 43 | 24 | 125 |
FWPRec has training costs comparable to sequential baselines and is consistently faster than Mamba4Rec. Its primary computational advantage remains gradient-free online adaptation rather than training speed.
Q3. How does KuaiRand's learned ID embedding setting relate to the frozen-embedding protocol?
As stated in Table 2 and Appendix A.2, KuaiRand has no item text in our setup. We therefore use learned ID embeddings in each model's native configuration. KuaiRand serves as an additional ID-based evaluation, while the other datasets provide representation-controlled comparisons using shared frozen content embeddings.
Q4. Can the authors report scaling to longer histories and wall-clock throughput?
We added Appendix E.5 (Update Latency and Throughput), extending the efficiency evaluation to 10,000 historical interactions on an RTX 4080 SUPER GPU. Each serving step incorporates one interaction and scores 100 candidates. Please see the results below.
Table 5: Median incremental serving latency (ms).
| Method | B=1, L=20 | B=1, L=10,000 | B=1,024, L=10,000 |
|---|---|---|---|
| GRU4Rec | 0.08 | 0.09 | 0.24 |
| LRURec | 0.24 | 0.28 | 0.34 |
| Mamba4Rec | 0.25 | 0.25 | 0.34 |
| SASRec | 0.25 | 2.04 | 38.20 |
| HSTU | 0.42 | 0.43 | OOM |
| FWPRec (dense) | 0.31 | 0.31 | 0.67 |
B denotes batch size and L denotes history length.
FWPRec maintains constant latency as history length increases, achieving approximately 1.5 million updates per second with dense memory at batch size 1,024. These results directly demonstrate history-independent serving costs, complementing the persistent-memory analysis in Section 4.4.
Response to Reviewer 3
We thank the reviewer for the constructive feedback. We have added experiments addressing comparison fairness, statistical reliability, recommendation performance, and computational efficiency. We clarify the experimental protocol below.
Responses to Critical Weaknesses
1. Comparison fairness regarding signed feedback.
To investigate whether FWPRec's gains come from access to negative feedback, we added Appendix E.6 (Signed-Input Sequential Baselines), extending HSTU to accept signed interactions through learned feedback embeddings. Please see the results below.
Table 1: Comparison with positive-only and signed-input HSTU (test nDCG@10).
FWPRec remains stronger on MIND and EB-NeRD despite HSTU receiving signed feedback and using up to 100 historical interactions, compared with FWPRec's 20. The news recommendation gains therefore cannot be explained solely by access to negative feedback.
2. Statistical reliability.
We added Appendix E.2 (Experimental Results with Statistics), reporting results across three training seeds and paired hierarchical bootstrap analysis (Appendix A.5).
On MIND, FWPRec achieves nDCG@10 of 0.4048 ± 0.0015, compared with 0.4013 ± 0.0018 for Mamba4Rec and 0.4007 ± 0.0028 for HSTU. Both differences are distinguishable from zero under paired bootstrap and seed-level confidence intervals. These results support a small but statistically reliable improvement rather than an effect of random variation.
3. Mixed ranking performance and comparison with news recommenders.
As discussed in Section 4.2, FWPRec targets lightweight preference adaptation rather than replacing sequential recommendation architectures. To further investigate its performance, Appendix C.3 (Effect of Enhanced Writer Inputs) introduces feedback-type and temporal features, improving nDCG@10 on four datasets, including Amazon (0.2496 → 0.2825), MovieLens-1M (0.2421 → 0.2800), and KuaiRand (0.4858 → 0.5613).
We also added Appendix E.4 (Comparison with News Recommendation Models), evaluating NRMS (Wu et al., 2019) and LSTUR (An et al., 2019) on the same test impressions. Please see the results below.
Table 2: Comparison with established news recommendation models (test nDCG@10).
FWPRec achieves comparable performance to NRMS on MIND and outperforms both models on EB-NeRD, despite using frozen item representations rather than jointly trained news encoders. These results support the effectiveness of FWPRec within the representation-reuse setting studied in our paper.
4. Computational efficiency.
We added Appendix E.5 (Update Latency and Throughput), measuring incremental serving latency for histories of 20–10,000 interactions. FWPRec maintains constant latency of 0.31 ms per user and 0.67 ms per batch of 1,024 users, achieving approximately 1.5 million user updates per second. In contrast, SASRec's batch latency increases from 0.46 ms to 38.2 ms as history length grows from 20 to 10,000.
As discussed in Section 4.4, recurrent models also support constant-size states. FWPRec's advantage is combining history-independent updates with learned associative corrections while retaining existing item representations and retrieval infrastructure.