FWPRec lab notes

The idea

Research question and current position

The original paper proposes a reciprocal user-matching model in which a user's stated look-for representation is stable and fast-weight memory adjusts it using recent interactions. Because suitable reciprocal user-user datasets were not available, this repository studies the same mechanism under conventional user-to-item candidate ranking and removes retrospective reciprocal matching.

The current defensible position is:

FWPRec is an explicit bounded associative adaptation mechanism. It reuses cacheable item representations and updates user-specific fast state without changing global parameters. It offers a conditional quality/state/computation trade-off, with particularly encouraging evidence under explicit signed feedback. It is not a universal replacement for sequential recommendation, nor is dense FWPRec universally smaller or faster than a cached recurrent model.

Do not reposition the paper as “FWPRec beats sequence models.” Frozen MIND and KuaiRand tests reject that story. The most promising scientific contribution is the separation of a stable slow intent/profile anchor from learned, gradient-free fast associative updates, together with careful mechanism and state/accuracy analysis.

Proposed NeurIPS abstract

Modern sequential recommenders repeatedly encode interaction histories to track evolving user preferences, creating a tension between adaptation quality and online computation. We propose FWPRec, a lightweight alternative that decomposes user state into a stable preference anchor and a rapidly updated associative memory. The anchor may represent an explicit interest, search query, or cached long-term profile, while fast-weight matrices incorporate newly observed positive and negative feedback through gradient-free updates. Item representations and the slow anchor are reusable across requests; adaptation therefore requires neither parameter updates nor repeated sequence encoding. We evaluate FWPRec on real interaction data from MIND, EB-NeRD, KuaiRand, and Amazon Reviews using shared candidate sets, frozen semantic item embeddings, chronological splits, and strong recurrent, attention-based, state-space, and non-sequential baselines. FWPRec provides competitive ranking and adaptation performance, including improvements over strong sequential models in several controlled settings, while retaining bounded user state and incremental updates. Ablations show that learned value construction and the separation between a stable anchor and fast preference shifts are important. They also reveal an important limitation: historical averages are imperfect substitutes for explicit intent and do not improve every domain. FWPRec thus offers a scientifically grounded accuracy–adaptation–cost trade-off for recommendation settings where continually reconstructing large sequential user models is undesirable.

This abstract is intentionally conditional. Before submission, its efficiency phrasing should be checked against the dense-state and cached-GRU results in the checkpoint efficiency report.

How the model works

Method snapshot

Let q_base be a stable user interest, search query, or causal slow profile. For event l, slow write networks produce a key, value, and gate from the base state, item embedding, feedback sign/action, and time features. Positive and negative delta-rule memories are updated independently:

M_l = decay * M_{l-1}
      + gate_l * (v_l - M_{l-1} k_l) k_l^T.

The adapted query is a normalized base plus positive correction minus a weighted negative correction:

q_l = Norm(q_base
           + gain_pos * M_pos q_base
           - alpha_neg * gain_neg * M_neg q_base).

The matrices are per-user transient tensors, not learned parameters. Gradients train the slow write networks; an observed interaction updates fast state without an optimizer step. See model.py, memory.py, and the original mathematical discussion.

Implemented factors include:

• independent versus compact shared-trunk write networks;
• dense versus bounded low-rank factor-ring memory;
• dual, positive-only, and single-signed memory layouts;
• learned key and value residuals;
• fixed, mean-content, attention, learned-user, and hybrid anchors;
• optional evolving fast state in each write input;
• query-age or invariant inter-event time features; and
• batch reconstruction and incremental session-state execution.

The canonical configuration schema is configs/models/fwprec.yaml. Paper experiments use their checked-in resolved YAML settings rather than assuming these defaults.

Evidence hierarchy

Never promote development or diagnostic values into a final-test table. In particular, do not use synthetic fixtures, retired manipulated preference shifts, learned-ID MIND cold-start smoke runs, or deleted random-weight timing measurements as scientific evidence. The governing real-data policy is REAL_DATA_ADAPTATION_PROTOCOL.md.

Frozen final-test results

Final-test evidence

MIND-small: primary controlled test

All controlled neural models use the same frozen normalized 64-dimensional Qwen article vectors, candidates, split, and dataset seed. Neural results use model seeds 42/43/44. The frozen standalone claim required FWPRec to be within 1% of SASRecProfile overall and superior on complete-event H=3 adaptation. Both requirements failed.

Standalone FWPRec's overall effect is -2.97% relative to SASRecProfile, with a 95% interval of [-3.49%, -2.40%]. Its adaptation difference is -0.01353 with a 95% interval of [-0.01772, -0.00933]. SASRecFWP is overall non-inferior, but its adaptation interval crosses zero. Signed EMA is at least as compelling as the hybrid FWP control, so hybrid benefits cannot be uniquely attributed to matrix memory.

The authoritative report, support counts, conventional metrics, native-news baselines, and exact statistical decisions are in MIND_FINAL_TEST_RESULTS.md. The frozen plan is FINAL_STATISTICAL_ANALYSIS_PLAN.md.

KuaiRand: negative final result for standalone leadership

FWPRec's prespecified comparison with GRU4Rec failed overall non-inferiority and did not establish adaptation superiority. HSTU is descriptively strongest on every reported endpoint.

Retain this as a valid failed claim. Do not optimize a new KuaiRand variant against the already observed test set.

Amazon Video Games: strongest positive final evidence

Amazon uses explicit signed ratings, frozen Qwen product vectors, chronological splits, 100 candidates, and seeds 42/43/44. FWPRec's joint comparison with LRURec passed both overall non-inferiority and negative-rich superiority.

Against LRURec, the overall relative effect is +8.719% with 95% interval [+7.115%, +10.418%], and the negative-rich difference is +0.02727 with interval [0.01814, 0.03622]. This supports signed-feedback specialization, not universal adaptation superiority.

The complete KuaiRand/Amazon tables and locked analyses are in CROSS_DATASET_FINAL_TEST_RESULTS.md. Do not rerun or revise their frozen hypotheses in response to these outcomes.

Efficiency and persistent-state evidence

The checkpoint-backed EB-NeRD audit uses 200 unchanged consecutive real single-click transitions and a trained seed-42 checkpoint on an RTX 4080 SUPER.

Dense FWP is bounded in history length but is neither smaller nor faster than a cached compact GRU at dimension 64. It is modestly faster than full SAS/HSTU reconstruction in this implementation, but those models were not evaluated with optimized KV caches. Wall-clock evidence is secondary.

The original reconstructed FWP checkpoint has only 89.5% one-step top-1 agreement with cached updates because its anchor, recent window, and query-age features change during reconstruction. A fixed anchor with invariant inter-event features and aligned horizon 50 reaches 100% top-1 agreement and score MAE 9.74e-8, but its seed-42 NDCG@10 falls to 0.51519. Thus exact bounded sessions are implemented, but hard resets have a quality cost. The best-quality fixed-anchor model and the exact-persistence model are currently different variants.

Read EBNERD_CHECKPOINT_EFFICIENCY.md and EBNERD_PERSISTENT_FWPREC.md. The runner is real_data_efficiency.py.

Limitations and next steps

Known limitations and non-claims

• The current project does not evaluate the original reciprocal user-user matching and retrospective reranking component.
• A historical mean is only a proxy for the original explicit query/interest anchor. Amazon shows that this proxy can reduce full-model quality.
• Preference-shift detectors identify observed metadata-distribution changes, not ground-truth latent preference changes.
• MIND validation originally had only 25 complete H=3 events; its apparent adaptation advantage did not survive the much larger frozen test.
• Dense FWP state is large compared with cached GRU state and costs O(d^2) per update. “Bounded in history” does not mean “small in dimension.”
• Current exact persistent execution uses aligned resets and pays a ranking penalty. Unreset stateful training/evaluation is not yet implemented.
• Low-rank state reduces bytes but is not faster in the current unfused code.
• EB-NeRD results remain validation-only; do not describe them as held-out test confirmation.
• The baseline implementations are clean-room research reproductions, not byte-identical copies of upstream repositories.
• MIND/EB-NeRD native news encoders learn word embeddings from scratch and are not directly representation-matched to frozen Qwen models.
• Wall-clock measurements are hardware- and implementation-specific and must accompany analytical state and operation counts.

Recommended continuation order

1. Confirm the stable-anchor MIND result

Run seeds 43 and 44 for the already defined MIND fixed-anchor factorial without changing its seed-42-selected settings. Report all four rows, not only fixed anchor + FWP. This is the fastest test of whether the restored original slow/fast decomposition is robust.

Do not treat Amazon as a tuning target for inventing a favorable fixed proxy. Its negative transfer result should remain visible. A new Amazon anchor is scientifically justified only if its construction is motivated independently by the application, such as a prespecified longer pre-period signed profile.

2. Make persistent training match persistent serving

Implement a stateful objective that processes real user events in causal order, keeps a fixed exogenous/slow anchor, uses invariant write-time features, and truncates gradients without resetting fast state. Evaluate both ranking quality and cached/reconstructed semantics. Avoid synthetic transitions and artificial history manipulation.

The main target is to remove the hard-reset quality penalty while retaining exact incremental execution. Tensorize per-user update counters before making batched serving-throughput claims.

3. Re-evaluate the compact and low-rank variants under those semantics

Start from the shared-trunk writer because its parameter reduction is the most stable lightweight result. Compare dense, positive-only where appropriate, and rank-16 state. Report bytes, analytical operations, checkpoint-backed latency, and quality together. A smaller state is useful even if latency does not fall.

4. Obtain evidence with genuine exogenous anchors

The original hypothesis is best tested on a dataset with an explicit search query, stated interest, or session intent available before interaction. Keep that anchor fixed and compare:

anchor only
anchor + EMA
anchor + GRU update
anchor + FWP
strong full sequential model

This would distinguish the FWP mechanism from historical-anchor engineering and reconnect the user-to-item study to the original user-user motivation.

5. Transfer the real checkpoint efficiency audit

Repeat the EB-NeRD cached-state study on Amazon, where FWP has its strongest final quality evidence and signed memory is actually used. Include cached GRU, EMA, dense compact FWP, and rank-16 compact FWP under identical candidates and hardware. Do not claim comparison with an optimized Transformer cache unless one is actually implemented.

6. Update the paper only after the above evidence

Use the existing notebook/artifact pipeline. Preserve separate tables for:

• representation-controlled ranking;
• signed-feedback and adaptation endpoints;
• mechanism ablations;
• stable-anchor experiments; and
• quality/state/computation trade-offs.

Do not combine native news encoders, frozen-Qwen user-state controls, and different candidate protocols into a single causal ranking.