简体中文 · English

Knowledge–Geometry Decoupling: Refreshable Pretrained Transfer for Streaming Recommendation¶

Core mechanism reproduction. Defining operators run locally. Private data, large weights and dedicated production serving are not reproduced; paper results are separate.

Paper information¶

Field	Details
Paper	arXiv 2608.02738
Institution	Xiamen University / Shopee
First public release	2026-08-03 (arXiv v1)
Author code	Yes: official author code.
Local adapter or method key	`kgd`
Local reproduction code	`src/auto_research/reproductions/kgd/`

Original paper summary¶

Background and main changes¶

Topic: Pretraining and transfer for streaming recommendation. Adjacent clicks are not always useful dependencies. KGD uses BMTP to select collaboratively or semantically related future items. During transfer, freeze the refreshable knowledge encoder, read it with cross-attention, and write task geometry with ACR orthogonal to the anchor.

flowchart LR
    N0["Behavior sequence"] --> N1
    N1["BMTP multi-step supervision"] --> N2
    N2["Refreshable knowledge encoder"] --> N3
    N3["Read-only transfer and ACR"] --> N4
    N4["Ranking task"]

Key figure from the paper¶

Figure 2: Overall design and components. From the paper; copyright belongs to the authors. Click for the source.

Core equations¶

\(\mathcal L_{BMTP}=-\sum_{j\in\mathcal R_t}\log p(i_{t+j}\mid i_{\le t}),\quad r_{ACR}=r-\operatorname{proj}_{e}(r)\)

Offline and online results¶

The paper reports 4%–12% gains across eight public benchmarks over strong pretrained transfer baselines. Shopee homepage search shows GMV/user +1.75% and ad revenue +1.53%, followed by full rollout.

Local reproduction¶

Compare adjacent-item NTP with BMTP, frozen knowledge and orthogonal ACR on MovieLens-1M, using full-catalog ranking.

auto-research reproduce --paper kgd --dataset-dir data --seed 42

See stable metrics. No checkpoint is committed.

Comparison: Matched data split and budget, with the paper-specific mechanism disabled for control and enabled in kgd. Relative change is in public-seed42.json; cross-paper percentages are not comparable.

Scope¶

Local results verify execution and direction, not reproduction at paper scale.
Private features, traffic and serving are unavailable; paper online values are references only.
Evolve-compatible architectures are registered as candidates. Serving-only methods remain runnable standalone adapters, not trainable genomes.