Knowledge-Geometry Decoupling: Refreshable Pretrained Transfer for Streaming Recommendation
The paper is withdrawn, with arXiv citing data-sharing and privacy rules tied to industrial co-authors.
The abstract described KGD, a recommender-system method meant to keep pretrained behavioral knowledge separate from task-specific adaptation. It claimed 4-12% gains on eight public benchmarks, a 90-day production-stream advantage, and deployment at Shopee. The listed live A/B test claimed GMV per user rose 1.75% and ad revenue rose 1.53%. HF Daily Papers' note
The abstract described KGD, a recommender-system method meant to keep pretrained behavioral knowledge separate from task-specific adaptation. It claimed 4-12% gains on eight public benchmarks, a 90-day production-stream advantage, and deployment at Shopee. The listed live A/B test claimed GMV per user rose 1.75% and ad revenue rose 1.53%. HF Daily Papers' note
score 4