TWIN V2: Scaling Ultra-Long User Behavior Sequence Modeling for Enhanced CTR Prediction at Kuaishou

简体中文

Fidelity: core mechanism. Paper conclusions, local public-data experiments, and omitted work are reported separately.

Paper information

Field	Details
Paper	CIKM 2024
Organization	Kuaishou
First public release	2024-07-23 (arXiv v1)
Authors' code	The paper did not supply official or author code when checked on 2026-08-09.
Adapter / method	twin-v2
Local reproduction code	src/auto_research/reproductions/twin_v2/

Original paper

Motivation and method

Offline hierarchical clustering compresses millions of lifetime behaviors into virtual items with cluster-size information. Online, GSU retrieves relevant clusters and ESU fine-ranks with cluster-aware target attention.

flowchart LR
 A["Public input"] --> B["TWIN V2 core mechanism"]
 B --> C["Matched-budget training or execution"]
 C --> D["Public evaluation and diagnostics"]

Key figure from the paper

Figure 2: Overview of the paper's method and key components. From the original paper; copyright remains with the authors.

Core equations

\[ a_c=q^\top k_c-\tfrac12\log |c|,\quad h=\sum_{c\in\operatorname{TopK}(a)}\sum_{i\in c}\alpha_i v_i. \]

Paper results

Kuaishou reports watch time +0.672%/+0.800%/+0.728% across three scenarios, serving the main traffic of around 400 million DAU.

Local reproduction

Local comparison: The shared transition + content baseline and TWIN-V2 treatment differ only in the core mechanism. ndcg_at_10 rises from 0.0354 to 0.0433, or +22.45% relative.

auto-research reproduce --paper twin-v2 --dataset-dir data --seed 42

See the fixed metrics.

Reproduction boundaries

Full-catalog MovieLens-100K ranking exercises paper-related candidate-aware retrieval, hierarchical compression, or distillation. No private logs or production serving are used; online A/B comes only from the paper and must remain separate from local relative changes.