Counteracting Duration Bias in Video Recommendation via Counterfactual Watch Time¶

Fidelity: core mechanism. Paper results, public-data measurements, and omissions are reported separately.

Paper information¶

Field	Details
Paper	arXiv 2406.07932
Organization	Kuaishou Technology / Renmin University of China
First public release	2024-06-12 (arXiv v1)
Authors' code	Public implementation
Adapter / method	`cwm`
Local reproduction code	`src/auto_research/reproductions/cwm/`

Original paper¶

Motivation and method¶

Use counterfactual watch-time estimation to remove video-duration bias, distinguishing observed watch time from potential utility under a common duration intervention.

flowchart LR
 A["Public input or history"] --> B["CWM core path"]
 B --> C["Paper-specific state or objective"]
 C --> D["Matched-budget evaluation"]

Key figure from the paper¶

Figure 4: Overview of the paper's method and key components. From the original paper; copyright remains with the authors.

Core equations¶

\[ \hat y(d_0)=\mathbb E[Y\mid X,do(D=d_0)],\quad s=\hat y(d_0)-\lambda\,\operatorname{bias}(D). \]

Paper results¶

Kuaishou mean watch time +2.9%, video views +2.5%, and CTR +0.3%.

Local reproduction¶

Local comparison: The baseline is the shared transition + content scorer. Adding only the CWM core mechanism changes NDCG@10 by -17.89% relative.

MovieLens-100K, 260 users and 420 items, seed 42: NDCG@10 0.0354 → 0.0291 (-17.89%).

auto-research reproduce --paper cwm --dataset-dir data --seed 42
auto-research evolve --model rankmixer --dataset movielens-100k --direction "组合 cwm 与已安装论文算子" --generations 2 --population 4

See the fixed metrics.

Reproduction boundaries¶

Full-catalog MovieLens-100K exercises the paper's core state, objective, or inference path. No private features, production traffic, or online serving are used. Paper A/B results are cited separately and local numbers do not represent the original large model, data, traffic, or custom kernels.