Contextual Distillation Model for Diversified Recommendation¶

简体中文

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

Paper information¶

Field	Details
Paper	arXiv 2406.09021
Organization	Kuaishou Technology
First public release	2024-06-13 (arXiv v1)
Authors' code	The paper did not provide official or author code when checked on 2026-08-09.
Adapter / method	`cdm`
Local reproduction code	`src/auto_research/reproductions/cdm/`

Original paper¶

Motivation and method¶

A controllable MMR teacher generates lists balancing relevance and diversity. Distill contextual marginal gains into a student suitable for low-latency serving.

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

Key figure from the paper¶

Figure 1: Overall workflow, stages, and data flow. From the original paper; copyright remains with the authors.

Core equations¶

\[ y_t=\arg\max_i[s(i)-\lambda\max_{j<t}\operatorname{sim}(i,j)],\quad\mathcal L=\|f_\theta-y_{teacher}\|^2. \]

Paper results¶

Kuaishou main-app watch time rises +0.406% and clustering coefficient falls -0.957%.

Local reproduction¶

Local comparison: The shared transition + content scorer is the baseline. Adding only the cdm core mechanism yields +1.25% relative NDCG@10.

MovieLens-100K, 260 users and 420 items, seed 42: NDCG@10 moves from 0.0354 to 0.0358 (+1.25%) under the matched baseline and treatment.

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

See the fixed metrics.

Reproduction boundaries¶

Full-catalog MovieLens-100K executes the core state, objective, or inference path. No private company features, production traffic, or online service is used. Paper A/B measurements are distinct from local results, which do not represent the original model scale, private data, traffic, or custom kernels.