PLE: Progressive extraction of shared and task-specific information

简体中文

Actual shared/task-specific experts, CGC gates and separate heads; not Tencent Video scale.

Paper information

Field	Details
Paper	RecSys 2020 paper
Institution	Tencent.
First public release	2020-09-22 (ACM RecSys 2020)
Author code	No official author code linked by the paper as checked on 2026-08-09.
Local adapter or method key	ple
Local reproduction code	src/auto_research/reproductions/ple/

Paper method and evidence

MMoE shares all experts across tasks, which can entangle common and conflicting signals. PLE splits shared and task-specific groups; layered CGC gates progressively extract both kinds of information and reduce seesaw effects.

flowchart LR
  X[Shared input] --> E[Shared and task-specific experts]
  E --> G[Task-specific CGC gates]
  G --> H[Separate task heads]

Original paper figure

The original figure shows progressive extraction. From the original paper; copyright remains with the authors.

Core equation

\[ f^k(x)=\sum_{i\in E_k\cup E_s}g_i^k(x)f_i(x). \]

Reported results

The paper reports outperforming MMoE on public and Tencent Video tasks and describes deployment. Included as a user-approved classic exception.

Local reproduction

MovieLens-100K two-task entire-space protocol. Compare matched MMoE with two task-specific groups plus shared experts; each task gates only its own group and the shared group. Three-seed mean task AUC +1.34%; CTR 0.56882→0.57982, conversion 0.56009→0.56418. Metrics. Run auto-research reproduce --paper ple --dataset-dir data.