PLE Progressive Layered Extraction

PLE (Progressive Layered Extraction, proposed by Tencent PCG at RecSys 2020 Best Paper) is a breakthrough multi-task recommendation architecture that eliminates Negative Transfer and Seesaw Phenomena; instead of indiscriminately sharing all experts across all tasks like MMoE, PLE explicitly bifurcates the expert pool into two distinct classes: 1) Task-Specific Experts (dedicated exclusively to a single objective, e.g., CTR-only or CVR-only); 2) Shared Experts (capturing universal commonalities); using multi-layer cascading Customized Gate Control (CGC) blocks, PLE progressively separates task-specific parameters from shared parameters layer-by-layer.

Heavy ranking across multi-objective entertainment feeds (Tencent Video, WeChat Channels, TikTok) optimizing Click, Like, Share, Complete View, and Purchase.

MMoE suffers severe 'Seesaw Phenomena' (improving one task degrades another) when task correlations are heterogeneous; PLE's physical isolation of task-specific and shared experts mathematically guarantees conflict-free multi-task representation learning.

Diagram the routing flow of task-specific vs shared experts across single-layer CGC and multi-layer progressive cascades in PLE?

Compare Shared-Bottom, Cross-Stitch, MMoE, and PLE in robustness against negatively correlated tasks (high click, zero completion)?

Why does cascading extraction (shallow layers extracting generic representations, deep layers isolating specific tasks) align with deep representation learning?

What are the parameter budget controls and distributed training VRAM optimizations when scaling PLE to 5-8 simultaneous targets?

How to combine multiple predicted probabilities ($p\text{CTR}, p\text{CVR}, p\text{Finish}$) into a unified ranking score in final serving?