Shopee Live Personalized CTR Optimization via Calibration‑Based Meta‑Learning

This article introduces Shopee's personalized click‑through‑rate (CTR) method built on calibration‑based meta‑learning.

Shopee Live is an in‑app live streaming scenario whose recommendation pipeline consists of recall, coarse‑ranking, fine‑ranking, and re‑ranking; the focus of this work is the fine‑ranking stage and CTR optimization.

The live recommendation aims to increase positive feedback such as watch duration, interactions, and orders while reducing negative signals like short plays. Modeling targets include content‑consumption metrics (CTR, stay time) and e‑commerce metrics (add‑to‑cart, order); this work concentrates on the content‑consumption targets.

Model Evolution – Calibration4CVR : Proposed in 2018, Calibration4CVR uses representation calibration to share neurons between CTR and CVR layers, calibrating the CVR layer from the CTR layer.

Model Evolution – CBMR (Calibration‑based meta‑Rec) : The baseline Wide&Deep model lacks per‑sample personalization. CBMR adds meta‑learning modules EmbCB and MlpCB that generate sample‑adaptive parameters for the embedding and MLP layers, achieving a “one‑sample‑one‑model” effect. Experiments show AUC improvement of 1.9%, +4.87% total watch time, and +13.91% orders.

EmbCB Optimization : Initial attempts with SeNet (sparse feature mean‑pooling + concat) produced zero weights due to ReLU activation. Switching to sigmoid, then ELU, and adjusting hidden‑layer initialization did not resolve the issue. Replacing the fixed pooling vector with a learnable vector and using linear activation increased weight variance; expanding shared vectors from 1 to 2 gave a 0.15% AUC gain, and to 16 gave 0.36% gain. Adding ReLU non‑linearity, widening hidden layers to 128 further raised AUC by 0.5%.

MlpCB Optimization : Inspired by PPNet, the original design kept userid/hostid separate, limiting performance. Making all features trainable in both the main and gate networks, and converting parallel multi‑gate layers to serial connections, increased AUC by 1.25% while keeping parameter growth manageable.

Multi‑Task Modeling – MultiCBMR : Extends CBMR to a multi‑task setting (e.g., CTR and stay‑time). EmbCB provides sample‑specific embedding weights, while MlpCB supplies task‑specific MLP parameters. Expanding MlpCB influence to every expert layer and task layer further improves AUC; sharing EmbCB across experts keeps parameter count reasonable.

Multi‑Scene Modeling – MultiCBMR (Scene) : Applies the same architecture across multiple live scenes and user cohorts, yielding consistent gains in offline AUC and online metrics.

In summary, the calibration‑based meta‑learning framework substantially outperforms traditional Wide&Deep models in Shopee Live's fine‑ranking, demonstrating the effectiveness of per‑sample adaptive modeling for CTR and related metrics.