Edge‑Cloud Integrated Recommendation for Xianyu "Mario" Feature

Background: Xianyu’s feed pages (home, search results, detail page) rely heavily on recommendation algorithms. With increasing mobile compute power, part of the processing can be moved to the client, offering real‑time response, resource savings, and richer data collection.

Advantages of on‑device computation:

1. Real‑time: feature extraction and scoring can be completed within ~1 s on the device.

2. Resource saving: distributing workload to devices reduces cloud compute pressure.

3. Multi‑dimensional data: fine‑grained user actions and privacy‑sensitive data can be consumed locally without uploading.

Mario Feature: an innovative UI element on the Xianyu home feed. When a user clicks a product card, the client requests a query and four recommended items from the backend, displaying them in a compact card to boost click‑through and conversion.

Problem: Initial launch showed lower CTR for Mario’s first‑ and second‑level clicks compared with regular items.

Analysis: User satisfaction after clicking a product can be inferred from dwell time and subsequent actions (e.g., viewing specs, reading reviews, chatting with the seller). Only high‑satisfaction clicks should trigger Mario to avoid wasteful impressions.

Solution: Deploy a two‑stage CTR model:

• CTR (edge): real‑time decision on the device based on fine‑grained user behavior whether to request/show the Mario module.

• CTR (cloud): determines which items to display when the module is triggered.

The edge model runs entirely on the client using Alibaba’s edge‑compute container, producing a satisfaction score that gates Mario’s appearance.

Results: The edge‑cloud hybrid approach reduced server request volume by 28.0%, increased Mario’s CTR by 31.1%, and lifted conversion rate by 10% while keeping the overall system minimally invasive.