Intelligent Publishing Solution for Xianyu C2C Product Structuring

Background Xianyu, a C2C marketplace, allows users to list items by entering basic information, but achieving product structuring is challenging because users lack motivation to provide detailed attributes.

Solution Options Two offline approaches were considered: an algorithmic association method and a social activity‑based method, both suffering from long latency and lack of user confirmation. A manual association approach during publishing was also examined but shifted too much cost to users.

The proposed Intelligent Publishing solution automatically links a newly posted item to a matching product in the Taobao/Tmall catalog, providing structured information without extra user effort.

Business Logic The workflow focuses on three steps: (1) subject detection using AI on the device, (2) real‑time intelligent recognition and guidance to capture core information, and (3) result feedback where users confirm the suggested match.

Architecture Design The system is divided into three layers: UI presentation, logical processing, and framework core. It combines Flutter for cross‑platform UI, Java/Objective‑C for business logic, and C++ for performance‑critical AI modules.

Technical Challenges include ensuring real‑time recognition during publishing and maximizing matching success rates by fully exploiting AI capabilities.

Key Advantages are: (1) mobile AI enabled by AliNN, (2) access to the world’s largest product database (Taobao & Tmall), and (3) strong AI expertise from Alibaba DAMO Academy.

The architecture incorporates a flexible pipeline that can reorder or add modules, supporting both Android and iOS implementations.

Algorithmic Architecture The core algorithm matches items using multi‑frame predictions and tightly integrates algorithmic decisions with user interaction, prompting additional captures when a single frame is insufficient.

Results Real‑time performance shows no perceptible lag for users, and multi‑frame recognition improves matching accuracy by about 20% over single‑frame methods.

Outlook The intelligent publishing feature will launch in September, initially for video listings, with future extensions to images and activity scenarios, aiming to further increase product structuring rates on Xianyu.

Author: 意境 – Source: Xianyu Tech public account