Intelligent Publishing Solution for Xianyu C2C Platform

Background Xianyu is a typical C2C idle‑goods trading platform where users can freely publish items by entering name, price, stock, etc. While this freedom drives activity, it also creates challenges for product structuring.

Product structuring is difficult because users lack motivation to provide detailed attributes. The goal is to enable low‑cost structuring without heavy back‑office solutions.

Solution Options

1. Offline scheme : algorithmic association and social‑based activities. It can link items to similar products or tag attributes, but suffers from long data‑return cycles and lack of user confirmation.

2. Manual association : guide users during publishing to tag attributes or link similar items. Simple but shifts all cost to users, risking drop‑off.

The optimal answer is the Intelligent Publishing scheme: automatically associate a newly posted item with a product in the Taobao/Tmall catalog, inheriting its structured information.

Business Logic

The pipeline consists of three steps: (1) Subject focus – AI identifies the main object captured by the user; (2) Intelligent recognition & guidance – real‑time detection guides the user to capture core information; (3) Result feedback & user confirmation – after capture, the system presents matched items for user selection.

Architecture Design

Technical challenges include real‑time recognition during publishing and maximizing AI‑driven matching success. Xianyu leverages three advantages: mobile AI (AliNN), the massive Taobao/Tmall product database, and Alibaba Damo Academy’s AI capabilities.

The system uses a three‑layer architecture: UI presentation layer, logic processing layer, and framework layer (core modules). Implementation spans Flutter (UI), Java/OC (logic), and C++ (performance‑critical algorithms).

Key design considerations:

Utilize cross‑platform Flutter for UI consistency.

Deploy AI models on‑device to reduce latency and network dependence.

Compress uploaded images to ~10 KB, limiting traffic.

Flexible pipeline orchestration to adjust module order and add features.

Encrypt images to protect user privacy.

Algorithmic Architecture

The core algorithm matches the posted item to similar products. It upgrades single‑frame prediction to multi‑frame prediction and tightly integrates algorithmic feedback with user interaction.

When a frame is insufficient for confident prediction, the system prompts the user for additional shots, iterating until a reliable result is obtained.

Effectiveness

Real‑time performance shows no perceptible lag or frame drops. Multi‑frame recognition improves matching accuracy by about 20% over single‑frame.

Outlook

Intelligent publishing will be released in September, initially for video posting, later expanding to images and activities. Future vision includes fully automated item information extraction (tags, price, condition) with only user confirmation required.