Xianyu Product Understanding and Quality Item Selection Framework

Background

Xianyu has accumulated hundreds of millions of online items, but the quality varies, harming buyer experience. The goal is to introduce product understanding to select high‑quality items and improve the shopping experience.

Pain Points

Current Xianyu items suffer from:

Seller issues: many personal sellers are offline for long periods, preventing transactions.

Product issues: low‑quality goods, traffic‑driving listings, duplicate postings.

Price issues: false pricing and low cost‑performance.

Supply‑demand imbalance: large mismatches across categories.

We address these four dimensions to understand items and perform selection.

Product Attribute Dimension

New‑product attributes : brand, SKU/SPU, model, etc.

Second‑hand attributes : depreciation information (e.g., purchase date, wear, scratches) varies by category.

Coverage of these attributes is currently low; we will improve accuracy via user guidance and algorithmic association. We also filter out pornographic, low‑quality, traffic‑driving, and duplicate items.

Supply‑Demand Relationship

The classic supply‑demand curve defines equilibrium price (P) and quantity (Q). Xianyu shows many imbalances, e.g., excess women’s clothing vs. fast‑moving phones. Analyzing supply‑demand per category, brand, and depreciation helps alleviate overstock and guide replenishment.

Price Attribute

Buyers seek low price and high cost‑performance. By analyzing new and second‑hand attributes we can assess an item’s value and competitiveness, informing pricing decisions.

Seller Attribute

Metrics include activity rate, response rate, sell‑through rate, dispute rate, and positive feedback. Since Xianyu relies on personal sellers, offline or unresponsive sellers block transactions, making seller attributes a critical factor.

Implementation

We derive four indices—product information, supply‑demand, price, and seller—based on the identified factors. These indices are positively correlated with primary KPIs such as exposure‑to‑payment rate and sales velocity.

Regression analysis and weighting produce a "quality item score". The resulting score distribution is shown in the accompanying pie chart.

Quality Item Selection

How to select quality items?

We define objectives and constraints, e.g., maximize quality score while limiting total items to 20 million, ensuring category balance, and keeping dispute rate ≤ 0.3 %. This forms a combinatorial optimization problem whose solution yields the quality item pool.

Data Back‑testing

Back‑testing over a month compares quality items’ primary metrics against market averages. Consistent outperformance validates the selection method.

Simulation System

Beyond back‑testing past performance, a simulation predicts future impact of the quality‑item strategy, leveraging elementary, intermediate, and advanced knowledge (e.g., demographic preferences, activity‑specific effects, generalized purchase‑activity relationships).

Results

Deploying the quality‑item selection in new‑user first‑purchase and major promotion scenarios increased click‑through rates by up to 20 % and interaction‑payment rates by up to 100 % (exact figures omitted for data security).

Outlook

We will continue to refine the product‑layering system, collaborate with operations to enrich the product knowledge base, guide users, and expand the pool of high‑quality items, further enhancing user experience.