How to Build an End‑to‑End Marketplace Recommendation System: Product, Algorithms & Implementation

58转转 Overview

58转转 is a real‑person second‑hand goods trading platform under 58.com.

Designing the Recommendation Product Framework

Home page recommendation : extract user profile (age, gender, category preferences) and recommend items on the homepage.

Item detail page recommendation : "Customers also bought" and "You may also like" style related‑item suggestions.

Nearby recommendation : increase weight of geographic location, including frequent areas such as home and office.

Search recommendation : handle synonyms, near‑synonyms, misspellings, pinyin and suggest prompts like "Did you mean …?".

Recall recommendation : after user exits, use an RFM model to push coupons or messages for retention.

RFM Model Explained

RFM scores users by Recency (most recent purchase time), Frequency (purchase count in recent period) and Monetary (spending amount in recent period) to obtain a weighted intent score.

Implementing Recommendation Strategies

User Profiling

Combine registration data, historical behavior (purchase, favorite, like, share, comment, browse), and WeChat data to build a user characteristic profile, e.g.:

Age → recommend baby products, 3C items, etc.

Gender → recommend baby, beauty, health items.

Phone model → recommend mobile accessories.

Active time → push messages during active periods.

Category preference → recommend related categories.

Region → enable nearby recommendation.

Item Profiling

Items are unstructured; profiling is done per category (e.g., shoes) using keyword frequency and manual review, capturing attributes such as type, material, size, season, gender, price, shipping.

Tagging and Personalized Recommendation

Assign tags to both users and items.

Collect all item IDs a user has interacted with (purchase, favorite, like, share, comment, browse) into a set.

Aggregate tags of those items and weight them to compute user‑tag preferences.

Score each item by the weighted sum of its tags against the user’s tag preferences.

Exclude items already interacted with and rank the rest by score.

For search recommendation, add search‑condition weighting; for nearby recommendation, increase geographic weight.

Note: personalization accuracy depends on historical behavior data; for new users, fallback to “hot” items.

Classification‑Based Recommendation

Model the purchase decision as a binary (0/1) classification problem or probability estimation. Steps:

Prepare training data containing user, item, and purchase label, ensuring coverage of all users, tags, and items.

Build a classification model.

Predict purchase probability for each user‑item pair.

Rank items by probability and recommend the top ones.

Collaborative Filtering Recommendation

Construct a user‑item preference matrix from interaction data (purchase, favorite, like, share, comment, browse). Then:

Prepare the sparse matrix as training data.

Train a system‑filtering model.

Generate a personalized item list for each user.

Be aware that new items suffer from data sparsity in collaborative filtering and may need complementary strategies.

Note: This article was recorded by 58转转 recommendation lead Wang Hongquan during a technical exchange with 58转转, with minutes taken by Shen Jian.