Interest-Based Live Stream Recommendation System for Xianyu

Background: After launching Xianyu live streaming, the biggest challenge was growth. Analysis showed distinct interest stages between short‑view and long‑view audiences.

Goal: Within three weeks, build a precise targeting platform to boost conversion UV for top live rooms and improve conversion rates.

Solution Overview: Combine operational expertise, BI analysis, and algorithms to understand users, anchors, and live rooms, then match interest groups with live streams.

User Understanding: Collect user features (basic profile, search, browse, post, transaction, interaction, interest tags). Features are computed offline, normalized, and stored for crowd selection.

Interest Tag Generation (Option 1): Structure keywords from behavior text, deduplicate, tokenize, and score. Convert operational keyword groups into DSL queries. Execute DSL to dump users matching keywords and assign interest tags. Perform crowd selection (intersection/union/difference) to export final groups. Challenges: short development window (2‑3 weeks) and massive storage cost (~30 PB) made this option infeasible.

Alternative Offline Approach (Option 2): Process unstructured text offline, extract structured text, match keyword groups, and compute user‑anchor interest relations. This reduces storage cost and supports complex UDFs, though it offers lower real‑time capability.

Delivery Implementation: Offline pipeline generates relations, syncs to an online graph database for recommendation. Configuration lives in an RDB, synced hourly to the warehouse; offline jobs update partitions to keep data near‑real‑time.

Results: In less than three weeks the end‑to‑end platform was live. Top live rooms exceeded UV targets, and click‑through rates for trial placements increased dramatically, with some categories achieving multiple‑fold gains.

Future Outlook: Enrich interest‑tag features, enable self‑service tag creation, expand placement capabilities with multi‑dimensional A/B testing, and abstract the core pipeline to support other community and non‑community products, ultimately achieving content structuring and low‑cost operations.