Intelligent Traffic Distribution in 58 Local Services: Algorithmic Practices and System Optimization

58 Local Services comprises hundreds of sub‑categories and multiple innovative and international businesses, making intelligent traffic distribution a major challenge. By integrating search and recommendation scenarios, a unified traffic distribution system was built to deliver precise content to users.

The presentation covers four main topics: an introduction to the local services ecosystem, the characteristics and problems of traffic distribution on the main site, the solutions implemented for the main site, and a summary with future outlook.

1. Business Background – The local services business, formerly known as the "Yellow Pages," now includes over 200 industries (e.g., beauty, food, housekeeping, education) and three major product lines: main site, to‑home, and to‑store/e‑commerce. Traffic is organized through a partner network that reshapes traditional distribution channels.

2. Main‑Site Traffic Flow – Users arrive via app/PC/M‑terminal, browse list pages, view details, and initiate contact via phone or in‑app chat. Key interaction data include impressions, clicks, calls, chats, and comments.

3. Challenges – Severe information homogeneity, complex user segmentation (anonymous, new, low‑activity), varied decision cycles (short‑term vs. long‑term), and a massive, heterogeneous label space (>200 categories, >10 recommendation scenes) make accurate targeting difficult.

4. Solutions

Information Structuring – Introduced structured publishing templates, standardized service attributes, and intelligent pricing to increase content distinctiveness. Designed and applied label systems (industry, generic, and application tags) and built a label‑mining pipeline that extracts candidate words, merges synonyms, and resolves industry ambiguities.

Knowledge Structuring – Created a hierarchical tag taxonomy by clustering tags into dimensions using vector similarity, then expanded to scene‑specific tags. This enabled cross‑category tag discovery and supported tag‑based retrieval, recommendation, and similarity search.

User Intent Modeling – Developed multi‑task learning models that share a common user representation while optimizing for multiple objectives (CTR, CVR, Call/UV). User vectors are generated from recent behavior sequences via embedding, Bi‑LSTM, and attention mechanisms, with incremental updates to capture evolving intent.

Layered Optimization – Optimized recall by converting query intent into a tree‑structured representation and using vector‑based retrieval; enhanced ranking with multi‑objective learning; applied real‑time intent weighting to adjust tag importance based on recent actions; and implemented dynamic UI adjustments (titles, images, summaries) to reflect current user intent.

System Integration – Unified search, recommendation, and summarization components through shared data assets (user profiles, behavior logs, content items) and a component registry, reducing duplication, enabling model reuse, and simplifying configuration for different scenarios.

5. Summary & Outlook – Future work includes re‑ranking after user interactions, cross‑category recommendations (e.g., moving → cleaning), group‑order recommendations, periodic service suggestions, closed‑loop feedback from offline conversions, social‑network‑based distribution via partners, and profit‑driven traffic allocation. The goal is to continuously refine the algorithmic pipeline to better match users with the vast array of local services.