Intelligent Transaction System Construction for Halu Carpool

On July 22, 2022, the GIAC Global Internet Architecture Conference was held in Shenzhen, where senior algorithm expert Wang Fan from Halu gave a keynote titled “Construction of Halu Carpool Intelligent Transaction System”. The talk covered three main parts: business background, smart applications, and methodological summary.

Business Background – Halu started with shared bicycles and launched the carpool service more than three years ago. The service relies on algorithmic driver‑passenger matching to provide trustworthy rides. The product flow includes passenger order creation, order push to drivers, driver acceptance, and order completion, forming a funnel where many users drop off at each stage.

The complexity of the problem stems from a highly variable market environment (different cities, weather, competition), a three‑party ecosystem (drivers, passengers, platform), numerous decision points (cancellations, complaints), and a multifaceted goal system (completion volume, revenue, driver income, passenger retention, complaint rate).

Problem Decomposition – The process was broken down into order matching, pricing & marketing, and arbitration. Each sub‑problem has its own input factors (market environment, driver/passenger density, budget) and business outcomes (completion volume, revenue, retention, complaint rate).

Definition of Smart Applications – Based on the decomposition, four algorithmic applications were identified: intelligent matching, intelligent pricing, intelligent marketing, and intelligent arbitration. A value‑analysis framework was used to prioritize projects, with intelligent matching receiving the most investment.

Order Matching – The carpool scenario differs from on‑demand rides in timing, driver constraints, price, and user experience, making matching more challenging. The intelligent matching system combines offline features (demographics, historical routes) and real‑time features (click behavior, live trajectory) in a recommendation pipeline consisting of recall, ranking, and rule‑engine stages.

The architecture was compared with advertising recommendation systems, highlighting the need for real‑time computation, large candidate sets, and heavy map service usage.

Algorithm Evolution – The system evolved through three stages: (1) early rule‑based and simple models (linear, LightGBM); (2) deepening with Wide&Deep, DeepFM, and xDeepFM models; (3) customized models tailored to Halu’s data characteristics, including real‑time sequence models, multi‑task models, and “smart route” models. These iterations yielded over a 20% increase in overall order volume.

The next article will detail the ranking model, pricing, arbitration, and methodological lessons learned.