Market Mechanisms and Control Measures in Ele.me Food Delivery Recommendation Algorithms

This article, based on a talk by Ele.me R&D Director Ma Yao at the 2018 AI Pioneer Conference, explains how the company’s food‑delivery recommendation platform uses market mechanisms and control measures to maximize traffic efficiency while maintaining a healthy ecosystem.

1. Ele.me Delivery Recommendation Business Model – Search and recommendation serve as the core traffic entry, handling millions of orders daily across various scenarios such as home page, search, activity venues, and order pages, covering not only meals but also fruits, fresh produce, medicine, and flowers.

2. Platform Goals and Positioning – The platform balances three roles: users (traffic), merchants (supply), and the platform itself. Objectives include providing a good decision experience for users, delivering precise traffic to merchants, and ensuring platform revenue and ecosystem health.

3. Special Challenges in Food Delivery Recommendation – Challenges include post‑order delivery latency, peak‑time traffic spikes, resource constraints (restaurants, riders), and regional traffic distribution, all of which require sophisticated real‑time optimization.

4. Market Mechanisms in the Recommendation Algorithm – Efficiency is pursued through deep data mining, precise recall, and intelligent ranking. The system decomposes goals such as GMV into sub‑models (click, conversion, order value) and continuously evolves from conversion‑rate optimization to visit‑rate, GMV, and finally real‑paid GMV.

5. System Architecture – The stack includes big‑data processing (Spark, Hadoop), deep‑learning frameworks (TensorFlow), distributed services (EKV), and real‑time computation platforms. Feature engineering combines manual and model‑generated features across user, merchant, and context dimensions.

6. Model Evolution and Ranking – The ranking pipeline progressed from rule‑based methods to linear models (LR), then to non‑linear models, GBDT, and deep learning (Wide&Deep, DeepFM). Online learning with XGBoost + FTRL and real‑time engines further adapts to rapid behavior changes.

7. Control Measures – To address long‑term ecosystem goals, the platform employs multi‑model fusion, user‑lifecycle analysis, regional customization, and multi‑armed bandit algorithms (Thompson sampling, Epsilon‑Greedy, UCB, LinUCB) for traffic reallocation, new‑store support, and exploration‑exploitation balance.

8. Practical Results – Experiments show significant improvements such as 70% higher exposure efficiency for new stores and 30% increase in new‑store conversion rates.

Author Introduction – Ma Yao, Ele.me R&D Director responsible for search, recommendation, ranking, and advertising technologies, with prior experience at Meituan‑Dianping and Tencent.

Recruitment Notice – Ele.me is hiring senior algorithm engineers, experts, and architects for search/recommendation/advertising/user growth, inviting interested candidates to contact Ma Yao via email.