How AI Predicts Food Delivery Times: Insights from Alibaba’s KDD 2020 Study

Introduction

Recently, Alibaba’s local life smart logistics team’s paper “Order Fulfillment Cycle Time Estimation for On‑Demand Food Delivery” was accepted as an oral presentation at KDD 2020 Applied Data Science Track (SIGKDD, CCF A‑class conference, 5.8% acceptance).

Problem Overview

OFCT (Order Fulfillment Cycle Time) estimation is more complex than typical ETA problems because it involves supply‑demand relationships, unknown restaurant preparation time, and uncertain rider behavior. The paper introduces the problem to the academic community and proposes an effective solution.

Model Overview

By decomposing the delivery process, we analyze differences from other ETA tasks and explain influential features. Features are fed into a deep neural network, with latent vectors for restaurants, users, and riders to enhance prediction. A novel post‑processing neural operator improves convergence speed and accuracy. The model is deployed on Ele.me serving millions of users daily.

Key Features

User: order‑to‑delivery time for each user.

Restaurant: time from order acceptance to food ready.

Rider: time from order acceptance to delivery, including multi‑order pickups.

Platform: order assignment and route planning.

Difference from Standard ETA

Standard ETA predicts travel time from departure to destination, e.g., ride‑hailing. OFCT predicts the whole interval from user order to rider‑handed delivery, involving additional factors such as restaurant location, preparation time, and dispatch system.

Feature Engineering

Features include spatial IDs (user area, restaurant, city, grid), temporal attributes (hour, weekday), order size (item count, price), and derived supply‑demand ratios and completion rates. High order price often indicates larger or heavier items, leading to longer fulfillment.

Supply‑demand imbalance affects fulfillment length; higher ratios increase average wave length and time. Restaurant pending order count reflects busy status, causing longer waiting times.

Rider Information Utilization

Riders are grouped by grid cells; familiarity with local restaurants and neighborhoods improves efficiency. Features such as rider‑to‑restaurant distance and current order load are encoded and processed with an attention mechanism.

Similar‑Order ETA Estimation

Historical orders are used as memory; K‑nearest neighbor search finds similar past orders, and weighted average of their real delivery segment times provides an additional feature.

Handling Long‑Tail Data

The regression model struggles with right‑skewed long‑tail distribution. A post‑processing neural operator rescales predictions based on mean (μ) and standard deviation (σ) of true values, improving convergence and accuracy. An adaptive Box‑Cox inverse transform further enhances long‑tail fitting.

Experimental Results

The proposed model (encoding + prediction) outperforms the existing online baseline: MAE reduced by 9.8%, user complaint rate reduced by 19.3%, demonstrating clear user‑experience gains.

Conclusion and Future Work

The model is successfully deployed in production. Future research will focus on improving restaurant preparation time modeling and developing a multi‑task framework for simultaneous preparation‑time and fulfillment‑time prediction.