MRGRP: Empowering Courier Route Prediction in Food Delivery Service with Multi-Relational Graph Conference: Companion Proceedings of the ACM Web Conference 2025; April 28-May 2, 2025; Sydney, NSW, Australia Companion Proceedings of the ACM Web Conference 2025 (WWW Companion ’25), April 28-May 2, 202

Chang Liu Huan Yan ∗ Hongjie Sui, Haomin Wen, Yuan Yuan, Yuyang Han Hongsen Liao Xuetao Ding Jinghua Hao, Yong Li ∗ · 2025

Read MRGRP: Empowering Courier Route Prediction in Food Delivery Service with Multi-Relational Graph Conference: Companion Proceedings of the ACM Web Conference 2025; April 28-May 2, 2025; Sydney, NSW, Australia Companion Proceedings of the ACM Web Conference 2025 (WWW Companion ’25), April 28-May 2, 202 by Chang Liu Huan Yan ∗ Hongjie Sui, Haomin Wen, Yuan Yuan, Yuyang Han Hongsen Liao Xuetao Ding Jinghua Hao, Yong Li ∗ (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-11999.txt.

Introduction

Liu and a team including Yan, Sui, and Wen guess food fleet routes from a web of task links in this twenty twenty-five paper. Food fleets must call each rider's next stops to pair tasks well and keep waits low. Hand rules use few task traits and miss how riders pick for themselves.

Inside, tasks are cast as nodes in a web tied by space, time, and pick-drop links. A graph-aware net reads that web to learn how tasks sway each rider's picks. A route builder then weighs rider traits and live gap cues as it lays the route. Tests on past fleet data from towns of all sizes top the chart. On a live fleet hub, the tool beat the hand rule by a clear grade.

Catalog record

CREATORChang Liu Huan Yan ∗ Hongjie Sui, Haomin Wen, Yuan Yuan, Yuyang Han Hongsen Liao Xuetao Ding Jinghua Hao, Yong Li ∗
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE56 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-11999

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

MRGRP: Empowering Courier Route Prediction in Food Delivery Service with Multi-Relational Graph Conference: Companion Proceedings of the ACM Web Conference 2025; April 28-May 2, 2025; Sydney, NSW, Australia Companion Proceedings of the ACM Web Conference 2025 (WWW Companion ’25), April 28-May 2, 2025, Sydney, NSW, Australia DOI: 10.1145/3701716.3715250 ISBN: 979-8-4007-1331-6/2025/04 Thanks: ∗ * Huan Yan and Yong Li are corresponding authors. CCS: Computing methodologies Planning for deterministic actions Chang Liu Huan Yan ∗ Hongjie Sui, Haomin Wen, Yuan Yuan, Yuyang Han Hongsen Liao Xuetao Ding Jinghua Hao, Yong Li ∗ Abstract . Instant food delivery has become one of the most popular web services worldwide due to its convenience in our daily lives. A fundamental problem in this service scenario is the accurate prediction of courier routes, which is essential for optimizing task dispatch and thereby improving delivery efficiency. It not only increases the satisfaction of both couriers and users but also drives higher profitability for the platform. The deployed heuristic prediction approach of the platform accounts for only limited human-selected task features and neglects couriers’ preferences, resulting in sub-optimal performances. Moreover, existing learning-based methods fail to adequately explore the diverse factors influencing couriers’ decision-making behaviors and intricate relationships among factors. To meet the urgent need for a powerful route prediction approach that benefits millions of couriers, users, and the platform itself, we propose a M ulti- R elational G raph-based R oute P rediction (MRGRP) method, which enables fine-grained modeling of the correlations among tasks influencing couriers’ decision-making and achieves accurate prediction. We encode spatial and temporal proximity, along with the pickup-delivery relationships of tasks, into a multi-relational graph, then design a GraphFormer architecture to capture these complex correlations. Furthermore, we introduce a route decoder that leverages information about couriers as well as dynamic distance and time contexts for dynamic prediction. It also utilizes existing route solutions as a reference to find better outcomes. Experimental results demonstrate that our proposed model attains state-of-the-art performance in route prediction on offline data from cities of diverse scales. We then deploy our model on the Meituan Turing online platform, where it significantly surpasses the existing deployed heuristic algorithm and achieves a high route prediction accuracy of 0.819, which is required for the quality and satisfaction of couriers and users in the instant food delivery service. We have open-sourced our code at https://github.com/tsinghua-fib-lab/MGRoute . † † cc-license: by Figure 1. The overall framework of MRGRP. 1. Introduction Instant food delivery services have become an indispensable aspect of contemporary daily life, primarily due to the convenience they offer in consumption ( Lan et al. 2020 ; de Araujo and Etemad 2021 ; Huang et al. 2020 ) . After customers submit orders, the service platform communicates the order details to the restaurant and assigns the delivery task to a courier. Once the courier accepts the assignment, they collect the prepared food at the restaurant and subsequently deliver it to the customer’s location. Existing analysis ( Gao et al. 2021 ; Li and Phillips 2018 ; Wen et al. 2021 ) indicate significant deviations between the actual routes of couriers and the recommended routes by the platform. Therefore, one of the fundamental problems in this service scenario is the accurate prediction of courier routes, which is crucial for downstream tasks, including arrival time estimation and task dispatching. Effective route prediction facilitates delivery efficiency and task allocation, thereby improving the working experiences of couriers and increasing user retention. Ultimately, these advancements contribute to greater profitability for the platform. Extensive research and numerous initiatives have been undertaken in the field to address this issue. Rule-based methods, such as TimeRank and DisGreedy, predict routes based solely on distance and time, thereby neglecting the intricate features associated with tasks. The heuristic approach TSFH deployed by Meituan, China’s largest food delivery service provider, considers only a limited set of human-selected features and overlooks couriers’ preferences. Consequently, these methods result in sub-optimal performance and fail to meet the stringent requirements for downstream tasks. With advancements in learning techniques, significant efforts have been made to leverage these methods to capture diverse task features, thereby enhancing prediction performance. For instance, FDNET ( Gao et al. 2021 ) utilizes a deep architecture to process task features and integrates a combination of LSTM and attention network to predict future routes. Graph2Route ( Wen et al. 2022a ) introduces a dynamic graph neural network to capture task correlations. Although these methods have demonstrated improved prediction performance, there are still three key challenges that require further consideration: • Inadequate exploration of the multifaceted correlations and dependencies among tasks. During route selection, it is crucial to simultaneously consider various aspects of task dependencies and correlations, such as spatial-temporal proximity and sequential order. For example, after completing a pick-up task, a courier may face three subsequent delivery tasks: one that is spatially proximate but not urgent, another that is distant but requires immediate delivery, and a third with moderate distance and urgency, corresponding to the same order as the just-picked food. Accurately capturing and managing these diverse and complex factors presents a significant challenge for existing frameworks. As the number of tasks increases, their interdependencies become increasingly complex, necessitating the development of effective modules to capture these correlations and thereby facilitate route predictions. • Large solution search space in route prediction. Route prediction models are usually required to identify the most executable route from a vast feasible solution space that grows exponentially with the number of tasks. Such inherent complexity renders the route prediction problem exceptionally challenging. Consequently, there is a pressing need for the development of advanced frameworks capable of effectively managing the vast solution space of route predictions. • Inherent dynamism of problem conditions. The pick-up and delivery process is characterized by an inherent dynamism, where the problem context continuously evolves in response to the courier’s changing location and the passage of time. Such dynamism introduces complexities that cannot be overlooked during route prediction. For instance, the relative distances between tasks shift as the courier moves to a new location. Additionally, certain tasks that may have initially appeared non-urgent can escalate in priority as time progresses, demanding immediate attention. Ignoring this spatial and temporal dynamism can lead to sub-optimal routes and potential delays in service. To deal with the above challenges, we develop a powerful M ulti- R elational G raph-based R oute P rediction (MRGRP) method with an encoder-decoder architecture. To address the first challenge, we extract and establish multiple relationships among dispatched tasks by considering various factors that influence couriers’ decision-making and represent these relationships in the form of a multi-relational graph. Furthermore, we design a multi-relational graph encoder based on the GraphFormer architecture, which proficiently assimilates these intricate correlations among different tasks. To tackle the second challenge, we design a route decoder that leverages existing competitive route solutions as a reference to guide the prediction process. The decoder utilizes distances and the estimated remaining time of tasks to enhance dynamic route prediction, effectively narrowing the solution space and improving the prediction accuracy. Additionally, we introduce an auxiliary arrival time prediction procedure that estimates arrival times at each decoding step, thereby enabling the calculation of the remaining time for each task. By further integrating the

Read the complete text