When routing meets recommendation: Solving dynamic order recommendations problem in peer-to-peer logistics platforms

• Main Authors: ZHANG, Zhiqin, JOE, Waldy, ER, Yuyang, LAU, Hoong Chuin
• Format: text
• Language: English
• Published: Institutional Knowledge at Singapore Management University 2023
• Subjects: Crowdsourced delivery; Data-driven optimization; Recommendations system; Databases and Information Systems; Data Storage Systems
• Online Access: https://ink.library.smu.edu.sg/sis_research/8348; https://ink.library.smu.edu.sg/context/sis_research/article/9351/viewcontent/When_Routing_Meets_Recommendation.pdf
• Institution: Singapore Management University

Peer-to-Peer (P2P) logistics platforms, unlike traditional last-mile logistics providers, do not have dedicated delivery resources (both vehicles and drivers). Thus, the efficiency of such operating model lies in the successful matching of demand and supply, i.e., how to match the delivery tasks with suitable drivers that will result in successful assignment and completion of the tasks. We consider a Same-Day Delivery Problem (SDDP) involving a P2P logistics platform where new orders arrive dynamically and the platform operator needs to generate a list of recommended orders to the crowdsourced drivers. We formulate this problem as a Dynamic Order Recommendations Problem (DORP). This problem is essentially a combination of a user recommendation problem and a Dynamic Pickup and Delivery Problem (DPDP) where the order recommendations need to take into account both the drivers’ preference and platform’s profitability which is traditionally measured by how good the delivery routes are. To solve this problem, we propose an adaptive recommendation heuristic that incorporates Reinforcement Learning (RL) to learn the parameter selection policy within the heuristic and eXtreme Deep Factorization Machine (xDeepFM) to predict the order-driver interactions. Using real-world datasets, we conduct a series of ablation studies to ascertain the effectiveness of our adaptive approach and evaluate our approach against three baselines - a heuristic based on routing cost, a dispatching algorithm solely based on the recommendation model and one based on a non-adaptive version of our proposed recommendation heuristic - and show experimentally that our approach outperforms all of them.