Skip to content

Papers

Fulfillment & Last-Mile

Enhancing Real-Time Order Pooling with Skilled Couriers: A 5.3% Boost in Efficiency

The paper "Harvesting Efficient On-Demand Order Pooling from Skilled Couriers" introduces a novel framework, the Skilled Courier Delivery Network (SCDN), to enhance real-time order pooling in on-demand food delivery (OFD) services. By leveraging graph representation learning and skilled courier data, the SCDN improves the efficiency of order assignments, reducing detours and overtime risks.

Original source: arXiv

Enhancing Real-Time Order Pooling with Skilled Couriers: A 5.3% Boost in Efficiency

Ask SCI.AI about this story

📋 本文要点

The paper "Harvesting Efficient On-Demand Order Pooling from Skilled Couriers" introduces a novel framework, the Skilled Courier Delivery Network (SCDN), to enhance real-time order pooling in on-demand food delivery (OFD) services. By leveraging graph representation learning and skilled courier data, the SCDN improves the efficiency of order assignments, reducing detours and overtime risks.

Paper: Harvesting Efficient On-Demand Order Pooling from Skilled Couriers: Enhancing Graph Representation Learning for Refining Real-time Many-to-One Assignments

Authors: Yile Liang, Jiuxia Zhao et al.

Published: 2024-06-20

Venue: arXiv preprint

Source: https://arxiv.org/abs/2406.14635

Research Background

The recent surge in on-demand food delivery (OFD) services has transformed how consumers access meals, but it also presents significant challenges for platforms in terms of efficient order assignment. This section explores the problem, its importance, the industry context, and the shortcomings of prior approaches.

The Problem and Its Importance

The rapid growth of OFD services, driven by technological advancements and consumer demand, has led to an increase in the number of orders that need to be delivered within tight time constraints. In 2023, Meituan Waimai, one of China’s leading OFD platforms, handled over 70 million orders daily, involving 6.24 million couriers. The key challenge is to efficiently pool multiple orders for simultaneous delivery, thereby reducing costs and improving service reliability. However, the complexity and real-time nature of order assignment make this task extremely difficult. Unreasonable order pooling can lead to detours and prolonged delivery times, negatively impacting the experiences of all stakeholders.

The importance of efficient order pooling cannot be overstated. For consumers, timely and reliable deliveries are crucial for maintaining satisfaction. For merchants, ensuring that their products are delivered fresh and on time is essential for customer retention. For couriers, efficient routes and fewer detours mean higher earnings and better working conditions. For the platform, optimizing order pooling leads to reduced operational costs and increased profitability. Therefore, finding a solution to this problem is critical for the sustainability and growth of OFD platforms.

Industry Context and Shortcomings of Prior Approaches

In the OFD industry, platforms like Meituan Waimai act as intermediaries, connecting consumers, merchants, and couriers. The goal is to balance the needs of these stakeholders while optimizing efficiency. Prior research on order pooling has primarily focused on traditional warehouse management, where time constraints are more relaxed. These methods, which typically operate on a scale of minutes or even hours, are not well-suited for the urgency required in OFD, where decisions must be made within seconds.

Recent studies have attempted to address this issue by using geographical proximity and promised delivery times, but these criteria-based batching rules limit the scope for consolidation. Other approaches, such as exact algorithms and iterative clustering, suffer from high computational complexity and are impractical for real-time processing. For example, an exact algorithm proposed in [31] assumes perfect information about the arrival of orders, which is unrealistic in real-world scenarios. Iterative clustering methods, as seen in [10, 11], achieve order consolidation but are computationally intensive and hinder real-time processing.

Additionally, the “last mile” offline environment, characterized by unforeseen road closures, natural obstacles, and pandemic-related lockdowns, further complicates order pooling decisions based on coarse data and limited awareness. OFD platforms often lack the detailed spatiotemporal data needed to make informed decisions, leading to suboptimal order pooling and reduced delivery efficiency.

Key Findings

The paper presents several key findings that demonstrate the effectiveness of the Skilled Courier Delivery Network (SCDN) in enhancing real-time order pooling. Each finding is supported by method principles, key design/algorithm logic, experimental setup, and concrete evidence.

Graph Modelling: Constructing the SC Delivery Network

The SCDN framework constructs a delivery network from skilled courier (SC) route sequences, with flow units (FUs) as nodes linked by SC behavior sequences. An FU is a directed vector from pick-up areas of interest (AOIs) to delivery AOIs. The network is formulated as an attributed multiplex heterogeneous network (AMHEN), with FU nodes featuring multiple attributes for temporal and spatial information, and links representing two different types of courier behaviors: pick-up and delivery. This approach captures the shared experiences of SCs, integrating all FU sequences into a unified yet heterogeneous graph. The use of rich temporal and spatial information, such as average historical order amount and delivery distance, enhances the learning accuracy.

To illustrate, consider a scenario where a courier picks up orders from three different locations and delivers them to three different destinations. The pick-up and delivery sequences form a sequence of FUs, which are then integrated into the AMHEN. This integration allows the system to learn from the collective behavior of skilled couriers, identifying patterns and potential efficiencies that can be leveraged for future order pooling.

Experimental results show that the SCDN framework significantly improves the quality of order pooling. For instance, the total MD score of the many(order)-to-one(courier) assignment (MOA) problem is improved by 5.3%, indicating more efficient order assignments with reduced detours and overtime risks. Additionally, the newly-built mode cuts the average incremental pick-up time for couriers during noon peak hours by 51%. The system also boosts courier efficiency by 45-55% during noon peak hours, while upholding the timely delivery commitment.

Learning Algorithm: Enhanced Attributed Heterogeneous Network Embedding (EATNE)

Based on GATNE, an effective graph representation learning (GRL) method for AMHEN, the authors derive an enhanced attributed heterogeneous network embedding (EATNE) approach tailored for OFD. The EATNE approach includes three key enhancements:
1. **Region-Congregated Negative Sampling:** To improve algorithm performance, a region-congregated negative sampling mechanism is proposed, which outperforms traditional randomized negative sampling. This mechanism ensures that negative samples are drawn from the same region as the positive samples, leading to more meaningful embeddings. Specifically, the region-congregating mechanism reduces the noise in the training process, resulting in embeddings that better capture the local context and relationships between FUs.
2. **Customized Margin Ranking Loss:** Instead of using cross-entropy, a customized margin ranking loss is employed to refine embedding quality. This loss function encourages the model to learn embeddings that better capture the relationships between FUs, leading to more accurate predictions. The margin ranking loss helps in distinguishing between similar and dissimilar FUs, thereby improving the overall quality of the embeddings.
3. **Cold Start Mitigation Mechanism:** To address dispersed order distribution and limited FU coverage in SC behaviors, a cold start mitigation mechanism is built using geographic information to generate embeddings for previously unseen FUs, thus broadening coverage. This mechanism ensures that even in regions with limited historical data, the system can still generate meaningful embeddings. The cold start mitigation mechanism uses geographic features to infer the likely characteristics of new FUs, allowing the system to make more informed decisions even in data-scarce regions.

Experimental results show that the EATNE approach significantly improves the quality of FU embeddings. For instance, the total MD score of the MOA problem is improved by 5.3%, indicating more efficient order assignments with reduced detours and overtime risks. Additionally, the newly-built mode cuts the average incremental pick-up time for couriers during noon peak hours by 51%. Furthermore, the system boosts courier efficiency by 45-55% during noon peak hours, while upholding the timely delivery commitment.

MOA Search Space Refinement and OFD Application

Utilizing FU embeddings, the SCDN refines the order combination and courier recall mechanisms within Meituan’s dispatch system, facilitating superior real-time order pooling. The SCDN refines order structure profiles and pinpoints scale-effect hotspots within the MOA’s vast search space, uncovering independent and small-scale subspaces for thorough and high-quality order pooling. The new delivery mode developed through SCDN enhances courier efficiency without compromising service reliability.

Online tests reveal that the SCDN significantly improves order pooling. The total MD score of the MOA problem is improved by 5.3%, and the average incremental pick-up time for couriers during noon peak hours is reduced by 51%. Furthermore, the system boosts courier efficiency by 45-55% during noon peak hours, while upholding the timely delivery commitment.

To achieve these improvements, the SCDN framework uses scalable low-dimension vector calculations instead of massive and time-consuming PDRP computations. This approach effectively prunes the MOA problem’s search space, shown in Figure 3, and extracts small-scale and isolated subspaces promising for high-quality order consolidation results, facilitating real-time, effective order pooling. The scalability of the SCDN is particularly important during peak hours when the number of orders and couriers is at its highest. By reducing the computational load, the SCDN enables the dispatch system to handle large volumes of orders efficiently, ensuring that couriers are assigned the most optimal routes in real-time.

Limitations

While the SCDN framework demonstrates significant improvements in real-time order pooling, it also has several limitations and areas for further improvement. This section enumerates these limitations, their impact, and possible mitigations.

Limited Generalizability

The SCDN framework is primarily tested and deployed in the context of Meituan Waimai, a specific OFD platform in China. The unique characteristics of this platform, such as the large number of orders and couriers, may not be representative of other OFD platforms or regions. This limitation impacts the generalizability of the findings. For example, smaller platforms or those operating in less densely populated areas may not see the same level of improvement. To mitigate this, future research could test the SCDN in different OFD platforms and regions to validate its effectiveness across various contexts. Additionally, the framework could be adapted to account for local factors such as traffic patterns, road conditions, and consumer preferences.

Dependency on Skilled Courier Data

The SCDN relies heavily on the availability and quality of skilled courier (SC) data. If the platform does not have a sufficient number of skilled couriers or if the data is incomplete, the performance of the SCDN may be compromised. This dependency can be addressed by developing robust data collection and validation mechanisms to ensure the availability and quality of SC data. For instance, platforms could implement more rigorous training programs for couriers and use incentives to encourage participation in data collection. Additionally, incorporating data from less experienced couriers and using transfer learning techniques could help improve the robustness of the framework. Transfer learning can enable the system to leverage knowledge from similar tasks or regions, making it more adaptable to different environments.

Computational Complexity in Large-Scale Deployments

Although the SCDN reduces the computational complexity of the MOA problem, it still requires significant computational resources, especially in large-scale deployments. This can be a challenge for platforms with limited computational infrastructure. For example, in big cities during peak hours, the number of orders and couriers can be overwhelming, requiring substantial computational power to process in real-time. To mitigate this, the framework could be optimized further to reduce computational requirements. Techniques such as parallel processing and distributed computing could be employed to handle large-scale deployments more efficiently. Cloud-based solutions could also provide the necessary computational power, allowing platforms to scale their operations without significant upfront investment in hardware.

Practical Implications

The SCDN framework offers several practical implications for supply-chain and AI practitioners. This section outlines at least three concrete scenarios, decisions, and implementation paths for leveraging the SCDN in real-world applications.

Enhanced Real-Time Decision-Making

The SCDN framework enables OFD platforms to make more informed and efficient real-time decisions regarding order pooling and courier assignments. By leveraging the rich spatiotemporal data and skilled courier behaviors, platforms can dynamically adjust their order pooling strategies to optimize delivery efficiency and reduce detours. This can be implemented by integrating the SCDN into the existing dispatch systems, allowing for real-time updates and adjustments based on current conditions and courier availability.

For example, during peak hours, the system can automatically identify and prioritize high-quality order combinations, ensuring that couriers are assigned the most efficient routes. This not only reduces delivery times but also improves the overall experience for couriers, who can complete more deliveries in less time. Platforms can also use the insights from the SCDN to fine-tune their dispatch algorithms, continuously improving the accuracy and efficiency of their order pooling processes. The real-time decision-making capabilities of the SCDN can be particularly beneficial in dynamic environments where conditions change rapidly, such as during unexpected weather events or sudden surges in order volume.

Improved Courier Experience and Retention

By reducing the average incremental pick-up time and minimizing detours, the SCDN framework enhances the overall experience for couriers. This can lead to higher job satisfaction and better retention rates. Platforms can use the insights from the SCDN to design incentive programs and training initiatives that further support couriers, ensuring they have the tools and knowledge to perform their tasks more efficiently. For instance, platforms can offer bonuses for completing high-efficiency deliveries and provide real-time feedback to couriers, helping them improve their performance.

Additionally, the SCDN can be used to create personalized training programs for couriers, based on their individual performance and the specific challenges they face. By providing targeted training and support, platforms can help couriers develop the skills they need to succeed, leading to a more motivated and productive workforce. This, in turn, can lead to better service for consumers and higher satisfaction for all stakeholders. Improved courier retention can also reduce the costs associated with recruiting and training new couriers, further contributing to the overall efficiency and profitability of the platform.

Scalable and Adaptive Solutions for Different Markets

The SCDN framework can be adapted to different markets and regions, providing scalable solutions for OFD platforms operating in diverse environments. By customizing the framework to account for local factors such as traffic patterns, road conditions, and consumer preferences, platforms can achieve optimal order pooling and delivery outcomes. This can be implemented by conducting localized data analysis and fine-tuning the SCDN parameters to match the specific needs of each market.

For example, in urban areas with heavy traffic, the SCDN can be adjusted to prioritize routes that avoid congestion, while in rural areas, it can focus on minimizing travel distances. Continuous monitoring and evaluation of the SCDN’s performance can help identify areas for further improvement and adaptation. Platforms can also use the SCDN to experiment with different order pooling strategies, testing and refining their approaches to find the most effective solutions for their specific contexts. The adaptability of the SCDN makes it a versatile tool for OFD platforms looking to optimize their operations in a wide range of environments, from densely populated urban centers to more sparsely populated rural areas.

Source: https://arxiv.org/abs/2406.14635

Ask SCI.AI Finished reading? Continue with SCI.AI. Explore the related policy, route, company and historical context. Continue asking
Stay ahead of supply chain shifts

One email a week: the stories that matter and new risk events — in English, free.

Prefer RSS? Articles Risk events

Welcome Back!

Login to your account below

Create New Account!

Fill the forms below to register

Retrieve your password

Please enter your username or email address to reset your password.

Scan to share via WeChat

Open WeChat and scan the QR code to share

QR Code

Add New Playlist