Introduction
In the complex ecosystem of "last-mile" logistics, terminal stations serve as the critical interface between the supply chain and the end customer. Predicting the performance of these stations—measured through metrics like on-time delivery, pick-up speed, and customer complaint rates—is essential for operational efficiency. However, traditional predictive models often treat these metrics in isolation, failing to account for their deep interdependencies.
As shown in the figure above, service metrics are not static; they exhibit significant temporal fluctuations and correlations. For instance, a push for higher delivery speeds might inadvertently lead to a spike in complaints due to mishandled parcels. Furthermore, stations are highly heterogeneous, with performance patterns influenced by local demographics and geography. To address these challenges, the authors propose AdaService, a framework designed for the joint prediction of multiple service capabilities.
The AdaService Framework
AdaService addresses the limitations of siloed prediction by integrating high-order relationship modeling with multi-task learning. The architecture is built upon three technical pillars:
1. Multi-faceted Hypergraph Construction
Unlike standard graphs that model pairwise relationships, hypergraphs can represent groups of related entities. AdaService constructs these hypergraphs by first using an LSTM-based autoencoder to generate "performance embeddings" for each station. Stations are then clustered into hyperedges based on both their physical locations and their "semantic" similarity—grouping stations that behave similarly even if they are geographically distant (e.g., stations in different cities serving similar commercial districts).
2. Multi-faceted Hypergraph Convolution Network (MHCN)
The MHCN processes these complex groupings in two stages. First, Station Convolution aggregates features within each hyperedge to capture local and semantic context. Second, a Hyperedge Attention mechanism dynamically weights the importance of different service capabilities for a specific station at any given time, allowing the model to adapt to changing operational conditions.
3. Adaptive Multi-faceted Estimation
The final stage employs a Mixture-of-Experts (MoE) architecture for multi-task learning. Instead of a single monolithic network, multiple specialized "expert" sub-networks learn different operational patterns. A "gating" network then determines the optimal combination of these experts for each specific task, such as predicting the complaint rate versus the delivery rate. This allows the model to capture the nuanced trade-offs between competing metrics.
Experimental Evaluation
The researchers evaluated AdaService using a massive real-world dataset from JD Logistics, encompassing 770,000 records from 8,000 terminal stations across China over a one-month period. The dataset included both static features (GPS, area type) and dynamic operational data (order volume, courier availability).
Key Results
AdaService significantly outperformed baseline models like XGBoost, STGCN, and DyHCN across all metrics:
- Complaint Rate Prediction: Achieved the most dramatic improvement, with a 39.62% gain in accuracy.
- On-time Delivery: Improved accuracy by 18.98%.
- On-time Pick-up: Improved accuracy by 9.30%.
Beyond statistical accuracy, the model's practical utility was demonstrated through a courier re-scheduling simulation. Using AdaService's predictions to guide resource allocation resulted in a 2.11% improvement in real-world delivery rates without a corresponding increase in customer complaints.
Insights and Future Directions
A key insight from the study is the importance of "semantic neighborhoods." Stations serving high-density office environments in different cities often share more operational characteristics than a residential station just a few miles away. The hypergraph approach effectively captures these non-local functional similarities.
Additionally, the model showed higher sensitivity and value in mid-sized cities (e.g., Hefei) compared to mega-cities like Beijing. In highly optimized mega-cities, operations are often more stable, whereas mid-sized cities face more volatile resource constraints where accurate co-prediction provides a greater competitive advantage.
Future work will focus on testing the generalizability of AdaService in other service-oriented domains, such as food delivery and ride-hailing, and further refining the model's sensitivity to historical data windows and hypergraph clustering parameters.
Summary updated