Research map / Probabilistic and scientific modeling / Molecules and proteins
Adaptive Multi-Faceted Service Capabilities Co-Prediction for Nationwide Terminal Stations in Logistics
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AdaService, a hypergraph-based multi-task framework that jointly estimates on-time delivery, on-time pick-up, and complaint rates for nationwide logistics terminal stations.
Problem
Existing studies estimate logistics station service capabilities using isolated metrics (on-time delivery, pick-up, complaints), ignoring their interrelationships and heterogeneity across stations.
Approach
Encodes stations as nodes in a multi-faceted hypergraph built from autoencoder-LSTM performance embeddings clustered by K-means, applies hypergraph convolution with attention to get station capability representations, and uses an MoE multi-task head with learned task weights to co-predict the three metrics.
Key results
- AdaService improves estimation accuracy for on-time delivery, on-time pick-up, and complaint rates by up to 18.98%, 9.30%, and 39.62% over state-of-the-art baselines.
- Delivery accuracy of 0.96 vs 0.91 for DyHCN; MAE 0.010 vs 0.016.
- Cross-station courier re-scheduling simulation improved on-time delivery rate by 2.11%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AdaService instead of the strongest baseline | Nationwide terminal stations · Nationwide terminal stations | MSE: 0.74±0.013 -> 0.70±0.017 | Helped |
| use AdaService instead of the strongest baseline | Nationwide terminal stations · Nationwide terminal stations | MAE: 0.014±0.005 -> 0.010±0.003 | Helped |
| use AdaService instead of the strongest baseline | Nationwide terminal stations · Nationwide terminal stations | MAPE (%): 0.44±0.2 -> 0.39±0.2 | Helped |
| use AdaService instead of the strongest baseline | Nationwide terminal stations · Nationwide terminal stations | Delivery Accuracy: 0.93±0.008 -> 0.96±0.012 | Helped |
| use AdaService instead of the strongest baseline | Nationwide terminal stations · Nationwide terminal stations | Pick-up Accuracy: 0.46±0.011 -> 0.49±0.015 | Helped |
| use AdaService instead of the strongest baseline | Nationwide terminal stations · Nationwide terminal stations | Complaints Accuracy: 0.73±0.014 -> 0.77±0.015 | Helped |
| replace MHCN with current-state single-station MLP encoding (ablation) | AdaService · Nationwide terminal stations | MAE: 0.010±0.003 -> 0.026±0.002 | Hurt |
| replace MHCN with station LSTM encoding (ablation) | AdaService · Nationwide terminal stations | MAE: 0.010±0.003 -> 0.022±0.001 | Hurt |
| replace MHCN with neighboring-station STGCN encoding (ablation) | AdaService · Nationwide terminal stations | MAE: 0.010±0.003 -> 0.015±0.002 | Hurt |
| replace adaptive MoE with hard parameter sharing (ablation) | AdaService · Nationwide terminal stations | MAE: 0.010±0.003 -> 0.031±0.001 | Hurt |
| replace adaptive MoE with Task Relation-based MTL (ablation) | AdaService · Nationwide terminal stations | MAE: 0.010±0.003 -> 0.023±0.001 | Hurt |
| increase historical data length beyond l=2 (ablation) | AdaService · Nationwide terminal stations | MAE: 0.010±0.003 -> 0.016±0.001 (l=3), 0.021±0.002 (l=4), 0.023±0.002 (l=5) | Hurt |
Surprising findings
- Multi-task learning added to LSTM lowered pick-up accuracy (0.36 to 0.30), showing MTL is not uniformly beneficial across baselines.
Limitations
- Evaluated on a single proprietary dataset from one logistics company over one month.
- Performance is sensitive to historical window length, convolution depth, and distance threshold.
Experimental setup
Datasets: proprietary logistics station records from a leading Chinese logistics company (770,000 records, 8,000 stations)
Benchmarks: proprietary logistics station dataset
Metrics: MSE MAE MAPE accuracy (delivery, pick-up, complaints)
Baselines: Historical Average XGBoost MLP LSTM STGCN DyHCN MTL
Builds on: DyHCN (dynamic hypergraph convolutional network) multi-task learning STGCN K-means
Closest papers by meaning
- CyPortQA: Benchmarking Multimodal Large Language Models for Cyclone Preparedness in Port Operation (AAAI 2026 · similarity 0.78)
- Data Disparity and Temporal Unavailability Aware Asynchronous Federated Learning for Predictive Maintenance on Transportation Fleets (AAAI 2024 · similarity 0.78)
- Spatial-Temporal Heterogenous Graph Contrastive Learning for Microservice Workload Prediction (AAAI 2025 · similarity 0.77)
- SHIELD: Multi-task Multi-distribution Vehicle Routing Solver with Sparsity and Hierarchy (ICML 2025 · similarity 0.77)
- Why Has Predicting Downstream Capabilities of Frontier AI Models with Scale Remained Elusive? (ICML 2025 · similarity 0.77)
- Many-Objective Multi-Solution Transport (ICLR 2025 · similarity 0.77)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.