ADAPT-FUSE: Sınır Ötesi E-Ticaret Teslimat Risklerini Tahmin Etmek İçin Hibrit Bir Yapay Zeka Çerçevesi

Özet (abstract)

Cross-border e-commerce (CBEC) is among the fastest-growing areas of global trade, yet predicting on-time delivery remains difficult owing to fragmented logistics networks, customs variability, and demand uncertainty. Most existing methods rely on single-model architectures that do not jointly capture the spatial, temporal, and relational dependencies of supply-chain data. We propose ADAPT-FUSE (Adaptive Deep Attention Prediction Transformer with Fuzzy Unified Stacking Ensemble), a five-layer hybrid framework that integrates Temporal Convolutional Networks, Bidirectional LSTM encoders, Multi-Head Self-Attention Transformers, Graph Neural Networks, and Fuzzy Inference Systems. These extractors are combined through an adaptive gated fusion mechanism, tuned by dual-objective Bayesian and Genetic Algorithm optimization, and aggregated by a calibrated stacking ensemble. Evaluation uses a fully synthetic benchmark of 5000 CBEC transactions across eight major Chinese provinces (2021–2024), with marginal distributions calibrated to publicly available aggregate trade and logistics statistics; no proprietary records or human-participant data were used. The dataset, generator, model code, seeds, and raw result logs are released for full reproducibility. ADAPT-FUSE attains the best accuracy (0.8940), macro recall (0.8504), and macro F1-Score (0.8543) among eight evaluated models, with macro precision (0.8585) and AUC (0.9441) level with the strongest baseline; five-fold stratified cross-validation gives accuracy 0.8752 ± 0.0095. The margin over the strongest baseline is not statistically significant (McNemar p = 0.5044), and ablation shows the fuzzy-inference and recurrent branches carry the most weight while the attention and graph branches contribute little at this data scale. ADAPT-FUSE offers a competitive, fully reproducible architecture for delivery-risk prediction that shifts the operating point toward better minority-class recall. Because the benchmark is synthetic and the observed margin is not statistically significant, we frame the contribution as an architecture rather than a performance breakthrough; validation on real operational data is required before deployment. Future work will extend the framework to multi-modal data such as satellite imagery and real-time weather feeds.

Yazarların özeti; kaynağından alınmıştır. Discover Artificial Intelligence, 2026 · DOI ↗