ResNeXt-SqueezeViTNet: A Ubiquitous Sensing Framework for Predictive Risk Assessment in Smart City Logistics Systems

Modern smart city logistics operate in highly unpredictable environments shaped by congestion, fluctuating demand, and diverse IoT sensing conditions. These factors produce complex, imbalanced, and temporally inconsistent data that challenge traditional predictive modeling approaches. To address these limitations, this study introduces ResNeXt-SqueezeViTNet, a hybrid residual, compressed, and attention-driven framework designed for multi-label logistics risk prediction. The architecture integrates ResNeXt grouped convolutions for structural feature extraction, SqueezeNet modules for compact multi-scale representation, and a Vision Transformer attention layer to capture long-range temporal and contextual patterns. A tailored preprocessing and feature-engineering pipeline strengthens the learning process. Clustered Adaptive Synthetic Balancing (CASB) improves minority-class representation, while the Adaptive Composite Feature Selector (ACFS) identifies critical attributes using weighted statistical divergence, density-guided enhancement, and correlation-based pruning. Additional feature enrichment from AAS and ART further enhances the input space. Experiments were conducted on 400,000 logistics records from Southern California, using strict temporal separation to avoid data leakage. The model achieved 99.0% accuracy, 98.4 F1-score, and 0.96 AUC, consistently surpassing strong baseline methods such as SafeLogFL, Hybrid TCN Transformer, FedFormer TS, and Swin GNN FusionNet. Comprehensive ablation studies, statistical analyses, and efficiency evaluations confirm the robustness and real-time suitability of the proposed framework for intelligent logistics risk forecasting in smart city environments.