Information Theory-Based Feature Extraction for Transportation Mode Detection using Federated Ensemble Learning

Transportation mode detection is a key enabler of intelligent mobility systems, supporting real-time traffic management, personalized travel services, and urban planning. However, achieving high accuracy while balancing computational efficiency and user privacy remains a critical challenge, known as the efficiency-privacy-accuracy trilemma. This work introduces a novel information-theoretic approach integrating entropy and complexity metrics into centralized and federated learning architectures. Our approach addresses performance issues in traditional federated learning by leveraging the structural properties and patterns within time-series data. The results demonstrate that the proposed approach outperforms conventional methods in MLP, Federated MLP, and federated ensemble learning, achieving accuracy improvements of 3.8%, 4.31%, and 5.54%, respectively. These results enable practical applications such as privacy-preserving transportation services and data-efficient urban planning.