Feature importance from novel dataset for on-device transport mode detection

Urbanization has increased reliance on intelligent transportation systems, with transport mode detection playing a pivotal role. Transport mode detection leverages sensor data from mobile devices to classify transportation modes, enabling applications such as optimized route planning and autonomous ticketing. However, existing datasets for developing accurate transport mode detection solutions often lack diversity in devices, participants, and transportation modes, limiting their generalizability. Feature selection in transport mode detection research also remains inconsistent, relying heavily on domain-specific experimentation. This study introduces a novel dataset addressing these limitations by including data from 101 participants across 57 unique device models, including both Android and iOS, and covering 10 transportation modes. Additionally, we present an ensemble-based framework for feature evaluation and reduction, which systematically ranks features in a generic and transferable manner. Evaluation shows that models trained using features ranked by the proposed framework achieve up to a 75% reduction in feature size while maintaining competitive accuracy, enabling efficient, on-device transport mode detection solutions. The framework also identifies the most impactful sensors and aggregation functions, offering insights transferable across diverse algorithms and applications.