TRAVEL MODE IDENTIFICATION WITH RESIDUAL NEURAL NETWORKS AND FEATURE-WISE ATTENTION

Abstract

Understanding travel behavior is essential for improving transportation systems and supporting sustainable and efficient mobility. The current study introduces a method leveraging deep learning to automatically detect travel modes such as walking, biking, driving, bus, and train using GPS data collected from mobile devices. The proposed method analyzes raw GPS trajectory data to identify patterns in movement behavior and classify them into spec transportation modes. The system is built on a neural network that incorporates residual connections and a Feature-Wise Attention mechanism, allowing it to focus on key spatiotemporal signals such as speed, acceleration, bearing, and jerk. This design reduces reliance on manual feature engineering and improves the model's adaptability across diverse travel scenarios. The framework was trained and evaluated on the GeoLife GPS dataset, which includes a variety of real-world trips from multiple users. The model demonstrated strong performance in classifying all five target modes and showed robustness across trips of varying lengths and characteristics. The model provides additional insights by indicating which features play a key role in distinguishing between different transportation modes. This research overs a practical tool for analyzing travel behavior using GPS data and overs meaningful insights for transportation planners, mobility researchers, and policymakers. The system is also applicable to automated fare collection in public transit networks. It supports reports to reduce congestion, enhance public transit services, and promote intelligent transportation systems and sustainable mobility.