Class-subspace learning: a SHAP-driven framework for class-aware transportation mode detection

Accurate transportation mode detection (TMD) from mobile crowdsourced data is fundamental for effective urban planning, yet it presents significant analytical challenges. The ubiquity of GPS-enabled devices has enabled the development of machine learning models for this task. However, these models conventionally apply a single, globally selected feature set uniformly across all transportation modes, which can mask the distinct feature requirements of individual classes. This study introduces a class-aware classification framework that represents each transportation mode through its own SHAP-derived feature subspace, capturing mode-specific feature relevance within a single unified model. We evaluate the framework using two real-world GPS trajectory datasets, employing thorough preprocessing and spatial filtering. We enrich the feature space by extracting time-domain characteristics and frequency-domain features obtained using the Fast Fourier Transform. The proposed framework achieves weighted ROC–AUC scores of up to 78.8% and 89.5% on the respective datasets, performing comparably to established global feature selection baselines while additionally exposing the distinct discriminative feature signatures of each transportation mode. These findings indicate that class-aware modelling offers a promising and interpretable alternative to globally optimised strategies and merits further investigation within broader multi-class classification settings.