In this study, we introduce MIAS-427, one of the largest and most comprehensive inertial datasets for in-air signature recognition, comprising 4270 multivariate signals. This dataset addresses a critical gap in the field by providing a robust foundation for advancing research in cognitive computation and biometric authentication. Leveraging embodied cognition theory, we propose a novel feature selection approach using dimension-wise Shapley Value analysis, which uncovers the intrinsic relationship between human motoric preferences and device-specific sensor data. Our methodology includes a thorough statistical analysis with domain descriptors and DTW algorithms, alongside a comparative evaluation of seven deep-learning models on both the MIAS-427 and smartwatch datasets. The FCN and InceptionTime models achieved remarkable accuracies of 98% and 97.73% on MIAS-427 and smartwatch data, respectively. Notably, our analysis revealed that [Formula: see text] and [Formula: see text] contributed the most (12.82%) and least (8.71%) for the smartwatch, while [Formula: see text] and [Formula: see text] contributed the most (15.63%) and least (7.26%) for MIAS-427, highlighting significant dimension compatibility variations across devices. This research not only provides a valuable dataset for the community but also offers novel insights into human motoric behavior, paving the way for the development of more effective cognitive computation models.
Keywords: Dataset; Deep learning; Feature selection; Inertial data; Online signature.
© 2025. The Author(s).