SATIC: Self-Adaptive Temporal Inverse Covariance Clustering for Driving Style Recognition

Automatic recognition of driving behavior at the state and driver levels is important for personalized driver assistance, safety assessment, and actuarial modeling. However, TICC-style inverse covariance clustering relies on globally fixed regularization and switching penalties, which may be unreliable under cluster imbalance and insufficiently responsive to local changes in nonstationary driving behavior. This study proposes SATIC, a self-adaptive temporal inverse covariance clustering framework that integrates adaptive model estimation with adaptive temporal segmentation. In the M-step, a vectorized Newton solver for sparse inverse covariance estimation is combined with cluster-specific regularization. This design reduces solver runtime from 12.7 s to 7.2 s, and the iteration count from approximately 850 to 10–30. In the E-step, equal-weight fusion of continuity coefficients over 0.3–2.0 s produces time-varying switching penalties without trainable fusion parameters. Under an idealized squared-loss formulation, a preselected constant switching weight can incur minimax regret of Ω(T), whereas online gradient descent attains O(log T). The linear regret bound indicates that the cumulative excess loss of the fixed strategy may grow with sequence length and that its average regret need not vanish. In contrast, the logarithmic bound yields vanishing average regret for the online adaptive strategy. This result motivates an adaptive switching penalty that tracks local temporal variations. SATIC was evaluated against four baselines on OCSLab, UAH-DriveSet, NGSIM, and a real-vehicle dataset. On NGSIM, vehicle-level comparisons favored SATIC for the Calinski–Harabasz and Davies–Bouldin indices after false-discovery-rate correction. The raw p-values were 0.022 and 0.015, with rank-biserial correlations of 0.51 and 0.55, respectively. Additional experiments on the real-vehicle dataset evaluated its applicability under practical driving conditions, where its overall performance remained comparable to the baselines and varied across drivers. Exploratory analyses further related the inferred states to physics-defined driving categories and identified three interpretable driver groups. Overall, SATIC provides an interpretable adaptive framework for analyzing nonstationary driving sequences, although the magnitude of its empirical benefits varies across datasets, evaluation metrics, and drivers.