Context-Aware Neural SDEs for Robust Irregular Time-Series Classification

Irregularly observed time series models are often trained on relatively clean data but deployed under degraded observation quality caused by sensor dropouts and missingness. We study the train-clean/test-degraded setting for continuous-time classification. We propose Context-Aware Neural Stochastic Differential Equations (CA-NSDE), which combine a stable linear-noise SDE backbone with time-varying gating over four complementary information sources. We evaluate the model on 30 UEA/UCR benchmark datasets under increasing MCAR missingness at test time. Across this setting, CA-NSDE achieves the highest degraded-condition accuracy among compared methods and obtains lower multiclass Brier scores across the reported missingness levels. These results indicate that stable stochastic dynamics together with context-aware source fusion support robust continuous-time classification under train-clean/test-degraded missingness shift.