CADENCE: A Confidence-Adaptive Dual-Expert Network for Fast and Accurate Time Series Classification
CADENCE introduces a CPU-native dual-expert architecture for fast time series classification. It splits representation learning into two specialized pathways: one using convolutional features and another using distributional interval approximations across kinematic and spectral data. A confidence-adaptive router dynamically switches between these experts, achieving high accuracy (0.8864) much faster than complex monolithic methods.