As subtle manifestations of human affect and cognitive processes, micro-gestures (MGs) hold critical potential for advancing emotion-aware human-computer interaction systems. However, progress in MG recognition remains critically impeded by the absence of well-annotated, balanced datasets that faithfully preserve the transient nonverbal cues. To address this challenge, we propose a novel Discrete Wavelet Transform Mix (DWTMiX) Augmentation based Contrastive Learning framework for skeletal micro-gesture recognition. Specifically, to precisely model the multiscale feature of MGs, the discrete wavelet transform is introduced to decompose the original MG signals into high-frequency and low-frequency components. Afterwards, we propose a cross-sample high-frequency shuffling and recombination technique, which constructs challenging contrastive sample pairs through high-frequency feature exchange and low-frequency base fusion. Finally, this enhancement strategy is incorporated into a contrastive learning framework, tailored for skeleton-based MG recognition. Comprehensive evaluations on three datasets, iMiGUE, SMG and MA-52 demonstrate the superior performance of our proposed method.