ContraLSP is a novel method for explaining time series data by introducing counterfactual samples and locally sparse perturbations. The method uses contrastive learning to generate uninformative perturbations that maintain the distribution of the data. It also incorporates sample-specific sparse gates to generate binary-skewed and smooth masks that capture temporal trends and select salient features. Empirical studies on both synthetic and real-world datasets show that ContraLSP outperforms state-of-the-art models in explaining time series data. The method addresses challenges such as distribution shift and generalization errors by using a smooth constraint and temporal trend function. ContraLSP is evaluated on various datasets, including synthetic and real-world clinical tasks, and demonstrates superior performance in terms of feature importance, mask entropy, and other metrics. The method is effective in identifying important features and providing interpretable explanations for time series models.ContraLSP is a novel method for explaining time series data by introducing counterfactual samples and locally sparse perturbations. The method uses contrastive learning to generate uninformative perturbations that maintain the distribution of the data. It also incorporates sample-specific sparse gates to generate binary-skewed and smooth masks that capture temporal trends and select salient features. Empirical studies on both synthetic and real-world datasets show that ContraLSP outperforms state-of-the-art models in explaining time series data. The method addresses challenges such as distribution shift and generalization errors by using a smooth constraint and temporal trend function. ContraLSP is evaluated on various datasets, including synthetic and real-world clinical tasks, and demonstrates superior performance in terms of feature importance, mask entropy, and other metrics. The method is effective in identifying important features and providing interpretable explanations for time series models.