The paper introduces ContraLSP, a novel method for explaining multivariate time series data. ContraLSP addresses the challenges of identifying important locations and complex temporal patterns in time series data by incorporating counterfactual samples and sample-specific sparse gates. The method uses contrastive learning to generate uninformative perturbations while maintaining the distribution of the original data. Sample-specific sparse gates are employed to create binary-skewed and smooth masks that capture temporal trends and select salient features. Empirical studies on synthetic and real-world datasets demonstrate that ContraLSP outperforms state-of-the-art models in explaining time series data, showing significant improvements in explanation quality. The method is particularly effective in handling heterogeneous samples and complex temporal dynamics, making it a valuable tool for enhancing the transparency and interpretability of time series models in various fields.The paper introduces ContraLSP, a novel method for explaining multivariate time series data. ContraLSP addresses the challenges of identifying important locations and complex temporal patterns in time series data by incorporating counterfactual samples and sample-specific sparse gates. The method uses contrastive learning to generate uninformative perturbations while maintaining the distribution of the original data. Sample-specific sparse gates are employed to create binary-skewed and smooth masks that capture temporal trends and select salient features. Empirical studies on synthetic and real-world datasets demonstrate that ContraLSP outperforms state-of-the-art models in explaining time series data, showing significant improvements in explanation quality. The method is particularly effective in handling heterogeneous samples and complex temporal dynamics, making it a valuable tool for enhancing the transparency and interpretability of time series models in various fields.