Abstract:Perturbation-based time series interpretability methods reveal model decision-making mechanisms by perturbing input sequences and observing changes in model outputs, thereby identifying the sensitivity to specific time steps or features. Although existing methods improve interpretability via learnable perturbations, they often rely on global perturbations and neglect the correlations among local features. To address this limitation, a time series interpretability method based on adaptive patches and information entropy gating is proposed. Specifically, an adaptive patch module is introduced to strengthen the local correlations of perturbations, while the GRU is replaced with an iTransformer encoder to more effectively capture temporal dependencies. Additionally, an information entropy metric is employed to weight and aggregate feature importance, further improving interpretability and prediction accuracy. Experimental results indicate that the proposed method outperforms existing baselines on multiple time series interpretability tasks, effectively improving feature identification accuracy, precisely identifying key factors influencing prediction results, and significantly enhancing model transparency and reliability.