Traditional deep learning models have shown remarkable performance in side-channel analysis, yet their inherent black-box nature limits the understanding of their internal decision-making processes. In this paper, we propose SG-SCA, an interpretable deep learning side-channel analysis model. The model introduces a novel ANOVA F-statistic-based differential evaluation method to select shapelets with high representational power, enabling high-quality data transformation that effectively replaces the original side-channel data. Furthermore, by integrating the graph attention mechanism, SG-SCA leverages the graph structure to capture temporal dependencies in the data, enhancing both the model’s performance and interpretability. Experimental results on the ASCAD-fix dataset with masking protection and the AES-HD dataset with parallel processing protection show that SG-SCA provides concise and accurate interpretations via attention matrix visualization and weighted betweenness centrality analysis. This work offers a new and transparent deep learning model for side-channel analysis, facilitating security evaluation and guiding the design of more robust cryptographic implementations.