<p>Skin cancer represents a major global health concern, making early and accurate diagnosis important for improving clinical outcomes. However, dermoscopic images often exhibit subtle visual differences among lesion categories, which presents challenges for automated classification. This study presents a hierarchical feature multi-contrastive learning framework for skin cancer classification. The framework employs a multi-granularity feature extraction strategy through three complementary pathways—global, center-local, and peripheral-local—to capture both overall lesion structures and localized lesion characteristics. An attention erasure module is introduced to suppress highly activated regions during training, encouraging the model to explore additional discriminative cues within dermoscopic images. In addition, a supervised contrastive loss formulation is adopted that incorporates adaptive feature center updates, dynamic temperature scaling, and a normalized center shift regularization term to stabilize representation learning and improve discrimination between visually similar samples. Experiments conducted on publicly available datasets, including ISIC-2017, HAM10000, and ISIC-2019, show competitive performance across multiple evaluation metrics such as accuracy, recall, precision, and F1-score, with classification accuracy exceeding 98% on the ISIC-2019 dataset. These results suggest that combining multi-path spatial representations with contrastive learning provides an effective representation learning strategy for dermoscopic image classification. The proposed framework may also support future investigations on multimodal data integration, cross-center evaluation, model compression, and interpretability in automated skin lesion analysis.</p>