Bladder cancer (BC) is a significant and prevalent malignant tumor of the urinary system, with accurate and efficient diagnosis remaining a critical challenge. Multiple Instance Learning (MIL) has shown promise in analyzing histopathological Whole Slide Images (WSIs) by enabling weakly supervised localization of critical regions for cancer diagnosis. However, BC’s benign-malignant classification poses challenges due to the high heterogeneity of histological subtypes within both categories, requiring precise mapping to binary classifications. To address this, we propose a multi-view attention framework to enhance WSI representation. This includes a multi-view masked attention mechanism to avoid redundant feature capture, a diversity learning constraint to ensure comprehensive representation, and a dual-granularity supervised contrastive learning strategy to improve inter-class discriminability. Experiments on a BC WSI dataset demonstrate that our method effectively distinguishes benign and malignant categories, significantly improving classification performance metrics. The dataset, code and model weights will be available to assist in clinical decision-making.