Abstract: ‍ ‍The brain network derived from resting-state functional Magnetic Resonance Imaging （rs-fMRI） provides valuable insights into the functional integration of the brain， enhancing our understanding of brain cognition and the underlying mechanisms of various diseases. In recent years， deep learning algorithms， particularly graph neural networks， have shown significant promise in assisting with the diagnosis of brain diseases due to their ability to adapt to the inherent topology of brain networks. However， existing graph neural network methods often rely on a static representation of brain network topology， leading to limited performance based on the quality of the graph input. Additionally， the high cost of rs-fMRI data acquisition restricts dataset size， which hampers the advancement and application of deep learning technologies. To address these challenges， this paper proposes a supervised contrastive learning algorithm designed for multi-view brain networks. This approach retains the traditional static topology view of brain networks while also incorporating a semantic view， wherein the graph structure is dynamically adjusted during model training. This modification aims to capture the rich semantic information present within the brain network for specific tasks. Furthermore， we introduce a cross-attention module that learns joint representations between multiple views， enhancing the ability to represent features effectively. Finally， we construct a supervised contrastive learning mechanism based on multi-view fusion. This paired sample-level learning approach not only expands the training dataset to some extent but also aids in extracting more discriminative features at the individual level， leading to more reliable disease diagnoses. Experiments conducted on two publicly available neuroimaging datasets for brain diseases demonstrate the effectiveness of the proposed multi-view brain network supervised contrastive learning algorithm.