<p>In the age of big data, multi-view clustering has garnered significant focus for its capability to harness both the diversity and commonality across different views. However, challenges such as noisy and redundant features, along with the difficulty in capturing cross-view relationships, have constrained its performance. To tackle these issues, we introduce a framework which builds a tensor by embedding features from multiple views and refines this tensor through rotation, thereby enhancing the capture of cross-view relationships. We also propose an anchor representation strategy with a regularization term to preserve local geometric structures while reducing computational complexity. Additionally, two self-weighted operations are performed: one assigns adaptive weights to individual features to build robust graphs, while the other adjusts graph weights based on their stability. Extensive empirical results on six multi-view datasets show SWARTR consistently outperforms state-of-the-art baselines in clustering accuracy, robustness, and scalability.</p>