<p>The objective of clustering is to group data points into distinct categories based on their varying representations of characteristics. Two well-known clustering methods are K-means clustering and spectral clustering. K-means clustering partitions the data points by minimizing the squared distance between each sample and the cluster prototype. In contrast, spectral clustering divides data points by learning the spectral embedding, which is achieved through the eigendecomposition of the Laplacian matrix of a graph. To leverage the strengths of both methods for multiview clustering, we introduce a novel approach called Dual Nonnegative and Orthogonal Reconstruction (DNOR) multi-view clustering. This method simultaneously decomposes the original data and the similarity graph to learn various features. The advantages of DNOR are as follows: (1) It thoroughly explores the information within the original data and the local geometric structure, yielding diverse features that enhance clustering performance. (2) It utilizes these diverse features to link the original data with the similarity graph, thus uncovering hidden structural information through the interplay among different views. (3) It addresses noise in the original data by incorporating concepts from symmetric spectral clustering, ensuring that the decomposition results of both the original data and the similarity graph are closely aligned. To demonstrate the effectiveness of our proposed DNOR method, we conducted extensive experiments on seven datasets. These experiments have shown that our method’s clustering effect is better than that of the nine state-of-the-art approaches.</p>