Existing incomplete multi-view clustering algorithms have made some progress, however, they have certain limitations: 1) Most existing methods fail to accurately uncover the latent information of missing data, especially when handing varying missing rate across different views; 2) They cannot preserve both global and local structures of data; 3) They fail to effectively extract high-order correlations and complementary information between ones. To address these issues, the paper proposes an unbalanced incomplete multi-view clustering algorithm based on missing-value learning and structure preservation. The algorithm maps the original high-dimensional data to a low-dimensional space using linear projection, and combines a completion matrix based on neighbornood relationships and an integrity constraint to fill in the missing values, ensuring that the filled missing values closely match the original structural features of the data. Additionally, the algorithm employs subspace clustering techniques, hyper-Laplacian regularization, and the weighted tensor Schatten-$p $ norm, to simultaneously capture the global and local structure and high-order correlation. Experimental results with ten benchmark algorithms on eight simulated incomplete multi-view datasets with various missing values show that the performance of the proposed algorithm is superior to the comparison algorithms.