The existing multi-view graph learning methods are mainly based on the premise that the data has good completeness, and do not fully consider the learning problem on incomplete data caused by element missing. To address this issue, this paper proposes a multi-view graph learning method with incomplete data. On the one hand, the method puts the data reconstruction and graph learning into the unified framework within view, which learns the view specific neighbor relationship among samples from the reconstructed data to compensate for the influence of data missing on data distribution. On the other hand, in order to preserve the two-dimensional structure of the neighbor graph, tensor analysis is introduced to globally construct a multi-view based fusion graph learning constraint to further capture the high-order potential correlations hidden in multiple views. The proposed framework optimizes the data reconstruction, view-specific graph and fusion graph learning alternatively, which benefits each other during iterations and effectively improve learning ability on incomplete data. The proposed graph learning method is applied to two kinds of incomplete data spectral clustering experiments. The experimental results demonstrate that our proposed method outperforms the existing mainstream multi-view graph learning methods on multiple evaluations and robustness.