This paper proposes a novel method called Sample-level Cross-view Similarity Learning (SCSL) for Incomplete Multi-view Clustering (IMVC). The main challenge in IMVC is constructing a complete similarity matrix when data is missing across views. Existing methods often fail to account for missing relationships between samples, leading to incomplete or abnormal similarity matrices. SCSL addresses these issues by constructing a complete similarity matrix that considers relationships between samples across all views, not just within individual views. It projects all samples into a common dimension and constructs a complete similarity matrix based on both inter-view and intra-view relationships. Additionally, it introduces a consistent latent representation learning module to enhance projection reliability and incorporates a graph Laplacian regularization term to improve the quality of the similarity matrix. The proposed SCSL method includes three main components: sample-level cross-view similarity learning, consensus representation learning, and graph Laplacian regularization. The sample-level cross-view similarity learning component constructs a complete similarity matrix by considering relationships between samples across all views. The consensus representation learning component enhances the reliability of projections by learning a consistent latent representation. The graph Laplacian regularization term connects the latent representation with the similarity matrix, enabling a global exploration of relationships between samples and further improving the quality of the learned complete similarity matrix. The method is evaluated on six benchmark datasets, demonstrating its effectiveness in processing incomplete multi-view clustering tasks. The results show that SCSL outperforms existing IMVC methods in terms of clustering accuracy, normalized mutual information, and purity. The method is also shown to be robust to parameter variations and converges efficiently. The proposed SCSL method provides a comprehensive solution to the challenges of incomplete multi-view clustering by constructing a complete similarity matrix and leveraging graph Laplacian regularization to enhance the quality of the similarity matrix.This paper proposes a novel method called Sample-level Cross-view Similarity Learning (SCSL) for Incomplete Multi-view Clustering (IMVC). The main challenge in IMVC is constructing a complete similarity matrix when data is missing across views. Existing methods often fail to account for missing relationships between samples, leading to incomplete or abnormal similarity matrices. SCSL addresses these issues by constructing a complete similarity matrix that considers relationships between samples across all views, not just within individual views. It projects all samples into a common dimension and constructs a complete similarity matrix based on both inter-view and intra-view relationships. Additionally, it introduces a consistent latent representation learning module to enhance projection reliability and incorporates a graph Laplacian regularization term to improve the quality of the similarity matrix. The proposed SCSL method includes three main components: sample-level cross-view similarity learning, consensus representation learning, and graph Laplacian regularization. The sample-level cross-view similarity learning component constructs a complete similarity matrix by considering relationships between samples across all views. The consensus representation learning component enhances the reliability of projections by learning a consistent latent representation. The graph Laplacian regularization term connects the latent representation with the similarity matrix, enabling a global exploration of relationships between samples and further improving the quality of the learned complete similarity matrix. The method is evaluated on six benchmark datasets, demonstrating its effectiveness in processing incomplete multi-view clustering tasks. The results show that SCSL outperforms existing IMVC methods in terms of clustering accuracy, normalized mutual information, and purity. The method is also shown to be robust to parameter variations and converges efficiently. The proposed SCSL method provides a comprehensive solution to the challenges of incomplete multi-view clustering by constructing a complete similarity matrix and leveraging graph Laplacian regularization to enhance the quality of the similarity matrix.