<p>Multi-view clustering has become a research hotspot in the field of graph data mining in recent years. However, in practical applications, data may be incomplete, resulting in missing values or noise in certain views, which poses significant challenges for multi-view clustering. Existing methods for incomplete multi-view clustering still face two major challenges: First, most current approaches either simply discard missing views or rely on naive imputation techniques (e.g., mean filling), which inevitably introduce noise and distort the underlying data structure. This leads to suboptimal feature extraction as the corrupted representations cannot effectively capture cross-view relationships. Second, while existing methods attempt to learn consistent representations across views, they typically enforce rigid consistency constraints without preserving view-specific discriminative information. This oversimplified treatment fails to leverage the complementary nature of multi-view data, resulting in degraded clustering performance, especially under high missing rates. To address these issues, this paper proposes a novel method Cross-View Attention with Adversarial Learning for Incomplete Multi-view Clustering (CVAL). Specifically, we design a cross-view attention mechanism that dynamically captures global dependencies between views through adaptive feature fusion, enabling robust representation learning even in the presence of missing data. To enhance the discriminability of learned features, we introduce an adversarial learning module to regularize the latent space to follow the realistic data distribution. In addition, we propose a K-nearest-neighbor based cyclic graph constraint that simultaneously preserves local geometric structure and cross-view consistency. To validate the effectiveness of the proposed method, we conducted extensive experimental evaluations on several widely used multi-view datasets. Experimental results demonstrate that CVAL achieves significant advantages in clustering performance.</p>