This paper proposes a novel method called Adaptive Feature Imputation with Latent Graph for Deep Incomplete Multi-View Clustering (AGDIMC) to address the challenges of incomplete multi-view clustering. Traditional methods suffer from inaccurate imputation of missing data and failure to utilize latent graph structures for clustering. AGDIMC incorporates view-specific deep encoders to capture embedded features and constructs partial latent graphs on complete data to preserve topological information. It uses an adaptive imputation layer to impute missing data based on cross-view soft cluster assignments and global cluster centroids. The method also introduces an adaptive imputation strategy based on global pseudo-labels and local cluster assignments to reduce the negative impact of low-quality imputed samples and cluster structure discrepancies. Experimental results on multiple real-world datasets show that AGDIMC outperforms existing methods in clustering performance. The method leverages latent graphs to capture topological information and uses a graph embedding constraint to preserve neighbor relationships. The framework includes a GCN module for latent graph exploration and an adaptive imputation module to refine imputed features. The method also incorporates a self-supervised clustering layer to enhance clustering performance. The algorithm is optimized through initialization and fine-tuning stages, with the adaptive imputation module dynamically improving clustering results. The method is evaluated on three datasets, demonstrating its effectiveness in handling incomplete multi-view data. The results show that AGDIMC achieves superior clustering performance compared to existing state-of-the-art methods.This paper proposes a novel method called Adaptive Feature Imputation with Latent Graph for Deep Incomplete Multi-View Clustering (AGDIMC) to address the challenges of incomplete multi-view clustering. Traditional methods suffer from inaccurate imputation of missing data and failure to utilize latent graph structures for clustering. AGDIMC incorporates view-specific deep encoders to capture embedded features and constructs partial latent graphs on complete data to preserve topological information. It uses an adaptive imputation layer to impute missing data based on cross-view soft cluster assignments and global cluster centroids. The method also introduces an adaptive imputation strategy based on global pseudo-labels and local cluster assignments to reduce the negative impact of low-quality imputed samples and cluster structure discrepancies. Experimental results on multiple real-world datasets show that AGDIMC outperforms existing methods in clustering performance. The method leverages latent graphs to capture topological information and uses a graph embedding constraint to preserve neighbor relationships. The framework includes a GCN module for latent graph exploration and an adaptive imputation module to refine imputed features. The method also incorporates a self-supervised clustering layer to enhance clustering performance. The algorithm is optimized through initialization and fine-tuning stages, with the adaptive imputation module dynamically improving clustering results. The method is evaluated on three datasets, demonstrating its effectiveness in handling incomplete multi-view data. The results show that AGDIMC achieves superior clustering performance compared to existing state-of-the-art methods.