In practice, multi-view multi-label classification often faces the dual challenge of missing views and labels. Existing methods typically avoid redundant computations by simply masking missing items, which neither recovers missing view information nor provides effective supervision for unknown labels, leading to underutilized information. To address this, we propose the Dynamic Graph-Guided Imputation Network (DGGI-Net), which centers on high-quality view completion. Specifically, DGGI-Net adaptively constructs intra-view similarity graphs and aggregates available neighborhood features on these graphs to impute missing views, thereby restoring more complete representations and reducing cross-view bias. Concurrently, we design a dual-path fusion strategy that aggregates original and imputed views into two separate branches for classification learning, preserving reliable information while mitigating cross-view contamination. Furthermore, we introduce the Category-Adaptive Pseudo-Label Augmentation (CAPLA) mechanism. Based on the empirical class distribution of known labels, CAPLA sets dual thresholds for each category and discards low-confidence instances, generating high-confidence pseudo-labels for unlabeled instances, alleviating supervision sparsity while keeping the pseudo-label distribution consistent with the empirical class distribution under class imbalance. Experiments on five public datasets demonstrate that our method consistently outperforms strong baseline methods.
Mao et al. (Fri,) studied this question.