Alleviating Observation Bias via Causal-Invariant Meta-Learning for Unbalanced Incomplete Multi-view Clustering

In incomplete multi-view clustering, unbalanced missingness is prevalent, where different views exhibit significantly varying missing rates, causing severe observation bias. This imbalance poses two core challenges: models develop serious learning biases by over-relying on low-missing-rate views while neglecting high-missing-rate ones, and cross-view data recovery becomes extremely difficult due to sparse training samples in highly missing views, leading traditional generation methods into a "data starvation" dilemma. Existing methods either naively assume low-missing-rate views as high-quality or lack effective debiasing mechanisms, showing limited performance under imbalance. To address this, we propose the Causal-Invariant Meta-Learning Network (CIMLN). It employs a meta-learning paradigm to transfer knowledge across views, using complete samples as support sets to guide generation for highly missing views. Meanwhile, it incorporates a causal inference framework with counterfactual reasoning and adversarial intervention strategies to eliminate spurious dependencies on observation patterns, learning causally invariant clustering structures. These modules synergistically optimize to ensure generated representations possess both feature fidelity and clustering discriminability. Extensive experiments on benchmarks demonstrate the effectiveness of CIMLN.