Adversarial Fair Incomplete Multi-View Clustering

Incomplete multi-view clustering gains increasing attention by considering the view-missing issue in practical applications. However, existing methods impute the shared representation from a sample's unmissing views, which neglects the inter-view distribution difference and result in semantic bias. Additionally, to ensure model fairness, most methods sacrifices clustering performance and are difficult to trade off the relationships between clustering and fairness. In this paper, we propose a novel Adversarial Fair Incomplete Multi-View Clustering (AFIMVC). To accurately learn the shared representation, we leverage contextual information from viewcomplete representations to enhance view-incomplete ones with Cross-Sample Attention mechanism. Besides, we introduce feature-level and cluster-level alignment constraints to exploit consistent and discriminative information. Finally, we develop an adaptive fairness disentangled learning module, which is composed of adversarial disentangled mechanism and adaptive weight learning. The adversarial disentangled mechanism is implemented by an adversarial network with a gradient reversion. They make the shared representation independent of sensitive attributes and dynamically trade off clustering accuracy and fairness. Extensive experiments on three benchmark datasets demonstrate the superiority of the proposed method compared with the state-of-the-art methods.