Adversarial incomplete multi-view clustering with adaptive contrastive learning.

Abstract

Incomplete multi-view data frequently arises in scenarios where certain views are unavailable due to sensor malfunctions or data corruption. In recent years, deep incomplete multi-view clustering has gained significant attention for its capacity to address such challenges. However, most existing deep learning frameworks struggle to effectively balance the contributions of different views. In addition, their performance often deteriorates when they overly depend on the quality of features from specific views. To overcome these limitations, we propose a novel incomplete multi-view clustering method, termed A²CLN, which integrates adaptive contrastive learning with an adversarial learning network. Specifically, we design an adaptive contrastive learning module that dynamically adjusts the contrastive learning parameters based on the significance of the shared information within each view. This module enables the extraction of shared information from the available views while simultaneously preserving complementary features, thereby optimizing the clustering structure. Furthermore, a generative adversarial network is incorporated to enhance the quality of latent feature representations through adversarial training, leading to improved clustering performance. Extensive clustering experiments conducted on six multi-view datasets demonstrate that A²CLN outperforms existing deep incomplete multi-view clustering methods.