Abstract
Information bottleneck (IB) theory aims to compact coding while retaining task-relevant information, but the lack of prior knowledge in clustering poses challenges to its application. Besides, since most methods fail to emphasize clustering security, empty clusters caused by excessive preference can lead to significant performance degradation. Motivated by these, a scalable framework is proposed, named Information Bottleneck Inspired Balanced Multiview Clustering (IB2MC), which boasts clustering security, optimization synergy, and tuning simplicity. To begin with, the in-depth discussion on unsupervised IB theory reveals the balance principle via maximizing label entropy, which preserves the uncertainty of cluster distribution to overcome bias or neglect for clusters. On this basis, compressed descriptors are expected to evolve toward discriminability, thus facilitating label inference from them at the same time. Inspired by this, our method proposes dynamic cosine graph learning for label transmission, and achieves seamless label extraction to strengthen dual-level consistency. Moreover, the representation alignment under independent constraint can enhance cluster separability, while the trace ratio for self-balance clustering can further avoid extra hyperparameter tuning. In this way, the time complexity of our method is linear with sample size, while no pre- or post-processing is required for joint optimization. To demonstrate the effectiveness of our method, seventeen state-of-the-art methods are chosen as baselines and our method ranks first in average performance across all twelve real-world data sets.