<p>Incomplete multi-view clustering (IMC) is a challenging task in real-world applications. Its critical issue is to learn a reliable consensus representation to describe the cluster structure of complex incomplete multi-view data stably. To use the high-order correlations across different views, tensor learning (TL) has played an important role in IMC methods. However, existing TL-IMC methods commonly focus on the consistency of consensus representation, while ignoring the discriminative ability or diversity. This degrades the reliability of consensus representation. To ensure the consistency, discriminative ability and diversity of consensus representation at the same time, this paper proposes a novel TL-IMC method based on enhanced view-feature learning and balanced consensus principle (EVL-BCP). Specifically, EVL-BCP utilizes the fact that consensus representation is centralized from view-features. Firstly, EVL leverages both the intra-view low-rankness and inter-view low-rankness of graph tensor to strengthen the discriminative ability of all view-features, which facilitates to achieve the discriminative ability of consensus representation. Secondly, BCP integrates variance maximization mechanism and centralized constraint to seek for a stronger consistency between view-features in a diversity-induced partitionable view-subspace, which facilitates to introduce the consistency and diversity of consensus representation. Lastly, by aid of a shared feature subspace, EVL and BCP are coupled with each other to further exploit the coupling encouragements of discriminative ability, diversity and consistency. Abundant experiments are conducted on various multi-view datasets with several missing-view scenarios. The results demonstrate the superiority of EVL-BCP.</p>