Incomplete Multi-View Clustering With Joint Partition and Graph Learning

Incomplete multi-view clustering (IMC) aims to integrate the complementary information from incomplete views to improve clustering performance. Most existing IMC methods try to fill the incomplete views or directly learn a common representation based on matrix factorization or subspace learning. The former may introduce useless and/or even noisy information especially for data with a large missing rate. The latter relies on the initialization and ignores the data structures. To address these issues, we propose a novel Joint Partition and Graph (JPG) learning method for IMC. JPG can be formulated by two key components: unified partition space learning and consensus graph learning. The partition space is more robust to noise and the graph learning helps uncover the data structures. Specifically, JPG iteratively constructs local incomplete graph matrices, generates incomplete base partition matrices, stretches them to produce a unified partition matrix, and employs it to learn a consensus graph matrix. For efficiency, JPG adaptively allocates a large weight to the stretched base partition that is close to the unified partition, determines parameters, and imposes a low-rank constraint on graphs. Finally, the clusters can be obtained directly from the consensus graph. Experimental results on several benchmark datasets demonstrate the effectiveness and superiority of JPG over the state-of-the-art baselines.