Clustering Ensemble via Cluster-wise Optimization Graph Learning
Huan Zhang, Liang Du · 2021
Clustering is an important research direction of unsupervised learning. With the development of research, clustering algorithms emerge in endlessly, and they all exert significant clustering effects in specific scenarios.However, in the face of complex practical problems, single clustering algorithm is often limited to its inherent clustering scene, and the emergence of clustering ensemble algorithm greatly alleviates this problem. At present, there are a large number of clustering ensemble methods, but the clustering results can not guarantee to meet the block diagonal property, resulting in incorrect clustering. In this paper, we propose a clustering ensemble method of cluster level fusion.First, because the reliability of different clustering results is inconsistent, we use different weights to measure the reliability of the similarity matrix formed by the base clustering results. Second, we use the block diagonal property as a priori to ensure that the similarity matrix is partitioned.Finally, we use symmetric non-negative matrix factorization on the final consistent similarity matrix to obtain the final low-dimensional non-negative clustering label matrix. Experiments show that our method is better than several most state-of-the-art baseline methods on different data sets.