关键词: 半监督聚类, 谱聚类, 成对约束, 相似度矩阵, 自动聚类, 网络故障检测

Abstract: Focusing on the problem of inadequate use of priori knowledge and the problem that the number of clusters is required in most existing algorithms in network fault detection, a new semi-supervised automatic clustering algorithm that combines propagating pairwise constraints information and determining the number of clusters automatically is proposed. By learning a new similarity measure function to satisfy the constraints, and improving the NJW algorithm, automatic spectral clustering is done on the non-standardized Laplacian matrix eigenvector to improve the clustering performance. The experiments based on the UCI standard data sets and network measured data sets show that the proposed algorithm is more accurate in clustering than the comparative algorithms, and can meet the actual needs of the network fault detection.

Key words: semi-supervised clustering, spectral clustering, pairwise constraints, similarity matrix, automatic clustering, network fault detection