Dissimilarity-based indicator graph learning for clustering

Lin Yuan

doi:10.1016/j.neucom.2023.126887

2023-12-01

PRE

DOI：10.1016/j.neucom.2023.126887

原文链接

原文求助

摘要

En 中文

Although learning a similarity graph with a 0-1 value is regarded as an important issue for clustering tasks, this topic has been scarcely reported in the literature since its NP-hard. Along these lines, this issue is addressed in the work by introducing a novel concept. Concretely, the indicator graph clustering (IGC) method was proposed. Based on the dissimilarity and coordinate descent method, IGC is able to learn an indicator graph with a 0-1 value, which is just the similarity graph with a 0-1 value. Meanwhile, the clustering result can be directly obtained by the indicator graph. From the experimental results on several datasets, it was verified that IGC performs faster than k-means, especially on high-dimensional datasets. Furthermore, IGC was extended to a multi-view version, named indicator graph clustering for multi-view (IGCMV). Our model fused the clustering indicator matrices from different views, which can be obtained by IGC, into the final clustering indicator matrix by a projection matrix. To the best of our knowledge, this work is the first attempt to learn a consensus indicator graph from indicator graphs of all views. In addition different from most of the existing multi-view clustering algorithms, the proposed IGCMV enables the production of the final clustering result, as well as each view's clustering result simultaneously. This is helpful for measuring the quality of each view. IGCMV can be efficiently optimized by an alternative iteration strategy. Extensive experiments were also carried out on several benchmark multi-view datasets demonstrating the effectiveness of the developed algorithm to the state-of-the-art multi-view clustering algorithms.

Keyword:

Similarity graph

Multi-view data

Indicator matrix

Clustering

Coordinate descent method