Artificial Intelligence Review Искусственный интеллект

From multi-view data features to clusters: a unified approach

Dornaika, Fadi
Springer Nature · 2023-12-01 · DOI 10.1007/s10462-023-10616-y

Аннотация

Clustering data from multiple sources or views has become an important issue in real-world applications. Graph-based methods take advantage of graphs that encode the local and global structure of the data. Although graph-based methods provide good clustering performance, they need to estimate the view graphs or the consensus graph from the raw data in a separate step. Their performance may be affected by the noisy graphs. To overcome this limitation and promote end-to-end multi-view clustering solutions, this paper presents two end-to-end multi-view clustering solutions starting from the data or their kernel representations. The first proposed solution is based on a single objective function that allows the joint estimation of the graph of each view, the consensus graph, the spectral projection matrices for all views, the soft clustering assignments, and the weights of each view. The second solution uses a different objective function where the links and constraints for the soft clustering assignment matrix use the consensus graph matrix and the consensus spectral projection matrix. Both methods enforce the mutual similarity of each graph and provide a direct clustering result without any additional step. The two proposed methods are tested on several real image and text datasets, which prove their superiority.

Читать у издателя →