A rank-based framework through manifold learning for improved clustering tasks

• Autor(a) principal: Rozin, Bionda [UNESP]
• Data de Publicação: 2021
• Outros Autores: Pereira-Ferrero, Vanessa Helena [UNESP], Lopes, Leonardo Tadeu [UNESP], Guimarães Pedronette, Daniel Carlos [UNESP]
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://dx.doi.org/10.1016/j.ins.2021.08.080; http://hdl.handle.net/11449/229448
• topic: Clustering; Manifold learning; Ranking; Similarity learning

The relevance of diversified data preprocessing approaches for improving clustering tasks is remarkable. Once the effectiveness is direct impacted by feature representation and similarity definition, considerable attention from the research community has been drawn to this direction. More recently, rank-based manifold learning methods have been successfully explored in unsupervised similarity learning for retrieval scenarios. Such methods consider the underlying dataset manifold to compute a new similarity measure, which increases the separability of data from distinct classes. In this paper, a rank-based framework for clustering tasks is proposed based on contemporary manifold learning methods. A flexible model is employed, where ranking structures are the representation of similarity information among data samples. Subsequently, is made the exploration of unsupervised similarity learning. It is also possible to compute more effective similarity measures and clustering results. To assess the effectiveness of the proposed framework was conducted a comprehensive experimental evaluation. The tests involved various public image datasets, considering different manifold learning and clustering methods. The quantitative experiments take into consideration comparisons with traditional and recent state-of-the-art clustering approaches.