A Hybrid Visualization Approach to Perform Analysis of Feature Spaces

• Autor(a) principal: Júnior, Wilson Estécio Marcílio [UNESP]
• Data de Publicação: 2020
• Outros Autores: Eler, Danilo Medeiros [UNESP], Garcia, Rogério Eduardo [UNESP], Correia, Ronaldo Celso Messias [UNESP], Silva, Lenon Fachiano [UNESP]
• Tipo de documento: Artigo de conferência
• Idioma: eng
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://dx.doi.org/10.1007/978-3-030-43020-7_32; http://hdl.handle.net/11449/201829
• topic: Explainability; Explainable artificial intelligence; Feature space; Interpretability; Visual analytics

In this paper, we propose a hybrid visualization by combining a projection based approach with star plot visualization to inspect feature spaces. While the projection based visualization is used to depict the instances similarities from high-dimensional spaces onto a bi-dimensional space, the star plot visual metaphor enables inspection of features (attributes) relationship. By inspecting feature spaces, analysts can assess their quality and analyze which features contribute for the formation of clusters. To validate our proposal, we demonstrate how to improve feature spaces to generate more cohesive clusters, as well as how to analyze deep learning features of distinct Convolutional Neural Network (CNN) architectures.