Authors:
Mateus Espadoto
1
;
Francisco Caio M. Rodrigues
1
and
Alexandru C. Telea
2
Affiliations:
1
Institute of Mathematics and Statistics, University of São Paulo, Brazil, Johann Bernoulli Institute, University of Groningen and The Netherlands
;
2
Johann Bernoulli Institute, University of Groningen and The Netherlands
Keyword(s):
Machine Learning, Dimensionality Reduction, Image-based Visualization.
Related
Ontology
Subjects/Areas/Topics:
Abstract Data Visualization
;
Computer Vision, Visualization and Computer Graphics
;
General Data Visualization
;
High-Dimensional Data and Dimensionality Reduction
;
Information and Scientific Visualization
;
Visual Data Analysis and Knowledge Discovery
Abstract:
Visualizing decision boundaries of modern machine learning classifiers can notably help in classifier design, testing, and fine-tuning. Dense maps are a very recent method that overcomes the key sparsity-related limitation of scatterplots for this task. However, the trustworthiness of dense maps heavily depends on the underlying dimensionality-reduction (DR) techniques they use. We design and perform a detailed study aimed at finding the best DR techniques to use when creating trustworthy dense maps, by studying a large collection of 28 DR algorithms, 4 classifiers, and 2 datasets from a real-world challenging classification problem. Our results show how one can pick suitable DR algorithms to create dense maps that help understanding classifier behavior.