Features and Metric from a Classifier Improve Visualizations with Dimension Reduction

The goal of this work is to improve visualizations by using a task-related metric in dimension reduction. In supervised setting, metric can be learned directly from data or extracted from a model fitted to data. Here, two model-based approaches are tried: extracting a global metric from classifier parameters, and doing dimension reduction in feature space of a classifier. Both approaches are tested using four dimension reduction methods and four real data sets. Both approaches are found to improve visualization results. Especially working in classifier feature space is beneficial for showing possible cluster structure of the data.