Project: Task-guided Dimensionality Reduction
Description
The typical definition of dimensionality reduction is that points are mapped to the visual space, preserving the original pairwise distance between all points (global techniques) or between points belonging to the same neighborhood (local techniques). In light of this definition, the precision of the distance preservation sought by any technique is equally important regardless of the data instances' relevance for a specific data analysis task; that is, existing algorithms and techniques usually do not consider the data analysis focus when mapping the data from the original to the visual space.
This project aims to adapt existing techniques to improve the distance preservation of certain data instances or groups of instances, optimizing the preservation of such instances at the cost of reducing the mapping quality of the others (unimportant instances). Different scenarios could benefit from that, from protein folding analysis to multi-media collection organization and visualization.
Details
- Supervisor
-
Fernando Paulovich
- Interested?
- Get in contact