Full text
| |
| Author(s): |
Robson Carlos da Motta
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | São Carlos. |
| Institution: | Universidade de São Paulo (USP). Instituto de Ciências Matemáticas e de Computação (ICMC/SB) |
| Defense date: | 2014-10-13 |
| Examining board members: |
Maria Cristina Ferreira de Oliveira;
Jesús Pascual Mena Chalco;
Carla Maria Dal Sasso Freitas;
Alípio Mário Guedes Jorge;
Luis Gustavo Nonato
|
| Advisor: | Maria Cristina Ferreira de Oliveira; Alneu de Andrade Lopes |
| Abstract | |
Multidimensional projections are valuable tools to generate visualizations that support exploratory analysis of a wide variety of complex high-dimensional data. Many examples are found in the literature of visual data analysis tasks that employ projections to explore, for instance, text, image, network and sensor data. Nonetheless, dierent projection techniques applied to a particular data set, or even alternative parameterizations of a single technique, can produce very distinct outcomes, as techniques adopt different strategies to reduce data dimensionality. Few resources are available to support assessing projection quality and, in general, existing solutions focus on specific properties. Thus, a broader assessment typically requires considerable human effort. In this work we introduce a framework to compute projection evaluation measures that focus on neighborhoods and clusters. To elaborate this framework we conducted (i) an experimental study to better understand how users perceive projections and (ii) an investigation of possible data representations capable of favoring the identification of neighborhoods and clusters. The observations resulting from the experimental study have been considered to propose and validate a novel graph data model, called Extended Minimum Spanning Tree (EMST), which captures data properties shown to be consistent with the observations by the participants in the study. The EMST graph has been validated by means of two comparative studies conducted to identify neighborhoods and clusters in multidimensional data. Under this framework, five novel measures of projection quality are introduced, two of them to assess properties related to the visual separation of classes, and three to assess the preservation of data properties in the original space, in particular the preservation of class separation, the preservation of neighborhoods and the preservation of groups. The quality measures have been applied to projections of synthetic data sets, favoring their understanding and interpretation, and also to projections of real data sets, illustrating their potential applicability in real scenarios. The newly introduced neighborhood preservation measures are also compared with existing methods in order to illustrate their differences and similarities. (AU) | |
| FAPESP's process: | 09/03306-8 - Complex Networks for Visual Mining of Documents Collections |
| Grantee: | Robson Carlos da Motta |
| Support Opportunities: | Scholarships in Brazil - Doctorate |