Visual Analysis of Feature Selection for Data Mining Processes.
Humberto Razente, Fabio Jun Takada Chino, Maria Camila N. Barioni, Agma J. M. Traina, Caetano Traina · Anais · 2004
The amount of data collected in the last decades has become a source of valuable information, allowing organizations to improve their competitiveness. However, the associated data analysis processes – transforming it into useful information – have became a hard work. In many cases, the data is composed of many items and many dimensions of interest, turning their comprehension an awkward process. The elimination of correlated features may reduce the complexity of the analysis techniques. The visual comparison of the results supplied by dimensionality reduction techniques also allows better understanding of the results and may lead to the discovery of correlations among the features. This work presents a new technique named Vertical Data Splitting Visualization, which allows overlapping different mappings of the same multidimensional dataset into a three-dimensional Euclidean space. It allows a visual observation of existent correlations among features. The experiments showed that it is linearly scalable on both number of features and number of tuples mapped and is fast, allowing the interaction with datasets of hundreds of thousands tuples.