| Summary: | Multidimensional projection techniques are important tools employed in data set exploration and data mining tasks. The data set instances are described in a multidimensional space and projection techniques can be employed to reduce the data set dimensionality and to aid the visualization of instances relations in a computer screen. Usually, the whole multidimensional space is projected, i.e., if it is composed by distinct feature spaces they are handled as a unique feature space. This work proposes an alternative approach dealing with multidimensional spaces as distinct feature spaces, so multidimensional projections can reduce the dimensionality of each feature space into unidimensional spaces and be visualized by a scatter plot -- each unidimensional space will be associated with an axis. Our approach was compared with the traditional way that projects the whole multidimensional space (feature spaces) into the bi-dimensional space. Experiments with different data sets were performed to evaluate which approach better preserves the groups cohesion on the projected space, revealing our approach with good results.
|
|---|
