Dimensionality reduction as the available method to overcome the “curses of dimensionality” has attracted wide attention. However, the pervious studies treat the visualization and the subsequent classification performance separately. In this case, we do not know whether there is a underlying relationship (i.e., direct proportion) between visualization and the followed classification performance. In this paper we compare several dimensionality reduction techniques on three different types of data sets: 1) Benchmark, 2) Image, and 3) Text data. Specifically, to intuitively evaluate the quality of the dimension reduced data, the visualization analysis is carried out, in which we use a covariance matrix related criteria to quantify the information content of the data. Moreover, we also consider the classification accuracy in different latent spaces (i.e., dimensions) as another performance criterion to further analysis the relationship between the information content and the subsequent classification task. The experimental results show that there is no direct proportion relationship between the information content and the further classification performance.
CITATION STYLE
Song, X., Cai, Z., & Xu, W. (2014). Dimensionally reduction: An experimental study. Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 8933, 380–393. https://doi.org/10.1007/978-3-319-14717-8_30
Mendeley helps you to discover research relevant for your work.