Principal Curves
Principal curves are one of the nonlinear generalizations of principal components. They were first defined by Trevor Hastie and Werner Stuetzle as "self-consistent" smooth curves which pass through the "middle" of a d-dimensional probability distribution or data cloud. The figure on the right shows a 2-dimensional data set and its principal curve.
Principal curves are related to several other unsupervised learning methods, including
- Vector quantization
- Principal component analysis
- Self-organizing maps
- Generative topographic mapping
- Autoassociative neural networks
This set of pages is a review of the area, including the theory of principal curves, algorithms to find principal curves of data sets, and some practical applications. It also features the Java implementation of the Polygonal Line Algorithm.
If you have any comments on the contents of these pages, please feel free to email to me (kegl A iro D umontreal D ca).
来源:http://www.iro.umontreal.ca/~kegl/research/pcurves/