Initialising Exploratory Projection Pursuit Networks

C. García-Osorio (Spain) and C. Fyfe (UK)


Artificial neural networks, exploratory projection pursuit, data visualization, weights initialisation.


We have previously developed three artificial neural net work methods of performing exploratory projection pursuit (EPP). EPP finds a low-dimensional linear projection of a high dimensional data set. A user can search for structure in the low-dimensional projection by eye. However, the projections found are typically very dependent on the ini tial conditions of the parameters (weights) in our artificial neural networks. In this study we initialise the EPP net works with values taken from Andrews' Curves and show that the resulting networks converge very fast and reliably to the optimal projections.

Important Links:

Go Back