An unsupervised dimensionality-reduction technique

A new procedure for variable selection, which runs in two steps, is introduced. First, an unsupervised and very fast variable selection procedure is applied: a parameter that accounts for the correlation between the features available is computed and, only near 20% of initial variables (those that are less collinear) are retained for further selection. Then, a fine-tuning selection based on a deterministic method (stepwise) coupled to a simple probabilistic neural network is conducted on the variable subset that resulted from the first selection step. The method is demonstrated using a database consisting of vapors of ethanol, acetone and toluene and their binary mixtures (120 variables). Vapors can be simultaneously identified and quantified with a 95.83% success rate and the time needed for the whole process is about 5 minutes in a Pentium 4 PC platform. Being unsupervised, the fast variable selection method applies generally, even in aroma analysis problems where category discovery is an issue. This is illustrated by applying the method to mixture analysis using direct mass spectrometry