I'm experimenting with using wavelets for feature extraction from a
data time series and then classifying based on the extracted features
into one of two categories. My data set consists of 75% sample of
Type I and 25% of Type II. I've tried both a Feed Forward ANN and
LVQ. Neither approach have produced good results on out of training
sample sets, i.e., low identification rate of type II.
I would appreciate any advice on the following:
1. Preprocessing - Should I preprocess the wavelet coefficient
inputs in any manner? (I'm compressing the data time series to reduce
the number of wavelet coefficients presented to the network)
2. Training - Matlab does not appear to support
early-stopping/validation for LVQ. I would appreciate any advice on
training parameters for the LVQ to improve generalization. (I do use
early stopping for the FF ANN)
Thanks is advance.