Wavelet Decomposition of Signal and Feature Selection by LASSO for Pattern Recognition

́ Abstract. There is searched the balance between an increase of pattern recognition risk and a decrease of a model size. The experiments are performed for noisy signals, decomposed in wavelet bases. Wavelet representation of signals, i.e. representation by wavelet coefficients called signal features, constitutes the full model. The presented feature selection method is based on the Lasso algorithm (Least Absolute Shrinkage and Selection Operator). The aim of the experiment is to find an optimal model size and investigate the relations between the risk, the number of signal features and the noise level. A new criterion of feature selection is proposed that minimizes both the risk and the number of signal features. The experimental risk of classification is analysed for all possible reduced by Lasso models and for several values of noise levels.