Feature weighting and selection using a hybrid approach based on Rademacher complexity model selection

This study proposes a hybrid feature weighting and selection model for reducing the system dimensionality, improving the classification accuracy. The hybrid selection model is tuned by means of genetic algorithms, where the involved evaluation uses the Rademacher complexity using the k-nearest neighbors classifier. This approach simultaneously minimizes the feature number and training error and provides information about the relevance of each feature. The model was tested on artificial databases as well as by using features extracted from cardiac signals. The used ECG records for ischemic detection correspond to the E-STT database and the used heart sound database for cardiac murmur detection corresponds to phonocardiographic (PCG) records assembled in the National University of Colombia. The classification error result in the ischemic detection was 1.3% with 50.7% of dimensionality reduction rate, while in the cardiac murmur detection was 6.9% with 87.3% of dimensionality reduction rate.