A Practical Study on Feature Selection Methods in Pattern Recognition: Examples of Handwritten Digits and Printed Musical Notation

In the article we present a practical study on methods for numerical feature selection. We compare quality of classification models built on different sets of features. In particular, we consider the problem of handwritten digits recognition and printed musical notation recognition. We apply a suite of index-based and wrapper methods for feature selection. Experiments show that both on regular data set of handwritten digits and on imbalanced data set of printed musical notation we can easily find a subset of more or less 20 features that will assure a highly accurate classification model.