Iterative Forward Selection Method Based on Cross-validation Approach and Its Application to Infant Cry Classification

摘要

Feature selection is an important step in building an effective and efficient classifier. In this paper, we propose an iterative forward selection method (IFSM) inspired by the cross-validation approach. IFSM ranks all variables by a score called the discriminant power score (DPS). DPS represents the discriminant ability of each variable, which is estimated using different training datasets through cross-validation. In this paper, IFSM is applied to an infant cry dataset. The following results were derived from the experiment: 1) the variable with a higher DPS is effective for building a better classification model, and 2) IFSM shows a drastic improvement in classification accuracy compared to non-iterative IFSM. We believe that IFSM is a promising method in the field of pattern recognition.