Supervised and unsupervised machine learning for gender identification through hand's anthropometric data

The goal of this study is to determine the best gender identifiers from the hand anthropometric measurements. Five algorithms are used and their performances quantified. The first algorithm is based on computing distances of test subjects to pre-computed masculine/feminine mean characteristics. Then, the k-nearest neighbours, the K-means algorithms, the linear and the quadratic discriminant techniques are applied to segregate males and females. To select the relevant attributes, the recursive feature elimination and the stepwise regression methods are used. All these methods are leading to high accuracy rates of genders recognition. However, the linear and quadratic discriminant methods are the most accurate. Breadth and circumference features are better than the length features in identifying the gender. The palm and the thumb are the parts of the hand with the highest rate of gender recognition. Breadths of the index and the thumb and the palm circumference are the best individual identifiers.