Abductive Learning Ensembles for Hand Shape Identification

A novel method is presented for hand shape identification based on abductive machine learning. We developed several models and investigated their performance on raw hand shape data for 20 and 40 participants in the form of three different categories of geometric measurements: twelve finger features, two palm features, and three whole hand features. Performance was compared when using each category of features separately and when combining them together. Moreover, we describe two novel and more effective approaches using an ensemble of three abductive networks combined at either the score level or the decision level. The effect of doubling the number of participants from 20 to 40 was studied as well. The ensemble approach achieved overall identification accuracies of 100 and 98.3333 % for the 20-participant and 40-participant datasets, respectively. This compares favorably with other learning approaches tried on the same datasets, including decision trees, support vector machines, and rule-based classifiers.