Exploring the underlying structure of haptic-based handwritten signatures using visual data mining techniques

Abstract

In this paper, multidimensional and time-varying haptic-based handwritten signatures are analyzed within a visual data mining paradigm while relying on unsupervised construction of virtual reality spaces using classical optimization and genetic programming. Specifically, the suggested approaches make use of nonlinear transformations to map a high dimensional feature space into another space of smaller dimension while minimizing some error measure of information loss. A comparison between genetic programming and classical optimization techniques in the construction of visual spaces using large haptic datasets, is provided. In addition, different distance functions (used in the nonlinear mapping procedure between the original and visual spaces) are examined to explore whether the choice of measure affects the representation accuracy of the computed visual spaces. Furthermore, different classifiers (Support Vector Machines (SVM), k-nearest neighbors (k-NN), and Naive Bayes) are exploited in order to evaluate the potential discrimination power of the generated attributes. The results show that the relationships between the haptic data objects and their classes can be appreciated in most of the obtained spaces regardless of the mapping error. Also, spaces computed using classical optimization resulted in lower mapping errors and better discrimination power than genetic programming, but the later provides explicit equations relating the original and the new spaces.