Haptic handwritten signatures: the effect of deconcentrated dissimilarities on manifold extraction

The use of a haptic-based handwritten signatures has an intrinsic biometric nature and an important potential in user identification/authentication because it incorporates tactile information. However, in order to exploit this potential for constructing decision systems, it is necessary to gain an appropriate understanding of the internal structure of the data, which in relational representations tend to be very highly dimensional. Most machine learning techniques i) are affected by the curse of dimensionality, ii) use algorithms involving distances (usually Euclidean), but in high dimensional spaces they suffer from the concentration phenomenon. This paper explores the behavior of different strategies for distance deconcentration of haptic data when used for nonlinear unsupervised mappings into low dimensional spaces. An aposteriori use of class information shows that deconcentration transformations improve class cohesion and separation, which can improve the performance of machine learning algorithms.