The Impact of Segmentation and Overlapping in Feature Extraction for Biometric Human Authentication Systems

Abstract

One of the arduous challenges in Machine Learning is how to extract features withenough information that will simplifythe learning process of classificationmodels;therefore,leading to better predictionsand human interpretations. We investigated the impact of segmentationand overlapping techniques used to extract features from accelerometer data to optimize the performance of Machine Learning models designed for Biometric User Authentication via walking patterns. Results showed that bigger segmentations were beneficial to the individual performance of the features and detrimental for systems fed with a set of features. Also, there was no evidence found supporting the increase in the overall performance of the system by usingthe method of overlapping. Finally, via a brute-force feature selection algorithm, we achieved a 71% classification accuracy (with 10/34 features) vs. 64% (with 34 features), regardless of the system’sconfiguration meaning that key features hold more weight than mere segmentationand overlapping methods. KeyTerms⎯Acceleration, Biometric Human Authentication, Feature Extraction, Supervised Learning.