Resumen de Approximate ensemble methods for physical activity recognition applications
Abstract The main interest of this thesis focuses on computational methodologies able to reduce the degree of complexity of learning algorithms and its application to pyshical activity recognition.
Random Projections will be used to reduce the computational complexity in Multiple Classifier Systems. A new boosting algorithm and a new one-class classifcation methodology have been developed. In both, random projections are used for reducing the dimensionality of the problem and for generating diversity, exploiting in this way the benefits that ensembles of classifiers provide in term of performances and stability.
Moreover, the new one-class classification methodology, based on an ensemble strategy able to approximate a multidimensional convex-hull, has been proved to overperform state-of-the-art one-class classification methodology.
The practical focus of the thesis is towards Physical Activity Recognition. A new hardware platform for wearable computing application has been developed and used for collecting data of activities of daily living allowing to study the optimal features set able to successful classify activities.
Based on the classification methodologies developed and the study conducted on physical activity classification, a machine learning architecture capable to provide a continuous authentication mechanism for mobile-devices users has been worked out, as last part of the thesis. The system, based on a personalized classifier, states on the analysis of the characteristic gait patterns typical of each individual ensuring an unobtrusive and continuous authentication mechanism.
