Robust unsupervised learning using kernels

This thesis aims to study deep connections between statistical robustness and machine learning techniques, in particular, the relationship between some particular kernel (the Gaussian kernel) and the robustness of kernel-based learning methods that use it. This thesis also presented that estimating the mean in the feature space with the RBF kernel, is like doing robust estimation of the mean in the data space with the Welsch M-estimator. Based on these ideas, new robust kernel to machine learning algorithms are designed and implemented in the current thesis: Tukey’s, Andrews’ and Huber’s robust kernels which each one corresponding to Tukey’s, Andrews’ and Huber’s M-robust estimator, respectively. On the one hand, kernel-based algorithms are an important tool which is widely applied to different machine learning and information retrieval problems including: clustering, latent topic analysis, recommender systems, image annotation, and contentbased image retrieval, amongst others. Robustness is the ability of a statistical estimation method or machine learning method to deal with noise and outliers. There is a strong theory of robustness in statistics; however, it receives little attention in machine learning. A systematic evaluation is performed in order to evaluate the robustness of kernel-based algorithms in clustering showing that some robust kernels including Tukey’s and Andrews’ robust kernels perform on par to state-of-the-art algorithms

• Autores: Gallego Mejia, Joseph Alejandro
• Fecha de publicación: 2017
• Institución: Universidad Nacional de Colombia
• Repositorio: Universidad Nacional de Colombia
• Idioma: spa
• OAI Identifier: oai:repositorio.unal.edu.co:unal/59937
• Acceso en línea: https://repositorio.unal.edu.co/handle/unal/59937; http://bdigital.unal.edu.co/57770/
• Palabra clave: 62 Ingeniería y operaciones afines / Engineering; Machine Learning; Dimensionality Reduction; Unsupervised Learning; Kernel Learning Approach; Robust Statistics; Welsch Estimator; Aprendizaje de máquina; Reducción de la dimensionalidad; Aprendizaje con métodos de Kernel; Estadística robusta; Eestimadorde Welsch.
• Rights: openAccess
• License: Atribución-NoComercial 4.0 Internacional