An investigation into new kernels for categorical variables
Year: 2013
Type: dissertation
Abstract: Kernel-based methods first appeared in the form of support vector\nmachines. Since the first Support Vector Machine (SVM) formulation in\n1995, we have seen how the number of proposed kernel\nfunctions has quickly grown, and how these kernels have approached a\nwide range of problems and domains. The most common and direct\napplications of these methods are focused on continuous numeric data,\ngiven that SVMs at the end involves the solution of an optimization problem. Additionally, some kernel functions have been oriented to more\nsymbolic data, in problems like text analysis, or hand-written digits\nrecognition. But surprisingly, there is a gap in the area of kernel\nfunctions devoted to handle datasets with qualitative variables. One of the\nmost common practices to overcome this lack consists on recoding the\nsource qualitative information, making them suitable for applying numeric\nkernel functions.\nThis thesis presents the development of new kernel functions that can\nbetter model symbolic information presented as categorical variables, in a\ndirect way, and without the need of data preprocessing methods. The\nproposition is based on the use of probabilistic information (probability\nmass distribution) to compare the different modalities of a variable.\nAdditionally, the idea is formulated through a modular schema, combining a\nset of components to obtain the kernel functions, facilitating the\nmodification and extension of single components.\nThe experimental results suggest an slightly improvement with respect\nto traditional kernel functions, in the accuracy obtained on classification\nproblems. This progress is clearer on datasets with known probabilistic\nstructure.
Authors: Villegas García, Marco Antônio
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W652184952
