Dominating sets of the Gabriel Graph : na approach for one-class and online learning classifiers Gabriel graph

• Autor(a) principal: Wagner José de Alvarenga Júnior
• Data de Publicação: 2022
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Repositório Institucional da UFMG
• Texto Completo: http://hdl.handle.net/1843/47406
• topic: Online learnig; Gabriel graph; Dominating set; KDE; RBF; One-class classifier; SlFN; Engenharia elétrica; Aprendizado do computador; Teoria dos grafos

Online Learning of non-stationary data streams is characterized by changes in the data generating function, which may impact the predictive performance of a model. Therefore, classifiers capable of adapting to such situations constitute a viable solution. Generally, such models rely on hyperparameters that need to be previously configured. A different task that also presents issues, concerning setting of hyperparameters, is the learning with one-class classifiers, in which the information from only one class is used to establish a decision boundary. The main proposal of this thesis is to use the structural information from a data set to define classifiers, in the two Learning Paradigms previously discussed. This goal is achieved by exploring the fact that an Independent Dominating Set, when induced from the Gabriel graph, tends by definition to result in a subset of dominating points, with representative characteristic of the original set. Thus, an Independent Dominating Set algorithm that neither requires setting hyperparameters nor the use of any optimization method to find a solution is proposed, as well as an online updating procedure for the Gabriel graph. These two methods are used to define the hypeparameters of models based on Radial functions: a KDE estimator for the online scenario and an RBF network as a one-class classifier. This graph dominance approach results in an appropriate and distributed number of Radial functions, in the input domain, and a stable radius that cover the training points and leads to a classifier with appropriate Capacity. An algorithm based on the independent dominating set of the Gabriel graph is also proposed to extract representative subsets from large data sets. This thesis also presents an online training method for a regularized SLFN network that continually maintains the learning process. The method uses an adaptive window to mitigate the impact of concept drifts. These methods were tested with synthetic data sets and with data from a real industrial process.