Árvore de decisão por agrupamento com DBSCAN aproximativo

Abstract

Machine learning aims to create systems that can learn from data, identify patterns, and make decisions with minimal human intervention. Generally, machine learning methods employ various geometric concepts involving points in multidimensional space. On the other hand, in the literature of algorithms for solving geometric concepts, there are numerous algorithms of this nature that involve randomized and approximate procedures. These procedures sometimes result in an improvement in expected time complexity and at other times in simplifying the implementation of these algorithms. This thesis presents a variant of the decision tree model, where new predicates are considered to partition the data. Instead of using only univariate predicates (associated with a single feature of the data), as is the case with the ordinary decision tree, the new variant also considers multi-feature predicates. Such predicates involve separating the data by their relevance to each of the groups produced by the unsupervised clustering model DBSCAN. The DBSCAN algorithm has a high computational time complexity to be directly integrated into decision trees. Therefore, the proposal involves using an approximate version with linear time complexity of the DBSCAN algorithm to mitigate this impact. The thesis not only proposes this but also conducts experiments on various machine learning benchmark datasets.