Árboles de decisión para grandes conjuntos de datos
Decision Trees are are among the most used supervised classification algorithms. Currently, there are several algorithms for building decision trees, however, just a few of these algorithms allow processing large datasets. Besides, those algorithms designed for processing large datasets have some re...
| Author: | |
|---|---|
| Format: | doctoral thesis |
| Status: | Versión aceptada para publicación |
| Publication Date: | 2010 |
| Country: | México |
| Institution: | Instituto Nacional de Astrofísica, Óptica y Electrónica |
| Repository: | Repositorio Institucional del INAOE |
| Language: | Spanish |
| OAI Identifier: | oai:inaoe.repositorioinstitucional.mx:1009/506 |
| Online Access: | http://inaoe.repositorioinstitucional.mx/jspui/handle/1009/506 |
| Access Level: | Open access |
| Keyword: | info:eu-repo/classification/Árboles de decisión/Decision trees info:eu-repo/classification/Bases de datos muy grandes/Very large databases info:eu-repo/classification/Inteligencia artificial/Artificial intelligence info:eu-repo/classification/Reconocimiento de patrones/Pattern recognition info:eu-repo/classification/cti/1 info:eu-repo/classification/cti/12 info:eu-repo/classification/cti/1203 |
| Summary: | Decision Trees are are among the most used supervised classification algorithms. Currently, there are several algorithms for building decision trees, however, just a few of these algorithms allow processing large datasets. Besides, those algorithms designed for processing large datasets have some restrictions, for example: spatial restrictions; the number of times that they have to scan the whole training set for building the decision tree; some algorithms only use a small subsample of the training set, but for obtaining this subsample they spend a lot of time, specially for large datasets; other algorithms use several parameters, which can be very difficult to determine by the user. For this reason, in this thesis we propose algorithms for building decision trees for large datasets, that solve the restrictions of the most recent algorithms in the state of the art, considering that the number of classes is lesser than the number of instances in the training set. The proposed algorithms use the whole training set for building the decision tree, without storing the whole training set in memory. In particular, in this thesis, we propose two algorithms for building multivariate decision trees for instances described by numeric attributes. The first algorithm uses all the attributes in the internal nodes of the decision tree. However, if the instances are described by a large number of attributes, the time needed for traversing the tree can be too long. For this reason, we propose a second algorithm, which uses splitting attribute subsets in the internal nodes. Although the previous algorithm generates multivariate decision trees using splitting attribute subsets, the time needed for traversing the decision tree can also be too long. Hence, in this thesis, we propose two algorithms for building univariate decision trees. The first one for instances described by numeric attributes, and the second for instances with mixed attributes. Based on the experimental results, we can conclude that our algorithms are faster than the most recent algorithms for building decision trees for large datasets, maintaining competitive accuracy. Therefore, the proposed algorithms are a good option for building decision trees for large datasets. |
|---|