作者: David Julian Lubinsky
DOI:
关键词: Entropy (information theory) 、 Recursive partitioning 、 Efficient algorithm 、 Decision tree 、 Statistics 、 Bivariate analysis 、 Mathematics
摘要: We extend the recursive partitioning approach to classifier learning use more complex splits at each decision node. To do this a new split criterion is derived. Efficient algorithms for finding optimal under linear, rectangular, and corner are presented. These allow trees model non-orthogonal structure result in that often perform better terms of reclassification accuracy than traditional methods, as well being significantly smaller. Second, we discuss lack consistency criteria such entropy Gini. no optimize accuracy, whereas predictive most important metric by which measured only leads consistent estimates thresholds. A modification standard tree growing algorithm ensures proposed shown give smaller with performance on number datasets. show might miss large amounts certain conditions can generate unbounded size.