The Automatic Construction of Multivariate Split Point Regression Trees: A Genetic Programming Approach

• Main Authors: Shih Ming Lin, 林世明
• Other Authors: Chao Chang Chiu
• Format: Others
• Language: en_US
• Published: 2002
• Online Access: http://ndltd.ncl.edu.tw/handle/70618057738330582399

碩士 === 元智大學 === 資訊管理學系 === 90 === Data mining is the automated search for interesting and useful relationships between attributes in database. In many of the best techniques (such as neural networks) yield little in terms of usable rules. In recent years, there has been considerable success in the use of genetic programming (GP) to evolve pattern recognizers. In this article we presents a GP-multivariate split point regression tree algorithm (called GPMRT) as an alternative to existing regression tree approaches. This is a reformed algorithm from previous research (called GPRT). It is using genetic programming and local linear regression to construct regression trees by genetic selection of features and Univariate split points, then using tree pruning and evaluation trying to find out an optimized regression tree. In our research, we focus on multivariate split points and introduce a MDL principle to balance accuracy and parsimony. We want to prove its efficiency of the splitter of the multi dimension regression tree is perfect than single dimension regression tree. The experiment results show that GPMRT is better than GPRT in training and testing phase.