A new inductive data mining method for automatic generation of decision trees from data (GPTree) is presented. Compared with other decision tree induction techniques that are based upon recursive partitioning employing greedy searches to choose the best splitting attribute and value at each node therefore will necessarily miss regions of the search space, GPTree can overcome the problem. In addition, the approach is extended to a new method (YAdapt) that models the original continuous endpoint by adaptively finding suitable ranges to describe the endpoints during the tree induction process, removing the need for discretisation prior to tree induction and allowing the ordinal nature of the endpoint to be taken into account in the models built. A strategy for further improving the predictive performance for previously unseen data is investigated that uses multiple decision trees, i.e., a decision forest, and a majority voting strategy to give predictions (GPForest). The methods were applied to QSAR (quantitative structure – activity relationships) modelling for eco-toxicity prediction of chemicals and to the analysis of a historical database for a wastewater treatment plant.
[1]
J. Dearden,et al.
Assessment and modeling of the toxicity of organic chemicals to Chlorella vulgaris: development of a novel database.
,
2004,
Chemical research in toxicology.
[2]
X Z Wang,et al.
Induction of decision trees using genetic programming for modelling ecotoxicity data: adaptive discretization of real-valued endpoints
,
2006,
SAR and QSAR in environmental research.
[3]
J. Ross Quinlan,et al.
Improved Use of Continuous Attributes in C4.5
,
1996,
J. Artif. Intell. Res..
[4]
Anita Young,et al.
Genetic Programming for the Induction of Decision Trees to Model Ecotoxicity Data
,
2005,
J. Chem. Inf. Model..