C. Fernandez, M.A. Vicente, O. Reinoso, A. Gil, and L. Paya (Spain)
Artificial intelligence, knowledge representation, decision trees, readability, feature selection.
A novel measure for the goodness of a decision tree gener ation algorithm is proposed. The goal is to show simulta neously the expected classification accuracy and tree com plexity, under different tunings of the algorithm. This in formation is relevant in order to decide which algorithm is more suitable for a certain application of decision trees. The proposed measure is represented as a behavior curve, defined as the monotonic upper envelope of the data cloud showing all possible results in the accuracy vs. complex ity representation. With slight modifications, such behav ior curve can be applicable to other classifiers, like rule lists, neural networks, etc. The proposed measure is used to study the effects of a previous feature selection step in the behavior of Quinlan’s C4.5 algorithm. Experimental results are shown with datasets belonging to a robot grasp application and also with datasets obtained from the UCI repository.
Important Links:
Go Back
