属性の識別能力の局所性を考慮した確率的決定木の構築

抄録

Generally, all the values of each attribute do not always work well in induction of any domain. It often causes poor performance of decision trees to handle thus values as ones of adequate discriminating Power. This paper presents a method of building probabilistic decision trees from continuous-valued attributes, considering locality of their discriminating powers. We cluster out the set of the training data into subsets, focussing on correlations among value of attribute and probabilities of identifications with each class. A set of each distribution of probability density of data, which is presumed from each subset, generate such branches corresponding to the level of the discriminating power and dealing with the noises in the attribute values of data stochastically. Empirical results compared with C4.5 shows some advantages, in applying them to real-world domain, diagnosis problem of image-processed data of cancer cells.