Abstract. Decision trees are widely used in different applications for problem solving and for knowledge representation. In the paper algorithms for decision tree constructing with bounds on complexity and precision are considered. In these algorithms different measures for time complexity of decision trees and different measures for uncertainty of decision tables are used. New results about precision of polynomial approximate algorithms for covering problem solving [1,2] show that some of considered algorithms for decision tree constructing are, apparently, close to unimprovable.
IntroductionIn 1983 the paper [4] was published (short variant was published in 1982 [3]) which contained bounds on time complexity of decision trees and algorithms for decision tree constructing (in this paper decision trees were named conditional tests). In algorithms different measures for time complexity of decision trees (depth, weighted depth and others) and different measures for uncertainty of decision tables were used. Bounds on precision for these algorithms were considered. These algorithms resemble on algorithms of J.R. Quinlan [10, 11] however they were proposed independently. Obtained results were published only in Russian. New bounds on precision of polynomial approximate algorithms for covering problem solving [1,2] show that some of considered in [4] algorithms are, apparently, close to unimprovable polynomial approximate algorithms for constructing decision trees with minimal depth. This paper contains a survey of some results from [3, 4] (definitions of complexity and uncertainty measures, description of algorithms for decision tree constructing, bounds on precision of these algorithms), some results from [5] (bounds on complexity of considered algorithms and bounds on precision of algorithms for one important class of decision tables), and reasons about closeness of some algorithms to unimprovable.Mentioned results are useful for obtaining bounds on time complexity of decision trees [7,8]. Also these results may be useful in data mining and knowledge discovery for decision tree constructing. The considered algorithms allow variations of complexity and uncertainty measures in broad bounds: the statements