Research On Decision Tree Models Under Different Decision-making Environments

With the continuous development of computer technology, data mining and machine learning techniques have been widely used in all areas of society. Data mining and machine learning is an important research topic in the decision tree classification. Traditional method of decision tree classification only cares about the accuracy of classification, did not take into account the fact that different misclassification would have different costs, cannot be applied to cost-sensitive environments. To solve the decision problem in cost-sensitive environments, cost-sensitive decision tree has been proposed on the basis of traditional decision trees. After nearly a decade of research and development, cost-sensitive decision tree learning achieved fruitful results. Different types of cost-sensitive learning technology have been proposed one after another, like misclassification costs, test costs, teacher costs, computation costs and intervention costs, etc. After that, heterogeneous cost-sensitive decision tree is presented.After cost-sensitive decision trees have been proposed, the cost sensitive issues can be solved well. In practice, however, any decision should take into account a number of factors. For example, the income sensitive issues need to take into account income factors, preference sensitive issues need to take into account preference factor. Traditional decision trees and cost-sensitive decision tree has struggled to adapt to these circumstances, we must continue to develop new theories and methods to meet the new challenges.Faced with these new problems and new challenges, this paper focuses on construction of decision trees separately in the cost-sensitive environments and preference-sensitive environments. In order to solve the cost-sensitive decision problem and preference-sensitive decision-making problems, this paper proposes a cost and benefits sensitive decision tree and preference-sensitive decision trees. Experiments demonstrated the feasibility and effectiveness of these methods. The main innovations of this paper are as follows:(1) Raised the issue of decision trees in complex decision-making environment. First of all, this paper introduces the basic theory of decision tree, answered "what is a decision tree" and "how to build a decision tree". And then summarizes the current decision tree algorithm, which can be divided into traditional decision trees (to get the smallest error rate as the ultimate goal) and cost-sensitive decision trees (to obtain a minimum cost for the target). Analyzes the shortcomings of traditional decision trees and cost-sensitive decision tree, and pointed out that "benefits-sensitive decision-making environment" and "preference-sensitive decision-making environment" bring new challenges to the decision tree.(2) In the benefits-sensitive decision-making environment, this paper proposes a new decision tree classifier based on the costs and benefits (CBDT). In the algorithm, we abandoned the traditional judging standards of node class marking, to switch to UCB (Unit cost-benefit) criterion, at the same time considering misclassification cost and correct classification benefits, eventually in order to get the maximum benefits and the small cost at the same time. We conducted a variety of experiments, including the "comparison of different algorithms", " comparison of different miss rate" and "significant difference analysis of UCB". Experimental results show that our approach is more effective and practical.(3) In the Preference-sensitive decision-making environment, this paper proposes a preference sensitive decision tree classification algorithm (PSDT). We analyzed the preference sensitive issues, define the preference degree and preference cost, and gives the concept of preference sensitive learning. Then, the article describes a method to determine the best preferences degree by self-adaption, designs attribute selection factor based on preference-sensitive, and constructs a preference cost sensitive decision tree model and pruning algorithm. We conducted a variety of experiments, including the "effects of preference degree and preference cost on PSDT ",’compare PSDT tree with C4.5 tree " and " verify the effectiveness of PSDT ". Experiments show that PSDT not only achieve a high accuracy prediction for the preferences class but also ensure a good overall accuracy. PSDT solved the preference sensitive issues and is more effective and practical.