A Study On Improving Cost-sensitive Learning Based On Decision Trees

Generally speaking, Machine Learning is the study of computers that improve automatically through experience. The influence of machine learning will be unprecedented if computers can indeed improve themselves by experience completely automatically on all tasks. Thus a great number of researchers and experts have studied very hard to make computers learn and improve by themselves since the first electronic computer ENIAC came into the world in 1946. So far algorithms for special learning tasks have been proposed (e.g., [58][59]) and theories about learning have been proven and published (e.g., [60][61]). In particular, learning functions from specific training data is a key problem in machine learning, and the Decision Tree is one of the most popular learning algorithms studied and used in real-world applications [1][10]. Decision Tree method is applied to many fields nowadays, for instance, classifying patients by different symptoms [58], classifying faults by different causes, and classifying loan applications by the data about the applicants. The key tasks of these problems are to classify cases into appropriate classes, so they are also named the Classification Problem [1].Traditional Classification Problem aims at achieving high accuracy of classification under the assumption that the costs are the same for different misclassification errors. But the hypothesis is not realistic for real-world applications. Taking a diagnoses system as an example, the costs are very different between classifying a patient into a healthy person and classifying a healthy person into a patient. To solve this kind of problems, machine learning researchers have proposed many so-called Cost-sensitive Learning methods. Because of the widespread applications in realities, Cost-sensitive Learning has become a very active research area in Machine Learning [20][51].In this thesis, we first introduce the main research areas in Machine Learning, and analyze the current domestic and international research efforts on Cost-sensitive Learning. Based on our analyses, we point out the advantages and the limitations of the existing Cost-sensitive Learning methods. To overcome these limitations, we propose several important new ideas and methods. Experimental results have proved that the proposed new methods are feasible and effective. The following are the main...