Research On Machine Learning Methods And Applications In Evaluating Fund Projects

The rapid advances of computer and network technology has been promoting theimprovement of social productivity. Human has become more and more eager torealize intelligence on machines, which stimulates the development of computationalintelligence (CI) and machine learning (ML). By using the principles of statistics,computational intelligence, systems engineering, and management sciences, thisdissertation tries to make a systematic study about the ML theories and methods,concept learning algorithms, and support vectors machine, and their applications inthe evaluation and management of fund projects. The main contents are as follows.1. The general framework of intelligent systems or knowledge systems isintroduced, and the relation between machine learning and computational intelligenceis analyzed. Then the concept, methods, and applications of machine learning arereviewed, and the idea of using machine learning methods to assist the evaluation andmanagement of fund projects is advanced.2. We analyze the concept of PAC learnability, and derive the inequality boundon the number of training examples for a consistent learner or an inconsistent learnerto learn target concepts. We describe the consistency of ERM learning process instatistical learning theory. The VC-dimension is used to build the inequality boundbetween the empirical risk and the real risk, and the SRM induction principle isdiscussed.3. We analyze the strategies and typical algorithms of concept rule learning. Anovel approach for concept learning is proposed based on constrained clustering(CLCC). The concept of conjunctive learnability is defined, and a heuristic methodis formed to cluster the positive instances into multiple divisions. A fast procedure isformulated to learn CNF rule based on entropy. Moreover, a post pruning procedure isdesigned to deal with the over-fitting problem. Then the CLCC is used to evaluatefund projects, and experiments results are analyzed and compared with existingalgorithms.4. We study the models of support vector machine (SVM), and analyze theprocedure of SMO algorithm. Then we propose a novel and concise method for theselection of candidate support vectors (SCSV) based on the structural information oftwo classes in input space by calculating the relative Euclidean distance of all samplesto the boundary of another class. We also propose an abnormal examples filtering(AEF) procedure to find abnormal examples or outliers that may give rise to thedistortion of structural information on the boundaries of two classes. Then AEF+SCVand SMO is integrated to learn SVM classification functions. This method is thenused to evaluate fund projects, and experiments are implemented in detail.5. By incorporating decision support systems and multi-agents technologies, wepropose the machine learning based decision support system for evaluation andmanagement of fund projects (PE&PRDSS). The workflow process, main componentsand framework for the evaluation and information retrieval of fund projects arepresented, and ID3, CLCC, and SVM methods are adopted to build the knowledgesystems for the evaluation and information retrieval of fund projects. The prototypemodel of PE&PRDSS is designed by using multi-agents technology, and the functionand control pattern are described.