Integration Theory And Methods Of Classification

Selective ensemble learning can not only improve the generalization ability, but also can save storage space and increase the speed of predict-tion. Since the concept of selective ensemble learning proposed in 2003, there have appeared more and more research results and practical appli-cation. Tracking domestic and international research frontiers, this work devotes to deep exploration and experiment in view of ordinal classify-cation as follows:(1) Make a more comprehensive and meticulous research of the cur-rent international and domestic popular classic selective ensemble learn-ing based on Chunxia Zhang’s work. This work analyzes the diverse methods comparatively and lists the advantages, disadvantages and range of application. Furthermore, predict selective ensemble learning’s future development direction. In addition, explore the relationship between the common used kappa statistic and generalization ability in theoretical ana-lysis and experiment observation, in order to design a more reasonable selective methods.(2) In addition to nominal variables, there may exsit a kind of order relation in interpretation and response variables. That is, values of vari-ables present preference like good, medium and bad. Sometimes we may encounter that the interpretation variables behave monotony relation with response variables, that is to say, samples with high values of interpret-tation variables should be assigned to high value response variable. To figure out this, ordinal classification is born. Various machine learning algorithms such as decision tree, neural network and perception are trans-formed to classify monotonously. Meanwhile it is employed in assessing clientetes, bidding decision making, medical efficiency and a number of areas. Ensemble methods aim to ordinal classfication gradually rise, while selective methods concerning this is rare. This work proposes wei-ghted kappa gain based algorithm to select ordinal classification classi- tiers, and performs well in simulation experiments and real problems.Finally the work of this paper is summed up and the direction of further worthy work is pointed out.