A Study On Improvement And Applications Of Random Forest Classification Algorithm

This thesis mainly studies the random forests(RFs)classification algorithm and its applications,in attempt to improve the construction process of random forest.RFs are one of the most effective approach in integration methods.It not only has less parameters and high operation rate,but also can avoid the over-fitting phenomenon and has good ability to tolerate noise.However,the classification performance of RF is usually affected by single base classifiers.In general,the higher classification capability of single classifier is,the better classification effect of the whole integration forest.At the same time,the difference between the base classifiers is another one important factor that affects the performance of RF,the stronger independence of the base classifier is,the better performance of the integrated algorithm.In this dissertation,we launched a research and discussion on random forest classification algorithm and its applications,mainly including researching on rotation forest algorithm based on discriminative locality alignment and weighting vote,studying on random forest algorithm based on multiple kernel learning as well as the application research of random forests in human white blood cell classification.The main work can be summarized as follows:1.In order to improve the generalization performance of the RF,we devise a new rotation forest ensemble method to increase the diversity of each tree in the forests,which is in view of feature extension and transformation.Lastly,a weighting vote for base classifiers is applied to integrate the final ensemble results for the purpose of improving classification performance.Experimental results conducting with UCI classification datasets and face recognition databases demonstrate that proposed algorithm has higher recognition accuracy than other methods.2.This dissertation applies principal component analysis and linear discriminant analysis to transform the features into two different rotation spaces,and then a multiple kernel support vector machine(MKLSVM)is selected as a base learner for the construction of random forest.The proposed ensemble forest based on the multiple kernel support vector machine has higher classification accuracy than traditional integrated classification algorithm and support vector machine,through the validation on the classification of UCI datasets.3.A classification algorithm is put forward for human peripheral white blood cells(WBCs)in view of computer image processing and artificial intelligence.According to the characteristics of the nucleus and cytoplasm of five types of WBC,pairwise rotation invariant co-occurrence local binary pattern feature on texture and the integral invariant shape feature on morphology are exacted from segmented cell images besides the usual nuclear-cytoplasmic ratio and circularity features.Then all these features are combined together and normalized.Finally,a random forest is applied to classify those five types of WBC.Some experiments show that the proposed classification algorithm has a better recognition accuracy than some other existing classification methods for WBCs.